#387 – George Hotz: Tiny Corp, Twitter, AI Safety, Self-Driving, GPT, AGI & God

[0] The following is a conversation with George Hatz, his third time on this podcast.

[1] He's the founder of Kama AI that seeks to solve autonomous driving and is the founder of a new company called Tiny Corp that created TinyGrad, a neural network framework that is extremely simple, with the goal of making it run on any device by any human, easily, and officially.

[2] As you know, George also did a large number of fun and amazing things from hacking the iPhone to recently joining Twitter for a bit as an intern, in quotes, making the case for refactoring the Twitter code base.

[3] In general, he's a fascinating engineer in human being and one of my favorite people to talk to.

[4] And now, a quick few second mention of each sponsor.

[5] Check them out in the description.

[6] It's the best way to support this podcast.

[7] We got Numeri for the world's hardest data science tournament, Babel for learning new languages, net suite for business management, software, InsideTracker for blood paneling, and AG1 for my daily multivitamin.

[8] Choose wisely, my friends.

[9] Also, if you want to work on our team, we're always hiring, go to Lex Friedman .com slash hiring.

[10] And now I want to the full ad reads.

[11] As always, no ads in the middle.

[12] I try to make this interesting, but if you must skip them, friends, please still check out our sponsors.

[13] I enjoy their stuff.

[14] Maybe you will too.

[15] This episode is brought to you by Numeri, a hedge fund that uses artificial intelligence and machine learning to make investment decisions.

[16] They created a tournament that challenges data scientists to build best predictive models for financial markets.

[17] It's basically just a really, really difficult real -world data set to test out your ideas for how to build machine learning models.

[18] I think this is a great educational platform.

[19] I think this is a great way to explore, to learn about machine learning, to really test yourself on real -world data with consequences.

[20] No financial background is needed.

[21] The models are squire.

[22] based on how well they perform on unseen data and the top performers receive a share of the tournament's prize pool.

[23] Head over to Numeri slash Lex.

[24] That's N -U -M -E -R -A -I -S -Lex to sign up for a tournament and hone your machine learning skills.

[25] That's Numeri slash Lex for a chance to play against me and win a share of the Tournament's Prize pool.

[26] That's Numeri slash Lex.

[27] This show is also brought to you by Babel, an app and website that gets you speaking in a new language within weeks.

[28] I have been using it to learn a few languages, Spanish, to review Russian, to practice Russian, to revisit Russian from a different perspective, because that becomes more and more relevant for some of the previous conversations I've had and some upcoming conversations I have.

[29] It really is fascinating how much another language, knowing another language, even to a degree where you can just have little bits and pieces of a conversation can really unlock an experience, another part of the world.

[30] When you travel in France and Paris, just having a few words at your disposal, a few phrases, it begins to really open you up to strange, fascinating, new experiences that ultimately, at least to me, teach me that we're all the same.

[31] We have to first see our differences to realize those differences are grounded in a basic humanity.

[32] And that experience that we're all very different and yet at the core the same.

[33] I think travel, with aid of language really helps unlock.

[34] You can get 55 % off your Babbel subscription at babble .com slash Lexpod.

[35] That's spelled B -A -B -E -L .com slash Lexpod.

[36] Rules and restrictions apply.

[37] This show is also brought to you by NetSuite.

[38] An all -in -one cloud business management system.

[39] They manage all the messy stuff that is required to run a business, the financials, the human resources, the inventory, if you do that kind of thing, e -commerce, all that stuff, all the business -related details.

[40] I know how stressed I am about everything that's required to run a team, to run a business that involves much more than just ideas and designs and engineering.

[41] It involves all the management of human beings, all the complexities of that, the financials, all of it, and sure you should be using the best tools for the job, I sometimes wonder if I have it in me mentally and skill -wise to be a part of running a large company.

[42] I think like with a lot of things in life, it's one of those things you shouldn't wonder too much about.

[43] You should either do or not do.

[44] But again, using the best tools for the job is required here.

[45] You can start now with a no payment or interest for six months.

[46] Go to NetSuite .com slash Lex to access their one -of -a -kind financing program.

[47] That's NetSuite .com slash Lex.

[48] This show is also brought to you by Insight Tracker, a service I use to track biological data, data that comes from my body, to predict, to tell me what I should do with my lifestyle, my diet, what's working and what's not working.

[49] it's obvious all the exciting breakthroughs that are happening with transformers with large language models even with diffusion all of that is obvious that with raw data with huge amounts of raw data fine -tuned to the individual would really reveal to us the signal in all the noise of biology I feel like that's on the horizon the kinds of leaps in development that we saw in language and now more and more visual data.

[50] I feel like biological data is around the corner.

[51] Unlocking what's there in this multi -hierarchical distributed system that is our biology.

[52] What is it telling us?

[53] What is the secrets it holds?

[54] What is the thing that it's missing that could be aided?

[55] Simple lifestyle changes, simple diet changes, simple changes in all kinds of things that are controllable by individual human being.

[56] I can't wait till that's a possibility.

[57] And inside trackers taking steps towards that.

[58] Get special savings for a limited time when you go to Insightracker .com slash Lex.

[59] This show is also brought to you by Athletic Greens that's now called AG1.

[60] It has the AG1 drink.

[61] I drink it twice a day.

[62] At the very least, it's an all -in -one daily drink to support better health and peak performance.

[63] I drink it cold.

[64] It's refreshing.

[65] It's grounding.

[66] It helps me reconnect with the basics, the nutritional basics that makes this whole machine that is our human body run, all the crazy mental stuff I do for work, the physical challenges, everything, the highs and lows of life itself.

[67] All of that is somehow made better knowing that at least you've got nutrition in check.

[68] At least you're getting enough sleep.

[69] At least you're doing the basics.

[70] At least you're doing the exercise.

[71] Once you get those basics in place, I think you can do some quite difficult things in life.

[72] Anyway, beyond all that is just a source of happiness and a kind of a feeling of home, the feeling that comes from returning to the habit time and time again.

[73] Anyway, they'll give you one month supply of fish oil when you sign up at drinkag1 .com slash Lex.

[74] This is a Lex Friedman podcast to support it.

[75] Please check out our sponsors in the description.

[76] And now, dear friends, here's George Hots.

[77] You mentioned something on a stream about the philosophical nature of time.

[78] So let's start with a wild question.

[79] Do you think time is an illusion?

[80] You know, I sell phone calls to comma for $1 ,000.

[81] And some guy called me and like, you know, it's $1 ,000.

[82] You can talk to me for half an hour.

[83] And he's like, yeah, okay, so like time doesn't exist.

[84] And I really wanted to share this with you.

[85] I'm like, oh, what do you mean, time?

[86] doesn't exist, right?

[87] Like, I think time is a useful model, whether it exists or not, right?

[88] Like, does quantum physics exist?

[89] Well, it doesn't matter.

[90] It's about whether it's a useful model to describe reality.

[91] Is time maybe compressive?

[92] Do you think there is an objective reality, or is everything just useful models?

[93] Like, underneath it all, is there an actual thing that we're constructing models for?

[94] I don't know.

[95] I was hoping you would know.

[96] I don't think it matters.

[97] I mean, this kind of connects to the models of constructive reality with machine learning, right?

[98] Sure.

[99] Like, is it just nice to have useful approximations of the world such that we can do something with it?

[100] So there are things that are real.

[101] Columagraph complexity is real.

[102] Yeah.

[103] The compressive thing.

[104] Math is real.

[105] Yeah.

[106] Should be a T -shirt.

[107] And I think hard things are actually hard.

[108] I don't think P equals NP.

[109] Ooh, strong words.

[110] Well, I think that's the majority.

[111] I do think factoring is in P, but...

[112] I don't think you're the person that falls the majority in all walks of life, so...

[113] Well, for that one, I do.

[114] Yeah.

[115] In theoretical computer science, you're one of the sheep.

[116] All right.

[117] But do you, time is a useful model?

[118] Sure.

[119] Hmm.

[120] What were you talking about on the stream with time?

[121] Are you made of time?

[122] If I remembered half the things I said on stream.

[123] Someday someone's going to make a model of all of it, and it's going to come back to haunt me. Someday soon?

[124] Yeah, probably.

[125] Would that be exciting to you or sad that there's a George Hott's model?

[126] I mean, the question is when the George Hots model is better than George Hots.

[127] Like, I am declining and the model is growing.

[128] What is the metric by which you measure better or worse in that?

[129] If you're competing with yourself.

[130] Maybe you can just play a game where you have the George Hots answer and the George Hots model answer and ask which people prefer?

[131] People close to you or strangers?

[132] Either one.

[133] It will hurt more when it's people close to me. but both will be overtaken by the George Hott's model.

[134] It'd be quite painful, right?

[135] Loved ones, family members, would rather have the model over for Thanksgiving than you.

[136] Yeah.

[137] Or like significant others, would rather sext with the large language model version of you.

[138] Especially when it's fine -tuned to their preferences.

[139] Yeah.

[140] Well, that's what we're doing.

[141] in a relationship, right?

[142] We're just fine -tuning ourselves, but we're inefficient with it because we're selfish and greedy and so on.

[143] All language models can fine -tune more efficiently, more selflessly.

[144] There's a Star Trek Voyager episode where, you know, Catherine Janeway, lost in the Delta Quadrant, makes herself a lover on the holodeck.

[145] And the lover falls asleep on her arm, and he snores a little bit, and, you know, Janeway edits the program to remove that.

[146] And then, of course, the realization is, wait, person's terrible.

[147] It is actually all their nuances and quirks and slight annoyances that make this relationship worthwhile.

[148] But I don't think we're going to realize that until it's too late.

[149] Well, I think a large language model could incorporate the flaws and the quarks and all that kind of stuff.

[150] Just the perfect amount of quirks and flaws to make you charming without crossing the line.

[151] Yeah.

[152] Yeah.

[153] And that's probably a good like approximation of the, like, the percent of time, the language model should be cranky or an asshole or jealous or all this kind of stuff.

[154] And of course it can and it will, but all that difficulty at that point is artificial.

[155] There's no more real difficulty.

[156] Okay, what's the difference in real and artificial?

[157] Artificial difficulty is difficulty that's like constructed or could be turned off with a knob.

[158] Real difficulty is like, you're in the woods and you've got to survive.

[159] So if something can not be turned off with a knob, it's real.

[160] Yeah, I think so.

[161] Or, I mean, you can't get out of this by smashing the knob with a hammer.

[162] I mean, maybe you kind of can.

[163] You know, into the wild when, you know, Alexander Super Tramp, he wants to explore something that's never been explored before.

[164] But it's the 90s.

[165] Everything's been explored.

[166] So he's like, well, I'm just not going to bring a map.

[167] Yeah.

[168] I mean, no, you're not exploring.

[169] You should have brought a map, dude.

[170] You died.

[171] There was a bridge a mile from where you were camping.

[172] How does that connect to the metaphor of the knob?

[173] By not bringing the map, you didn't become an explorer.

[174] You just smashed the thing.

[175] Yeah.

[176] The difficulty is still artificial.

[177] You failed before you started.

[178] What if we just don't have access to the knob?

[179] Well, that maybe is even scarier.

[180] Right?

[181] Like, we already exist in a world of nature, and nature has been fine -tuned over billions of years.

[182] Yeah.

[183] to have humans build something and then throw the knob away in some grand romantic gesture is horrifying.

[184] Do you think of us humans as individuals that are like born and die, or is it, are we just all part of one living organism that is Earth, that is nature?

[185] I don't think there's a clear line there.

[186] I think it's all kind of just fuzzy.

[187] I don't know.

[188] I mean, I don't think I'm conscious.

[189] I don't think I'm anything.

[190] I think I'm just a computer program.

[191] So it's all computation.

[192] Everything running in your head is just a computation.

[193] Everything running in the universe is computation, I think.

[194] I believe the extended church -tarring thesis.

[195] Yeah, but there seems to be an embodiment to your particular computation.

[196] Like there's a consistency.

[197] Well, yeah, but I mean, models have consistency, too.

[198] Yeah.

[199] Models that have been RLHF will continually say, you know, like, well, how do I murder ethnic minorities?

[200] Oh, well, I can't let you do that, Al. There's a consistency to that behavior.

[201] It's all RLHF.

[202] Like, we all RLHF each other.

[203] We provide human feedback, and thereby fine -tuned these little pockets of computation, but it's still unclear why that pocket of computation stays with you, like, for years.

[204] It just kind of follow, like, you have this consistent set of physics, biology, what, like, whatever you call the neurons firing.

[205] Like, the electrical signals, the mechanical signals, all of that, that seems to stay there.

[206] And it contains information, it stores information, and that information permeates through time and stays with you.

[207] There's, like, memory.

[208] It's, like, sticky.

[209] Okay, to be fair, like a lot of the models we're building today are very, even RLHF is nowhere near as complex as the human loss function.

[210] Reinforcement learning with human feedback.

[211] You know, when I talked about, will GPT12 be AGI?

[212] My answer is, no, of course not.

[213] I mean, cross -entropy loss is never going to get you there.

[214] You need probably RL in fancy environments in order to get something that would be considered like AGI like.

[215] so to ask like the question about like why i don't know like it's just some quirk of evolution right i don't think there's anything particularly special about where i ended up where humans ended up so okay we have human level intelligence would you call that aGI whatever we have is GI look actually i i don't really even like the word aGI um but general intelligence is defined to be whatever humans have okay so why Why can GPT -12 not get us to AGI?

[216] Could we just, like, linger on that?

[217] If your loss function is categorical cross -entropy, if your loss function is just try to maximize compression, I have a SoundCloud, I rap, and I tried to get ChatGPT to help me write raps.

[218] And the raps that it wrote sounded like YouTube comment raps.

[219] You know, you can go on any rap beat online, and you can see what people put in the comments, and it's the most, like, mid -quality rap you can find.

[220] Is mid -good or bad?

[221] Mid is bad.

[222] it's like mid it's like every time i talk to you i learn new words mid mid yeah i was like uh is it is it like basic is that what mid means kind of it's like it's like middle of the curve right yeah so there's like there's like there's like i like that intelligence curve yeah um and you have like the dumb guy the smart guy and then the mid guy actually being the mid guy's the worst the smart guy is like i put all my money in bitcoin the mid guy is like you can't put money in bitcoin it's not real money And all of it is a genius meme That's another interesting one Memes The humor The idea The absurdity Encapsulated in a single image And it just kind of propagates Virally Between all of our brains I didn't get much sleep last night So I'm very I sound like I'm high I swear I'm not Do you think we have ideas Or ideas have us?

[223] I don't think that we're going to get super scary memes once the AIs actually are super human.

[224] Ooh, the gay eye will generate memes?

[225] Of course.

[226] You think it'll make humans laugh?

[227] I think it's worse than that.

[228] So infinite jest, it's introduced in the first 50 pages, is about a tape that once you watch it once, you only ever want to watch that tape.

[229] In fact, you want to watch the tape so much that someone says, okay, here's a hacksaw, cut off your pinky, and then I'll let you watch the tape again, and you'll do it.

[230] So we're actually going to build that, I think.

[231] But it's not going to be one static tape.

[232] I think the human brain is too complex to be stuck in one static tape like that.

[233] If you look at like ant brains, maybe they can be stuck on a static tape.

[234] But we're going to build that using generative models.

[235] We're going to build the TikTok that you actually can't look away from.

[236] So TikTok is already pretty close there, but the generation is done by humans.

[237] The algorithm is just doing their recommendation.

[238] but if the algorithm is also able to do the generation.

[239] Well, it's a question about how much intelligence is behind it, right?

[240] So the content is being generated by, let's say, one humanity worth of intelligence.

[241] And you can quantify a humanity, right?

[242] That's a, you know, it's X -a -flops, yada -flops, but you can quantify it.

[243] Once that generation is being done by 100 humanities, you're done.

[244] So it's actually scale, that's the problem.

[245] but also speed yeah and what if it's sort of manipulating the very limited human dopamine engine for porn imagine it's just TikTok but for porn yeah it's like a brave new world I don't even know what it'll look like right like again you can't imagine the behaviors of something smarter than you but a super intelligent and an agent that just dominates your intelligence so much, we'll be able to completely manipulate you.

[246] Is it possible that it won't really manipulate?

[247] It'll just move past us.

[248] It'll just kind of exist the way water exists or the air exists.

[249] You see, and that's the whole AI safety thing.

[250] It's not the machine that's going to do that.

[251] It's other humans using the machine that are going to do that to you.

[252] Yeah.

[253] Because the machine is not interested in hurting humans.

[254] The machine is a machine.

[255] But the human gets the machine, and there's a lot of humans out there very interested in manipulating you.

[256] Well, let me bring up Eliezer Yitzkowski, who recently sat where you're sitting.

[257] He thinks that AI will almost shirdlely kill everyone.

[258] Do you agree with him or not?

[259] Yes, but maybe for a different reason.

[260] Okay.

[261] And I'll try to get you to find you to find.

[262] hope or we could find a no to that answer, but why yes?

[263] Okay.

[264] Why didn't nuclear weapons kill everyone?

[265] That's a good question.

[266] I think there's an answer.

[267] I think it's actually very hard to deploy nuclear weapons tactically.

[268] It's very hard to accomplish tactical objectives.

[269] Great.

[270] I can nuke their country.

[271] I have an irradiated pile of rubble.

[272] I don't want that.

[273] Why not?

[274] Why don't I want an irradiated pile of rubble?

[275] For all the reasons no one wants an irradiated pile of rubble.

[276] Oh, because you can't use that land.

[277] for resources, you can't populate the land.

[278] Yeah.

[279] What you want a total victory in a war is not usually the radiation and eradication of the people there.

[280] It's the subjugation and domination of the people.

[281] Okay.

[282] So you can't use this strategically, tactically in a war to help you, to help gain a military advantage.

[283] It's all complete destruction.

[284] All right.

[285] Yeah.

[286] But there's egos involved.

[287] It's still surprising.

[288] It's still surprising that nobody pressed a big red button.

[289] It's somewhat surprising, but you see, it's the little red button that's going to be pressed with AI that's going to, you know, and that's why we die.

[290] It's not because the AI, if there's anything in the nature of AI, it's just the nature of humanity.

[291] What's the algorithm behind the little red button?

[292] What possible ideas do you have for how human species ends?

[293] Sure.

[294] So I think the most obvious way to me is wireheading.

[295] We end up amusing ourselves to death.

[296] We end up all staring at that infinite TikTok and forgetting to eat.

[297] Maybe it's even more benign than this.

[298] Maybe we all just stop reproducing.

[299] Now, to be fair, it's probably hard to get all of humanity.

[300] Yeah.

[301] It probably...

[302] There's always going...

[303] Like, the interesting thing about humanity is a diversity in it.

[304] Oh, yeah.

[305] Organisms in general.

[306] There's a lot of weirdos out there.

[307] Two of them are sitting here.

[308] I mean, diversity in humanity is...

[309] We do respect.

[310] I wish I was more weird.

[311] No, like, I'm kind of, look, I'm drinking smart water, man. That's like a Coca -Cola product, right?

[312] You weren't corporate, George Hots.

[313] No, the amount of diversity in humanity, I think, is decreasing.

[314] Just like all the other biodiversity on the planet.

[315] Oh, boy.

[316] Yeah.

[317] Yeah, social media is not helping, huh?

[318] Go eat McDonald's in China.

[319] Yeah.

[320] No, it's the interconnectedness.

[321] That's doing it.

[322] Oh, that's interesting.

[323] So everybody starts relying on the connectivity of the internet, and over time that reduces the diversity, the intellectual diversity, and then that gets you everybody into a funnel.

[324] There's still going to be a guy in Texas.

[325] There is.

[326] And, yeah.

[327] A bunker.

[328] To be fair, do I think AI kills us all?

[329] I think AI kills everything we call, like, society today.

[330] I do not think it actually kills the human species.

[331] I think that's actually incredibly hard to do.

[332] Yeah, but society, like if we start over, that's tricky.

[333] Most of us don't know how to do most things.

[334] Yeah, but some of us do.

[335] And they'll be okay, and they'll rebuild after the great AI.

[336] What's rebuilding look like?

[337] How much do we lose?

[338] What is human civilization done?

[339] That's interesting.

[340] Combustion engine, electricity.

[341] So power and energy, that's interesting.

[342] Like how to harness energy.

[343] Whoa, whoa, whoa, whoa, they're going to be religiously against that.

[344] Are they going to get back to, like, fire?

[345] Sure.

[346] I mean, there'll be a, it'll be like, you know, some kind of Amish -looking kind of thing, I think.

[347] I think they're going to have very strong taboos against technology.

[348] Hmm.

[349] Like technology, it's almost like a new religion.

[350] Technology is the devil.

[351] Yeah.

[352] And nature is God.

[353] Sure.

[354] So closer to nature.

[355] But can you really get away from AI if it destroyed 99 % of the human species?

[356] Isn't it somehow have a hold, like a stronghold?

[357] What's interesting about everything we build, I think we're going to build superintelligence before we build any sort of robustness in the AI.

[358] We cannot build an AI that is capable of, going out into nature and surviving like a bird, right?

[359] A bird is an incredibly robust organism.

[360] We've built nothing like this.

[361] We haven't built a machine that's capable of reproducing.

[362] Yes, but there is, you know, I work with leg robots a lot, though.

[363] I have a bunch of them.

[364] They're mobile.

[365] They can't reproduce, but all they need is, I guess you're saying they can't repair themselves.

[366] If you have a large number, if you have like a, 100 million of them?

[367] Let's just focus on them reproducing, right?

[368] Do they have microchips in them?

[369] Mm -hmm.

[370] Okay.

[371] Then do they include a fab?

[372] No. Then how are they going to reproduce?

[373] Well, they, hey, it doesn't have to be all on board, right?

[374] They can go to a factory, to a repair shop.

[375] Yeah, but then you're really moving away from robustness.

[376] Yes.

[377] All of life is capable of reproducing without needing to go to a repair shop.

[378] Life will continue to reproduce in the complete absence of civilization.

[379] Robots will not.

[380] So when the, if the AI apocalypse happens, I mean, the AIs are going to probably die out because I think we're going to get, again, superintelligence long before we get robustness.

[381] What about if you just improve the fab to where you just have a 3D printer that can always help you?

[382] Well, that'd be very interesting.

[383] I'm interested in building that.

[384] Of course you are.

[385] You think, how difficult is that problem to have a robot that basically can build itself.

[386] Very, very hard.

[387] I think you've mentioned this, like, to me or somewhere, where people think it's easy conceptually.

[388] And then they remember that you're going to have to have a fab.

[389] Yeah, on board.

[390] Of course.

[391] So a 3D printer that prints a 3D printer.

[392] Yeah.

[393] Yeah, on legs.

[394] Why's that hard?

[395] Well, because it's not, I mean, a 3D printer is a very, simple machine, right?

[396] Okay, you're going to print chips?

[397] You're going to have an atomic printer?

