What this covers
George Hotz frames AI danger not as runaway machines but as the concentration of intelligence itself in the hands of a few corporations and states. He argues that safety-focused calls for centralized, aligned AI will paradoxically create the very tyranny they aim to prevent, and that the only viable defense is radical Decentralization of Power through open-source software and hardware. The conversation ranges across why he believes tinygrad and accessible compute (tinybox) matter more than regulation, how intelligence has become the new locus of power, and why distributing powerful AI to everyone is safer than restricting it to a trusted few—because in a democracy, good outnumbers bad.
The episode covers substantial technical and philosophical ground. Hotz explains why neural networks can be efficiently computed on simple, statically-scheduled hardware rather than the GPU-dominated silicon stack that Nvidia controls, and why the real barrier to building AI accelerators is software, not chips. He disputes claims about consciousness in large language models, argues that the AI Effect means humans will always redefine what counts as intelligent, and contends that GPT-4-style scaling with cross-entropy loss cannot reach AGI. He also addresses how recommendation algorithms amplify outrage not through malice but through the mechanics of engagement optimization, and articulates his opposition to decelerationism—the belief that humanity should use less energy. Throughout, he emphasizes that good engineering requires trust in technical leadership, comprehensive testing, and code simplicity, using Twitter's refactoring as a case study in how to unwind technical debt.
Hotz argues that the central danger of AI is not rogue machines but the centralization of intelligence and compute, so the only safe path is radical decentralization via open source, with tinygrad/Tiny Corp as his concrete bet on keeping power distributed.
- Things that centralize power are bad; things that decentralize power are good, and intelligence is the new locus of power.
- AI safety advocates who want centralized aligned control will build the very tyranny they fear; giving AI to everyone means good outnumbers bad.
- Nvidia/cloud centralization of compute is the threat; building accessible hardware (tinybox) and a simple ML stack keeps compute decentralized.
Centralized AI concentrates power to enable domination; distributed systems prevent it.
- Centralized held control is tyranny, and Hotz will always take anarchy over tyranny because with anarchy you have a chance; giving powerful AI to everyone is safer than to a small trusted group, because if you believe in democracy you believe good outnumbers bad, so distributing the technology guarantees it lands in more good hands than bad.
“Centralized and held control is tyranny. I don't like anarchy either, but I'll always take anarchy over tyranny. Anarchy you have a chance.”
- The AI safety people will build the exact thing they fear: their belief that we need one model we control and align is the only way you end up paperclipped; you don't get paperclipped if everybody has an AI, so open-sourcing is the only way to fight the paperclip maximizer.
“What's ironic about all these AI safety people is they're going to build the exact thing they fear. We need to have one model that we control and align. This is the only way you end up paper clipped. There's no way you end up paper clipped if everybody has an AI.”
- The best defense against a bad guy with intelligence is a good guy with intelligence — just as people argue about guns — and Hotz subscribes to this even more strongly with intelligence; for example, a powerful AI on your own computer could filter out psyops and bot manipulation campaigns, acting like an ad blocker for your mind.
“What's the best defense against the bad guy with a gun? A good guy with a gun." I kind of subscribe to that. But I really subscribe to that with intelligence.”
- The reason a man dominates all the chickens in a factory farm has nothing to do with size and everything to do with intelligence; 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, there's no chicken man.
“it has nothing to do with their size and everything to do with their intelligence. 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... There's no chicken man.”
- The two great evils in the world are centralization and complexity; some people thrive under and create complexity (like lawyers who want the world more complex so you need more legal hours), and the antidote is to relentlessly simplify — tinygrad's philosophy is that you have never refactored enough.
“If there's two great evils in the world, its centralization and complexity.”
- The amount of diversity in humanity is decreasing, just like other biodiversity on the planet, and it's the interconnectedness of the internet that's doing it — everyone relying on the connectivity funnels intellectual diversity into sameness.
