George Hotz

George Hotz, also known as geohot, created the deep-learning framework tinygrad and founded tiny corp to make AI compute more accessible. His work combines a compact tensor framework with the compiler, runtimes, and drivers needed to execute models on GPUs. He describes the company's mission as “to commoditize the petaflop.” Tinygrad also connects to his earlier work at comma.ai: it became the framework running openpilot's driving model.[1][2][5][3]

Hotz started tinygrad in October 2020 as a toy project to learn about neural networks. He initially imposed a thousand-line limit to force economy, then removed it when the constraint became “a little code golfy.” His aim was readable machinery with little boilerplate, rather than short code at any cost. As he put it in his June 2023 interview, “I'm trying to build something that manages complexity.”[2][1]

In May 2023, Hotz described tinygrad's intermediate representation as twelve operations, expressing matrix multiplication and convolution through multiplication, sums, and movement operations such as reshape and permute. It supported fixed indexing—x[3]—while excluding data-dependent indexing—x[y]. His technical argument was that predictable computation and memory accesses allow a compiler to plan execution more thoroughly; passing that work into general, Turing-complete kernels sacrifices that visibility.[2]

Lazy execution enables operation fusion. Rather than immediately dispatching each tensor operation, tinygrad can retain operations in a local graph and combine them. Hotz illustrated this with A × B + C: a fused kernel reads the three inputs, performs the multiplication and addition in registers, and writes the result to memory. That avoids writing an intermediate A × B result and reading it back for a second kernel. He qualified laziness as a best-effort execution choice, rather than a guarantee that every operation remains deferred.[1]

Hotz also made execution inspection and correctness central to the developer experience. In 2023, he described graph output showing the kernels actually dispatched to the GPU, and command-line reports exposing each kernel's FLOPs, memory accesses, and efficiency. Tests compared small tinygrad and PyTorch examples in both forward and backward passes. His standard for continuous integration was concrete: “Like, I want that green check to mean I can merge this, right?” If a bug passed those checks, he considered fixing the tests a top priority.[1]

Simplicity did not mean immediate speed leadership. In the June 2023 interview, Hotz reported that tinygrad was about five times slower than PyTorch on NVIDIA, while about twice as fast as Qualcomm's library for openpilot's GPU workload. He described PyTorch's developers as very smart people making different tradeoffs, including choices suited to NVIDIA performance and a much larger engineering effort.[1]

Hotz announced tiny corp in a May 24, 2023 post titled “the tiny corp raised $5.1M.” His strategy was to build software for existing hardware before attempting a custom AI chip. In the June interview, he said seeing Sam Altman testify before Congress had strengthened his commitment: he feared restrictions on access to machine-learning compute and wanted an organization that would make chips available to anybody.[2][1]

AMD hardware was the initial target, but driver reliability shaped the practical limits of that plan. Although his funding post proposed a framework, runtime, and driver, Hotz subsequently distinguished feasible user-space runtime work from the much larger task of rewriting a kernel driver. He reported that AMD demonstration programs crashed his computer, that he emailed CEO Lisa Su, and that subsequent contact with AMD engineers produced a prerelease ROCm driver that fixed the kernel panic. The experience reinforced his view that open-source development requires responsive maintainers as well as published code.[2][1]

For tinybox, Hotz's proposed local AI computer, the engineering problem extended beyond choosing GPUs. In 2023, he described fitting six large cards into a machine, preserving PCIe interconnect bandwidth, keeping noise low enough for use under a desk, and operating from one wall outlet. Limiting GPU power was one of his design choices for meeting those constraints.[1]

In his December 2025 retrospective, Hotz credited both his own five years of work and years contributed by others. He described a stack containing a frontend, graph compiler, runtimes, and drivers, with work beginning to remove LLVM so that driving AMD GPUs would require only pure Python dependencies. He also reported an AMD contract targeting MI350X participation in MLPerf for Llama 405B training; this was a stated target, rather than an announced completed benchmark result.[5]

Hotz made high-quality tinygrad pull requests the route to employment at tiny corp. His May 2023 announcement called for in-person work in San Diego, but by the June interview he favored a remote company. He said technical screens and programming challenges gave him little predictive information compared with actual project contributions. His proposed progression was to solve bounties, then propose useful bounties, then join full-time with equity. In December 2025, he continued to describe hiring through repository contributions and a largely self-directed job.[2][1][5]

Hotz's 2023 view of AI coding emphasized tools that amplify a programmer's work. He argued that expecting a model to produce a working program in one pass ignored how humans program through iteration. He wanted models to discuss answers and help generate the fastest program meeting a specification—a task closely related to tinygrad's compiler work. In explaining who directs whom, he used the idea of an “API line,” while explicitly acknowledging that the concept was not his invention.[1]

By February 2026, Hotz framed personal learning as the long-term reason for local AI hardware. He argued that shared, fixed-weight models favor cloud batching, while meaningful learning for each user or organization could favor local computation. His proposed tinybox would update model weights through interactions with its owner. He explicitly described that capability as many years away; the contemporary focus remained large language-model training and generic infrastructure for driving GPUs.[6]

Before tinygrad, Hotz worked on unlocking consumer devices and later on driving models. In August 2007, he published an iPhone-unlocking method that allowed the phone to work on another network. Computerworld reported that he traded the second phone he unlocked to CertiCell founder Terry Daidone for a Nissan 350Z and three 8GB iPhones. He planned to give those phones to people who had helped with the unlock.[9][7]

His PlayStation 3 security work led to litigation with Sony Computer Entertainment America. Sony alleged that his published security information and software violated federal law; Hotz denied wrongdoing. An April 11, 2011 joint statement announced settlement, following an agreement in principle on March 31, and stated that he had consented to a permanent injunction.[8]

Hotz recalled entering AI through the Hutter Prize around 2014, when he tried to build compression models of Wikipedia. He connected that problem to language modeling and came to understand neural networks as compressors. In 2023, he said language models had progressed further than he expected, while endorsing Rich Sutton's “Bitter Lesson” against relying on hand-engineered solutions.[1]

At comma.ai, which Hotz founded, his model research was distinct from collaborators' vehicle-control work. Riccardo Biasini credited himself with developing the automated lateral and longitudinal controls for comma's first self-driving car after joining in 2016. In September 2018, Hotz moved from CEO to Head of Research Team and promoted Biasini to CEO. Hotz's announced remit was machine-learning models for openpilot's driving decisions and driver-attention tracking.[7]

In October 2022, Hotz announced that he was taking time away from comma. He distinguished the satisfaction of solving technical problems from the planning and patience needed to operate a maturing company: “It’s no longer a race car, it’s a boat.” He credited the people capable of running that organization and described a prospective Tiny Corporation as a hobby. His subsequent commitment to tiny corp and compute accessibility marked a change from that earlier characterization.[4][1]