George Hotz on tiny corp, tinygrad, Open-Source AI Safety, Self-Driving, and Twitter

High-Level Overview

In his third appearance on the Lex Fridman Podcast, programmer and hacker George Hotz, founder of comma.ai and the new tiny corp, ranges across nearly everything he's working on and thinking about. He explains tinygrad, his deliberately tiny neural network framework, and the "tinybox," a home deep learning computer meant to keep AI compute out of the hands of a few gatekeepers. He gives an update on comma's approach to self-driving, recounts his brief, much-discussed internship at Twitter, and lays out his contrarian view of AI safety: the real danger isn't machines going rogue but humans centralizing control of them.

The tone is fast, irreverent, and funny, with Hotz tossing out provocative one-liners ("I don't think I'm conscious") while Fridman pushes back and probes. The conversation swings between deep technical detail (computation paradigms, kernel fusion, refactoring legacy code), philosophy (consciousness, God, the meaning of life), and cultural riffs on AI girlfriends, video games, and the internet's lost golden age.

The Discussion Explores:

• Time, consciousness, and why Hotz calls himself "just a computer program"
• Hotz's version of AI doom: not killer robots but humans amusing themselves to death and bad actors using AI
• AI companions, loneliness, and what happens to relationships and identity online
• tiny corp's mission to decentralize compute, tinygrad's minimalist design, and the tinybox
• comma.ai's learned "human policy" for driving and the race with Tesla
• Why open source is Hotz's answer to AI safety, and his critique of centralized control
• What he learned at Twitter about code bases, incentives, and refactoring
• Programming with AI tools, video games, God as a creator, and effective accelerationism

Deep Dive into Main Ideas

Useful Models, Consciousness, and the Silicon Stack

Fridman opens by asking whether time is an illusion. Hotz answers pragmatically: "I think time is a useful model," and whether something "exists matters less than whether it describes reality usefully. Some things are real to him, like math, Kolmogorov complexity, and the hardness of hard problems (he doesn't think P equals NP). He jokes that one day there will be a George Hotz model better than George Hotz, and that relationships are just humans fine tuning ourselves," inefficiently.

On consciousness he is blunt: "I don't think I'm conscious." He adds, "I think I'm just a computer program." He doesn't deny the feeling of eating a Chicago deep dish pizza, but calls consciousness "a word that atheists use for souls" and says people want to believe so badly that they'd think a rock with a sad face drawn on it is sad. He thinks today's large language models are less conscious than a chicken. He describes two stacks of life: the biological stack, which starts with reproduction, and the silicon stack, which is improving far faster (he contrasts nearly identical presidents in 1980 and 2020 with wildly different computers) but still can't reproduce itself. Building a machine that can fabricate its own chips, he says, is enormously hard.

How AI Might End Society, and AI Companions

Asked whether he agrees with Eliezer Yudkowsky that AI will kill everyone, Hotz says yes, "but maybe for a different reason." Nuclear weapons didn't end the world because they are hard to use tactically, but AI offers a "little red button. His most likely scenario is wireheading: We end up amusing ourselves to death, staring at an infinite, AI-generated TikTok, a nod to the addictive tape in Infinite Jest. He says AI would likely kill everything we call like society today, not the human species, and that survivors might develop strong taboos against technology. Crucially, he argues, It's not the machine that's going to do that. It's other humans using the machine."