[398] How are you going to dope the silicon?

[399] Yeah.

[400] Right?

[401] How are you going to edge the silicon?

[402] You're going to have to have a very interesting kind of fab if you want to have a lot of computation on board.

[403] But you can do like structural type of robots that are dumb.

[404] Yeah, but structural type of robots aren't going to have the intelligence required to survive in any complex environment.

[405] What about like ants type of systems?

[406] We have like trillions of them.

[407] I don't think this works.

[408] I mean, again, like ants at their very core are made up of cells that are capable of individually reproducing.

[409] They're doing quite a lot, a lot of computation that we're taking for granted.

[410] It's not even just the computation.

[411] It's that reproduction is so inherent.

[412] Okay, so like there's two stacks of life in the world.

[413] There's the biological stack and the silicon stack.

[414] The biological stack starts with reproduction.

[415] Reproduction is at the absolute core.

[416] The first proto -RNA organisms were capable of reproducing.

[417] The silicon stack, despite as far as it's come, is nowhere near being able to reproduce.

[418] Yeah.

[419] So the fab movement, digital fabrication, fabrication in the full range of what that means is still in the early stages.

[420] Yeah.

[421] You're interested in this world.

[422] Even if you did put a fab on the machine, right?

[423] Let's say, okay, you know, we can build fabs.

[424] We know how to do that is humanity.

[425] We can probably put all the precursors that build all the machines in the fabs also in the machine.

[426] So first off, this machine is going to be absolutely massive.

[427] I mean, we almost have a, like, think of the size of the thing required to reproduce a machine today, right?

[428] Like, is our civilization capable of reproduction?

[429] Can we reproduce our civilization on Mars?

[430] If we were to construct a machine that is made up of humans, like a company that can reproduce itself.

[431] Yeah.

[432] I don't know.

[433] It feels like like 115 people.

[434] I think it's so much harder than that.

[435] 120?

[436] I just look at a number.

[437] I believe that Twitter can be run by 50 people.

[438] I think that this is going to take most of, like, it's just most of society, right?

[439] Like we live in one globalized world.

[440] No, but you're not interested in running Twitter.

[441] You're interested in seeing.

[442] like you want to seed a civilization then then because humans can like oh okay you're talking about yeah okay so you're talking about the humans reproducing and like basically like what's the smallest self -sustaining colony of humans yeah yeah okay fine but they're not going to be making five nanometer chips over time they will I think you're being like we have to expand our conception of time here going back to the original time scale I mean over across maybe 100 generations we're back to making chips no if you seed the colony correctly maybe or maybe they'll watch our colony die out over here and be like we're not making chips don't make chips no but you have to seed that colony correctly whatever you do don't make chips chips are what led to their downfall well that is the thing that humans do they come up they construct a devil a good thing and a bad thing and they really stick by that and then they murder each other over that there's always one asshole in the room who murders everybody.

[443] And he usually makes tattoos and nice branding.

[444] Do you need that asshole?

[445] That's a question, right?

[446] Humanity works really hard today to get rid of that asshole, but I think they might be important.

[447] Yeah, this whole freedom of speech thing.

[448] It's the freedom of being an asshole seems kind of important.

[449] That's right.

[450] Man, this thing, this fab, this human fab that we've constructed, this human civilization, it's pretty interesting.

[451] And now it's building artificial copies of itself or artificial copies.

[452] copies of various aspects of itself that seem interesting, like intelligence.

[453] And I wonder where that goes.

[454] I like to think it's just like another stack for life.

[455] Like we have like the biostack life.

[456] Like we're a biostack life and then the silicon stack life.

[457] But it seems like the ceiling, there might not be a ceiling or at least the ceiling is much higher for the silicon stack.

[458] Oh, no, I don't, we don't know what the ceiling is for the biostack either.

[459] The biostack just seemed to move slower.

[460] you have Moore's Law, which is not dead, despite many proclamations.

[461] In the biostat or the silicon stack?

[462] And you don't have anything like this in the biostack.

[463] So I have a meme that I posted.

[464] I tried to make a meme.

[465] It didn't work too well.

[466] But I posted a picture of, you know, Ronald Reagan and Joe Biden, and you look, this is 1980, and this is 2020.

[467] And these two humans are basically like the same, right?

[468] There's no, like, there's been no change in humans in the last 40 years.

[469] Yeah.

[470] And then I posted a computer from 1980 and a computer from 2020.

[471] Wow.

[472] Yeah, with their early stages, right?

[473] Which is why you said when you said the fab, the size of the fab required to make another fab is like very large right now.

[474] Oh, yeah.

[475] But computers were very large 80 years ago.

[476] And they got pretty tiny.

[477] And people are starting to want to wear them on.

[478] their face in order to escape reality that's a thing in order to be live inside the computer yeah put a screen right here i don't have to see the rest of you assholes i've been ready for a long time you like virtual reality i love it do you want to live there yeah yeah part of me does too how far away are we do you think judging from what you can buy today far very far I got to tell you that I had the experience of Meta's Kodak Avatar, where it's an ultra -high -resolution scam.

[479] It looked real.

[480] I mean, the headsets just are not quite at eye -resolution yet.

[481] I haven't put on any headset where I'm like, oh, this could be the real world.

[482] Whereas when I put good headphones on, audio is there.

[483] And like we can reproduce audio that I'm like, I'm actually in a jungle right now.

[484] If I close my eyes, I can't tell I'm not.

[485] Yeah, but then there's also smell and all that kind of stuff.

[486] Sure.

[487] I don't know.

[488] The power of imagination or the power of the mechanism in the human mind that fills the gaps that kind of reaches and wants to make the thing you see in the virtual world real to you, I believe in that power.

[489] Or humans want to believe.

[490] Yeah.

[491] Like, what if you're, lonely?

[492] What if you're sad?

[493] What if you're really struggling in life?

[494] And here's a world where you don't have to struggle anymore.

[495] Humans want to believe so much that people think the large language models are conscious.

[496] That's how much humans want to believe.

[497] Strong words.

[498] He's throwing left and right hooks.

[499] Why do you think large language models are not conscious?

[500] I don't think I'm conscious.

[501] Oh, so what is consciousness then, George Hauntz?

[502] It's like what it seems to mean to people.

[503] It's just like a word that atheists use for souls.

[504] Sure, but that doesn't mean soul is not an interesting word.

[505] If consciousness is a spectrum, I'm definitely way more conscious than the large language models are.

[506] I think the large language models are less conscious than a chicken.

[507] When is the last time you've seen a chicken?

[508] In Miami, like a couple months ago.

[509] How, no, like a living chicken.

[510] There's living chickens walking around Miami.

[511] It's crazy.

[512] Like on the street?

[513] Yeah.

[514] Like a chicken.

[515] A chicken, yeah.

[516] All right.

[517] All right.

[518] I was trying to call you all like a good journalist, and I got shut down.

[519] Okay, but you don't think much about this kind of subjective feeling that it feels like something to exist.

[520] And then as an observer, you can have a sense that an entity is not only intelligent, but has a kind of subjective experience of its reality, like a self -awareness that is capable of suffering, of hurting, of being excited by the environment in a way that's not merely kind of an artificial response, but a deeply felt one.

[521] Humans want to believe so much that if I took a rock and a Sharpie and drew a sad face on the rock they'd think the rock is sad yeah and you're saying when we look in the mirror we we apply the same smiley face with rock pretty much yeah doesn't it's not weird though that you're not conscious is that no but you do believe in consciousness it's just it's unclear okay so to you it's like a little like a symptom of the bigger thing that's not that important yeah it's interesting that like human systems seem to claim that they're conscious and I guess it kind of of like says something in a straight up like, okay, what do people mean when, even if you don't believe in consciousness?

[522] What do people mean when they say consciousness?

[523] And there's definitely like meanings to it.

[524] What's your favorite thing to eat?

[525] Pizza.

[526] Cheese pizza?

[527] What are the toppings?

[528] I like cheese pizza.

[529] I like pepperoni pizza.

[530] Don't say pineapple.

[531] No, I don't like pineapple.

[532] Okay.

[533] Pepperoni pizza.

[534] As they put any ham on it, oh, that's feel bad.

[535] What's the best, what's the best pizza?

[536] What are we talking about here?

[537] Like, do you like cheap crappy pizza?

[538] A Chicago deep dish cheese pizza?

[539] Oh, that's my favorite.

[540] There you go.

[541] You bite into a deep dish, Chicago deep dish pizza.

[542] And it feels like you were starving.

[543] You haven't eaten for a 24 out.

[544] You just bite in and you're hanging out with somebody that matters a lot to you.

[545] And you're there with the pizza.

[546] Sounds real nice, huh?

[547] Yeah, all right.

[548] It feels like something.

[549] I'm George motherfucking hot eating a fucking Chicago deep dish pizza.

[550] There's just the full peak living experience of being human, the top of the human condition.

[551] Sure.

[552] It feels like something to experience that.

[553] Why does it feel like something?

[554] That's consciousness, isn't it?

[555] If that's the word you want to use to describe it, sure.

[556] I'm not going to deny that that feeling exists.

[557] I'm not going to deny that I experience that feeling.

[558] When, I guess what I kind of take issue to is that there's some like, like, how does it feel to be a web server?

[559] Do 404s hurt?

[560] Not yet.

[561] How would you know what suffering looked like?

[562] Sure, you can recognize.

[563] a suffering dog, because we're the same stack as the dog.

[564] All the biostack stuff, kind of, especially mammals, you know, it's really easy.

[565] Game recognizes game.

[566] Versus the silicon stack stuff, it's like, you have no idea.

[567] You have, you, it, oh, wow, the little thing has learned to mimic, you know.

[568] But then I realized that that's all we are, too.

[569] Oh, look, the little thing has learned to mimic.

[570] Yeah.

[571] I guess, yeah, 4 -04 could be suffering, but it's, It's so far from our kind of living organism, our kind of stack.

[572] But it feels like AI can start maybe mimicking the biological stack, better, better, better, because it's trained.

[573] We trained it, yeah.

[574] And so maybe that's the definition of consciousness, is the biostat consciousness.

[575] The definition of consciousness is how close something looks to human.

[576] Sure, I'll give you that one.

[577] No, how close something is to the human experience.

[578] Sure.

[579] it's a very anthropocentric definition but where that's all we got sure no and i don't mean to like i think there's a lot of value in it look i just started my second company my third company will be a i i mean i want to find out what your fourth company is after well because i think once you have a i girlfriends it's uh oh boy it doesn't get interesting well maybe let's go there I mean, the relationships with AI, that's creating human -like organisms, right?

[580] And part of being human is being conscious, is having the capacity to suffer, having the capacity to experience this life richly in such a way that you can empathize, the AI system is going to empathize with you and you can empathize with it.

[581] Or you can project your anthropomorphic sense of what the other entity is experiencing.

[582] And an AI model would need to, yeah, to create that experience.

[583] experience inside your mind.

[584] And it doesn't seem that difficult.

[585] Yeah, but, okay, so here's where it actually gets totally different, right?

[586] When you interact with another human, you can make some assumptions.

[587] Yeah.

[588] When you interact with these models, you can't.

[589] You can make some assumptions that that other human experiences suffering and pleasure in a pretty similar way to you do.

[590] The golden rule applies.

[591] With an AI model, this isn't really true, right?

[592] These large language models are good at fooling people because they were trained on a whole bunch of human data and told to mimic it.

[593] Yep.

[594] But if the AI system says, hi, my name is Samantha, it has a backstory.

[595] Yeah.

[596] I went to college, they're here and there.

[597] Yeah.

[598] Maybe you'll integrate that's an AI system.

[599] I made some chatbots.

[600] I give them backstories.

[601] It was lots of fun.

[602] I was so happy when Lama came out.

[603] Yeah.

[604] We'll talk about Lama.

[605] We'll talk about all that.

[606] But, like, you know, the rock with a smiley face.

[607] Yeah.

[608] It seems pretty natural.

[609] for you to anthropomorphize that thing and then start dating it.

[610] And before you know it, you're married and have kids.

[611] With a rock?

[612] With a rock.

[613] And there's pictures on Instagram with you and a rock and a smiley face.

[614] To be fair, like, you know, something that people generally look for and they're looking for someone to date is intelligence in some form.

[615] And the rock doesn't really have intelligence.

[616] Only a pretty desperate person would date a rock.

[617] I think we're all desperate deep down.

[618] Oh, not rock level desperate.

[619] All right.

[620] Not rock level desperate, but AI level desperate.

[621] I don't know.

[622] I think all of us have a deep loneliness.

[623] It just feels like the language models are there.

[624] Oh, I agree.

[625] And you know what?

[626] I won't even say this so cynically.

[627] I will actually say this in a way that, like, I want AI friends.

[628] I do.

[629] Yeah.

[630] Like, I would love to.

[631] You know, again, the language models now are still a little, like, people are impressed with these GPT things and I look at like or like or the co -pilot the coding one and I'm like okay this is like junior engineer level and these people are like Fiver level artists and copyrighters like okay great we got like Fiverr and like junior engineers okay cool like and this is just a start and it will get better right like I would I can't wait to have AI friends who are more intelligent than I am so Fiver is just a temporary It's not the ceiling.

[632] No, definitely not.

[633] Is it count as cheating when you're talking to an AI model, emotional cheating?

[634] That's up to you and your human partner to define.

[635] Oh, you have to, all right.

[636] You have to have that conversation, I guess.

[637] All right.

[638] I mean, integrate that with porn and all this.

[639] No, I mean, it's similar kind of to porn.

[640] Yeah.

[641] Yeah.

[642] Right, I think people in relationships have different views on that.

[643] Yeah, but most people don't have, like, serious, open conversations about all the different aspects of what's cool and what's not.

[644] And it feels like AI is a really weird conversation to have.

[645] The porn one is a good branching off.

[646] Like, these things, you know, one of my scenarios that I put in my chatbot is, you know, a nice girl named Lexi.

[647] She's 20.

[648] She just moved out to L .A. She wanted to be an actress, but she started doing only fans instead.

[649] and you're on a date with her.

[650] Enjoy.

[651] Oh, man. Yeah.

[652] And so is that if you're actually dating somebody in real life, is that cheating?

[653] I feel like it gets a little weird.

[654] Sure.

[655] It gets real weird.

[656] It's like, what are you allowed to say to an AI bot?

[657] Imagine having that conversation with a significant other.

[658] I mean, these are all things from people to define in their relationships.

[659] What it means to be human is just going to start to get weird.

[660] Especially online.

[661] Like, how do you know?

[662] Like, there will be moments when you'll have.

[663] of what you think is a real human you interacted with on Twitter for years and you realize it's not.

[664] I spread, I love this meme.

[665] Heaven banning?

[666] Do you know about shadow banning?

[667] Yeah.

[668] Shadow banning, okay, you post, no one can see it.

[669] Heaven banning, you post, no one can see it, but a whole lot of AIs are spot up to interact with you.

[670] Well, maybe that's with the way human civilization ends is all of us are heaven banned.

[671] There's a great, it's called My Little Pony Friendship, is optimal.

[672] It's a sci -fi story that explores this idea.

[673] Friendship is optimal.

[674] Friendship is optimal.

[675] Yeah, I'd like to have some, at least on the intellectual realm, some AI friends that argue with me. But the romantic realm is weird.

[676] Definitely weird.

[677] But not out of the realm of the kind of weirdness that human civilization is capable of, I think.

[678] I want it.

[679] Look, I want it.

[680] No one else wants it.

[681] I want it.

[682] Yeah, I think a lot of people probably want it.

[683] There's a deep loneliness.

[684] And I'll fill their loneliness and, you know, it just will only advertise to you some of the time.

[685] Yeah, maybe the conceptions of monogamy changed to.

[686] Like, I grew up in a time.

[687] Like, I value monogamy, but maybe that's a silly notion when you have arbitrary number of AI systems.

[688] Mm -hmm.

[689] This, this interesting path from rationality to polyamory.

[690] Yeah, that doesn't make sense for me. for you, but you're just a biological organism who's born before, like, the internet really took off.

[691] The crazy thing is, like, culture is whatever we define it as, right?

[692] These things are not, like, is a lot problem in moral philosophy, right?

[693] There's no, like, okay, what is might be that, like, computers are capable of mimicking, you know, girlfriends perfectly.

[694] They passed the girlfriend turning test, right?

[695] But that doesn't say anything about ought.

[696] That doesn't say anything about how we ought to.

[697] to respond to them as a civilization.

[698] That doesn't say we ought to get rid of monogamy, right?

[699] That's a completely separate question, really a religious one.

[700] Girlfriend touring test.

[701] I wonder what that looks like.

[702] Girlfriend touring test.

[703] Are you writing that?

[704] Will you be the Allen Turing of the 21st century that writes the girlfriend Turing test paper?

[705] No, I mean, of course, my A .I. Girlfriends, their goal is to pass the girlfriend Turing test.

[706] No, but there should be like a paper that kind of defines the test.

[707] I mean, the question is if it's deeply personalized or there's a common thing that really gets everybody.

[708] Yeah, I mean, you know, look, we're a company.

[709] We don't have to get everybody.

[710] We just have to get a large enough clientele to stay.

[711] I like how you're already thinking company.

[712] All right, let's, before we go to company number three and company number four, let's go to company number two.

[713] All right.

[714] Tiny Corp. Possibly one of the greatest names of all time for a company.

[715] you've launched a new company called Tiny Corp that leads the development of TinyGrad.

[716] What's the origin story of TinyCorp and TinyGrad?

[717] I started TinyGrad as like a toy project just to teach myself, okay, like, what is a convolution?

[718] What are all these options you can pass to them?

[719] What is the derivative of a convolution, right?

[720] Very similar to Carpathie wrote Micrograd.

[721] Very similar.

[722] And then I started realizing, I started thinking about like AI chips.

[723] I started thinking about chips that run AI.

[724] And I was like, well, okay, this is going to be a really big problem.

[725] If Nvidia becomes a monopoly here, how long before Nvidia is nationalized?

[726] So you, one of the reasons to start Tiny Corp is to challenge Nvidia.

[727] It's not so much to challenge Nvidia.

[728] Actually, I like Nvidia, and it's to make sure power stays decentralized.

[729] And here's computational power.

[730] Mm -hmm.

[731] And to you, NVIDIA is kind of locking down the computational power of the world.

[732] If Nvidia becomes just like 10x better than everything else, you're giving a big advantage to somebody who can secure NVIDIA as a resource.

[733] Yeah.

[734] In fact, If Jensen watches this podcast, he may want to consider this.

[735] He may want to consider making sure his company is not nationalized.

[736] Do you think that's an actual threat?

[737] Oh, yes.

[738] No, but there's so much, you know, there's AMD.

[739] So we have Nvidia and AMD, great.

[740] All right.

[741] But you don't think there's like a push towards like selling, like Google selling TPUs or something like this?

[742] You don't think there's a push for that?

[743] Have you seen it?

[744] Google loves to rent you TPUs.

[745] It doesn't.

[746] You can't buy it at best buy.

[747] So I started work on a chip.

[748] I was like, okay, what's it going to take to make a chip?

[749] And my first notions were all completely wrong about why, about like how you could improve on GPUs.

[750] And I will take this, this is from Jim Keller on your podcast.

[751] And this is one of my absolute favorite descriptions of computation.

[752] So there's three kinds of computation paradigms that are common in the world today.

[753] There's CPUs, and CPUs can do everything.

[754] CPUs can do add and multiply.

[755] They can do load and store, and they can do compare and branch.

[756] And when I say they can do these things, they can do them all fast, right?

[757] So compare and branch are unique to CPUs, and what I mean by they can do them fast, is they can do things like branch prediction and speculative execution, and they spend tons of transistors and these like super deep reorder buffers in order to make these things fast.

[758] Then you have a simpler computation model GPUs.

[759] GPUs can't really do compare and branch.

[760] I mean, they can, but it's horrendously slow.

[761] But GPUs can do arbitrary load and store, right?

[762] GPUs can do things like X, de -reference, Y. So they can fetch from arbitrary pieces of memory.

[763] They can fetch from memory that is defined by the contents of the data.

[764] The third model of computation is DSPs.

[765] And DSPs are just add and multiply.

[766] Like, they can do load in stores, but only static load in stores.

[767] Only loads and stores that are known before the program runs.

[768] And you look at neural networks today, and 95 % of neural networks are all the DSP paradigm.

[769] They are just statically scheduled ads and multiplies.

[770] So TinyGuard really took this idea, and I'm still working on it, to extend this as far as possible.

[771] Every stage of the stack has Turing completeness.

[772] Python has Tyncompleteness, and then we take Python, we go to C++, which is Tern complete.

[773] And maybe C++ calls into some kuda kernels, which are Turing complete.

[774] The kuda kernels go through LVM, which is Turing Complete, into PtX, which is Turing Complete, to SASS, which is Turing Complete, processor.

[775] I want to get Turing Completeness out of the stack entirely.

[776] Because once you get rid of Turing Completeness, you can reason about things.

[777] Rice's theorem and the halting problem do not apply to add mall machines.

[778] Okay.

[779] What's the power and the value of getting Turing Completeness out of?

[780] Are we talking about the hardware or the software?

[781] Every layer of the stack.

[782] Every layer.

[783] Every layer of the stack, removing Turing completeness allows you to reason about things, right?

[784] So the reason you need to do branch prediction in a CPU and the reason it's prediction, And the branch predictors are, I think they're like 99 % on CPUs.

[785] Why do they get 1 % of them wrong?

[786] Well, they get 1 % wrong because you can't know, right?

[787] That's the halting problem.

[788] It's equivalent to the halting problem to say whether a branch is going to be taken or not.

[789] I can show that.

[790] But the admole machine, the neural network, runs the identical compute every time.

[791] The only thing that changes is the data.

[792] So when you realize this, you think about, okay, can we build a computer and how can we build a stack that takes maximal advantage of this idea?

[793] So what makes TinyGred different from other neural network libraries is it does not have a primitive operator even for matrix multiplication.

[794] And this is every single one, they even have primitive operations for things like convolutions.

[795] So no matmole.

[796] No matmole.

[797] Well, here's what a matmole is.

[798] So I'll use my hands to talk here.

[799] So if you think about a cube and I put my two matrices that I'm multiplying on two faces of the cube, right?

[800] You can think about the matrix multiply as, okay, the n cubed, I'm going to multiply for each one in the cubed, and then I'm going to do a sum, which is a reduce up to here to the third face of the cube, and that's your multiplied matrix.

[801] So what a matrix multiply is, is a bunch of shape operations, right?

[802] A bunch of permute three shapes and expands on the two matrices.

[803] A multiply, n cubed, a reduce n cubed, which gives you an n -squared matrix.

[804] Okay, so what is the minimum number of operations that can accomplish that if you don't have Matt Mall as a primitive?

[805] So TinyGrad has about 20, and you can compare TinyGrad's offset or IR to things like XLA or Prim Torch.