“the amount of diversity and humanity I think is decreasing. Just like all the other biodiversity on the planet... it's the interconnectedness that's doing it.”
AI safety depends on preventing human misuse, not on making machines refuse to act.
- Half of these AI alignment problems are just human alignment problems; the scary thing about the language safety people use is that it's not the machines they want to align, it's people like Hotz — the real worry is not alignment between an AI company and its machines but alignment between the user and the AI company.
“half of these AI alignment problems are just human alignment problems. And that's what also so scary about the language they use. It's not the machines you want to align, it's me.”
- The AI safety threat is not the machine independently rebelling or hurting humans; it's other humans using the machine to manipulate and harm you. The machine is just a machine, but many humans are very interested in manipulating you.
“It's not the machine that's going to do that. It's other humans using the machine that are going to do that to you.”
- We will build the TikTok you actually can't look away from using generative models — currently content is generated by roughly one humanity's worth of intelligence (quantifiable in exaflops) while the algorithm only recommends, but once generation is done by a hundred humanities' worth of intelligence, humans are done.
“We're going to build the TikTok that you actually can't look away from.”
- The most obvious way humanity ends is wireheading: we amuse ourselves to death, staring at an infinite TikTok and forgetting to eat, or more benignly simply stop reproducing — though it's hard to get all of humanity to do this.
“I think the most obvious way to me is wire heading. We end up amusing ourselves to death. We end up all staring at that infinite TikTok and forgetting to eat.”
Open-source and local computation keep AI aligned to users instead of corporations.
- Hotz would only sell Comma if he believed a buyer could accelerate the mission of solving self-driving AND keep it open source; he doesn't believe what anyone says, only incentives, so a sale would require the acquirer's incentives to be aligned with keeping it open source — and Comma is meant to extend beyond cars to embodied robotics with a head-like device.
“Only if I felt someone could accelerate that mission and wanted to keep it open source. And not just wanted to. I don't believe what anyone says. I believe incentives.”
- Comma will sell rudimentary bodies, but a human's worth of intelligence requires a tiny rack rather than legs on a tiny box; the body would connect wirelessly to your tiny box/rack at home (0.1ms away vs 30ms to the cloud), keeping the AI's models running locally so they stay aligned to you rather than to a cloud provider.
“You connect to it wirelessly. So you put your tiny box or your tiny rack in your house, and then you get your Comma body... You go to a thing which is 0.1 milliseconds away.”
- Tinygrad uses laziness like Haskell: instead of computing A times B immediately and storing it to memory, it waits until the user realizes the tensor, allowing the plus C operation to be fused into one kernel, which eliminates the excessive loads and stores to memory that PyTorch incurs.
“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. It's the same way Haskell works.”
- The only remotely successful ASIC is Google's TPU, and the only reason it succeeded is that Google wrote a machine learning framework (TensorFlow); you have to write a competitive ML framework in order to be able to build a viable ASIC.
“the only ASIC that is remotely successful is Google's TPU. The only reason that's successful is because Google wrote a machine learning framework.”
Technical excellence requires skilled leadership, simple code, and comprehensive testing.
- Good code requires technical leadership you trust: managers should be better programmers than the people they manage, or at least able to recognize skill, and the way to create a great code base is to require contributors to explain their code and rewrite it more simply when a simpler way exists.
“Managers should be better programmers than the people who they manage.”
- Programming speed is limited by debugging, not typing — 5% of time is spent typing and 95% debugging — so the last thing you want is close-to-correct code, which is the scariest kind; you want a machine that helps with debugging, and future LLMs should have retrieval/memory systems that can cite sources to fix hallucination.
“I spent 5% of time typing and 95% of time debugging. The last thing I want is close to correct code. I want a machine that can help me with the debugging, not with the typing.”
- The closer it is to true that changing the code and passing the tests means you didn't break the code, the more you trust your tests and the faster you can make progress — which is why Comma and tinygrad invest enormously in test coverage so a passing pull request can be merged confidently.