The two discuss AI relationships at length. Hotz, who has built chatbots with backstories, says half-jokingly that his third company will be AI girlfriends, and later says he wants AI friends smarter than he is. He admits that "all of us have a deep loneliness." He calls current AI Fiverr-level for artists and "junior engineer level" for coding, but says that's just the start. He shares the meme of "heaven banning, where a user is surrounded by AI accounts without knowing it, and notes that an AI can't truly offer the real difficulty that makes human relationships meaningful, only artificial difficulty that could be turned off with a knob."

tiny corp and tinygrad: Decentralizing Compute

Hotz started tinygrad as a toy project to teach himself, much like Andrej Karpathy's micrograd, then began thinking about AI chips. His worry: if Nvidia becomes a monopoly, "how long before Nvidia is nationalized?" He frames tiny corp as being about making sure power "stays decentralized." Drawing on a description from Jim Keller, he explains three computation paradigms: CPUs (which can compare and branch quickly), GPUs (arbitrary loads and stores), and DSPs (just static adds and multiplies). About 95% of neural networks, he says, fit the simplest DSP paradigm, so tinygrad aims to remove Turing completeness from the stack so that everything can be reasoned about.

tinygrad has no primitive for matrix multiplication; it builds everything from about 25 operations in four types (unary, binary, reduce, and movement ops), compared with roughly 250 for PrimTorch and about 2,000 kernels in PyTorch. He likens it to RISC versus CISC. Laziness lets tinygrad fuse operations and cut memory traffic, and it shows every kernel, flop, and byte used. The problem it solves is porting new ML accelerators, since many AI chip startups built good chips with terrible software. His predicted order of operations: first build a PyTorch-level stack for Nvidia and AMD with far fewer lines, then build an accelerator. tinygrad already runs comma's driving model on Qualcomm GPUs.

He's critical of AMD's software, saying their drivers crashed when he ran their demo apps in a loop, though he had a positive exchange with them and told them "just dumping the code on GitHub is not open source. Open source is a culture." He dislikes the cloud because of "who owns the off switch." The tinybox, priced at $15,000, will pack almost a petaflop of compute, over 100 gigabytes of GPU RAM, and six GPUs into a quiet machine that runs off a single wall outlet and can run a 65-billion-parameter LLaMA model in real time. He's also planning a "tiny rack" powered at car-charger level.

comma.ai and the Learned Driving Policy

Hotz says comma solved the core problem "years ago: building a model that outputs a human policy for driving, rather than hand-coding it as companies like Waymo and Cruise do. The next step is essentially Drive GPT: a learned simulator trained on driving data to predict the next frame, conditioned on actions, in which they can then run reinforcement learning. The reward isn't how human-like the driving is but whether a human would disengage, which he calls asking exactly the right question": what makes customers happy.

He says Tesla has always been about two years ahead and will probably stay there "because they're not doing anything wrong." comma has about 5,000 daily active users, a 40-petaflop compute cluster (he proposes 20 petaflops as "one person" of compute), is profitable, and builds its boards in-house in San Diego. The comma three costs $1,499 with a 30-day money-back guarantee. He says comma's mission is to solve self-driving cars while shipping intermediaries, not to sell a million boxes, and hints at robot bodies powered by a tinybox in your home.

Open Source as AI Safety

Hotz's central political thesis: "there's things that centralize power and they're bad and there's things that decentralize power and they're good." He argues that AI safety advocates who want one controlled, aligned model "are going to build the exact thing they fear," and that the way to avoid a paperclip maximizer is for everyone to have AI. He accuses OpenAI of using safety to hype GPT-2 and of refusing to disclose basic facts about GPT-4. "Half of these AI alignment problems are just human alignment problems," he says.

Fridman pushes back that language models could lower the barrier to violence, such as building a bomb or hiring a hitman. Hotz counters that anyone serious enough to cause harm could already find the information, and that the answer is to "Give intelligence to everybody." He'd respect a position that nobody should have AI, but not one where only trusted authorities do. "I'm not worried about alignment between AI company and their machines. I'm worried about alignment between me and AI company." His practical hope is personal AI "firewalls that protect your mind from manipulation, like an ad blocker for propaganda. He says he found 95% of people good in the hacking world and 95% bad in business, and that I will always take anarchy over tyranny." He praises Mark Zuckerberg for open-sourcing LLaMA and predicts Google won't be the top web page by the end of the decade, betting on a startup or Microsoft's Bing.