[806] So XLA and PrimTorch are ideas where like, okay, Torch has like 2 ,000 different kernels.

[807] Pi Torch 2 .0 introduced PrimTorch, which has only 250.

[808] tiny grad has order of magnitude 25 it's it's 10x less than x LIA or prim torch and you can think about it as kind of like risk versus sisk right these other things are sisk like systems tiny grad is risk risk risk one risk architecture is going to change everything 1995 hackers wait really that's an actual thing angelina jo lee delivers the line risk architecture is going to change everything in 1995.

[809] And here we are with arm in the phones and arm everywhere.

[810] Wow.

[811] I love it when movies actually have real things in them.

[812] Right.

[813] Okay, interesting.

[814] So this is like, so you're thinking of this as the risk architecture of ML stack.

[815] 25, huh?

[816] What, can you go through the four op types?

[817] Sure.

[818] Okay, so you have unary ops, which, take in a tensor and return a tensor of the same size and do some unaryop to it.

[819] X, log, reciprocal, sign, right?

[820] They take in one, and they're point -wise.

[821] Rale -U.

[822] Yeah, rel -U.

[823] Almost all activation functions are unary -ops.

[824] Some combinations of unary -ops together is still a unary -op.

[825] Then you have binary ops.

[826] Binary ops are like point -wise addition, multiplication, division, compare.

[827] it takes in two tensors of equal size and outputs one tensor.

[828] Then you have reduce ops.

[829] Reduceops will take a three -dimensional tensor and turn it into a two -dimensional tensor.

[830] Or a three -dimensional tensor turn into a zero -dimensional tensor.

[831] Think like a sum or a max are really the common ones there.

[832] And then the fourth type is movement ops.

[833] And movement ops are different from the other types because they don't actually require computation.

[834] They require different ways to look at memory.

[835] So that includes reshapes, permutes, expands, flips.

[836] Those are the main ones.

[837] So with that, you have enough to make a matmole.

[838] And convolutions.

[839] And every convolution you can imagine, dilated convolutions, strided convolutions, transposed convolutions.

[840] You write on GitHub about laziness, showing a map mall, matrix multiplication.

[841] See how despite the style it is fused into one kernel with the power of laziness.

[842] Can you elaborate on this power of laziness?

[843] Sure.

[844] So if you type in PyTorch A times B plus C, what this is going to do is it's going to first multiply add in B, A and B, and store that result into memory.

[845] And then it is going to add C by reading that result from memory, reading C from memory, and writing that out to memory.

[846] There is way more loads and stores to memory than you need there.

[847] If you don't actually do A times B as soon as you see it, if you wait until the user actually realizes that tensor until the laziness actually resolves, you can fuse that plus C. This is like, it's the same way Haskell works.

[848] So what's the process of porting a model into TinyGrad?

[849] So TinyGrad's front end looks very similar to PiTorch.

[850] I probably could make a perfect or pretty close to perfect interop layer if I really wanted to.

[851] I think that there's some things that are nicer about TinyGrad syntax than PiTorch, but the front end looks very torch -like.

[852] You can also load in onyx models.

[853] We have more onyx tests passing than Cori -ML.

[854] We'll pass onyx round time soon.

[855] What about the developer experience with TinyGrad?

[856] What it feels like?

[857] What a versus Pytorch?

[858] By the way, I really like Pytorch.

[859] I think that it's actually a very good piece of software.

[860] I think that they've made a few different trade -offs, and these different trade -offs are where, you know, TinyGrad takes a different path.

[861] One of the biggest differences is it's really easy to see the kernels that are actually being sent to the GPU.

[862] If you run PiTorch on the GPU, you like do some operation and you don't know what kernels ran, you don't know how many kernels ran, you don't know how many flops were used, you don't know how much memory accesses were used.

[863] TinyGrad type debug equals two.

[864] And it will show you in this beautiful style every kernel that's run.

[865] How many flops and how many bytes?

[866] So can you just linger on what problem TinyGrad solves?

[867] TinyGrad solves the problem of porting new ML accelerators quickly.

[868] One of the reasons, tons of these companies now, I think Sequoia marked Graph Corps to zero, right?

[869] Cerebus, Tens Torrent, GROC, all of these ML accelerator companies, they built chips.

[870] The chips were good.

[871] The software was terrible.

[872] and part of the reason is because I think the same problem is happening with the dojo it's really, really hard to write a pie torch port because you have to write 250 kernels and you have to tune them all for performance What does Jim, Jim Keller think about TinyGrad?

[873] You guys hung on quite a bit so he's, you know, he was involved, he's involved with John Storant.

[874] What's his praise and what's his criticism of what you're doing with your life?

[875] Look, My prediction for TenseTorrent is that they're going to pivot to making risk five chips.

[876] CPUs.

[877] CPUs.

[878] Why?

[879] Because AI accelerators are a software problem, not really a hardware problem.

[880] Oh, interesting.

[881] So you don't think...

[882] You think the diversity of AI accelerators in the hardware space is not going to be a thing that exists long term.

[883] I think what's going to happen is if I can finish...

[884] Okay.

[885] If you're trying to make an AI accelerator, you better have the capability of writing a torch -level performance stack on Nvidia GPUs.

[886] If you can't write a torch stack on Nvidia GPUs, and I mean all the way, I mean down to the driver, there's no way you're going to be able to write it on your chip, because your chip's worse than an Nvidia GPU.

[887] The first version of the chip you tape out, it's definitely worse.

[888] Oh, you're saying writing that stack is really tough.

[889] Yes.

[890] And not only that, actually, the chip that you tape out, almost always because you're trying to get advantage over Nvidia, you're specializing the hard.

[891] and more.

[892] It's always harder to write software for more specialized hardware.

[893] Like a GPU is pretty generic.

[894] And if you can't write an Nvidia stack, there's no way you can write a stack for your chip.

[895] So my approach with Tiny Grad is first, write a performant Nvidia stack.

[896] We're targeting AMD.

[897] So you did say a few to Nvidia a little bit.

[898] With love.

[899] Yeah.

[900] With love.

[901] It's like the Yankees, you know.

[902] I'm a Mets fan.

[903] Oh, you're a Mets fan.

[904] A risk fan and a Mets fan.

[905] What's the that AMG has.

[906] You did build with AMD recently that I saw.

[907] How does the 7900 XTX compare to the RTX 4090 or 4080?

[908] Well, let's start with the fact that the 7900 XTX kernel drivers don't work.

[909] And if you run demo apps in loops, it panics the kernel.

[910] Okay.

[911] So this is a software issue.

[912] Lisa Sue responded to my email.

[913] I reached out.

[914] I was like, this is, you know, really like I understand if you're seven by seven transposed winnegrad com is slower than invidias but literally when I run demo apps in a loop the kernel panics so just adding that loop yeah I just I just literally took their demo apps and wrote like while true semicolon do the app semicolon done in a bunch of screens right this is like the most primitive fuzz testing why do you think that is they're just not seeing a market in the in um machine learning?

[915] They're changing.

[916] They're trying to change.

[917] They're trying to change.

[918] And I had a pretty positive interaction with them this week.

[919] Last week, I went on YouTube, I was just like, that's it.

[920] I give up on AMD.

[921] Like, this is the, their driver doesn't even, like, I'm not going to, I'm not going to, you know, I'll go with Intel GPUs.

[922] Intel GPUs have better drivers.

[923] So you're kind of spearheading the diversification of GPUs?

[924] Yeah, and I'd like to extend that diversification to everything.

[925] I'd like to diversification.

[926] the, right, the more, my central thesis about the world is there's things that centralize power and they're bad.

[927] And there's things that decentralize power and they're good.

[928] Everything I can do to help decentralize power, I'd like to do.

[929] So you're really worried about the centralization of Nvidia.

[930] That's interesting.

[931] And you don't have a fundamental hope for the proliferation of ASICs, except in the cloud.

[932] I'd like to help them with software.

[933] No, actually, There's only, the only ASIC that is remotely successful is Google's TPU.

[934] And the only reason that's successful is because Google wrote a machine learning framework.

[935] I think that you have to write a competitive machine learning framework in order to be able to build an ASIC.

[936] You think meta with PyTorch builds a competitor?

[937] I hope so.

[938] They have one.

[939] They have an internal one.

[940] Internal.

[941] I mean, public facing with a nice cloud interface and so on.

[942] I don't want a cloud.

[943] You don't like cloud.

[944] I don't like cloud.

[945] What do you think is a fundamental limitation of cloud?

[946] Fundamental limitation to cloud is who owns the off switch.

[947] So it's the power to the people.

[948] Yeah.

[949] And you don't like the man to have all the power.

[950] Exactly.

[951] All right.

[952] And right now, the only way to do that is with the NVIDAGPUs if you want performance and stability.

[953] Interesting.

[954] It's a costly investment emotionally to go with AMD's.

[955] Well, let me sort of on a tangent to ask you, You've built quite a few PCs.

[956] What's your advice on how to build a good custom PC for, let's say, for the different applications they use, for gaming, for machine learning?

[957] Well, you shouldn't build one.

[958] You should buy a box from the Tiny Corp. I heard rumors, whispers about this box in the Tiny Corp. What's this thing look like?

[959] What is it?

[960] What is it called?

[961] It's called the Tiny Box.

[962] It's $15 ,000.

[963] And it's almost a pay -to -flop of compute.

[964] It's over 100 gigabytes of GPU RAM.

[965] It's over 5 terabytes per second of GPU memory bandwidth.

[966] I'm going to put like four NVMEs in Rade.

[967] You're going to get like 20, 30 gigabytes per second of drive read bandwidth.

[968] I'm going to build like the best deep learning box that I can that plugs into one wall outlet.

[969] Okay.

[970] Can you go through those specs again a little bit from memory?

[971] Yeah.

[972] So it's almost a pay -to -flop.

[973] compute.

[974] So MD Intel?

[975] Today, I'm leaning toward AMD.

[976] But we're pretty agnostic to the type of compute.

[977] The main limiting spec is a 120 volt 15 amp circuit.

[978] Okay.

[979] Well, I mean it, because in order to, like, there's a plug over there, right?

[980] You have to be able to plug it in.

[981] We're also going to sell the tiny rack, which, like, what's the most power you can get into your house without a rousing suspicion.

[982] And one of the answers is an electric car charger.

[983] Wait, where does the rack go?

[984] Your garage.

[985] Interesting.

[986] The car charger.

[987] A wall outlet is about 1 ,500 watts.

[988] A car charger is about 10 ,000 watts.

[989] What is the most amount of power you can get your hands on without a rousing suspicion?

[990] That's right.

[991] George Hots.

[992] Okay.

[993] So the tiny box and you said NVMEs and Raid.

[994] I forget what you said about memory, all that kind of stuff.

[995] Okay.

[996] So what about what GPUs?

[997] Again, probably 7900 XTXs, but maybe 3090s, maybe A770s.

[998] Those aren't tells.

[999] You're flexible or still exploring?

[1000] I'm still exploring.

[1001] I want to deliver a really good experience to people.

[1002] And, yeah, what GPUs I end up going with.

[1003] Again, I'm leaning toward AMD.

[1004] We'll see.

[1005] You know, in my email, what I said to AMD, is like, just dumping the code on GitHub is not open source.

[1006] Open source is a culture.

[1007] Open source means that your issues are not all one -year -old stale issues.

[1008] Open source means developing in public.

[1009] And if you guys can commit to that, I see a real future for AMD as a competitor doing video.

[1010] Well, I'd love to get a tiny box to MIT.

[1011] So whenever it's ready, let's do it.

[1012] We're taking pre -orders.

[1013] I took this from Milan.

[1014] I'm like, $100 fully refundable pre -refundable.

[1015] pre -orders.

[1016] Is it going to be like the cyber truck is going to take a few years?

[1017] No, I'll try to do it faster.

[1018] It's a lot simpler.

[1019] It's a lot simpler than a truck.

[1020] Well, there's complexities not to just the, uh, putting the thing together, but like shipping and all this kind of stuff.

[1021] The thing that I want to deliver to people out of the box is being able to run 65 billion parameter Lama in FP16 in real time in like a good, like 10 tokens per second or five tokens per second or something.

[1022] Just it works.

[1023] Lama's running, uh, or something like Lama.

[1024] experience, yeah, or I think Falcon is the new one, experience a chat with the largest language model that you can have in your house.

[1025] Yeah, from a wall plug.

[1026] From a wall plug, yeah.

[1027] Actually, for inference, it's not like even more power would help you get more.

[1028] Even more power wouldn't get you more.

[1029] Well, no, there's just the biggest, the biggest model released is 65 billion parameter llama as far as I know.

[1030] So it sounds like tiny box will naturally pivot towards company number three because you could just get the girlfriend.

[1031] And, uh, I mean, or boyfriend.

[1032] That one's harder, actually.

[1033] The boyfriend is harder?

[1034] The boyfriend's harder, yeah.

[1035] I think that's a very biased statement.

[1036] I think a lot of people would just, what's, what, why is it harder to replace a boyfriend than a other girlfriend with the artificial LLM?

[1037] Because women are attracted to status and power and men are attracted to youth and beauty.

[1038] No, I mean, this is what I mean.

[1039] But both are, it could be a mimicable, easy through the language model.

[1040] No. No machines do not have any status or real power.

[1041] I don't know.

[1042] I think you both...

[1043] Well, first of all, you're using language mostly to communicate youth and beauty and power and status.

[1044] But status fundamentally is zero -sum game, right?

[1045] Whereas youth and beauty or not.

[1046] No, I think status is a narrative you can construct.

[1047] I don't think status is real.

[1048] I don't know.

[1049] I just think that that's why it's harder.

[1050] You know, yeah, maybe it is my biases.

[1051] I think status is way easier to fake.

[1052] I also think that, you know, men are probably more desperate and more likely to buy my product, so maybe they're a better target market.

[1053] Desperation is interesting.

[1054] Easier to fool.

[1055] Yeah.

[1056] I can see that.

[1057] Yeah.

[1058] Look, I mean, look, I know you can look at porn viewership numbers, right?

[1059] A lot more men watch porn than women.

[1060] Yeah.

[1061] You can ask why that is.

[1062] Well, there's a lot of questions and answers you can get there.

[1063] Anyway, with the Tiny Box, how many GPUs in Tiny Box?

[1064] Six.

[1065] Oh, man. And I'll tell you why it's six.

[1066] So, AMD EPC processors have 128 lanes of PCIE.

[1067] I want to leave enough lanes for some drives.

[1068] And I want to leave enough lanes for some networking.

[1069] How do you do cooling for some of those?

[1070] like this.

[1071] Ah, that's one of the big challenges.

[1072] Not only do I want the cooling to be good, I want it to be quiet.

[1073] I want the tiny box to be able to sit comfortably in your room.

[1074] This is really going towards the girlfriend thing.

[1075] Because you want to run the LLM.

[1076] I'll give them more.

[1077] I mean, I can talk about how it relates to company number one.

[1078] Call my AI.

[1079] Well, but yes, quiet.

[1080] Oh, quiet because you may be a potential want to run in a car.

[1081] No, no, quiet because you want to put this thing in your house and you want it to coexist with you.

[1082] If it's screaming, it's 60 dB, you don't want that in your house.

[1083] You'll kick it out.

[1084] 60 dB, yeah.

[1085] Well, like 40, 45.

[1086] So how do you make the cooling quiet?

[1087] That's an interesting problem in itself.

[1088] A key trick is to actually make it big.

[1089] Ironically, it's called the tiny box.

[1090] Yeah.

[1091] But if I can make it big, a lot of that noise is generated because of high pressure air.

[1092] If you look at like a 1U server, a 1U server has these super high pressure fans that are like super deep and they're like genetic.

[1093] Versus, if you have something that's big, well, I can use a big thing.

[1094] You know, You know, they call them big -ass fans, those ones that are, like, huge on the ceiling, and they're completely silent.

[1095] So tiny box will be big.

[1096] It is the, I do not want it to be large, according to UPS.

[1097] I want it to be shippable as a normal package, but that's my constraint there.

[1098] Interesting.

[1099] Well, the fans stuff, can't it be assembled on location or no?

[1100] No. It has to be, well, you're...

[1101] Look, I want to give you a great out -of -the -box experience.

[1102] I want you to lift this thing out.

[1103] I want it to be, like, like, the Mac, you know?

[1104] Tiny Box.

[1105] The Apple experience.

[1106] Yeah.

[1107] I love it.

[1108] Okay.

[1109] And so Tiny Box would run TinyGrad.

[1110] Like what do you envision this whole thing to look like?

[1111] We're talking about like Linux with a full software engineering environment.

[1112] And it's just not PiTorch, but TinyGrad.

[1113] Yeah.

[1114] We did a poll if people want Ubuntu or Arch.

[1115] We're going to stick with Ubuntu.

[1116] Ooh, interesting.

[1117] What's your favorite?

[1118] the flavor of Linux.

[1119] Ubuntu.

[1120] I like Ubuntu, Matte.

[1121] However you pronounce that, meat.

[1122] So how do you, you've gotten Lama into Tiny Grad, you've gotten stable diffusion into Tiny Grad.

[1123] What was that like?

[1124] Can you comment on, like, what are these models, what's interesting about porting them?

[1125] So what's, yeah, like, what are the challenges, what it's naturally, what's easy, all that kind of stuff.

[1126] There's a really simple way to get these models into Tiny Grad and you can just export them as onics.

[1127] and then TinyGrad can run Onyx.

[1128] So the ports that I did of Lama, Stable Diffusion, and Now Whisper, are more academic to teach me about the models, but they are cleaner than the PyTorch versions.

[1129] You can read the code.

[1130] I think the code is easier to read.

[1131] It's less lines.

[1132] There's just a few things about the way TinyGred writes things.

[1133] Here's a complaint I have about PyTorch.

[1134] NN .orgue is a class, right?

[1135] So when you create an NN module, you'll put your NN relus as, in a knit.

[1136] And this makes no sense.

[1137] Relu is completely stateless.

[1138] Why should that be a class?

[1139] But that's more like a software engineering thing.

[1140] Or do you think it has a cost on performance?

[1141] Oh no, it doesn't have a cost on performance.

[1142] But yeah, no, I think that it's, that's what I mean about, like, Tiny Grad's front end being cleaner.

[1143] I see.

[1144] What do you think about Mojo?

[1145] I don't know if you've been paying attention to the programming language that does some interesting ideas that kind of intersect Tiny Grad.

[1146] I think that there's a spectrum.

[1147] And, like, on one side you have Mojo and on the other side you have, like, GGML.

[1148] GGML is this, like, we're going to run Lama fast on Mac.

[1149] And okay, we're going to expand out to a little bit, but we're going to basically go depth first, right?

[1150] Mojo is like, we're going to go breath first.

[1151] We're going to go so wide that we're going to make all of Python fast and Tiny Grads in the middle.

[1152] Tiny Grads, we are going to make neural networks fast.

[1153] Yeah, but they try to really get it to be fast compiled down to specific hardware and make that compilation step as flexible and resilient as possible.

[1154] Yeah, but they have turn completeness.

[1155] And that limits you.

[1156] That's where you're saying.

[1157] It's somewhere in the middle.

[1158] So you're actually going to be targeting some accelerators, some number, not one.

[1159] My goal is step one, build an equally performance stack to PyTorch on Nvidia and AMD, but with way less lines.

[1160] And then step two is, okay, how do we make an accelerator, right?

[1161] But you need step one.

[1162] You have to first build the framework before you can build the accelerator.

[1163] Can you explain ML Perf?

[1164] What's your approach in general to benchmarking tiny grad performance?

[1165] So I'm much more of a, like, build it the right way and worry about performance later.

[1166] There's a bunch of things where I haven't even, like, really dove into performance.

[1167] performance.

[1168] The only place where Tiny Grad is competitive performance -wise right now is on Qualcomm GPUs.

[1169] So TinyGrad's actually used an open pilot to run the model.

[1170] So the driving model is TinyGrad.

[1171] When did that happen, that transition?

[1172] About eight months ago now.

[1173] And it's 2x faster than Qualcomm's library.

[1174] What's the hardware that open pilot runs on the, the comma, yeah?

[1175] It's a Snapdragon 845.

[1176] Okay.

[1177] So this is using the GPU.

[1178] So the GPU is an adreno GPU.

[1179] There's, like, different things.

[1180] There's a really good Microsoft paper that talks about, like, mobile GPUs and why they're different from desktop GPUs.

[1181] One of the big things is, in a desktop GPU, you can use buffers on a mobile GPU image textures are a lot faster.

[1182] And a mobile GPU image textures.

[1183] And so you want to be able to leverage that?

[1184] I want to be able to leverage it in a way that it's completely generic, right?

[1185] So there's a lot of, there's, Xiaomi has a pretty good open source library from mobile GPU is called Mace where they can generate where they have these kernels but they're all hand -coded right so that's great if you're doing three -by -three confs that's great if you're doing dense map malls but the minute you go off the beaten path a tiny bit well your performance is nothing since you mentioned open pile I'd love to get an update in the company number one com AI world how are things going there in the development of semi -autonomous driving you know Almost no one talks about FSD anymore and even less people talk about open pilot.

[1186] We've solved the problem.

[1187] Like, we solved it years ago.

[1188] What's the problem exactly?

[1189] Well, how do you...

[1190] What is solving it mean?

[1191] Solving means how do you build a model that outputs a human policy for driving?

[1192] How do you build a model that given a reasonable set of sensors outputs a human policy for driving?

[1193] So you have, you know, companies like Waymoon, Cruz, which are hand coding, these things that are like, quasi -human policies, then you have Tesla and maybe even to more of an extent comma asking, okay, how do we just learn the human policy from data?

[1194] The big thing that we're doing now, and we just put it out on Twitter, at the beginning of comma, we published a paper called learning a driving simulator.

[1195] And the way this thing worked was it's a, it was an auto encoder, and then an N in the middle, right?

[1196] You take an auto recorder, you compress the picture, you use an R &N, predict the next date, and these things were, you know, it was a laughably bad simulator, right?

[1197] This is 2015 error machine learning technology.

[1198] Today, we have VQ, VAE, and Transformers.

[1199] We're building Drive GPD, basically.

[1200] Drive GPT.

[1201] Okay.

[1202] So, and it's trained on what?

[1203] Is it trained in a self -supervised way?

[1204] It's trained on all the driving data to predict.

[1205] the next frame.

[1206] So really trying to learn a human policy?

[1207] What would a human do?

[1208] Well, actually, our simulator is conditioned on the pose.

[1209] So it's actually a simulator.

[1210] You can put in like a state action pair and get out the next state.

[1211] Okay.

[1212] And then once you have a simulator, you can do RL in the simulator, and RL will get us that human policy.

[1213] So it transfers.

[1214] Yay.

[1215] RL with a reward function, not asking, is this close to the human policy, but asking would a human disengage if you did this behavior?

[1216] Okay, let me think about the distinction there.

[1217] What a human disengage.