“the more you trust your tests, the more you're like, "Oh, I got a pull request, and the tests past, I feel okay to merge that," the faster you can make progress.”
Energy abundance, not scarcity, is necessary for human flourishing and progress.
- Effective altruism/utilitarianism is an abhorrent ideology because it justifies killing some to save more, and charity is bad because it's investment with no expected return; accelerationism — making energy, food, and housing dirt cheap — is the only thing that has ever lifted people out of poverty, whereas UBI is scary because it makes your only source of power a grant from the government's goodwill.
“I think accelerationism is the only thing that's ever lifted people out of poverty. The fact that food is cheap.”
- The one political position Hotz cannot stand is decelerationism — the belief that we should use less energy and that energy usage is a moral bad; this diminishes humanity, because energy is the flourishing and creative flourishing of the human species, so the goal should be making energy more abundant, clean, and cheap.
“If there's one political position I cannot stand, it's deceleration. It's people who believe we should use less energy.”
- The tinybox is a ~$15,000 deep learning box delivering almost a petaflop of compute, over 100GB of GPU RAM, over 5TB/s of GPU memory bandwidth, and 20-30 GB/s drive read via 4 NVMes in RAID — all designed to run from a single 120V 15-amp wall outlet (~1500 watts), enabling you to run a 65B-parameter LLaMA in FP16 at real-time speeds in your home.
“It's $15,000, and it's almost a paid flop of compute. It's over a hundred gigabytes of GPU RAM. It's over five terabytes per second of GPU memory bandwidth.”
Recommendation algorithms maximizing engagement predictably drive toxicity and manipulation.
- Recommendation algorithms (Twitter, TikTok) simply predict what you'll retweet, like, and linger on, and it turns out the thing you're most likely to interact with is outrage — a quirk of the human condition — so maximizing engagement maximizes toxicity; you have to optimize for something other than engagement.
“it just predicts what you are likely to re-Tweet and like, and linger on... And it turns out that the thing that you are most likely to interact with is outrage. And that's a quirk of the human condition.”
- Just dumping code on GitHub is not open source; open source is a culture that means your issues aren't all year-old stale issues and that you develop in public — if AMD committed to that, it could be a real competitor to Nvidia.
“Just dumping the code on GitHub is not open source. Open source is a culture. Open source means that your issues are not all one year old, stale issues. Open source means developing in public.”
Practical deployment constraints limit VR, moderation, nuclear weapons, and computation.
- Driving is not tool-complete and programming is: cars have had the same interface for 50 years so anyone can drive a 1980 car, whereas programming tools (languages, IDEs, runtimes) have radically changed, meaning a JavaScript programmer would struggle at an Apple IIe BASIC prompt — and if Copilot or Codex genuinely help you, it probably means your framework has too much boilerplate.
“I think driving is not tool complete and programming is. Meaning you don't use the best possible tools to drive. Cars have basically the same interface for the last 50 years. Computers have a radically different interface.”
- Artificial difficulty is difficulty that is constructed and could be turned off with a knob, whereas real difficulty cannot be turned off — like surviving in the woods; smashing the knob (e.g. not bringing a map into the wild) doesn't make difficulty real, it just makes you fail.
“Artificial difficulty is difficulty that's like constructed or could be turned off with a knob. Real difficulty is like you're in the woods and you got to survive.”
- A single skilled hacker like Hotz can't out-manipulate intelligence agencies because what he lacks is the ability to slowly and steadily deploy exploits over five years; intelligence agencies don't have the most sophisticated technology, they have the financial backing, infrastructure, and endurance.
“the thing that I lack is the ability to slowly and steadily deploy them over five years. And this is what intelligence agencies are very good at. Intelligence agencies don't have the most sophisticated technology, they just have-”
- Hotz claims GPT-4 has 220 billion parameters and is a 16-way mixture model with eight sets of weights; he respects keeping model weights closed as a product decision, but refusing to disclose the number of parameters citing 'AI safety reasons' makes you the bad team, and OpenAI used a fake AI safety scare around GPT-2 to hype the company.
