← Originals
George Hotz

George Hotz

6 interviews

What they argue

George Hotz is an engineer and researcher who focuses on the structural reality of computation, machine learning infrastructure, and software architecture. He works in the territory of optimizing ML frameworks by reducing complexity and moving away from Turing-complete systems toward more predictable, high-performance designs.

  1. The Curse of Turing Completeness

    By making every layer of the software stack Turing complete, we inherit the Halting Problem, which makes it impossible to guarantee what a piece of code will do. Hotz argues that this approach is fundamentally flawed for building reliable software.

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

  2. The RISC-ML Design Principles

    Because neural networks are fundamentally static and predictable, we can achieve superior performance by stripping away the 'Turing-complete' overhead of current frameworks. This involves focusing only on the ADD/MUL operations required for neural networks.

    From Commoditizing the Petaflop — with George Hotz of the tiny corp (episode)

    • A related idea Andrej Karpathy — Leaky Abstractions in Neural NetworksUnlike standard software where APIs hide complexity, neural network components like Batch Norm or RNNs do not magically solve problems and can fail silently. Success…A Recipe for Training Neural Networks

  3. CISC vs. RISC in ML

    Current ML frameworks like PyTorch and XLA are analogous to CISC systems because they are designed to handle any possible computation. Hotz argues these are over-engineered and that we need a 'Reduced Instruction Set' to achieve true performance.

    From Ep 18: Petaflops to the People — with George Hotz of tinycorp (episode)

  4. The Anti-Turing Stack

    To build systems that are actually reason-capable and predictable, one must map out the entire stack and remove Turing completeness. This architectural shift is necessary to move past the hype of LLMs into structural reality.

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

    • A related idea Demis Hassabis — The AlphaZero ArchitectureTo build a system capable of complex reasoning, one should use a three-layer stack: a massive transformer for general knowledge (The Prior), a World Model, and a…Demis Hassabis — Scaling, superhuman AAIs, AlphaZero atop LLMs, AlphaFold
    • A related idea Aman Sanger — The Abstraction Stack ModelTo manage complex systems with AI, developers should use a mental model of navigating different levels of code. This includes viewing the system as a set of logical…#447 – Cursor Team: Future of Programming with AI

  5. Static Core Identification

    A framework should acknowledge that neural network operations are static loops. This principle is used to determine if a framework is over-engineered.

    From Ep 18: Petaflops to the People — with George Hotz of tinycorp (episode)

Every idea names the summary it came from. Written by a local model from this mind’s own work, never from memory.

Interviews 3

Open George Hotz in Originals