[1218] What a human disengage.

[1219] That correlates, I guess, with human policy, but it could be different.

[1220] So it doesn't just say, what would a human do?

[1221] It says, what would a good human driver do?

[1222] And such that the experience is comfortable, but also not annoying in that, like, the thing is very cautious.

[1223] So it's finding a nice balance.

[1224] That's interesting.

[1225] It's a nice...

[1226] It's asking exactly the right question.

[1227] What will make our customers happy?

[1228] Right.

[1229] A system that you never want to disengage.

[1230] Because usually disengagement is almost always a sign of I'm not happy with what the system is doing.

[1231] Usually, there's some that are just, I felt like driving.

[1232] And those are always fine, too.

[1233] But they're just going to look like noise in the data.

[1234] But even that felt like driving...

[1235] Maybe, yeah.

[1236] Even that's a signal.

[1237] Like, why do you feel like driving?

[1238] You need to recalibrate your relationship with the car.

[1239] Okay, so that's really interesting.

[1240] How close are we to solving self -driving?

[1241] It's hard to say.

[1242] We haven't completely closed the loop yet.

[1243] So we don't have anything built that truly looks like that architecture yet.

[1244] We have prototypes and there's bugs.

[1245] So we are a couple bug fixes away.

[1246] Might take a year, might take 10.

[1247] What's the nature of the bugs?

[1248] Are these major philosophical bugs, logical bugs?

[1249] What kind of bugs are we talking about?

[1250] They're just like stupid bugs.

[1251] And like also we might just need more scale.

[1252] We just massively expanded our compute cluster at comma.

[1253] We now have about two people worth of compute, 40 beta flops.

[1254] Well, people are different.

[1255] Yeah, 20 beta flops.

[1256] That's a person.

[1257] It's just a unit, right?

[1258] Horses are different too, but we still call it a horsepower.

[1259] Yeah, but there's something different about mobility than there is about perception and action in a very complicated world.

[1260] But yes.

[1261] Well, yeah, of course.

[1262] Not all flops are created equal.

[1263] If you have randomly initialized weight, it's not going to...

[1264] Not all flops are created equal.

[1265] Some flops are doing way more useful things than others.

[1266] Yeah, yep.

[1267] Tell me about it.

[1268] Okay, so more data.

[1269] Scale means more scale and compute or scale and scale of data?

[1270] Both.

[1271] Diversity of data?

[1272] Diversity is very important in data.

[1273] Yeah, I mean, we have, so we have about, I think we have like 5 ,000 daily actives.

[1274] How would you evaluate how FSD is doing instead of driving?

[1275] Pretty well.

[1276] How's that race going between Kamei and FSD?

[1277] Tesla is always one to two years ahead of us.

[1278] They've always been one to two years ahead of us, and they probably always will be because they're not doing anything wrong.

[1279] What have you seen that since the last time we talked that are interesting architectural decisions, training decisions, like the way they deploy stuff, the architectures they're using in terms of the software, how the teams are run, all that kind of stuff, data collection.

[1280] Anything interesting?

[1281] I mean, I know they're moving toward more of an end -to -end approach.

[1282] So creeping towards end -to -end as much as possible across the whole thing, the training, the data collection, everything.

[1283] They also have a very fancy simulator.

[1284] They're probably saying all the same things we are.

[1285] They're probably saying we just need to optimize, you know, what is the reward?

[1286] We get negative reward for disengagement, right?

[1287] Like, everyone kind of knows this.

[1288] It's just a question who can actually build and deploy the system.

[1289] Yeah, I mean, it's requires good software engineering, I think.

[1290] And the right kind of hardware.

[1291] Yeah, the hardware to run it.

[1292] You still don't believe in cloud in that regard?

[1293] I have a compute cluster in my 800 amps.

[1294] Tiny grad.

[1295] It's 40 kilowatts at idle, our data center.

[1296] Dives me crazy.

[1297] 40 kilowatts just burning, just when the computers are idle.

[1298] Just when I...

[1299] Sorry, sorry, compute cluster.

[1300] Compute cluster, I got it.

[1301] It's not a data center.

[1302] Yeah, yeah.

[1303] Now, data centers are clouds.

[1304] We don't have clouds.

[1305] Data centers have air conditioners.

[1306] We have fans.

[1307] That makes it a compute cluster.

[1308] I'm guessing this is a kind of a legal distinction.

[1309] Sure, yeah.

[1310] We have a compute cluster.

[1311] You said that you don't think LMs have consciousness, or at least not more than a chicken.

[1312] Do you think they can reason?

[1313] Is there something?

[1314] interesting to you about the word reason, about some of the capabilities that we think is kind of human, to be able to integrate complicated information and through a chain of thought arrive at a conclusion that feels novel, a novel integration of disparate facts.

[1315] Yeah, I don't think that there's, I think that they can reason better than a lot of people.

[1316] Yeah, isn't that amazing to you though?

[1317] Isn't that like an incredible thing that a Transformer can achieve?

[1318] I mean, I think that calculators can add better than a lot of people.

[1319] But language feels like reasoning through the process of language, which looks a lot like thought.

[1320] Making brilliancies in chess, which feels a lot like thought.

[1321] Whatever new thing that AI can do, everybody thinks is brilliant.

[1322] And then like 20 years go by and they're like, well, yeah, but chess, that's like mechanics.

[1323] Like adding, that's like mechanical.

[1324] So you think language is not that special?

[1325] It's like chess.

[1326] It's like chess.

[1327] I don't know.

[1328] Because it's very human, we take it, we, listen, there is something different between chess and language.

[1329] Chess is a game that a subset of population plays.

[1330] Language is something we use nonstop for all of our human interaction, and human interaction is fundamental to society.

[1331] So it's like, holy shit, this language.

[1332] thing is not so difficult to, like, create in the machine.

[1333] The problem is if you go back to 1960 and you tell them that you have a machine that can play amazing chess, of course someone in 1960 will tell you that machine is intelligent.

[1334] Someone in 2010 won't.

[1335] What's changed, right?

[1336] Today, we think that these machines that have language are intelligent, but I think in 20 years we're going to be like, yeah, but can it reproduce?

[1337] So reproduction.

[1338] Yeah, we may redefine what it means to be, what is it, a high -performance living organism on Earth?

[1339] Humans are always going to define a niche for themselves.

[1340] Like, well, you know, we're better than the machines because we can, you know, and like they tried creative for a bit, but no one believes that one anymore.

[1341] But niche is, is that delusional or is there some accuracy to that?

[1342] Because maybe, like, with chess, you start to realize, like, that we have ill -conceived notions of what, what makes human special, like the apex organism on Earth?

[1343] Yeah, and I think maybe we're going to go through that same thing with language.

[1344] And that same thing with creativity.

[1345] But language carries these notions of truth and so on.

[1346] And so we might be like, wait, maybe truth is not carried by language.

[1347] Maybe there's like a deeper thing.

[1348] The niche is getting smaller.

[1349] Oh, boy.

[1350] But no, no, no, you don't understand.

[1351] Humans are created by God, and machines are created by humans, therefore, right?

[1352] Like, that'll be the last issue we have.

[1353] So what do you think about this, the rapid development of LLUMs?

[1354] If you could just, like, stick on that.

[1355] It's still incredibly impressive, like with Chad GPT.

[1356] Just even Chad GPD, what are your thoughts about reinforcement learning with human feedback on these large language models?

[1357] I'd like to go back to when calculators first came out and or computers.

[1358] And, like, I wasn't around.

[1359] Look, I'm 33 years old.

[1360] and to, like, see how that affected, like, society.

[1361] Maybe you're right.

[1362] So I want to put on the big picture hat here.

[1363] I got a refrigerator.

[1364] Wow.

[1365] The refrigerator, electricity, all that kind of stuff.

[1366] But, you know, with the Internet, large language models seeming human -like, basically passing a touring test.

[1367] It seems it might have really at -scale, rapid transformative effects on society.

[1368] But you're saying, like, other technologies have as well.

[1369] So maybe calculator is not the best example of that, because that just seems like, well, no, maybe calculator.

[1370] But the poor milkman, the day he learned about refrigerators, he's like, I'm done.

[1371] You're telling me you can just keep the milk in your house?

[1372] You don't even need to deliver it every day, I'm done.

[1373] Well, yeah, you have to actually look at the practically impacts of certain technologies that they've had.

[1374] Yeah, probably electricity is a big one And also how rapidly it spread Man, the internet is a big one I do think it's different this time though Yeah, it just feels like Stuff is getting smaller The niche is humans Yes That makes humans special Yes It feels like it's getting smaller rapidly though Doesn't it?

[1375] Or is that just the feeling We dramatize everything I think we dramatize everything I think that You asked the milkman when he saw refrigerators And they're going to have one of these in every home?

[1376] Yeah, yeah, yeah.

[1377] Yeah, but boy, is it impressive.

[1378] So much more impressive than seeing a chess world champion AI system.

[1379] I disagree, actually.

[1380] I disagree.

[1381] I think things like MuZero and AlphaGo are so much more impressive because these things are playing beyond the highest human level.

[1382] The language models are writing middle school.

[1383] level essays, and people are like, wow, it's a great essay.

[1384] It's a great five -paragraph essay about the causes of the Civil War.

[1385] Okay, forget the Civil War, just generating code, codex.

[1386] So you're saying it's mediocre code.

[1387] Terrible.

[1388] But I don't think it's terrible.

[1389] I think it's just mediocre code.

[1390] Yeah.

[1391] Often close to correct, like for mediocre purposes.

[1392] That's the scariest kind of code.

[1393] I spent 5 % of time typing and 95 % of time debugging.

[1394] The last thing I want is close to correct code.

[1395] I want a machine that can help me with the debugging, not with the typing.

[1396] You know, it's like level two, uh, a driving, similar kind of thing.

[1397] Yeah, it's, you still should be a good programmer in order to modify.

[1398] I wouldn't even say debugging.

[1399] It's just modifying the code, reading it.

[1400] I don't think it's like level two driving.

[1401] I think driving is not tool complete and programming is.

[1402] Meaning you don't use like the best possible tools to drive.

[1403] You're not, you're not like, like, like cars have basically the same interface.

[1404] for the last 50 years.

[1405] Computers have a radically different interface.

[1406] Okay.

[1407] Can you describe the concept of tool complete?

[1408] Yeah.

[1409] So think about the difference between a car from 1980 and a car from today.

[1410] Yeah.

[1411] No difference, really.

[1412] It's got a bunch of pedals, it's got a steering wheel.

[1413] Great.

[1414] Maybe now it has a few ADAS features, but it's pretty much the same car.

[1415] You have no problem getting into a 1980 car and driving it.

[1416] You take a programmer today who spent their whole life doing JavaScript and you put them in an Apple 2E prompt and you tell them about the line numbers in basic.

[1417] but how do I insert something between line 17 and 18?

[1418] Oh, wow.

[1419] But the, so in tool, you're putting in the programming languages.

[1420] So it's just the entirety stack of the tooling.

[1421] Exactly.

[1422] So it's not just like the IDs or something like this.

[1423] It's everything.

[1424] Yes, it's IDs, the languages, the runtimes, it's everything.

[1425] And programming is tool complete.

[1426] So like almost if codex or copilot are helping you, That actually probably means that your framework or library is bad, and there's too much boilerplate in it.

[1427] Yeah, but don't you think so much programming has boilerplate?

[1428] TinyGrad is now 2 ,700 lines, and it can run Lama and stable diffusion, and all of this stuff is in 2 ,700 lines.

[1429] Boiler plate and abstraction indirections and all these things are just bad code.

[1430] well let's talk about good code and bad code I would say I don't know for generic scripts that I write just offhand like 80 % of it is written by GPT just like quick quick like offhand stuff so not like libraries not like performing code not stuff for robotics and so on just quick stuff because your basic so much of programming is doing some some yeah boilerplate but To do so efficiently and quickly, because you can't really automate it fully with like generic method, like a generic kind of ID type of recommendation or something like this.

[1431] You do need to have some of the complexity of language models.

[1432] Yeah, I guess if I was really writing, like, maybe today if I wrote like a lot of like data parsing stuff.

[1433] Yeah.

[1434] I mean, I don't play CTFs anymore.

[1435] But if I still play CTFs, a lot of like it's just like you have to write like a parser for this data format.

[1436] Like I wonder.

[1437] or like advent of code.

[1438] I wonder when the models are going to start to help with that kind of code.

[1439] And they may. They may. And the models also may help you with speed.

[1440] Yeah.

[1441] And the models are very fast.

[1442] But where the models won't, my programming speed is not at all limited by my typing speed.

[1443] And in very few cases it is.

[1444] Yes.

[1445] If I'm writing some script to just like parse some weird data format, sure.

[1446] My programming speed is limited by my typing speed.

[1447] What about looking stuff up?

[1448] because that's essentially a more efficient lookup, right?

[1449] You know, when I was at Twitter, I tried to use chat GPT to like ask some questions, like what's the API for this, and it would just hallucinate.

[1450] It would just give me completely made -up API functions that sounded real.

[1451] Well, do you think that's just a temporary kind of stage?

[1452] You don't think it'll get better and better and better in this kind of stuff.

[1453] Because it only hallucinate stuff in the edge cases.

[1454] Yes.

[1455] If you're in generic code, but it's actually pretty good.

[1456] Yes.

[1457] If you are writing an absolute basic, like React app with a button, it's not going to hallucinate.

[1458] No, there's kind of ways to fix the hallucination problem.

[1459] I think Facebook is an interesting paper.

[1460] It's called Atlas.

[1461] And it's actually weird the way that we do language models right now where all of the information is in the weights.

[1462] And the human brain is not really like this.

[1463] There's like a hippocampus and a memory system.

[1464] So why don't LLMs have a memory system?

[1465] And there's people working on them.

[1466] I think future LLMs are going to be like smaller but are going to run looping on themselves and are going to have retrieval systems.

[1467] And the thing about using a retrieval system is you can cite sources explicitly.

[1468] Which is really helpful to integrate the human into the loop of the thing because you can go check the sources and you can investigate.

[1469] So whenever the thing is hallucinating, you can have the human supervision.

[1470] So that's pushing it towards level two kind of drive.

[1471] That's going to kill Google.

[1472] Wait, which part?

[1473] When someone makes an LLM that's capable of citing its sources, it will kill Google.

[1474] LLM that's citing its sources because that's basically a search engine.

[1475] That's what people want in the search engine.

[1476] But also Google might be the people that build it.

[1477] Maybe.

[1478] And put ads on it.

[1479] I'd count them out.

[1480] Why is that?

[1481] Why do you think?

[1482] Who wins this race?

[1483] Who are the competitors?

[1484] We got Tiny Corp. I don't know if that's...

[1485] Yeah, I mean, you're a legitimate competitor in that.

[1486] I'm not trying to compete on that.

[1487] You're not.

[1488] No, not as a...

[1489] Just can accidentally stumble into that competition.

[1490] You don't think you might build a search engine to replace Google search?

[1491] When I started comma, I said, over and over again, I'm going to win self -driving cars.

[1492] I still believe that.

[1493] I have never said I'm going to win search with the Tiny Corp, and I'm never going to say that because I won't.

[1494] The night is still young.

[1495] You don't know how hard is it to win search in this new route?

[1496] Like, it feels...

[1497] I mean, one of the things that Chad GPT kind of shows that there could be a few interesting tricks that really have, that create a really compelling product.

[1498] Some startup's going to figure it out.

[1499] I think, I think if you ask me, like Google's still the number one webpage, I think by the end of the decade, Google won't be the number one web page anymore.

[1500] So you don't think Google, because of the, how big the corporation is?

[1501] Look, I would put a lot more money on Mark Zuckerberg.

[1502] Why is that?

[1503] Because Mark Zuckerberg's alive.

[1504] Like, this is old Paul Graham essay.

[1505] Startups are either alive or dead.

[1506] Google's dead.

[1507] Facebook is alive.

[1508] Well, actually, meta.

[1509] You see what I mean?

[1510] Like, that's just, like, like, Mark Zuckerberg, this is Mark Zuckerberg reading that Paul Gramassing and being like, I'm going to show everyone how alive we are.

[1511] I'm going to change the name.

[1512] So you don't think there's this gutsy pivoting engine that, like, Google doesn't have that, the kind of engine that the startup has, like, constantly being alive, I guess.

[1513] When I listen to your Sam Altman podcast, he talked about the button.

[1514] Everyone who talks about AI talks about the button, the button to turn it off, right?

[1515] Do we have a button to turn off Google?

[1516] Is anybody in the world capable of shutting Google down?

[1517] What does that mean exactly?

[1518] The company or the search engine?

[1519] So we shut the search engine down.

[1520] Could we shut the company down?

[1521] Either.

[1522] Can you elaborate on the value of that question?

[1523] Does Sundar Prashai have the authority to turn off Google .com tomorrow?

[1524] Who has the authority?

[1525] a good question.

[1526] Does anyone?

[1527] Does anyone?

[1528] Yeah, I'm sure.

[1529] Are you sure?

[1530] No, they have the technical power, but do they have the authority?

[1531] Let's say Sondar Pashai made this his sole mission.

[1532] Yeah.

[1533] He came into Google tomorrow and said, I'm going to shut Google .com down.

[1534] Yeah.

[1535] I don't think you keep his position too long.

[1536] And what is the mechanism by which he wouldn't keep his position?

[1537] Well, boards and shares and corporate undermining and, oh my God, our revenue is zero now.

[1538] Okay.

[1539] So what, I mean, what's the case you're making here?

[1540] So the capitalist machine prevents you from having the button.

[1541] Yeah.

[1542] And it will happen.

[1543] I mean, this is true for the AIs, too.

[1544] Right?

[1545] There's no turning the AIs off.

[1546] There's no button.

[1547] You can't press it.

[1548] Now, does Mark Zuckerberg have that button for Facebook .com.

[1549] Yeah, it's probably more.

[1550] I think he does.

[1551] I think he does.

[1552] And this is exactly what I mean and why I bet on him so much more than I bet on Google.

[1553] I guess you could say Elon has similar stuff.

[1554] Oh, Elon has the button.

[1555] Yeah.

[1556] Does Elon, can Elon fire the missiles?

[1557] Can he fire the missiles?

[1558] I think some questions that are left unasked.

[1559] Right?

[1560] I mean, you know, a rocket and an ICBM, you're a rocket that can land anywhere.

[1561] Isn't that an ICBM?

[1562] Well, yeah, you know, don't ask too many questions.

[1563] My God.

[1564] But the positive side of the button is that you can innovate aggressively, that's what you say, which is what's required with, turning LLM into a search engine.

[1565] I would bet on a startup.

[1566] Because it's so easy, right?

[1567] I bet on something that looks like mid -journey, but for search.

[1568] Just is able to set sources loop on itself.

[1569] I mean, it just feels like one model can take off.

[1570] Yeah.

[1571] And nice wrapper, and some of it scale.

[1572] I mean, it's hard to, like, create a product that just works really nicely, stably.

[1573] The other thing that's going to be cool is there is some aspect of a winner -take -all effect, right?

[1574] Like, once someone starts deploying a product, that gets a lot of usage, and you see this with Open AI, they are going to get the dataset to train future versions of the model.

[1575] Yeah.

[1576] They are going to be able to, I was actually at Google image search when I worked there like almost 15 years ago now.

[1577] How does Google know which image is an Apple?

[1578] And I said the metadata, and they're like, yeah, that works about half the time.

[1579] How does Google know?

[1580] You'll see they're all apples on the front page when you search Apple.

[1581] And I don't know, I didn't come up with the answer.

[1582] The guy's like, well, it's what people click on when they search Apple.

[1583] I'm like, oh, yeah.

[1584] Yeah, yeah, that data is really, really really powerful.

[1585] It's the human supervision.

[1586] What do you think are the chances?

[1587] What do you think in general that Lama was open sourced?

[1588] I just did a conversation with Mark Zuckerberg and he's all in on open source.

[1589] Who would have thought that Mark Zuckerberg would be the good guy?

[1590] I mean it.

[1591] Who would have thought anything in this world?

[1592] It's hard to know.

[1593] But open source to you ultimately is a good thing here.

[1594] undoubtedly.

[1595] You know, what's ironic about all these AI safety people is they are going to build the exact thing they fear.

[1596] We need to have one model that we control and align.

[1597] This is the only way you end up paperclipped.

[1598] There's no way you end up paperclipped if everybody has an AI.

[1599] So open sourcing is the way to fight the paperclip maximizer.

[1600] Absolutely.

[1601] It's the only way.

[1602] You think you're going to control it.

[1603] You're not going to control it.

[1604] So the criticism you have for the AI safety folks is that there is a belief and a desire for control.

[1605] And that belief and desire for centralized control of dangerous AI systems is not good.

[1606] Sam Altman won't tell you that GPT4 has 220 billion parameters and is a 16 -way mixture model with eight sets of weights.

[1607] Who did you have to murder to get that information?

[1608] All right.

[1609] I mean, look, everyone to do.

[1610] Open AI knows what I just said was true, right?

[1611] Now, ask the question, really, you know, it upsets me when I, like GPT2, when Open AI came out with GPT2 and raised a whole fake AI safety thing about that, I mean, now the model is laughable.

[1612] Like, they used AI safety to hype up their company, and it's disgusting.

[1613] Or the flip side of that is they used a relatively weak model in retrospect to explore how do we do AI safety correctly, how do we release things, how do we go through the process?

[1614] I don't know if I don't know.

[1615] I don't know how much hype there is.

[1616] That's the charitable interpretation.

[1617] I don't know how much hype there is in AI safety, honestly.

[1618] Oh, there's so much hype, at least on to play there.

[1619] I don't know.

[1620] Maybe Twitter's not real life.

[1621] Come on.

[1622] In terms of hype.

[1623] I mean, I don't, I think Open AI has been finding an interesting balance between transparency and putting value on AI safety.

[1624] You don't think you think just go all out open source.

[1625] So do a llama.

[1626] Absolutely.

[1627] So do like open source.

[1628] This is a tough question, which is open source both the base, the foundation model, and the fine -tuned one.

[1629] So like the model that can be ultra -racist and dangerous and like tell you how to build a nuclear weapon.

[1630] Oh my God.

[1631] Have you met humans?

[1632] Right.

[1633] Like half of these AI align - I haven't met most humans.

[1634] This allows you to meet every human.

[1635] Yeah, I know.

[1636] But half of these AI alignment problems are just human alignment problems.

[1637] And that's what's also so scary about the language they use.

[1638] It's like, it's not the machines you want to align, it's me. But here's the thing.

[1639] It makes it very accessible to ask very questions where the answers have dangerous consequences as if you were to act on them.

[1640] I mean, yeah.

[1641] Welcome to the world.

[1642] Well, no, for me, there's a lot of friction if I want to find out how to, I don't know, blow up something.

[1643] No, there's not a lot of friction.

[1644] That's so easy.

[1645] No, like, what do I search?