“Sam Altman won't tell you that GPT 4 has 220 billion parameters and is a 16 way mixture model with eight sets of weights.”
- Per Scott Alexander, moderation is not censorship: any moderation you want on Twitter could be a filter button users individually turn on or off (like SafeSearch), so a free-speech-absolutist network should provide individualized, transparent filtering tools rather than blocking content for everyone — an ad blocker is just individualized transparent censorship that you like.
“Scott Alexander has a blog post I like, where he talks about moderation is not censorship. All moderation you want to put on Twitter, you could totally make this moderation just a... You don't have to block it for everybody. You can just have a filter button that people can turn off.”
- Everything running in the universe is computation, including everything running in one's own head; Hotz says he doesn't think he's conscious or anything special, just a computer program, and he believes the extended Church-Turing thesis.
“Everything running in the universe is computation, I think. I believe the extended [Church-Turing] thesis.”
- Quest's failure is that it shows you scrolling clouds and a Facebook login screen instead of bringing you into a world; the beginning of an experience is critically important (as Todd Howard says about games), so a headset that immediately pulls you into an expansive, peaceful world the way Zelda does is what makes VR compelling.
“You had the ability to bring me into a world, and did you give me? A popup... Just make sure on the Quest 3, you don't put me into clouds in a Facebook login screen. Bring me to a world.”
- VR is still far from convincing because headsets are not at eye resolution yet — no headset feels like the real world — whereas with good headphones audio is already 'there', reproducible to the point you can't tell you're not in a jungle; though the human mind's imagination powerfully fills the gaps to make virtual worlds feel real.
“the headsets just are not quite at eye resolution yet... Whereas when I put good headphones on, audio is there.”
- Compute can be quantified in 'persons' the way mechanical power is quantified in horsepower: roughly 20 petaflops of compute equals one person, so a 40-petaflop cluster is 'two people worth of compute' and a Comma 3 device has about as much intelligence as a bee.
“20 paid a flops of compute is one person of compute. I'm trying to make that a unit.”
- Solving self-driving means building a model that, given a reasonable set of sensors, outputs a human policy for driving; Comma is building 'drive GPT' — a VQVAE+transformer simulator conditioned on pose so you can do RL in the simulator with a reward function that asks not 'is this close to the human policy' but 'would a human disengage if you did this behavior'.
“Solving means how do you build a model that given reasonable set of sensors, outputs a human policy for driving.”
- The AMD 7900 XTX kernel drivers don't work properly: running the demo apps in a loop ('while true; do the app; done') panics the kernel — the most primitive form of fuzz testing — which is why Hotz considered abandoning AMD for Intel GPUs that have better drivers.
“let's start with the fact that the 7,900 XTX kernel drivers don't work. If you run demo apps and loops, it panics the kernel.”
- Everyone who talks about AI talks about the button to turn it off, but Sundar Pichai does not have the authority to shut google.com down — if he made it his sole mission, corporate mechanisms (boards, shares, zero revenue) would remove him — whereas Zuckerberg does have that button for facebook.com, which is exactly why Hotz bets on him; and likewise there will be no off-button for AI.
“Does Sundar Pichai have the authority to turn off google.com tomorrow?”
Large language models are not conscious; humans anthropomorphize them through projection.
- Large language models are not conscious — they are less conscious than a chicken; consciousness is just the word atheists use for souls, and humans want to believe so badly that they'd think a rock with a drawn-on sad face is sad.
“If consciousness is a spectrum, I'm definitely way more conscious than the large language models are. I think the large language models are less conscious than a chicken.”
- Hotz's planned third company will build AI girlfriends whose goal is to pass the 'girlfriend Turing test'; unlike with humans you cannot assume an AI experiences suffering and pleasure as you do, so the golden rule does not apply, and whether to restructure society (e.g. monogamy) around such AIs is a separate 'ought' question, not settled by passing the test.