Lessons from Twitter

Hotz deleted his first Twitter account in 2010 because the algorithm rewards outrage. He took the internship because he believed in a principled commitment to free speech, though he says what emerged was "a monarchy deciding what to ban" instead of an oligarchy. He wants individualized, transparent filtering, such as a "no ad hominem" filter based on the hierarchy of argument, rather than top-down censorship.

His main technical recommendation to Elon Musk was "refactors before features." He ranks code quality at Facebook first, then Google, then Twitter, and maintains that 50 people could build and maintain Twitter, explaining that much of the hate comes from people who, if doing things were possible, would "have to do some soul searching." He found that Twitter's promotion system rewarded writing libraries others used, like an Nginx replacement nobody needed. At comma, engineers must explain their code, and if he sees a simpler way, they rewrite it. He calls comma a transparent dictatorship with free exit. His refactoring plan: first build tests you trust, especially a replay system like comma's "process replay" at the service boundaries, then replace large modules with simpler code as long as the tests pass. He says a refactor could take six months to a year and make feature development ten times faster, but admits he isn't the right person to run Twitter.

Programming, AI Tools, and the Future of Work

Hotz says he currently does none of his programming with GPT. "My programming speed is not at all limited by my typing speed," and he wants a tool that spots bugs, not one that writes "close to correct code." He loves mypy, the Python type checker, and wishes Python had a runtime flag to enforce types. He calls programming "tool complete" in a way driving isn't, and says if Copilot saves you lots of boilerplate, your framework is probably bad; tinygrad runs LLaMA and Stable Diffusion in about 2,700 lines. He predicts he'll use AI tools a little in two years and heavily in five, and that humanity's niche will keep shrinking, pointing to the sci-fi novel The Metamorphosis of Prime Intellect and its sequel. At tiny corp, he plans remote work where the only way to get a job is to contribute on GitHub, and he envisions AI agents submitting pull requests. He sees prompt engineering as the top of the abstraction stack and says a cloud AI will always be aligned with its company, "And that's why you have to buy a tiny box."

Games, God, Accelerationism, and Winning

Hotz names World of Warcraft as the greatest game for how it opened his mind in 2005, alongside Skyrim and GTA V, and is excited for worlds filled with intelligent characters. He criticizes Meta's Quest for greeting users with a login screen instead of a world. On God, he says: "I never really considered how absolutely like silly atheism is because like I create worlds." Game makers create worlds, so of course we were created, and he finds comfort in that.

He calls effective altruism "a terribly evil ideology," arguing that charity is less effective than building companies, and says "accelerationism is the only thing that's ever lifted people out of poverty." He opposes universal basic income and wants food, housing, and especially energy to be dirt cheap, saying the political position he can't stand is people who believe we should use less energy. Asked about Andrej Karpathy, whose teaching inspired tinygrad, Hotz praises him as someone who "just wants to teach you," while admitting he himself wants to prove he's smarter. His meaning of life is "to win," and the ultimate game, he says, is to stand eye to eye with God.

Conclusion

George Hotz's worldview ties together his many projects: power should be decentralized, code should be as simple as possible, and humans, not machines, are the real alignment problem. tinygrad and the tinybox are his bet on keeping AI compute in ordinary people's hands, comma is his long game on learned self-driving, and his Twitter stint reinforced his conviction that trustworthy tests and relentless refactoring beat piling on features. Whether you agree with his takes on AI safety, effective altruism, or consciousness, the episode is a lively tour through the mind of an engineer who wants to build things, simplify them, and win, fittingly closed by Lex Fridman with Einstein's line that everything should be made as simple as possible, but not simpler.

This is a summary of "#387 – George Hotz: Tiny Corp, Twitter, AI Safety, Self-Driving, GPT, AGI & God" from Lex Fridman Podcast, giving you the GIST so you can decide if you want the REST. Listen to the full episode.