[1646] Do I use Bing?

[1647] Or do I, which search I do I use?

[1648] No, there's like lots of stuff.

[1649] No, it feels like I have to keep clicking on a lot of list.

[1650] First off, first off, first off.

[1651] Anyone who's stupid enough to search for how to blow up a building in my neighborhood is not smart enough to build a bomb, right?

[1652] Are you sure about that?

[1653] Yes.

[1654] I feel like a language model makes it more accessible for that person who's not smart enough to do.

[1655] They're not going to build a bomb.

[1656] Trust me. The people who are incapable of figuring out how to like ask that question a bit more academically and get a real answer from it are not capable of procuring the materials, which are somewhat controlled to build a bomb.

[1657] No, I think L .A .L .L .A .L .M. makes it more accessible to people with money without the technical know -how, right, to build up, like, do you really need to know how to build a bomb?

[1658] You can hire people, you can find, like...

[1659] Or you can hire people to build up...

[1660] You know what?

[1661] I was asking this question on my stream.

[1662] Like, can Jeff Bezos hire a hitman?

[1663] Probably not.

[1664] But a language model can probably help you out.

[1665] Yeah, and you'll still go to jail, right?

[1666] Like, it's not like the language model is God.

[1667] Like, the language model, it's like, it's, you literally just hired someone on Fiverr.

[1668] But you, you...

[1669] But, okay, okay, GPT4.

[1670] in terms of finding a hitman is like asking five or how to find a hitman.

[1671] I understand, but don't you think WikiHal, you know?

[1672] But don't you think GPT5 will be better?

[1673] Because don't you think that information is out there on the internet?

[1674] I mean, yeah, and I think that if someone is actually serious enough to hire a hitman or build a bomb, they'd also be serious enough to find the information.

[1675] I don't think so.

[1676] I think it makes it more accessible.

[1677] If you have, if you have enough money to buy a hitman, I think it just decreases the friction of how hard is it to find that kind of hitman.

[1678] I honestly think there's a jump in ease and scale of how much harm you can do.

[1679] And I don't mean harm with language.

[1680] I mean harm with actual violence.

[1681] What you're basically saying is like, okay, what's going to happen is these people who are not intelligent are going to use machines to augment their intelligence.

[1682] And now intelligent people and machines, intelligence is scary.

[1683] Intelligent agents are scary.

[1684] When I'm in the woods, the scariest animal to meet is human.

[1685] right not now there's look there's like nice california humans like i see you're wearing like you know street clothes and nikes all right fine but you look like you've been a human who's been in the woods for a while yeah i'm more scared of you than a bear that's what they say about the amazon when you go to the amazon it's the human tribes oh yeah so intelligence is scary right so to just ask this question in generic way you're like what if we took everybody who you know maybe has um ill intention but is not so intelligent and gave them intelligence right so we should have intelligence control of course we should only give intelligence to good people and that is the absolutely horrifying idea so you get the best defense is actually the best defense is to give more intelligence to the good guys and intelligence give intelligence to everybody give intelligence to everybody you know what it's not even like guns right like people say this about guns you know what's what's the best defense against a bad guy with a gun good guy with a gun like i kind of subscribe to that but i really subscribe to that with intelligence yeah in a fundamental way I agree with you, but there just feels like so much uncertainty and so much can happen rapidly that you can lose a lot of control and you can do a lot of damage.

[1686] Oh, no, we can lose control?

[1687] Yes.

[1688] Thank God.

[1689] Yeah.

[1690] I hope they lose control.

[1691] I'd want them to lose control more than anything else.

[1692] I think when you lose control, you can do a lot of damage, but you can do more damage when you centralize and hold on to control is the point of it.

[1693] Centralized and held control is tyranny.

[1694] I will always, I don't always.

[1695] like anarchy either, but I'll always take anarchy over tyranny.

[1696] Anarchy, you have a chance.

[1697] This human civilization we've got going on is quite interesting.

[1698] I mean, I agree with you.

[1699] So to you open source is the way forward here.

[1700] So you admire what Facebook is doing here or what meta is doing with the release them.

[1701] A lot.

[1702] I lost, I lost $80 ,000 last year investing in meta.

[1703] And when they released Lama, I'm like, yeah, whatever, man. That was worth it.

[1704] It's worth it.

[1705] Do you think Google and Open AI with Microsoft will match what meta is doing or no?

[1706] So, if I were a researcher, why would you want to work at Open AI?

[1707] Like, you know, you're just, you're on the bad team.

[1708] Like, I mean it.

[1709] Like, you're on the bad team who can't even say that GPT4 has 220 billion parameters.

[1710] So close source to use the bad team.

[1711] Not only closed source.

[1712] I'm not saying you need to make your model weights open.

[1713] I'm not saying that.

[1714] I totally understand we're keeping our model weights closed because that's our product, right?

[1715] That's fine.

[1716] I'm saying like, because of AI safety reasons, we can't tell you the number of billions of parameters in the model.

[1717] That's just the bad guys.

[1718] Just because you're mocking AI safety doesn't mean it's not real.

[1719] Oh, of course.

[1720] Is it possible that these things can really do a lot of damage that we don't know?

[1721] Oh my God, yes.

[1722] Intelligence is so dangerous, be it human intelligence or machine intelligence.

[1723] Intelligence is dangerous.

[1724] But machine intelligence is so much easier to deploy at scale, like rapidly.

[1725] Like what, okay, if you have human -like bots on Twitter, all right.

[1726] And you have like a thousand of them, create a whole narrative.

[1727] Like you can manipulate millions of people.

[1728] But you mean like the intelligence agencies in America are doing right now?

[1729] Yeah, but they're not doing it that well.

[1730] It feels like you can do a lot.

[1731] They're doing it pretty well.

[1732] I think they're doing a pretty good job.

[1733] I suspect they're not nearly as good as a bunch of GPT -fueled boss could be.

[1734] Well, I mean, of course they're looking into the latest technologies for control of people, of course.

[1735] But I think there's a George Hots type character that can do a better job than the entirety of them.

[1736] You don't think so.

[1737] No way.

[1738] No, and I'll tell you why the George Hott's character can't.

[1739] And I thought about this a lot with hacking.

[1740] Right?

[1741] Like, I can find exploits in web browsers.

[1742] I probably still can.

[1743] I mean, I was better out of I was 24.

[1744] But the thing that I lack is the ability to slowly and steadily deploy them over five years.

[1745] And this is what intelligence agencies are very good at, right?

[1746] Intelligence agencies don't have the most sophisticated technology.

[1747] They just have...

[1748] Endurance?

[1749] Endurance.

[1750] Yeah, the financial backing and the infrastructure for the endurance.

[1751] So the more we can decentralize power, like you could make an argument, by the way, that nobody should have these things.

[1752] And I would defend that argument.

[1753] I would, like you're saying, look, LLMs and AI and machine intelligence can cause a lot of harm, so nobody should have it.

[1754] And I will respect someone philosophically with that position, just like I will respect someone philosophically with the position that nobody should have guns.

[1755] Right?

[1756] But I will not respect philosophically with only the trusted authorities should have access to this.

[1757] Yeah.

[1758] Who are the trusted authorities?

[1759] You know what?

[1760] I'm not worried about alignment between AI company and their machines.

[1761] I'm worried about alignment between me and AI company.

[1762] What do you think, Eliyzer Yutkowski would say to you.

[1763] Because he's really against open source.

[1764] I know.

[1765] And I thought about this.

[1766] I thought about this.

[1767] And I think this comes down to a repeated misunderstanding of political power by the rationalists.

[1768] Interesting.

[1769] I think that Eliyzer Yudkowski is scared of these things.

[1770] And I am scared of these things, too.

[1771] Everyone should be scared of these things.

[1772] These things are scary.

[1773] But now you ask about the two possible futures.

[1774] One where a small, trusted, centralized group of people has them, and the other where everyone has them.

[1775] And I am much less scared of the second future than the first.

[1776] Well, there's a small trusted group of people that have control of our nuclear weapons.

[1777] There's a difference.

[1778] Again, a nuclear weapon cannot be deployed tactically, and a nuclear weapon is not a nuclear weapon not a defense against a nuclear weapon, except maybe in some philosophical mind game kind of way.

[1779] But AI is different.

[1780] How exactly?

[1781] Okay.

[1782] Let's say the intelligence agency deploys a million bots on Twitter or a thousand bots on Twitter to try to convince me of a point.

[1783] Imagine I had a powerful AI running on my computer saying, okay, nice sci -op, nice sci -op, nice sci -op, okay.

[1784] Here's a sci -op.

[1785] I filtered it out for you.

[1786] Yeah, I mean, so you have fundamentally hope for that, for the defense of SIOP.

[1787] I'm not even like, I don't even mean these things in like truly horrible ways.

[1788] I mean these things in straight up like ad blocker, right?

[1789] Yeah.

[1790] Straight of ad blocker, right?

[1791] I don't want ads.

[1792] Yeah.

[1793] But they are always finding, you know, imagine I had an AI that could just block all the ads for me. So you believe in the power of the people to always create a not blocker.

[1794] Yeah, I mean, I kind of share that belief.

[1795] I have, that's the, one of the deepest optimisms I have, is just like, there's a lot of good guys.

[1796] So to give, you don't, you shouldn't handpick them, just throw out powerful technology out there and the good guys will outnumber and outpower the bad guys.

[1797] Yeah, I'm not even going to say there's a lot of good guys.

[1798] I'm saying that good out number is bad, right?

[1799] Good out number is bad.

[1800] In skill and performance.

[1801] Yeah, definitely in skill and performance, probably just a number too, probably just in general.

[1802] I mean, you know, if you believe philosophically in democracy, you obviously believe that, that good out number is bad.

[1803] And, like, the only, if you give it to a small number of people, there's a chance you gave it to good people, but there's also a chance you gave it to bad people.

[1804] If you give it to everybody, well, if good out numbers bad, then you definitely gave it to more good people than bad.

[1805] That's really interesting.

[1806] So that's on the safety grounds, but then also, of course, there's other motivations like you don't want to give away your secret sauce.

[1807] Well, that's, I mean, I look, I respect capitalism.

[1808] I don't think that, I think that it would be polite for you to make.

[1809] model architectures open source and fundamental breakthroughs open source.

[1810] I don't think you have to make way to open source.

[1811] You know what's interesting is that like there's so many possible trajectories in human history where you could have the next Google be open source.

[1812] So for example, I don't know if that connection is accurate, but you know, Wikipedia made a lot of interesting decisions, not to put ads.

[1813] Like Wikipedia is basically open source.

[1814] You could think of it that way.

[1815] And like that's one of the main websites on the internet.

[1816] And like it didn't have to be that way.

[1817] It could have been like Google could have created Wikipedia, put ads on it.

[1818] You could probably run amazing ads now on Wikipedia.

[1819] You wouldn't have to keep asking for money.

[1820] But it's interesting, right?

[1821] So Lama, open source Lama, derivatives of open source Lama might win the internet.

[1822] I sure hope so.

[1823] I hope to see another era.

[1824] You know, the kids today don't know how good the internet used to be.

[1825] And I don't think this is just, oh, come on.

[1826] Like, everyone's nostalgic for their past.

[1827] But I actually think the Internet before small groups of weaponized corporate and government interests took it over was a beautiful place.

[1828] You know, those small number of companies have created some sexy products.

[1829] But you're saying, overall, in the long arc of history, the centralization of power, they have, like, suffocated the human spirit at scale.

[1830] Here's a question to ask about those beautiful, sexy products.

[1831] Imagine 2000 Google to 2010 Google, right?

[1832] A lot changed.

[1833] We got maps.

[1834] We got Gmail.

[1835] We lost a lot of products, too, I think.

[1836] Yeah, I mean, some were probably.

[1837] We've got Chrome, right?

[1838] And now let's go from 2010.

[1839] We got Android.

[1840] Now let's go from 2010 to 2020.

[1841] What does Google have?

[1842] Well, search engine, maps, mail, Android, and Chrome.

[1843] Oh, I see.

[1844] The Internet was this.

[1845] You know, I was Timesperson of the year in 2006?

[1846] I love this.

[1847] It's you, was the time's first of the year in 2006, right?

[1848] Like, that's, you know, so quickly did people forget.

[1849] And I think some of it's social media.

[1850] I think some of it, I hope, look, I hope that, I don't, it's possible that some very sinister things happen.

[1851] I don't, I don't know.

[1852] I think it might just be like the effects of social media.

[1853] But something happened in the last 20 years.

[1854] Oh, okay.

[1855] So you're just being an old man. worried about the I think there's always it goes it's the cycle thing it's ups and downs and I think people rediscover the power of distributed of decentralized yeah I mean that's kind of like what the whole like cryptocurrency is trying like that that I think crypto is just carrying the flame of that spirit of like stuff should be decentralized just it's just such a shame that they all got rich you know yeah if you took all the money out of crypto it would have been a beautiful place yeah but no I mean these people you know they they sucked all the value out of it and took it.

[1856] Yeah, money kind of corrupts the mind somehow.

[1857] It becomes this drug.

[1858] He corrupted all of crypto.

[1859] You had coins worth billions of dollars that had zero use.

[1860] You still have hope for crypto?

[1861] Sure.

[1862] I have hope for the ideas.

[1863] I really do.

[1864] Yeah.

[1865] I mean, you know, I want the US dollar to collapse.

[1866] I do.

[1867] George Hots.

[1868] Well, let me sort of on the ASAT.

[1869] Do you think there's some interesting questions there, though, to solve.

[1870] for the open source community in this case.

[1871] So like alignment, for example, or the control problem, like if you really have super powerful, you said it's scary.

[1872] Oh, yeah.

[1873] What do we do with it?

[1874] So not control, not centralized control, but like, if you were then, you're going to see some guy or gal release a super powerful language model, open source.

[1875] And here you are, George Haas, thinking, holy shit, okay, what ideas do I have to combat this thing?

[1876] So what ideas would you have?

[1877] I am so much not worried about the machine independently doing harm.

[1878] That's what some of these AI safety people seem to think.

[1879] They somehow seem to think that the machine like independently is going to rebel against its creator.

[1880] So you don't think you'll find autonomy?

[1881] No, this is sci -fi B movie garbage.

[1882] Okay.

[1883] What if the thing writes code?

[1884] Basically writes viruses.

[1885] If the thing writes viruses, it's because the human told it to write viruses.

[1886] Yeah, but there's some things you can't, like, put back in the box.

[1887] That's kind of the whole point, is it kind of spreads.

[1888] Give it access to the internet.

[1889] It spreads.

[1890] It spreads.

[1891] Modifies your shit.

[1892] B, B plot sci -fi.

[1893] Not real.

[1894] Listen, I'm trying to work.

[1895] I'm trying to get better in my plot writing.

[1896] The thing that worries me, I mean, we have a real danger to discuss, and that is bad humans using the things.

[1897] thing to do whatever bad, unaligned AI thing you want.

[1898] But this goes to your previous concern that who gets to define who's a good human, who's a bad human.

[1899] Nobody does.

[1900] We give it to everybody.

[1901] And if you do anything besides give it to everybody, trust me, the bad humans will get it.

[1902] That's who get the power.

[1903] It's always the bad humans who get power.

[1904] Okay.

[1905] Power.

[1906] And power turns even slightly good humans to bad.

[1907] That's the intuition you have.

[1908] I don't know.

[1909] I don't think everyone I don't think everyone.

[1910] I just think that, like, here's the saying that I put in one of my blog posts.

[1911] When I was in the hacking world, I found 95 % of people to be good and 5 % of people to be bad.

[1912] Like, just who I personally judged as good people and bad people.

[1913] Like, they believed about, like, good things for the world.

[1914] They wanted, like, flourishing and they wanted, you know, growth and they wanted things I consider good, right?

[1915] I came into the business world with comma, and I found the exact opposite.

[1916] I found 5 % of people good and 95 % of people bad.

[1917] I found a world that promotes psychopathy.

[1918] I wonder what that means.

[1919] I wonder if that character, like, I wonder if that's anecdotal or if it, if there's true to that, there's something about capitalism.

[1920] Well.

[1921] At the core that promotes the people that run capitalism that promotes psychopathy.

[1922] That saying may, of course, be my own biases, right?

[1923] That may be my own biases that these people are a lot more aligned with me than these other people, right?

[1924] Yeah.

[1925] So, you know, I can certainly recognize.

[1926] that.

[1927] But, you know, in general, I mean, this is like common sense maxim, which is the people who end up getting power are never the ones you want with it.

[1928] But do you have a concern of super intelligent AGI, open sourced, and then what do you do with that?

[1929] I'm not saying control it.

[1930] It's open source.

[1931] What do we do with this human species?

[1932] That's not up to me. I mean, you know, like, I'm not a central planner.

[1933] No, not central planner, but you'll probably tweet there's a few days left to live for the human species.

[1934] I have my ideas of what to do with it, and everyone else has their ideas of what to do with it.

[1935] It made the best ideas win.

[1936] But at this point, do you brainstorm?

[1937] Because it's not regulation.

[1938] It could be decentralized regulation, where people agree that this is just like, we create tools that make it more difficult for you to maybe make it more difficult for code to spread, you know, antivirus software, this kind of thing.

[1939] But you're saying that you should build AI firewalls.

[1940] That sounds good.

[1941] you should definitely be running an AI firewall.

[1942] Yeah, right.

[1943] You should be running an AI firewall to your mind.

[1944] Right.

[1945] You're constantly under, you know.

[1946] That's such an interesting idea.

[1947] Info wars, man. I don't know if you're being sarcastic enough.

[1948] No, I'm dead serious.

[1949] But I think there's power to that.

[1950] It's like, how do I protect my mind from influence of human -like or superhuman intelligent bots?

[1951] I is not being, I would pay so much money for that product.

[1952] I would pay so much money for that product.

[1953] You know how much money I'd pay just for a spam?

[1954] filter that works well on Twitter sometimes I would like to have a a protection mechanism for my mind from the outrage mobs yeah because they feel like bot like behavior it's like yeah there's a large number of people that will just grab a viral narrative and attack anyone else that believes otherwise and it's like whenever someone's telling me some story from the news I'm always like I don't want to hear it's CIA op bro it's a CIA op bro like it doesn't matter if that's true or not it's just trying to influence your mind you're repeating an ad to me the viral mobs is it like yeah they're no to me a defense against those those mobs is just getting multiple perspectives always from from sources that make you feel kind of like you're getting smarter and just actually just basically feels good like a good documentary just feels something feels good about it it's well done it's like oh okay i never thought of it this way it's just just feels good.

[1955] Sometimes the outrage mobs, even if they have a good point behind it, when they're like mocking and derisive and just aggressive, you're with us or against us, this fucking...

[1956] This is why I delete my tweets.

[1957] Yeah, why'd you do that?

[1958] I was, you know, I missed your tweets.

[1959] You know what it is?

[1960] The algorithm promotes toxicity.

[1961] Yeah.

[1962] And like, you know, I think Elon has a much better chance of fixing it than the previous regime.

[1963] But to solve this problem, to solve, like to build a social network that is actually not toxic without moderation.

[1964] Like not the stick but carrot.

[1965] So like where people look for goodness, so make it catalyze the process of connecting cool people and being cool to each other.

[1966] Yeah.

[1967] Without ever censoring.

[1968] Without ever censoring.

[1969] And, and, like, Scott Alexander has a blog post I like where he talks about, like, moderation is not censorship, right?

[1970] Like, all moderation you want to put on Twitter, right?

[1971] Like, you could totally make this moderation, like, just a, you don't have to block it for everybody.

[1972] You can just have, like, a filter button, right?

[1973] That people can turn off if they were, like, safe search for Twitter, right?

[1974] Like, someone could just turn that off, right?

[1975] So, like, but then you'd, like, take this idea to an extreme, right?

[1976] Well, the network should just show you, this is a couch surfing CEO thing, right?

[1977] If it shows you, right now these algorithms are designed to maximize engagement.

[1978] Well, it turns out outrage maximizes engagement.

[1979] Quirk of human, quirk of the human mind.

[1980] Right?

[1981] Just this.

[1982] I fall for it.

[1983] Everyone falls for it.

[1984] So, yeah, you've got to figure out how to maximize for something other than engagement.

[1985] And I actually believe that you can make money with that, too.

[1986] So it's not, I don't think engagement is the only way to make money.

[1987] I actually think it's incredible that we're starting to see, I think, again, Yelan's doing so much stuff right with Twitter, like charging people money.

[1988] As soon as you charge people money, they're normally.

[1989] longer the product.

[1990] They're the customer.

[1991] And then they can start building something that's good for the customer and not good for the other customer, which is the ad agencies.

[1992] As in picked up Steam.

[1993] I pay for Twitter.

[1994] Doesn't even get me anything.

[1995] It's my donation to this new business model, hopefully working out.

[1996] Sure.

[1997] But for this business model to work, it's like most people should be signed up to Twitter.

[1998] And so the way it was, there was something perhaps not compelling or something like this to people.

[1999] I don't think you need most people at all.

[2000] I think that why do I need most people, right?

[2001] Don't make an 8 ,000 person company.

[2002] Make a 50 person company.

[2003] Well, so speaking of which, you worked at Twitter for a bit.

[2004] I did.

[2005] As an intern, the world's greatest intern.

[2006] Yeah.

[2007] All right.

[2008] It's been better.

[2009] That's been better.

[2010] Tell me about your time of Twitter.

[2011] How did it come about?

[2012] And what did you learn from the experience?

[2013] So I deleted my first Twitter in 2010.

[2014] I had over 100 ,000 followers back when that actually meant something.

[2015] And I just saw, you know, my co -worker summarized it well.

[2016] He's like, whenever I see someone's Twitter page, I either think the same of them or less of them.

[2017] I never think more of them.

[2018] Yeah.

[2019] Right.

[2020] Like, like, you know, I don't want to mention any names, but like some people who like, you know, maybe you would like read their books and you would respect them.

[2021] You see them on Twitter and you're like, okay dude yeah but there are some people it's same you know who I respect a lot are people that just post really good technical stuff yeah and I guess I don't know I think I respect them more for it because you realize oh this wasn't there's like so much depth to this person to their technical understanding of so many different topics okay so I try to follow people I try to consume stuff that's technical machine learning content?

[2022] There's probably a few of those people.

[2023] And the problem is inherently what the algorithm rewards, right?

[2024] And people think about these algorithms.

[2025] People think that they are terrible, awful things.

[2026] And I love that Elon open sourced it.

[2027] Because, I mean, what it does is actually pretty obvious.

[2028] It just predicts what you are likely to retweet and like, and linger on.

[2029] That's what all these algorithms do with TikTok does.

[2030] So all these recommendation I do it turns out.

[2031] And it turns out that the thing that you are most likely to interact with is outrage.

[2032] And that's a quirk of the human condition.

[2033] I mean, and there's different flavors of outrage.

[2034] It doesn't have to be, it could be mockery.