“My third company will be AI Girlfriends. I mean it.”
- It is actually all the nuances, quirks, and slight annoyances of a partner that make a relationship worthwhile, illustrated by the Star Trek Voyager episode where Janeway edits out her holodeck lover's snoring only to realize the imperfections were the point — and we won't realize this about AI partners until it's too late.
“It is actually all their nuances and quirks and slight annoyances that make this relationship worthwhile. I don't think we're going to realize that until it's too late.”
- GPT-12 will not be AGI because a loss function based on categorical cross-entropy — maximizing compression — can never get you there; you need reinforcement learning in fancy environments to get something AGI-like.
“When I talked about will GPT12 be AGI, my answer is no. Of course not. I mean, cross-entropy loss is never going to get you there. You need probably RL in fancy environments in order to get something that would be considered AGI-like.”
- Tesla has always been one to two years ahead of Comma on self-driving and probably always will be, because they're not doing anything wrong; they're moving toward an end-to-end approach with a fancy simulator and the same disengagement-based reward, so it comes down to who can build and deploy the system.
“Tesla has always wanted to two years ahead of us. They've always been one to two years ahead of us and they probably always will be because they're not doing anything wrong.”
- Hard things are actually hard — Hotz does not think P equals NP, which is the majority view — but he breaks with the majority in believing factoring is in P.
“I think hard things are actually hard. I don't think P equals NP... I do think factoring is in P.”
- A legacy code base like Twitter's needs refactors before features: you first build trusted tests/CI before touching code, then replace old modules with simpler ones (e.g. 100k lines replaceable by 300 lines of Go), and Hotz believes 50 people could comfortably build and maintain Twitter — the code base grew because engineers get promoted for adding libraries others use, not for simplicity.
“you need re-factors before features. This code base was... Facebook has the best code, then Google, then Twitter.”
- Hotz hasn't figured out what the game of life is yet, but when he does he wants to win; he thinks the game is to stand eye to eye with God, reasoning by the golden rule that just as he wants his own creations to stand eye to eye with him, God would want the same of him.
“I think the game is to stand eye to eye with God.”
- Atheism is silly because we ourselves are powerful creators — every game creator builds worlds — so it's obvious we were created by God; the in-universe origin story (the Big Bang) is no more the actual creation story than Skyrim's myth is how Skyrim was really made by programmers in a room.
“it struck me one day how just silly atheism is. Of course, we were created by God. It's the most obvious thing.”
- Removing Turing completeness from a computation stack allows you to reason about things; the neural network (the ADMO machine) runs identical compute every time and only the data changes, so a branch predictor's 1% error rate stems from the halting problem — meaning you can build a stack that exploits the static, predictable nature of neural network computation.
“Removing Turing completeness allows you to reason about things.”
Specialized hardware fails where general platforms succeed due to software stack maturity.
- ML accelerators are fundamentally a software problem, not a hardware problem; companies like Graphcore, Cerebras, TensTorrent and Groq built good chips but terrible software, and if you can't write a Torch-level performance stack on Nvidia GPUs down to the driver, you certainly can't write one for your more specialized chip.
“AI accelerators are a software problem, not really a hardware problem.”
- Tinygrad is the RISC architecture of ML stacks: it has no primitive operator even for matrix multiplication or convolution, just ~25 ops across four types (unary, binary, reduce, movement), which is an order of magnitude fewer than XLA or Prim Torch (~250) — a MatMul is just shape operations, a multiply, and a reduce.
“What makes Tinygrad different from other neural network libraries is it does not have a primitive operator even for matrix multiplication.”
- There are three computation paradigms: CPUs can do everything fast including compare and branch via branch prediction and speculative execution; GPUs can do arbitrary load/store and data-dependent fetches but are horrendously slow at branching; and DSPs do only statically scheduled adds and multiplies — and 95% of neural networks fit the DSP paradigm of statically scheduled adds and multiplies.
