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

Here is the English translation of the provided Chinese investment research notes.

At a Glance

George Hotz is the founder of comma-ai (autonomous driving) and Tiny Corp (TinyGrad framework). The main thread of this episode revolves around his radical views on the path of AI development: He firmly believes that the real threat to AI safety is not machines spiraling out of control, but humans using AI for centralized control; therefore, open source is the only correct defense strategy.

Decentralization vs. Centralization: The Power Struggle in the AI Era

George Hotz argues that the centralization of computing power (especially NVIDIA's monopoly) is the greatest current threat, and Tiny Corp's mission is to decentralize this power through an open software stack.

• Monopoly Risk: Hotz warns that if NVIDIA achieves a 10x advantage over its competitors in the AI chip space, its owners would gain immense geopolitical power, even facing the risk of "nationalization." Citing Jim Keller, he notes that 95% of current neural network operations fall under the "DSP paradigm" (statically scheduled addition and multiplication), providing a theoretical basis for designing simpler, more specialized hardware.
• TinyGrad's Philosophy: TinyGrad is a minimalist neural network framework with a core opset of only about 25 operations, far fewer than PyTorch's 2000 or PrimTorch's 250. Hotz compares it to a RISC architecture, simplifying inference and optimization by eliminating Turing completeness. He explains: "Once you get rid of Turing completeness, you can reason about things. Rice's theorem and the halting problem don't apply to finite state machines."
• Hardware Strategy: Hotz believes that AI accelerators are fundamentally a software problem. Any new chip company that cannot write a high-performance software stack for NVIDIA GPUs will not succeed on its own chips. TinyGrad's strategy is to first prove it can achieve equivalent performance on NVIDIA and AMD before considering designing a dedicated chip. He predicts TensTorrent will eventually pivot to the RISC-V CPU business.

Autonomous Driving: From Imitation to Reinforcement Learning

Hotz believes comma-ai has solved the core problem of "how to build a model that outputs a human driving policy," and the current focus is shifting from "imitation learning" to "reinforcement learning."

• Drive GPT: comma-ai is building a driving simulator called "Drive GPT." It uses a VQ-VAE and Transformer architecture to learn to predict the next frame in a self-supervised manner. The simulator can accept a "state-action" pair and output the next state.
• Reward Function: The key innovation is that instead of simply imitating human behavior, they use reinforcement learning (RL) to optimize a more direct reward function: "Would a human disengage because of this behavior?" The goal is to find a balance that is both comfortable and not overly cautious.
• Current Status and Bottlenecks: Hotz states the system is not yet fully closed-loop, with some "dumb bugs" and scaling issues. comma-ai's compute cluster has been expanded to about 40 petaflops (which he jokingly calls "two people's" worth of compute). He believes Tesla is 1-2 years ahead in autonomous driving, but "they aren't doing anything wrong."

AI Safety, Openness, and the Future of Humanity

Hotz is highly critical of the mainstream AI safety narrative, arguing its core is about "control," and that open source is the only way to prevent AI from being misused.

• The Real Threat: Hotz believes the way AI kills everyone is not through machines becoming conscious, but "humans using machines against other humans." He proposes the "wireheading" scenario—where humans are captivated by infinitely perfect personalized content (like "infinite TikTok") and eventually "entertained to death"—as a more realistic threat.
• The Case for Open Source: He refutes Eliezer Yudkowsky's view that AI will inevitably destroy humanity, arguing, "If you give AI to everyone, there are more good people than bad people, so the good people will win." He criticizes OpenAI for using "AI safety" as an opaque excuse and points out: "Half of these AI alignment problems are just human alignment problems. You don't want to align the machine; you want to align me."
• Computation and Life: Hotz proposes a distinction between the "biological stack" and the "silicon stack." The core of the biological stack is "replication," while the silicon stack cannot yet self-replicate. He believes superintelligence will arrive before true "robustness" (like a bird's ability to survive in nature) is achieved. Therefore, after AI destroys human civilization, it would itself perish due to a lack of robustness, while a small number of surviving humans would rebuild society with a strong taboo against technology.

Position Moves

Ticker	Guest Sentiment	Key Data
NVIDIA	Risk Warning	Monopoly risk, potential for nationalization; its software stack is the biggest barrier for competitors
AMD	Risk Warning / Neutral	7900 XTX drivers have kernel panic issues; Hotz is in communication with them, hoping to improve open-source culture
Tesla (FSD)	Neutral / Acknowledgment	1-2 years ahead of comma-ai, on the right track with its end-to-end approach
Meta (Llama)	Bullish	Open-sourcing Llama is a "heroic act," representing the correct path for AI development
OpenAI	Risk Warning / Criticism	Uses AI safety as an opaque excuse, is a "bad team"; GPT-4's parameter count is being concealed
Google	Risk Warning	A "dead" company, lacking innovation and "buttons"; its search position could be disrupted
Microsoft	Bullish	An "alive" company, performing best in open source and developer tools, likely to replace Google
TensTorrent	Prediction	Predicts it will pivot to the RISC-V CPU business, not AI accelerators

Judgments Worth Remembering

1. "The real threat to AI safety is not the machine, but other humans using the machine against you." — Hotz believes machines have no motive to harm humans, but humans do. Therefore, the only way to prevent AI from being misused is to give it to everyone, not let a few control it.

2. "Open source is the only way to prevent the 'paperclip maximizer'." — He argues that the attempt to centrally control powerful AI would precisely create the disaster it tries to avoid. Decentralizing power is the guarantee of safety.

3. "The core philosophy of TinyGrad is to eliminate Turing completeness." — By simplifying neural network operations to about 25 basic operations, TinyGrad makes reasoning, optimization, and porting to new hardware unprecedentedly simple, as theoretical limitations like the halting problem no longer apply.

4. "The ultimate reward function for autonomous driving is 'would a human disengage'." — comma-ai's Drive GPT project no longer merely imitates human driving but uses reinforcement learning to learn a driving strategy that makes human passengers comfortable and unwilling to take over.

5. "An AI chip company must first prove it can write a good software stack for NVIDIA." — Hotz asserts that AI accelerators are a software problem. If a company cannot write high-performance software for NVIDIA's general-purpose hardware, it will absolutely not succeed on its own more specialized chips.

6. "Silicon-based life lacks the core capability of the biological stack: self-replication." — He points out that all living things can self-replicate entirely without civilization's support, but no robot can. This means that even if AI destroys human civilization, it would itself perish due to a lack of robustness.

7. "Twitter's codebase needs a refactor, and the first step is to build a trusted test." — Hotz believes that before adding any new features, an offline, trustworthy testing system (similar to comma-ai's "process replay") must be established. This is a prerequisite for safely refactoring any legacy codebase.

8. "Effective altruism is a terrible ideology." — Hotz believes that pure utilitarianism (e.g., sacrificing the few to save the many) is dangerous, while "effective accelerationism" (lowering the cost of living through technology, such as energy and food) is the true way to improve human well-being.