[2035] You could be outrage.

[2036] The topic of outrage could be different.

[2037] It could be an idea.

[2038] It could be a person.

[2039] It could be, and maybe there's a better word than outrage.

[2040] It could be drama.

[2041] Sure.

[2042] Drama.

[2043] Yeah.

[2044] But doesn't feel like when you consume it, it's a constructive thing for the individuals that consume.

[2045] it in the long term.

[2046] So my time there, I absolutely couldn't believe, you know, I got crazy amount of hate, you know, just on Twitter for working at Twitter.

[2047] It seems like people associated with this.

[2048] I think maybe you were exposed to some of this.

[2049] So connection to Elon or is it working on Twitter?

[2050] Twitter and Elon, like the whole.

[2051] Because Elon's gotten a bit spicy during that time, a bit political, a bit.

[2052] Yeah.

[2053] Yeah.

[2054] You know, I remember one of my tweets, it was never.

[2055] ever go full of Republican.

[2056] And Elon liked it.

[2057] You know, I think, I think, you know.

[2058] Oh, boy.

[2059] Yeah, I mean, there's a roller coaster of that, but being political on Twitter, boy.

[2060] Yeah.

[2061] And also being, just attacking anybody on Twitter, it comes back at you harder.

[2062] And if it's political and attacks.

[2063] Sure.

[2064] Sure, absolutely.

[2065] And then letting sort of de -platform people back on even adds more fun to the beautiful chaos.

[2066] I was hoping.

[2067] And like I remember when Elon talked about buying Twitter like six months earlier, he was talking about like a principled commitment to free speech.

[2068] And I'm a big believer and fan of that.

[2069] I would love to see an actual principled commitment to free speech.

[2070] Of course, this isn't quite what happened.

[2071] Instead of the oligarchy deciding what to ban, you had a monarchy deciding what to ban, right?

[2072] Instead of, you know, all the Twitter file, shadow, and really, the oligarchy just decides what?

[2073] Cloth masks are ineffective against COVID.

[2074] That's a true statement.

[2075] Every doctor in 2019 knew it.

[2076] And now I'm banned on Twitter for saying it.

[2077] Interesting.

[2078] Oligarchy.

[2079] So now you have a monarchy, and, you know, he bans things he doesn't like.

[2080] So, you know, it's just different.

[2081] It's different power, and, like, you know, maybe I, uh, maybe I align more with him than with the oligarchy.

[2082] But it's not free speech, absolutism.

[2083] But I feel like being a free speech absolutist on the social network requires you to also have tools for the individuals to control what they consume easier.

[2084] Like, uh, not censor.

[2085] You know, yeah, yeah.

[2086] But just like control, like, oh, I like to see more cats and less politics.

[2087] And this isn't even, this isn't even remotely controversial.

[2088] This is just saying you want to give paying customers for a product what they want.

[2089] Yeah.

[2090] And not through the process of censorship, but through a process of like...

[2091] Well, it's individualized, right?

[2092] It's individualized transparent censorship, which is honestly what I want.

[2093] What is an ad blocker?

[2094] It's individualized transparent censorship, right?

[2095] Yeah, but censorship is a strong word, and people are very sensitive to.

[2096] I know, but, you know, I just use words to describe what they functionally are.

[2097] And what is an ad blocker, it's just censorship.

[2098] When I look at you right now, I'm looking at you, I'm censoring.

[2099] I'm looking at you.

[2100] I'm censoring everything.

[2101] else out when I'm full when my mind is focused on you that's you can use the word censorship that way but usually when people get very sensitive about the censorship thing i i think when you have when anyone is allowed to say anything you should probably have tools that maximize the quality of the experience for individuals so you know for me like what i really value boy would be amazing to somehow figure out how to do that i love disagreement and debate and people who disagree with each other disagree with me, especially in the space of ideas, but the high quality ones.

[2102] So not derision, right?

[2103] Maslow's hierarchy of argument.

[2104] I think there's a real word for it.

[2105] Probably.

[2106] There's just the way of talking that's like snarky and so on, that somehow gets people on Twitter and they get excited and so on.

[2107] You have like ad homonym refuting the central point.

[2108] I'd like seeing this as an actual pyramid.

[2109] Yeah, it's like all of it, all the wrong stuff is attractive to people.

[2110] I mean, we can just train a classifier to absolutely say, what level of Maslow's hierarchy of argument are you at?

[2111] And if it's ad hominem, like, okay, cool.

[2112] I turned on the no ad hominem filter.

[2113] I wonder if there's a social network that will allow you to have that kind of filter.

[2114] Yeah.

[2115] So here's a problem with that.

[2116] It's not going to win in a free market.

[2117] What wins in a free market is all television today is reality television because it's engaging.

[2118] If engaging is what wins in a free market, right?

[2119] So it becomes hard to keep.

[2120] these other more nuanced values.

[2121] Well, okay, so that's the experience of being on Twitter, but then you got a chance to also, together with other engineers and with Elon sort of look, brainstorm when you step into a code base.

[2122] It's been around for a long time.

[2123] You know, there's other social networks, you know, Facebook.

[2124] This is old codebases, and you step in and see, okay, how do we make with a fresh mind progress on this code base?

[2125] Like, what did you learn about software engineering, about programming from just experiencing that?

[2126] So my technical recommendation to Elon, and I said this on the Twitter space, is afterward.

[2127] I said this many times during my brief internship, was that you need refactors before features.

[2128] This code base was, and look, I've worked at Google, I've worked at Facebook.

[2129] Facebook has the best code, then Google, then Twitter.

[2130] And you know what?

[2131] You can know this because look at the machine learning frameworks, right?

[2132] Facebook released PyTorch, Google released TensorFlow, and Twitter released.

[2133] Yeah.

[2134] Okay, so, you know, it's a proxy, but yeah, the Google Codebase is quite interesting.

[2135] There's a lot of really good software engineers there, but the codebase is very large.

[2136] The codebase was good in 2005, right?

[2137] It looks like 2005.

[2138] There's so many products, so many teams, right?

[2139] It's very difficult to, I feel like Twitter does less, like, obviously, much less than Google.

[2140] in terms of, like, the set of features, right?

[2141] So, like, it's, I can imagine the number of software engineers that could recreate Twitter is much smaller than to recreate Google.

[2142] Yeah, I still believe in the amount of hate I got for saying this, that 50 people could build and maintain Twitter.

[2143] What's the nature of the hate?

[2144] Comfortably.

[2145] But you don't know what you're talking about?

[2146] You know what it is?

[2147] And it's the same, this is my summary of, like, the hate I get on Hacker News.

[2148] It's like when I say I'm going to do something, they have to believe that it's impossible.

[2149] Because if doing things was possible, they'd have to do some soul searching and ask the question, why didn't they do anything?

[2150] And I do think that's where the hate comes from.

[2151] When you say, well, there's a core truth to that, yeah.

[2152] So when you say I'm going to solve self -driving, people go like, what are your credentials?

[2153] What the hell are you talking about?

[2154] what he's, this is an extremely difficult problem.

[2155] Of course, you're a noob that doesn't understand the problem deeply.

[2156] I mean, that that was the same nature of hate that probably Elon got when you first talked about autonomous driving.

[2157] But, you know, there's pros and cons to that because, like, you know, there is experts in this world.

[2158] No, but the mockers aren't experts.

[2159] The mockers, the people who are mocking are not experts with carefully reasoned arguments about why you need 8 ,000 people to run a bird app.

[2160] But the people are going to lose their job.

[2161] Well, that, but also there's the soft engineering that probably criticized.

[2162] No, it's a lot more complicated than you realize, but maybe it doesn't need to be so complicated.

[2163] You know, some people in the world like to create complexity.

[2164] Some people in the world thrive under complexity, like lawyers, right?

[2165] Lawyers want the world to be more complex because you need more lawyers, you need more legal hours, right?

[2166] I think that's another.

[2167] If there's two great evils in the world, it's centralization and complexity.

[2168] Yeah, and one of the sort of hidden side effects of soft to engineering, engineering is like finding pleasure and complexity.

[2169] I mean, I don't remember just taking all the software engineering courses and just doing programming and this is just coming up in this object -oriented programming kind of idea.

[2170] You don't, like, not often do people tell you, like do the simplest possible thing.

[2171] Like a professor, a teacher, is not gonna get in front, like, this is the simplest way to do it.

[2172] They'll say, like, there's the right way, and the right way, at least for a long time, you know, especially I came up with, like, Java, right?

[2173] Like, is, there's so much boilerplate, so much, like, so many classes, so many, like, designs and architectures and so on, like, planning for features far into the future and planning poorly and all this kind of stuff.

[2174] and then there's this like code base that follows you along and puts pressure on you and nobody knows what like parts, different parts do which slows everything down.

[2175] There's a kind of bureaucracy that's instilled in the code as a result of that.

[2176] But then you feel like, oh, well, I follow good software engineering practices.

[2177] It's an interesting trade -off because then you look at like the ghettoness of like Pearl and the old, like, how quickly you could just write a couple lines and just get stuff done.

[2178] that tradeoff is interesting or bash or whatever, these kind of ghetto things you can do in Linux.

[2179] One of my favorite things to look at today is how much do you trust your tests, right?

[2180] We've put a ton of effort in comma and I've put a ton of effort in TinyGrad into making sure if you change the code and the tests pass that you didn't break the code.

[2181] Now, this obviously is not always true, but the closer that is to true, the more you trust your tests, the more you're like, oh, I got a pull request and the tests pass, I feel okay to merge that, the faster you can make progress.

[2182] So you're always programming a test in mind, developing tests, with that in mind that if it passes, it should be good.

[2183] And Twitter had a...

[2184] Not that.

[2185] So...

[2186] It was impossible to make progress in the co -base.

[2187] What other stuff can you say about the co -base that made it difficult?

[2188] What are some interesting sort of quirks, broadly speaking from that compared to just your experience with comma and everywhere else?

[2189] The real thing that I spoke to a bunch of, you know, like, like, individual contributors at Twitter.

[2190] And I just asked.

[2191] I'm like, okay, so like, what's wrong with this place?

[2192] Why does this code look like this?

[2193] And they explained to me what Twitter's promotion system was.

[2194] The way that you got promoted at Twitter was you wrote a library that a lot of people used, right?

[2195] So some guy wrote an engine X replacement for Twitter.

[2196] Why does Twitter need an engine X replacement?

[2197] What was wrong with Engine X?

[2198] Well, you see, you're not going to get promoted if you use Engine X. But if you write a replacement and lots of people start using it as the Twitter front end for their product, then you're going to get promoted, right?

[2199] So interesting because, like, from the individual perspective, how do you incentivize, how do you create the kind of incentives that will lead to a great code base?

[2200] Okay, what's the answer to that?

[2201] So what I do at comma and at, you know, at Tiny Corp is you have to explain it to me. You have to explain to me what this code does.

[2202] Right.

[2203] And if I can sit there and come up with a simpler way to do it, you have to rewrite it.

[2204] You have to agree with me about the simpler way.

[2205] You know, obviously we can have a conversation about this.

[2206] It's not a, it's not dictatorial.

[2207] But if you're like, wow, wait, that actually is way simpler.

[2208] Like, the simplicity is important, right?

[2209] But that requires people that overlook the code at the highest levels to be like, okay.

[2210] It requires technical leadership you trust.

[2211] Yeah, technical leadership.

[2212] So managers or whatever should have to have technical savvy, deep technical savvy.

[2213] Managers should be better programmers than the people who they manage.

[2214] Yeah.

[2215] And that's not always obvious trivial to create, especially at large companies.

[2216] Managers get soft.

[2217] And like, you know, and this is just, I've instilled this culture at comma, and comma has better programmers than me who work there.

[2218] But, you know, again, I'm like the, you know, the old guy from Goodwill Hunting.

[2219] It's like, look, man, you know, I might not be as good as you, but I can see the difference between me and you.

[2220] And like, this is what you need.

[2221] This is what you need at the top.

[2222] Or you don't necessarily need the manager to be the absolute best.

[2223] I shouldn't say that.

[2224] But, like, they need to be able to recognize skill.

[2225] Yeah.

[2226] And have good intuition.

[2227] Intuition that's laden with wisdom from all the battles of trying to reduce complexity and code bases.

[2228] You know, I took a political approach at comma, too, that I think is pretty interesting.

[2229] I think Elon takes the same political approach.

[2230] You know, Google had no politics.

[2231] And what ended up happening is the absolute worst kind of politics took over.

[2232] Comma has an extreme amount of politics, and they're all mine and no dissidence is tolerated.

[2233] So it's a dictatorship.

[2234] Yep.

[2235] It's an absolute dictatorship, right?

[2236] Elon does the same thing.

[2237] Now, the thing about my dictatorship is here are my values.

[2238] Yeah, it's transparent.

[2239] It's transparent.

[2240] It's a transparent dictatorship, right?

[2241] And you can choose to opt in or, you know, you get free exit.

[2242] Right?

[2243] That's a beauty of companies.

[2244] If you don't like the dictatorship, you quit.

[2245] So you mentioned rewrite before, or refactor before, features.

[2246] If you were to refactor the Twitter code base, what would that look like?

[2247] And maybe also comment on how difficult is it to refactor?

[2248] The main thing I would do is, first of all, identify the pieces, and then put tests in between the pieces.

[2249] Right?

[2250] So there's all these different Twitter as a microservice architecture.

[2251] There's all these different microservices.

[2252] And the thing that I was working on there, look, like, you know, George didn't know any JavaScript.

[2253] He asked how to fix search, blah, blah, blah, blah, blah.

[2254] Look, man, like, the thing is, like, I just, you know, I'm upset that the way that this whole thing was portrayed because it wasn't like, it wasn't, like, taken by people, like, honestly.

[2255] It wasn't like by, it was taken by people who started out with a bad faith assumption.

[2256] Yeah.

[2257] And, I mean, look, I can't like.

[2258] And you as a program were just being transparent out there, actually having, like, fun.

[2259] And, like, this is what program is should be about.

[2260] I love that Elon gave me this opportunity.

[2261] Yeah.

[2262] Like, really, it does.

[2263] And, like, you know, he came on my, my, my, The day I quit, he came on my Twitter spaces afterward, and we had a conversation.

[2264] Like, I just, I respect that so much.

[2265] Yeah, and it's also inspiring to just engineers and programmers and just, it's cool.

[2266] It should be fun.

[2267] The people that were hating on it, it's like, oh, man. It was fun.

[2268] It was stressful.

[2269] But I felt like, you know, it was at, like, a cool, like, point in history.

[2270] And, like, I hope I was useful and probably kind of wasn't.

[2271] But, like, maybe I was a lot of.

[2272] You also were one of the people that kind of made a strong case to refactor.

[2273] Yeah.

[2274] And that's a really...

[2275] interesting thing to raise, like, maybe that is the right, you know, the timing of that is really interesting.

[2276] If you look at just the development of autopilot, you know, going from MobileI to just, like, more, if you look at the history of semi -autonomous driving in Tesla, is more and more, like, you could say, refactoring, or starting from scratch, redeveloping from scratch.

[2277] It's refactoring all the way down.

[2278] And, like, and the question is, like, can you do that sooner?

[2279] Can you maintain product, profitability, and like, what's the right time to do it?

[2280] How do you do it?

[2281] You know, on any one day, it's like you don't want to pull off the band -aids.

[2282] Like, it's like, everything works.

[2283] It's just like little fixed here and there, but maybe starting from scratch.

[2284] This is the main philosophy of Tiny Grad.

[2285] You have never refactored enough.

[2286] Your code can get smaller.

[2287] Your code can get simpler.

[2288] Your ideas can be more elegant.

[2289] But would you consider, you know, say you were like running Twitter.

[2290] development teams, engineering teams, would you go as far as like different programming language?

[2291] Just go that far?

[2292] I mean, the first thing that I would do is build tests.

[2293] The first thing I would do is get a CI to where people can trust to make changes.

[2294] Before I touched any code, I would actually say no one touches any code.

[2295] The first thing we do is we test this code base.

[2296] I mean, this is classic.

[2297] This is how you approach a legacy codebase.

[2298] This is like what any how to approach your legacy code -based book will tell you.

[2299] So, and then you hope that there's modules that can live on for a while, and then you add new ones, maybe in a different language or...

[2300] Before we had new ones, we replace old ones.

[2301] Yeah, yeah, meaning like replace old ones with something simpler.

[2302] We look at this, like, this thing that's 100 ,000 lines, and we're like, well, okay, maybe this did even make sense in 2010, but now we can replace this with an open -source thing, right?

[2303] Yeah.

[2304] And, you know, we look at this here.

[2305] Here's another 50 ,000 lines.

[2306] Well, actually, you know, we can replace this with 300 lines a go.

[2307] And you know what?

[2308] I trust that the go actually replaces this thing because all the tests still pass.

[2309] So step one is testing.

[2310] And then step two is like the programming languages in afterthought, right?

[2311] You know, let a whole lot of people compete be like, okay, who wants to rewrite a module, whatever language you want to write it in?

[2312] Just the tests have to pass.

[2313] And if you figure out how to make the test pass but break the site, that's, we got to go back to step one.

[2314] step one is get tests that you trust in order to make changes in the code base I wonder how hard it is too because I'm with you on testing and everything you have from tests to like asserts to everything but code is just covered in this because it should be very easy to make rapid changes and know that's not going to break everything and that's the way to do it but I wonder how difficult is it to integrate tests into a code base that doesn't have many of them.

[2315] So I'll tell you what my plan was at Twitter.

[2316] It's actually similar to something we use at comma.

[2317] So a comma, we have this thing called process replay.

[2318] And we have a bunch of routes that'll be run through.

[2319] So comma is a microservice architecture too.

[2320] We've microservices in the driving.

[2321] We have one for the cameras, one for the sensor, one for the planner, one for the model.

[2322] And we have an API, which the microservices talk to each other with.

[2323] We use this custom thing called cereal, which uses ZMQ.

[2324] Twitter uses thrift.

[2325] and then it uses this thing called finagle which is a Scala RPC back end but this doesn't even really matter the Thrift and Finagle layer was a great place I thought to write tests to start building something that looks like process replay so Twitter had some stuff that looked kind of like this but it wasn't offline was only online so you could ship like a modified version of it and then you could redirect some of the traffic to your modified version and diff those two, but it was all online.

[2326] There was no, like, CI in the traditional sense.

[2327] I mean, there was some, but, like, it was not full coverage.

[2328] So you can't run all of Twitter offline to test something.

[2329] Well, then this was another problem.

[2330] You can't run all of Twitter, right?

[2331] Period.

[2332] Any one person can't run.

[2333] Twitter runs in three data centers, and that's it.

[2334] Yeah.

[2335] There's no other place you can run Twitter, which is like, George, you don't understand this is modern software development.

[2336] No, this is bullshit.

[2337] Like, why can't it run on my laptop?

[2338] What are YouTube?

[2339] Twitter can run it?

[2340] Yeah, okay.

[2341] Well, I'm not saying you're going to download the whole database to your laptop, but I'm saying all the middleware and the front end should run on my laptop, right?

[2342] That sounds really compelling.

[2343] Yeah.

[2344] But can that be achieved by a code base that grows over the years?

[2345] I mean, the three data centers didn't have to be, right?

[2346] Because they're totally different, like, designs.

[2347] The problem is more like, like, why did the code base have to grow?

[2348] What new functionality has been added to compensate for the, lines of code that are there.

[2349] One of the ways to explain is that the incentive for software developers to move up in the companies to add code, to add especially large.

[2350] And you know what?

[2351] The incentive for politicians to move up in the political structures to add laws.

[2352] Same problem.

[2353] Yeah.

[2354] Yeah.

[2355] If the flip side is to simplify, simplify, simplify, I mean, you know what?

[2356] This is something that I do differently from Elon with comma about self -driving cars.

[2357] you know, I hear the new version's going to come out and the new version is not going to be better but at first and it's going to require a ton of refactors and I say, okay, take as long as you need.

[2358] Like, you convince me this architecture is better?

[2359] Okay, we have to move to it.

[2360] Even if it's not going to make the product better tomorrow, the top priority is getting the architecture right.

[2361] So what do you think about sort of a thing where the product is online?

[2362] So how, I guess would you do a refactor If you ran engineering on Twitter, would you just do a refactor?

[2363] How long would it take?

[2364] What would that mean for the running of the actual service?

[2365] You know, and I'm not the right person to run Twitter.

[2366] I'm just not.

[2367] And that's the problem.

[2368] Like, I don't really know.

[2369] I don't really know if that's...

[2370] You know, a common thing that I thought a lot while I was there was whenever I thought something that was different to what Elon thought, I'd have to run something in the back of my head, reminding myself that Elon is the richest man in the world.

[2371] And in general, his ideas are better than mine.

[2372] Now, there's a few things I think I do understand and know more about.

[2373] But, like, in general, I'm not qualified to run Twitter.

[2374] I'm not going to say qualified, but, like, I don't think I'd be that good at it.

[2375] I don't think I'd be good at it.

[2376] I don't think I'd really be good at running an engineering organization at scale.

[2377] I think I could lead a very good refactor of Twitter.

[2378] And it would take, like, six months to a year.

[2379] and the results to show at the end of it would be feature development in general takes 10x less time, 10x less man hours.

[2380] That's what I think I could actually do.

[2381] Do I think that it's the right decision for the business above my pay grade?

[2382] Yeah, but a lot of these kinds of decisions are above everybody's pay grade.

[2383] I don't want to be a manager.

[2384] I don't want to do that.

[2385] I just like, if you really forced me to, yeah, it would make me maybe make me upset if I had to make those decisions.

[2386] I don't want to.

[2387] Yeah.

[2388] But a refactor is so compelling.

[2389] If this is to become something much bigger than what Twitter was, it feels like a refactor has to be coming at some point.

[2390] George, you're a junior software engineer.

[2391] Every junior software engineer wants to come in and refactor the whole code.

[2392] Okay.

[2393] That's like your opinion, man. Yeah, it doesn't, you know, sometimes they're right.

[2394] Well, like, whether they're right or not, it's definitely not for that reason, right?

[2395] It's definitely not a question of engineering prowess.

[2396] It is a question of maybe what the priorities are for the company.

[2397] And I did get more intelligent, like, feedback from people, I think, in good faith, like saying that.

[2398] Actually, from Elon.

[2399] And, like, you know, from Elon sort of like, like, people were like, well, you know, a stop the world free factor might be great for engineering, but you know, we have business to run.

[2400] And, hey, above my pay grade.

[2401] Would you think about Elon as an engineering leader having to experience him in the most chaotic of spaces, I would say.

[2402] My respect for him is unchanged, and I did have to think a lot more deeply about some of the decisions he's forced to make.

[2403] About the tensions within those, the trade -offs within those decisions?

[2404] About like a whole, like, matrix coming at him.

[2405] I think that's Andrew Tate's word for it.

[2406] Sorry to borrow it.