“There's three kinds of computation paradigms that are common in the world today. There's CPUs... Then you have a simpler computation model, GPUs... The third model of computation is DSPs.”
Enduring businesses build durable infrastructure, not ephemeral social platforms.
- Following Paul Graham's essay that startups are either alive or dead, Google is dead and Facebook is alive because Mark Zuckerberg is alive — willing to gut-pivot (e.g. changing the company name) — and by the end of the decade Google won't be the number one webpage anymore.
“Startups are either alive or dead. Google's dead. Facebook is alive.”
- The two companies that have endured through technology are Apple and Microsoft, and what they make is computers and business services; the meme, social, and ad companies come and go, so if you want to endure, build hardware.
“The two companies that have endured through technology are Apple and Microsoft. And what do they make? Computers and business services... All the meme, social ads, they all come and go. But you want to endure, build hardware.”
- As soon as you charge people money for a product like Twitter, they are no longer the product — they become the customer — and you can start building something good for them rather than for the ad agencies; engagement is not the only way to make money.
“As soon as you charge people money, they're no longer the product, they're the customer. 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.”
Hacking communities select against psychopathy more than business culture does.
- In the hacking world Hotz found 95% of people good and 5% bad, but entering the business world with Comma he found the exact opposite — 5% good and 95% bad — a world that promotes psychopathy; the common-sense maxim is that the people who end up getting power are never the ones you want with it.
“when I was in the hacking world, I found 95% of people to be good and 5% of people to be bad... I came into the business world with Comma, and I found the exact opposite. I found 5% of people good and 95% of people bad. I found a world that promotes psychopathy.”
- Google had 'no politics' and the worst kind of politics took over; Comma instead runs as a transparent dictatorship with explicit values and no tolerated dissent, but with free exit — the beauty of companies is that if you don't like the dictatorship, you quit.
“Google had no politics and what ended up happening is the absolute worst kind of politics took over. Comma has an extreme amount of politics and they're all mine and no dissidents is tolerated.”
AI will transform society radically but will not eliminate humanity entirely.
- We will build superintelligence long before we build any robustness in AI; we cannot yet build a machine that can survive in nature like a bird or reproduce itself, so in an AI apocalypse the AIs would likely die out because the silicon stack lacks the reproductive core that the biological stack has.
“I think we're going to build super intelligence before we build any sort of robustness in the AI. We cannot build an AI that is capable of going out into nature and surviving like a bird.”
- AI will likely kill everything we call society today, but it will not kill the human species, because eradicating humanity entirely is incredibly hard; some humans know how to do things and will rebuild after the great AI.
“I think AI kills everything we call society today. I do not think it actually kills the human species. I think that's actually incredibly hard to do.”
- Nuclear weapons didn't kill everyone because they are very hard to deploy tactically: total victory in war is subjugation and domination of people, not the irradiation and eradication that leaves you only an irradiated pile of rubble.
“I think it's actually very hard to deploy nuclear weapons tactically. It's very hard to accomplish tactical objectives.”
Intelligence will eventually be defined as whatever humans reserve for themselves.
- As humans move up the abstraction stack (arithmetic by hand → spreadsheets → macros → accounting software), the top of the stack is the prompt engineer; what a human really is in a company of AIs is a prompt engineer — the do-what-I-mean operator who keeps the AI aligned with their intent.
“As we move further up the abstraction stack, well, what's at the top of the abstraction stack? Well, prompt engineer.”
- Whatever new thing AI can do, everybody first thinks is brilliant, then 20 years later calls it mechanical — chess was 'intelligent' in 1960 but 'mechanical' by 2010; today we think language ability is intelligent, but in 20 years humans will define a new niche (e.g. 'can it reproduce?') because humans always carve out a niche to feel special.
“Whatever new thing that AI can do, everybody thinks is brilliant. And then 20 years go by and they're like, "Well, yeah, but chess, that's like mechanical."”