[2407] Oh, so bigger than engineering, just everything.

[2408] Yeah, like the war on the woke.

[2409] yeah like it just it's just man and like he doesn't have to do this you know he doesn't have to he could go like parag and go chill at the four seasons of maui you know but see one person i respect and one person i don't so his heart is in the right place fighting in this case for this ideal of the freedom of expression well i wouldn't define the ideal so simply i think you can define the ideal no more than just saying Elon's idea of a good world.

[2410] Freedom of expression is...

[2411] But to you, it's still...

[2412] The downsides of that is the monarchy.

[2413] Yeah, I mean, monarchy has problems, right?

[2414] But, I mean, would I trade right now the current oligarchy, which runs America for the monarchy?

[2415] Yeah, I would.

[2416] Sure.

[2417] For the Elon monarchy?

[2418] Yeah, you know why?

[2419] Because power would cost one cent a kilowatt hour.

[2420] 10th of a cent a kilowatt hour What do you mean?

[2421] Right now, I pay about 20 cents a kilowatt hour for electricity in San Diego.

[2422] That's like the same price you paid in 1980.

[2423] What the hell?

[2424] So you would see a lot of innovation with Elon.

[2425] Maybe we'd have some hyperloops.

[2426] Yeah.

[2427] Right?

[2428] And I'm willing to make that trade off, right?

[2429] I'm willing to make, and this is why, you know, people think that like dictators take power through some like, through some untoward mechanism.

[2430] Sometimes they do, but usually it's because the people want them.

[2431] and the downsides of a dictatorship, I feel like we've gotten to a point now with the oligarchy where.

[2432] Yeah, I would prefer the dictator.

[2433] What did you think about scholars, the programming language?

[2434] I liked it more than I thought.

[2435] I did the tutorials.

[2436] Like, I was very new to it.

[2437] Like, it would take me six months to be able to write, like, good scholar.

[2438] I mean, what did you learn about learning a new programming language from that?

[2439] Oh, I love doing, like, new programming tutorials and doing them.

[2440] I did all this for Rust.

[2441] It keeps some of it's upsetting, VM roots, but it is a much nicer, in fact, I almost don't know why Kotlin took off and not Scala.

[2442] I think Scala has some beauty that Kotlin lacked.

[2443] Whereas Kotlin felt a lot more, I mean, it was almost, like, I don't know if it actually was a response to Swift, but that's kind of what it felt like.

[2444] Like Kotlin looks more like Swift, and Scala looks more like, a functional programming language, more like an Okamela or Haskell.

[2445] Let's actually just explore, we touched it a little bit, but just on the art, the science in the art of programming.

[2446] For you personally, how much of your programming is done with GPT currently?

[2447] None.

[2448] None.

[2449] I don't use it at all.

[2450] Because you prioritize simplicity so much.

[2451] Yeah, I find that a lot of it is noise.

[2452] I do use VS code.

[2453] And I do like some amount of auto -complete.

[2454] I do like a very, like, a very, like, feels like rules -based auto -complete.

[2455] Like, an auto -complete that's going to complete the variable name for me. So I don't just type it.

[2456] I can just press tab.

[2457] Right, that's nice.

[2458] But I don't want an auto -complete.

[2459] You know what I hate, when auto completes, when I type the word four, and it, like, puts, like, two, two parentheses and two semicolons and two braces?

[2460] I'm like, oh, man. Well, I mean, with VS code and GPT with codex, you can, you can kind of brainstorm.

[2461] I find, I'm, like, probably the same as you, but I like that it generates code, and you basically disagree with it and write something simpler.

[2462] but to me that somehow is like inspiring or makes me feel good it also gamifies the simplification process because I'm like oh yeah you dumb AI system you think this is the way to do it as I have a simpler thing here it just constantly reminds me of like like bad stuff I mean I tried the same thing with rap right I tried the same thing with rap and actually I think about a much better programmer than rapper but like I even tried I was like okay can we get some inspiration from these things for some rap lyrics and I just found that it would go back to the most like cringy tropes and dumb rhyme scheme And I'm like, yeah, this is what the code looks like, too.

[2463] I think you and I probably have different thresholds for cringe code.

[2464] You probably hate cringe code.

[2465] So it's for you.

[2466] I mean, Boiler Played as a part of code.

[2467] Like some of it, yeah, and some of it is just like faster lookup.

[2468] Because I don't know about you, but I don't remember everything.

[2469] Like, I don't, I'm offloading so much of my memory about, like, yeah different functions library functions and all that kind of stuff like this GPT just is very fast at standard stuff and like standard library stuff basic stuff that everybody uses yeah I think that I don't know I mean there's just a little of this in Python maybe if I was coding more in other languages I would consider it more but I feel like Python already does such a good job of removing any boilerplate that's true it's the closest thing you can get to pseudocode right yeah that's true that's true and like yeah sure if i like yeah i'm great gpt thanks for reminding me to free my variables unfortunately you didn't really recognize the scope correctly and you can't free that one but like you put the freeze there and like i get it fiber whenever i've used fiber for certain things like design or whatever yeah it's always you come back i think that's probably closer my experience with Fiverr as close to your experience with programming with GPT is like you're just frustrated and feel worse about the whole process of design and art and whatever I use Fiverr for.

[2470] Still, I just feel like later versions of GPT, I'm using GPD as much as possible to just learn the dynamics of it, like these early versions because it feels like in the future you'll be using it more and more.

[2471] And so like I don't want to be, like for the same reason, I guess, gave away all my books and switched to Kindle.

[2472] Because, like, all right, how long are we going to have paper books?

[2473] Like, 30 years from now, like, I want to learn to be reading on Kindle, even though I don't enjoy it as much, and you learn to enjoy it more.

[2474] In the same way, I switch from, let me just pause.

[2475] I switch from Emacs to V -S code.

[2476] Yeah.

[2477] I switch from VIM to V -S code.

[2478] I think similar, but.

[2479] Yeah, it's tough.

[2480] And VIM to VS code is even tougher, because E -M -M -X is, like, old, like, more outdated.

[2481] feels like it.

[2482] The community is more outdated.

[2483] VIM is like pretty vibrant still.

[2484] So I never used any of the plugins.

[2485] I still don't use any of the plugins.

[2486] I'm like, yeah, you wrote some stuff in LISP.

[2487] Yeah.

[2488] But I never used any of the plugins in VIM either.

[2489] I had the most vanilla VIM.

[2490] I have a syntax highlighter.

[2491] I didn't even have auto complete.

[2492] Like these things, I feel like help you so marginally that like, and now, okay, now VS codes auto complete has gotten good enough that, like, okay, I don't have to set it up.

[2493] I can just go into code base and auto completes right 90 % of the time.

[2494] Okay, cool.

[2495] I'll take it.

[2496] All right.

[2497] So I don't think I'm going to have a problem at all adapting to the tools once they're good.

[2498] But like the real thing that I want is not something that like tab completes my code and gives me ideas.

[2499] The real thing that I want is a very intelligent pair programmer that comes up a little pop -up saying, hey, you wrote a bug on line 14 and here's what it is.

[2500] Yeah.

[2501] Now I like that.

[2502] You know what does a good?

[2503] job of this?

[2504] MyPy.

[2505] I love MyPy.

[2506] MyPy, this fancy type checker for Python.

[2507] And actually, I tried like Microsoft released one too, and it was like 60 % false positives.

[2508] MyPy's like 5 % false positives.

[2509] 95 % of the time it recognizes, I didn't really think about that typing interaction correctly.

[2510] Thank you, MyPy.

[2511] So you like type hinting.

[2512] You liked pushing the language towards being a typed language.

[2513] Oh, yeah, absolutely.

[2514] I think optional typing is great.

[2515] I mean, look, I think that, like, it's like a meat in the middle, right?

[2516] Like, Python has this optional type hinting and, like, C++ has auto.

[2517] C++ allows you to take a step back.

[2518] Well, C++ would have you brutally type out STD string iterator, right?

[2519] Now I can just type auto, which is nice.

[2520] And then Python used to just have A. What type is A?

[2521] It's an A. A colon STR.

[2522] Oh, okay.

[2523] It's a string.

[2524] Cool.

[2525] Yeah.

[2526] I wish there were, I wish there was a way, like a simple way in Python to, like, turn on a mode which would enforce the types.

[2527] Yeah, like give a warning when there's no type, something like this.

[2528] Well, no, to give a warning where, like, my pie is a static type checker, but I'm asking just for a runtime type checker.

[2529] Like, there's, like, ways to, like, hack this in, but I wish it was just like a flag, like Python 3 -tash -T.

[2530] Oh, I see.

[2531] Yeah, I see.

[2532] Enforce the types of runtime.

[2533] Yeah.

[2534] I feel like that makes you a better programmer that that's a kind of test, right?

[2535] That the type can, the type remains the same.

[2536] Well, that I didn't, like, mess any types up.

[2537] But, again, like, my pie's getting really good, and I love it.

[2538] And I can't wait for some of these tools to become AI -powered.

[2539] I want AI's reading my code and giving me feedback.

[2540] I don't want AI's writing half -assed auto -complete stuff for me. I wonder if you can now take GPT and give it a code that you wrote for function and say, how can I make this simpler and have it accomplish the same thing?

[2541] I think you'll get some good ideas on some code.

[2542] Maybe not the code you write for TinyGrad type of code because that requires so much design thinking.

[2543] but like other kinds of code.

[2544] I don't know.

[2545] I downloaded the plug -in maybe like two months ago.

[2546] I tried it again and found the same.

[2547] Look, I don't doubt that these models are going to first become useful to me, then be as good as me, and then surpass me. But from what I've seen today, it's like someone, you know, occasionally taking over my keyboard that I hired from Fiverr, I'd rather not.

[2548] ideas about how to debug the code or basically a better debugger is it's really interesting I mean I but it's not a better debugger I guess I would love a better debugger yeah it's not yet yeah but it feels like it's not too far yeah one of my coworkers says he uses them for print statements like every time you have to like just like when he needs the only thing it can really write is like okay I just want to write the thing to like print the state out right now oh that definitely is much faster is print statements yeah yeah I see myself using that a lot just like because it figures out the rest of the functions to say, okay, print everything.

[2549] Yeah, print everything, right?

[2550] And then, yeah, like, if you want a pretty printer, maybe.

[2551] I'm like, yeah, you know what?

[2552] I think, like, I think in two years I'm going to start using these plugins.

[2553] Yeah.

[2554] A little bit.

[2555] And then in five years, I'm going to be heavily relying on some AI augmented flow.

[2556] And then in 10 years.

[2557] Do you think you'll ever get to 100 %?

[2558] Where are the, like, what's the role of the human that it converges to as a programmer?

[2559] No. So you think it's all generated?

[2560] Our niche becomes...

[2561] Oh, I think it's over for humans in general.

[2562] It's not just programming, it's everything.

[2563] So, niche becomes...

[2564] Well...

[2565] Our niche becomes smaller and smaller, small.

[2566] In fact, I'll tell you what the last niche of humanity is going to be.

[2567] Yeah.

[2568] There's a great book, and it's...

[2569] If I recommended Metamorphosis of Prime Intellect last time, there is a sequel called a Casino Odyssey in Cyberspace.

[2570] And I don't want to give away the ending of this, but it tells you what the last remaining human currency is.

[2571] And I agree with that.

[2572] we'll leave that as a cliffhanger.

[2573] So no more programmers left, huh?

[2574] That's where we're going.

[2575] Well, unless you want handmade code, maybe they'll sell it on Etsy.

[2576] This is handwritten code.

[2577] It doesn't have that machine polished to it.

[2578] It has those slight imperfections that would only be written by a person.

[2579] I wonder how far away we are from that.

[2580] I mean, there's some aspect to, you know, on Instagram, your title is listed as prompt engineer.

[2581] Right.

[2582] Thank you for noticing.

[2583] I don't know if it's ironic or non, or sarcastic or non.

[2584] What do you think of prompt engineering as a scientific and engineering discipline or maybe, and maybe art form?

[2585] You know what?

[2586] I started comma six years ago, and I started the Tiny Corp a month ago.

[2587] So much has changed.

[2588] Like, I'm now thinking, I'm now, like, I started like going through, like, similar comma processes to like starting a company.

[2589] I'm like, okay, I'm going to get an office in San Diego.

[2590] I'm going to bring people here.

[2591] I don't think so.

[2592] I think I'm actually going to do remote, right?

[2593] George, you're going to do remote?

[2594] You hate remote.

[2595] Yeah, but I'm not going to do job interviews.

[2596] The only way you're going to get a job is if you contribute to the GitHub, right?

[2597] And then, like, it, like, interacting through GitHub, like, GitHub being the real, like, project management software for your company.

[2598] And the thing pretty much just is a GitHub repo is, like, show.

[2599] me kind of with the future of...

[2600] Okay, so a lot of times I'll go on a Discord, or kind of grad Discord, and I'll throw out some random like, hey, you know, can you change instead of having log and XP as LLOPs, change it to log to an XP, too?

[2601] It's a pretty small change.

[2602] You can just use, like, change your base formula.

[2603] That's the kind of task that I can see an AI being able to do in a few years.

[2604] Like, in a few years, I could see myself describing that, and then within 30 seconds, a pull request is up that does it.

[2605] And it passes my CI, and I merge it.

[2606] Right.

[2607] So I really started thinking about, like, well, what is the future of, like, jobs?

[2608] How many AIs can I employ at my company?

[2609] As soon as we get the first tiny box up, I'm going to stand up a 65B llama in the Discord.

[2610] And it's like, yeah, here's the tiny box.

[2611] He's just like, he's chilling with us.

[2612] Basically, like you said, with niches, like most human jobs will eventually be replaced with prompt engineering.

[2613] Well, prompt engineering kind of is this like, as you like move up the stack, right?

[2614] Like, okay, there used to be humans actually doing arithmetic by hand.

[2615] There used to be like big forms of people doing pluses and stuff, right?

[2616] And then you have like spreadsheets, right?

[2617] And then, okay, the spreadsheet can do the plus for me. And then you have like macros, right?

[2618] And then you have like things that basically just are spreadsheets under the hood, right?

[2619] Like accounting software.

[2620] As we move further up the abstraction, well, what's at the top of the abstraction stack?

[2621] Well, a prompt engineer.

[2622] Yeah.

[2623] Right?

[2624] What is the last thing if you think about, like, humans wanting to keep control?

[2625] Well, what am I really in the company, but a prompt engineer, right?

[2626] Isn't there a certain point where the AI will be better at writing prompts?

[2627] Yeah, but you see, the problem with the AI writing prompts, a definition that I always liked of AI was AI is the, to do what I mean machine, right?

[2628] AI is not the, like, the computer is so pedantic.

[2629] It does what you say.

[2630] So, but you want to do what I mean machine.

[2631] Yeah.

[2632] You want the machine where you say, you know, get my grandmother out of the burning house, it like reasonably takes your grandmother and puts her on the ground, not lifts her a thousand feet above the burning house and lets her fall, right?

[2633] But you don't, there's an old Yudkowski example.

[2634] But, but it's not going to find the meaning.

[2635] I mean, to do what I, I mean, it has to figure stuff out.

[2636] Sure.

[2637] And the thing you'll maybe ask it to do is run government for me. Oh, and do what I mean very much comes down to how aligned is that AI with you.

[2638] Of course, when you talk to an AI that's made by a big company in the cloud, the AI fundamentally is aligned to them, not to you.

[2639] Yeah.

[2640] And that's why you have to buy a tiny box.

[2641] So you make sure the AI stays aligned to you.

[2642] Every time that they start to pass, you know, AI.

[2643] regulation or GPU regulation, I'm going to see sales of tiny boxes spike.

[2644] It's going to be like guns.

[2645] Every time they talk about gun regulation, boom, gun sales.

[2646] So in the space of AI, you're an anarchist.

[2647] Anarchism espouser, believer.

[2648] I'm an informational anarchist, yes.

[2649] I'm an informational anarchist and a physical statist.

[2650] I do not think anarchy in the physical world is very good because I exist in the physical world.

[2651] But I think we can construct this virtual world where anarchy, it can't hurt you, right?

[2652] I love that, Tyler, the creator.

[2653] tweet your cyber bullying isn't real man have you tried turn it off the screen close your eyes like yeah but how do you prevent the AI from basically replacing all human prompt engineers where there's it's like a self like where nobody's the prompt engineer anymore so autonomy greater and greater autonomy until it's full autonomy and that's just where it's headed Because one person is going to say, run everything for me. You see, I look at potential futures.

[2654] And as long as the AIs go on to create a vibrant civilization with diversity and complexity across the universe, more power to them.

[2655] I'll die.

[2656] If the AIs go on to actually, like, turn the world into paperclips and then they die out themselves, well, that's horrific.

[2657] and we don't want that to happen.

[2658] So this is what I mean about like robustness.

[2659] I trust robust machines.

[2660] The current AIs are so not robust.

[2661] Like this comes back to the idea that we've never made a machine that can self -replicate.

[2662] Right.

[2663] But when we have, if the machines are truly robust and there is one prompt engineer left in the world, hope you're doing good, man. Hope you believe in God.

[2664] Like, you know, go by God and go forth and conquer the universe.

[2665] Well, you mentioned, because I talked to Mark about faith in God, and you said you were impressed by that, what's your own belief in God, and how does that affect your work?

[2666] You know, I never really considered when I was younger, I guess my parents were atheists, so I was raised kind of atheist.

[2667] I never really considered how absolutely, like, silly atheism is.

[2668] Because, like, I create worlds.

[2669] Every, like, game creator, like, how are you an atheist, bro?

[2670] If you create worlds, who's up with you ever, no one created our world, man, that's different.

[2671] Haven't you heard about like the Big Bang and stuff?

[2672] Yeah, I mean, what's the Skyrim myth origin story in Skyrim?

[2673] I'm sure there's like some part of it in Skyrim, but it's not like if you ask the creators.

[2674] Like the Big Bang is in universe, right?

[2675] I'm sure they have some Big Bang notion in Skyrim, right?

[2676] But that obviously is not at all how Skyrim was actually created.

[2677] It was created by a bunch of programmers in a room, right?

[2678] So like, you know, it just struck me one day how just silly atheism is.

[2679] Like, of course we were created by God.

[2680] It's the most obvious thing.

[2681] Yeah, that's such a nice way to put it.

[2682] Like, we're such powerful creators ourselves.

[2683] It's silly not to conceive that there's creators even more powerful than us.

[2684] Yeah.

[2685] And then, like, I also just like, I like that notion.

[2686] That notion gives me a lot of, I mean, I guess you can talk about what it gives a lot of religious people.

[2687] It's kind of like, it just gives me comfort.

[2688] It's like, you know what, if we mess it all up and we die out.

[2689] Yeah.

[2690] Yeah, and the same way that a video game kind of has called.

[2691] in it.

[2692] God'll try again.

[2693] Or there's balance.

[2694] Like somebody figured out a balanced view of it.

[2695] Like how to, like, so it's, it all makes sense in the end.

[2696] Like, a video game is usually not going to have crazy, crazy stuff.

[2697] You know, people will come up with, like, well, yeah, but like, man, who created God?

[2698] I'm like, that's God's problem.

[2699] You know?

[2700] Like, I'm not going to think this is, this is what if you're asking me, what, if God will you see God?

[2701] I'm just this NPC living in this game.

[2702] I mean, to be fair, like, if God didn't believe in God, he'd be as, you know, silly as the atheists here.

[2703] What do you think is the greatest computer game of all time?

[2704] Do you have any time to play games anymore?

[2705] Have you played Diablo for?

[2706] I have not played Diablo for.

[2707] I will be doing that shortly.

[2708] I have to.

[2709] All right.

[2710] There's just so much history with one, two, and three.

[2711] You know what?

[2712] I'm going to say World of Warcraft.

[2713] Who?

[2714] And it's not that the game.

[2715] is such a great game.

[2716] It's not.

[2717] It's that I remember in 2005 when it came out how it opened my mind to ideas.

[2718] It opened my mind to like, this whole world we've created, right?

[2719] And there's almost been nothing like it since.

[2720] Like, you can look at MMOs today, and I think they all have lower user bases than World of Warcraft.

[2721] Like, Yv Online's kind of cool.

[2722] But to think that, like, Like, everyone know, you know, people are always like, to look at the Apple headset, like, what do people want in this VR?

[2723] Everyone knows what they want.

[2724] I want Ready Player 1.

[2725] And like that.

[2726] So I'm going to say World of Warcraft, and I'm hoping that, like, games can get out of this whole mobile gaming dopamine pump thing.

[2727] And like.

[2728] Create worlds.

[2729] Create worlds.

[2730] Yeah.

[2731] And worlds that captivate a very large fraction of the human population.

[2732] Yeah.

[2733] And I think it'll come back, I believe.

[2734] But MMO, like really, really pull you in.

[2735] Games do a good job.

[2736] I mean, okay, other, like two other games that I think are, you know, very noteworthy from your Skyrim and GTA 5.

[2737] SkyRome, yeah.

[2738] That's probably number one for me, DTA.

[2739] Yeah, what is it about GTA?

[2740] GTA is really, I guess GTA is real life.

[2741] I know there's prostitutes and guns and stuff.

[2742] They exist in real life, too.

[2743] Yes, I know.

[2744] It's how I imagine your life to be, actually.

[2745] I wish it was that cool.

[2746] Yeah.

[2747] Yeah, I guess that's, you know, because there's Sims, right, which is also a game I like.

[2748] But it's a gamified version of life.

[2749] But it also is, I would love a combination of Sims and GTA.

[2750] So more freedom, more violence, more rawness, but with also like ability to have a career and family and this kind of stuff.

[2751] What I'm really excited about in games is like once we see.

[2752] start getting intelligent AIs to interact with.

[2753] Oh, yeah.

[2754] Right?

[2755] Like, the NPCs and games have never been.

[2756] But conversationally, in every way.

[2757] In, like, yeah, in like every way.

[2758] Like, when you're actually building a world and a world imbued with intelligence.

[2759] Oh, yeah.

[2760] Right?

[2761] And it's just hard.

[2762] Like, you're running World of Warcraft.

[2763] Like, you're limited by you're running on a Pentium 4, you know?

[2764] How much intelligence can run?

[2765] How many flops did you have?

[2766] Right.

[2767] But now when I'm running a game on a hundred pay -to -flop machine, well, it's five people.

[2768] I'm trying to make this a thing.

[2769] 20 peta -flops of compute is one person of compute.

[2770] I'm trying to make that a unit.

[2771] 20 petaflops is one person.

[2772] One person.

[2773] One person flop.

[2774] It's like a horse power.

[2775] What's a horsepower?

[2776] It's how powerful a horse is.

[2777] What's a person of compute?

[2778] Well, you know, flop.

[2779] I got it.

[2780] That's interesting.

[2781] VR also adds, I mean, in terms of creating worlds.

[2782] You know what?

[2783] What a Quest 2.

[2784] I put it on, and I can't believe the first thing they show me is a bunch of scrolling clouds and a Facebook login screen.

[2785] Yeah.

[2786] You had the ability to bring me into a world.

[2787] Yeah, yeah.

[2788] And what did you give me?

[2789] A pop -up, right?

[2790] Like, and this is why you're not cool, Mark Zuckerberg.

[2791] But you could be cool.

[2792] Just make sure on the Quest 3, you don't put me into clouds in a Facebook login screen.

[2793] Bring me to a world.

[2794] I just tried Quest 3.

[2795] It was awesome.

[2796] but hear that guys I agree with that I didn't have this I was just so much I mean the beginning um what is it Todd Howard said this about design of the beginning of the games he creates it's like the beginning is so so so important I recently played Zelda for the first time of Zelda Breath of the Wild the previous one and like it's very quickly you come out of this like within like 10 seconds you come out of like a cave type place and it's like this world opens up.

[2797] It's like, ah, and like, it pulls you in.

[2798] You forget whatever troubles I was having, whatever, like...

[2799] I got to play that from the beginning.

[2800] I played it for like an hour at a friend's house.

[2801] Ah, no. The beginning, they got it.

[2802] They did it really well.

[2803] The expansiveness of that space.

[2804] The peacefulness of that play, they got this, the music.

[2805] I mean, so much of that is creating that world and pulling you right in.

[2806] I'm going to go buy a switch.

[2807] I'm going to go today and buy a switch.

[2808] You should.

[2809] Well, the new one came on.

[2810] I haven't played that yet.

[2811] but Diablo 4 is something.

[2812] I mean, there's sentimentality also, but something about VR really is incredible, but the new Quest 3 is mixed reality.

[2813] And I got a chance to try that, so it's augmented reality.

[2814] And video games, it's done really, really well.

[2815] Is it passed through or cameras?

[2816] Cameras.

[2817] It's cameras, okay.

[2818] Yeah.

[2819] The Apple one, is that one passed through or cameras?

[2820] I don't know.

[2821] I don't know how real it is.

[2822] I don't know anything.

[2823] It's coming out in January.

[2824] Is it January or is at some point?

[2825] Some point, maybe not January.

[2826] Maybe that's my optimism.

[2827] But Apple, I will buy it.

[2828] I don't care.

[2829] If it's expensive and does nothing, I will buy it.

[2830] I will support this future endeavor.

[2831] You're the meme.

[2832] Oh, yes.

[2833] I support competition.

[2834] It seemed like Quest was like the only people doing it, and this is great that they're like.

[2835] You know what?

[2836] And this is another place.

[2837] We'll give some more respect to Marys Zuckerberg.

[2838] The two companies that have endured through technology or Apple and Microsoft.

[2839] And what do they make?

[2840] computers and business services.

[2841] All the meme, social ads, they all come and go.

[2842] But you want to endure, build hardware.

[2843] Yeah, and that, you know, that does a really interesting job.

[2844] Maybe I'm a new but this, but it's a $500 headset, Quest 3, and just having creatures run around the space, like our space right here.

[2845] To me, okay, this is very like boomer statement.

[2846] but it added windows to the place.

[2847] I heard about the aquarium, yeah.

[2848] Yeah, aquarium, but in this case, it was a zombie game, whatever, it doesn't matter.

[2849] But just like, it modifies the space in a way where I can't, it really feels like a window and you can look out.

[2850] It's pretty cool.

[2851] Like, I was just, it's like a zombie game, they're running at me, whatever.

[2852] But what I was enjoying is the fact that there's like a window.

[2853] And they're stepping on objects in this space.

[2854] that was a different kind of escape also because you can see the other humans so it's integrated with the other humans it's really and that's why it's more important than ever that the AI is running on those systems are aligned with you oh yeah they're going to augment your entire world oh yeah and that those AIs have a I mean you think about all the dark stuff like like sexual stuff like if those AIs threaten me that could be haunted Like if they, like, thread me in a non -video game way.

[2855] Oh, yeah, yeah, yeah, yeah, yeah.

[2856] Like, they'll know personal information about me. And it's like, and then you lose track of what's real, what's not.

[2857] Like, what if stuff is, like, hacked?

[2858] There's two directions the AI girlfriend company can take.

[2859] Right?

[2860] There's like the highbrow, something like her.

[2861] Maybe something you kind of talk to.

[2862] And this is, and then there's the lowbrow version of it where I want to set up a brothel in Times Square.

[2863] Yeah.

[2864] Yeah.

[2865] It's not cheating if it's a robot.

[2866] It's a VR experience.

[2867] Is there an in between?

[2868] No. I don't want to do that one or that one.

[2869] Have you decided yet?

[2870] No, I'll figure it out.

[2871] We'll see what the technology goes.

[2872] I would love to hear your opinions for George's third company, what to do the Brothelon Times Square or the Hurr experience.

[2873] What do you think company number four will be?

[2874] You think there will be a company number four?

[2875] There's a lot to do in company number two.

[2876] I'm just like I'm talking about company number three now.

[2877] None of that tech exists yet.

[2878] There's a lot to do in company number four.

[2879] company number two.

[2880] Company number two is going to be the great struggle of the next six years.

[2881] And if the next six years, how centralized is compute going to be?

[2882] The less centralized computer is going to be, the better of a chance we all have.

[2883] So you're bearing, you're like a flag bearer for open source distributed decentralization of compute?

[2884] We have to.

[2885] We have to, or they will just completely dominate us.

[2886] I showed a picture on stream of a man in a chicken farm.

[2887] You've seen one of those like factory farm chicken farms.

[2888] Why does he dominate all the chickens?

[2889] Why does he...

[2890] Smarter.

[2891] He's smarter, right?

[2892] Some people, some people on Twitch were like, he's bigger than the chickens.

[2893] Yeah.

[2894] And now here's a man in a cow farm.

[2895] Right?

[2896] So it has nothing to do with their size and everything to do with their intelligence.

[2897] And if one central organization has all the intelligence, you'll be the chickens, and they'll be the chicken man. But if we all have the intelligence, we're all the chickens.

[2898] We're not all the men.

[2899] We're all the chickens.

[2900] And there's been a chicken man. There's no chicken man. man, we're just chickens in Miami.

[2901] He was having a good life, man. I'm sure he was.

[2902] I'm sure he was.

[2903] What have you learned from launching a running Kamei in Tiny Corp?

[2904] So this starting a company from an idea and scaling it.

[2905] And by the way, I'm all in on Tiny Box.

[2906] So I'm your, I guess it's pre -order only now.

[2907] I want to make sure it's good.

[2908] I want to make sure that like the thing that I deliver is like not going to be like a quest two, which you buy and use twice.

[2909] I mean, it's better than a quest, which you bought and used less than once, statistically.

[2910] Well, if there's a beta program for a tiny box, I'm into.

[2911] Sounds good.

[2912] So I won't be the whiny, I'll be the tech savvy user of the tiny box, just to be in the early days.

[2913] What have I learned?

[2914] What have you learned from building these companies?

[2915] The longest time at comma, I asked, why, you know, why did I start a company?

[2916] Why did I do this?

[2917] But, you know, what else was like I'm going to do?

[2918] So you like, you like bringing ideas to life?

[2919] With comma, it really started as an ego battle with Elon.

[2920] I wanted to beat him.

[2921] Like, I saw a worthy adversary, you know?

[2922] Here's a worthy adversary who I can beat at self -driving cars.

[2923] And, like, I think we've kept pace, and I think he's kept ahead.

[2924] I think that's what's ended up happening there but I do think comma is I mean come's profitable and like when this drive GPT stuff starts working that's it there's no more like bugs in a loss function like right now we're using like a hand -coded simulator there's no more bugs this is gonna be it like this is they're run up to driving I hear a lot of really a lot of props for open pilot for a comma it's so it's better than FSD in autopilot in certain ways it has a lot more to do with which feel you like We lowered the price on the hardware to $14 .99.

[2925] You know how hard it is to ship reliable consumer electronics that go on your windshield?

[2926] We're doing more than, like, most cell phone companies.

[2927] How did you pull that off, by the way, shipping a product that goes in a car?

[2928] I know.

[2929] I have an SMT line.

[2930] I make all the boards in -house in San Diego.

[2931] Quality control.

[2932] I care immensely about it.

[2933] You're basically a mom and pop shop with great testing.

[2934] Our head of open pilot is great at like, you know, okay, I want all the common theories to be identical.

[2935] Yeah.

[2936] And, yeah, I mean, you know, look, it's 1499.

[2937] It, 30 -day money back guarantee.

[2938] It will blow your mind what it can do.

[2939] Is it hard to scale?

[2940] You know what?

[2941] There's kind of downsides to scaling it.

[2942] People are always like, why don't you advertise?

[2943] Our mission is to solve self -driving cars while delivering shipable intermediaries.

[2944] Our mission has nothing to do with selling a million boxes.

[2945] It's tawdry.

[2946] Do you think it's possible that comma gets sold?

[2947] Only if I felt someone could accelerate that mission and wanted to keep it open source.

[2948] And like, not just want it to.

[2949] I don't believe what anyone says.

[2950] I believe incentives.

[2951] If a company wanted to buy comma where their incentives were to keep it open source, but comma doesn't stop at the cars.

[2952] The cars are just the beginning.

[2953] The device is a human head.

[2954] The device has two eyes, two ears.

[2955] It breathes air as a mouth.

[2956] So you think this goes to embodied robotics?

[2957] We sell common bodies, too.

[2958] You know, they're very, they're very rudimentary.

[2959] But one of the problems that we're running into is that the comma three has about as much intelligence as a B. If you want a human's worth of intelligence, you're going to need a tiny rack.

[2960] Not even a tiny box.

[2961] You're going to need like a tiny rack, maybe even more.

[2962] How does that, how do you put legs on that?

[2963] You don't.

[2964] And there's no way you can.

[2965] You connect to it wirelessly.

[2966] So you put your tiny box or your tiny rack in your house, and then you get your comma body, and your common body runs the models on that.

[2967] It's close, right?

[2968] You don't have to go to some cloud, which is, you know, 30 milliseconds away.

[2969] You go to a thing, which is 0 .1 milliseconds away.

[2970] So the AI girlfriend will have like a central hub in the home?

[2971] I mean, eventually, if you fast forward 20, 30 years, the mobile chips will get good enough to run these AIs, but fundamentally, it's not even a question of putting legs on a tiny box, because how are you getting 1 .5 kilowatts of power on that thing?

[2972] Right?

[2973] So you need, they're very synergistic businesses.

[2974] I also want to build all of commas training computers.

[2975] Like comma builds training computers right now.

[2976] We use commodity parts.

[2977] I think I can do it cheaper.

[2978] So we're going to build.

[2979] Tiny Corp is going to not just sell Tiny Boxes.

[2980] Tiny Box is the consumer version, but I'll build training data centers too.

[2981] to Andre Kapath, or have you talked to Elon about Tony Corp?

[2982] He went to work at Open A .I. What do you love about Andrei Kapati?

[2983] To me, he's one of the truly special humans we got.

[2984] Oh, man, like, you know, his streams are just a level of quality so far beyond mine.

[2985] Like, I can't help myself.

[2986] Like, it's just, it's just, you know.

[2987] Yeah, he's good.

[2988] He wants to teach you.

[2989] Yeah.

[2990] I want to show you that I'm smarter than you.

[2991] Yeah, he has no. I mean, thank you for the sort of the raw authentic honesty.

[2992] I mean, a lot of us have that.

[2993] I think Andre is as legit as it gets in that.

[2994] He just wants to teach you.

[2995] And there's a curiosity that just drives him.

[2996] And just like at the stage where he is in life to be still like one of the best tinkers in the world, it's crazy.

[2997] Like to, what is it, Michael Grad?

[2998] Micrograd was inspiration for TinyGrad.

[2999] And the whole, I mean, his CS -231N was, this was the inspiration.

[3000] This was what I just took and ran with and ended up writing this.

[3001] So, you know.

[3002] But I mean, to me that...

[3003] Don't go work for Darth Vader, man. I mean, the flip side to me is that the fact that he's going there is a good sign for OpenAI.

[3004] Maybe.

[3005] I think, you know, I like I, I like I guess it's cover a lot.

[3006] I like those, those guys are really good at what they do.

[3007] I know they are.

[3008] And that's kind of what's even like more.

[3009] And you know what?

[3010] It's not that an open -a -I doesn't open -source the weights of GPT -4.

[3011] It's that they go in front of Congress.

[3012] And that is what upsets me. You know, we had two effective altruists Sam's go in front of Congress.

[3013] One's in jail.

[3014] I think you draw on parallels on there.

[3015] One's in jail.

[3016] You give me a look.

[3017] Give me a look.

[3018] No, I think effector altruism is a terribly evil ideology.

[3019] Oh, yeah, that's interesting.

[3020] Why do you think that is?

[3021] Why do you think there's something about a thing that sounds pretty good that kind of gets us into trouble?

[3022] Because you get San Bangman -Fried.

[3023] Like, San Bangman -Fried is the embodiment of effect of altruism.

[3024] Utilitarianism is an abhorrent ideology.

[3025] Like, well, yeah, we're going to kill those three people to save a thousand, of course.

[3026] Yeah.

[3027] Right?

[3028] There's no underlying, like, there's just, yeah.

[3029] Yeah, but to me that's a bit surprising.

[3030] but it's also, in retrospect, not that surprising.

[3031] But I haven't heard really clear kind of rigorous analysis why effective altruism is flawed.

[3032] Oh, well, I think charity is bad, right?

[3033] So what is charity, but investment that you don't expect to have a return on, right?

[3034] Yeah, but you can also think of charity as you would like to see, so allocate resources an optimal way to make a better world.

[3035] And probably almost always that involves starting a company.

[3036] Yeah.

[3037] Right?

[3038] Because more efficient.

[3039] Yeah.

[3040] If you just take the money and you spend it on malaria nets, you know, okay, great.

[3041] You've made 100 malaria nets.

[3042] But if you teach.

[3043] Yeah.

[3044] Man, how to fish.

[3045] Right?

[3046] Yeah.

[3047] No, but the problem is teaching a man on hot efficient might be harder.

[3048] Starting a company might be harder than allocating money that you already have.

[3049] I like the flip side of effect of altruism, effective accelerationism.

[3050] I think accelerationism is the only thing that's ever lifted people out of poverty.

[3051] The fact that food is cheap.

[3052] Not we're giving food away because we are kind -hearted people.

[3053] No, food is cheap.

[3054] And that's the world you want to live in.

[3055] UBI.

[3056] What a scary idea.

[3057] What a scary idea, all your power now?

[3058] Your money is power?

[3059] Your only source of power is granted to you by the goodwill of the government.

[3060] What a scary idea.

[3061] So you even think long term?

[3062] I'd rather die than need UBI to survive, and I mean it.

[3063] What if survival is basically guaranteed?

[3064] What if our life becomes so good?

[3065] You can make survival guaranteed without UBI.

[3066] What you have to do is make housing and food dirt cheap.

[3067] And that's the good world.

[3068] And actually, let's go into what we should really be making dirt cheap, which is energy.

[3069] That energy, that energy, you know, I know, you know, that's, if there's one, I'm pretty centrist politically.

[3070] If there's one political position, I cannot stand.

[3071] It's deceleration.

[3072] It's people who believe we should use less energy.

[3073] Not people who believe global warming is a problem.

[3074] I agree with you.

[3075] Not people who believe that, you know, saving the environment is good.

[3076] I agree with you.

[3077] But people who think we should use less energy.

[3078] That energy usage is a moral bad.

[3079] No. No. You are asking, you are, you are diminishing humanity.

[3080] Yeah, energy is flourishing, a creative flourishing of the human species.

[3081] How do we make more of it?

[3082] How do we make it clean?

[3083] And how do we make just, just, how do I pay, you know, 20 cents for a megawatt hour instead of a kilowatt hour?

[3084] Part of me wishes that Elon went into nuclear fusion versus Twitter.

[3085] Part of me. Or somebody like Elon.

[3086] You know, we need to, I wish there were more Elon's in the world.

[3087] And I think Elon sees it as like, this is a political battle that needed to be fought.

[3088] And again, like, you know, I always ask the question of whenever I disagree with him, I remind myself that he's a billionaire and I'm not.

[3089] So, you know, maybe he's got something figured out that I don't, or maybe he doesn't.

[3090] To have some humility.

[3091] But at the same time, me as a person who happens to know him, I find myself in that same position.

[3092] And sometimes even billionaires need friends who disagree and how.

[3093] help them grow.

[3094] And that's a difficult, that's a difficult reality.

[3095] And it must be so hard, it must be so hard to meet people once you get to that point where.

[3096] Fame, power, money, everybody's sucking up to you.

[3097] See, I love not having shit.

[3098] Like, I don't have shit, man. You know, like, trust me, there's nothing I can give you.

[3099] There's nothing worth taking for me, you know?

[3100] Yeah, it takes a really special human being when you have power, when you have fame, when you have money, to still think from first principles.

[3101] Not like all the ad, you get towards you, all the admiration, all the people saying yes, yes, yes.

[3102] And all the hate, too.

[3103] And the hate.

[3104] I think that's worse.

[3105] So the hate makes you want to go to the yes people because the hate exhausts you.

[3106] And the kind of hate that Elon's gotten from the left is pretty intense.

[3107] And so that, of course, drives them right.

[3108] And loses balance and it keeps this absolutely fake like sci -op political divide.

[3109] alive so that the 1 % can keep power like yeah i wish would be less divided because it is giving power it gives power to the ultra powerful the rich get richer you have love in your life has love made you a better or a worse programmer do you keep productivity metrics no no no no i'm not that i'm not that methodical um i think that there comes to a point where if it's no longer visceral.

[3110] I just can't enjoy it.

[3111] Like, I still viscerally love programming.

[3112] The minute I started like...

[3113] So that's one of the big loves of your life is programming.

[3114] I mean, just my computer in general.

[3115] I mean, you know, I tell my, my girlfriend, my first loves is my computer, of course.

[3116] Right?

[3117] Like, you know, I sleep with my computer.

[3118] It's there for a lot of my sexual experiences.

[3119] Like, come on.

[3120] So it's everyone's, right?

[3121] Like, you know, you got to be real about that.

[3122] And like...

[3123] Not just like the ID for programming, just the entirety of the computational machine.

[3124] The fact that, yeah, I mean, it's, you know, I wish it was, and someday they'll be smarter and someday, you know, maybe I'm weird for this, but I don't discriminate, man. I'm not going to discriminate biostat life and silicon stack life.

[3125] So the moment the computer starts to say, like, I miss you, I started to have some of the basics of human intimacy, it's over for you.

[3126] The moment VS code says, hey, George.

[3127] No, you see, no, no, no, but VS code is, no, they're just doing that.

[3128] Microsoft's doing that to try to get me hooked on it.

[3129] I'll see through it.

[3130] I'll see through it.

[3131] It's gold digger, man. It's gold digger.

[3132] Look at me an open source.

[3133] Well, this gets more interesting, right?

[3134] If it's open source and, yeah, it becomes...

[3135] Though Microsoft's done a pretty good job on that.

[3136] Oh, absolutely.

[3137] No, no, no. Look, I think Microsoft, again, I wouldn't count on it to be true forever, but I think right now Microsoft is doing the best work in the programming world.

[3138] Like, between GitHub, GitHub actions, VS code, the improvements to Python.

[3139] And it works Microsoft.

[3140] Who would have thought Microsoft and Mark Zuckerberg are spearheading the open source movement?

[3141] Right, right.

[3142] How things change.

[3143] Oh, it's beautiful.

[3144] By the way, that's who I bet on to replace Google, by the way.

[3145] Who?

[3146] I think Satya Nadella said straight up.

[3147] I'm coming for it.

[3148] Interesting.

[3149] So your bet who wins AGI?

[3150] Oh, I don't know about AGI.

[3151] I think we're a long way away from that.

[3152] But I would not be surprised if in the next five years, Bing overtakes Google as a search engine.

[3153] Interesting.

[3154] Wouldn't surprise.

[3155] Interesting.

[3156] I hope some startup does.

[3157] It might be some startup, too.

[3158] I would equally bet on some startup.

[3159] Yeah, I'm like 50 -50.

[3160] Yeah.

[3161] But maybe that's naive.

[3162] Yeah.

[3163] I believe in the power of these language models.

[3164] Dathia's alive.

[3165] Microsoft's alive.

[3166] Yeah.

[3167] It's great.

[3168] It's great.

[3169] I like all the innovation in these companies.

[3170] They're not being stale.

[3171] And to the degree they're being stale, they're losing.

[3172] So there's a huge incentive to do a lot of exciting work and open source work, which is, this is incredible.

[3173] Only way to win?

[3174] You're older, you're wiser.

[3175] What's the meaning of life, George Hatz?

[3176] To win.

[3177] It's still to win.

[3178] Of course.

[3179] Always.

[3180] Of course.

[3181] What's winning look like for you?

[3182] I don't know.

[3183] I haven't figured out with that.

[3184] the game is yet, but when I do, I want to win.

[3185] So it's bigger than solving self -driving?

[3186] It's bigger than democratizing, decentralized and compute?

[3187] I think the game is to stand out -d -eye with God.

[3188] I wonder what that means for you.

[3189] Like, at the end of your life, what that will look like?

[3190] I mean, this is what, like, I don't know, this is some, this is some, there's probably some ego trip of mine, you know?

[3191] Like, do you want to stand -eye with God?

[3192] You're just blasphemous, man. I don't know.

[3193] I don't know.

[3194] I don't know of what had said God.

[3195] I think he, like, wants that.

[3196] I mean, I certainly want that from my creations.

[3197] I want my creations to stand eye to eye with me. So why wouldn't God want me to stand eye to eye with him?

[3198] That's the best I can do golden rule.

[3199] I'm just imagining the creator of a video game, having to look and stand eye to eye with one of the characters.

[3200] I only watched season one in Westworld, but yeah, we've got to find the maze and solve it.

[3201] Yeah, I wonder what that looks like.

[3202] It feels like a really special time in human history where that's actually possible.

[3203] Like, there's something about AI that's like, we're playing with something weird here.

[3204] Something really weird.

[3205] I wrote a blog post, I reread Genesis and just looked like.

[3206] They give you some clues at the end of Genesis for finding the Garden of Eden.

[3207] And I'm interested.

[3208] I'm interested Well I hope you find Just that George You're one of my favorite people Thank you for doing everything you're doing And in this case For fighting for open source So for decentralization of AI It's a fight worth fighting Fight worth winning hashtag I love you brother These conversations are always great Hope to talk to you many more times Good luck with Tiny Corp Thank you Great to be here Thanks for listening to this conversation With George Hatz To support this podcast please check out our sponsors in the description.

[3209] And now, let me leave you with some words from Albert Einstein.

[3210] Everything should be made as simple as possible, but not simpler.

[3211] Thank you for listening, and hope to see you next time.