George Hotz Raises 5 Million For Tiny Corp
Legendary Hacker George Hotz is launching a new company
Created on May 25|Last edited on May 25
Comment
George Hotz, the pioneer behind Comma.ai, has carved out an impressive niche in the AI space. Comma.ai is a remarkable venture, aiming to revolutionize transportation by developing self-driving technology. The company's flagship product, Openpilot, is an open-source driving agent that currently supports over 60 car models. With thousands of daily active drivers, the technology has clocked in over millions of real-world miles, demonstrating its functionality and dependability.
Tinygrad
George Hotz, in his pursuit to uncover the intricacies of neural networks, developed Tinygrad in his spare time while also building Comma. Initially seen as a personal project, Tinygrad evolved into something much more—a minimalist yet effective deep learning library. Today, it has become an essential tool in the inference space, providing users with a simplified and lightweight platform to explore deep learning. With George's deep understanding of hardware/software integration, the project has grown in capability, and has amassed over 12 thousand stars on Github.
The Problem
Despite having robust hardware, Advanced Micro Devices, Inc. (AMD) has struggled to make its mark in the Machine Learning (ML) landscape. The root cause of this seems to be less about hardware capabilities and more about the shortcomings of its software. Users have reported a range of issues from kernel panics to problems with drivers and stability. AMD's open-source software platform, ROCm, intended to work seamlessly with popular ML frameworks such as PyTorch, has had its own share of obstacles. User experience suggests that the software doesn't always build out of the box, has potential for segmentation faults, and at times, produces incorrect results. In a nutshell, despite the potential of AMD's hardware, the software issues have greatly limited its adoption in the ML domain.
Enter Tiny Corp
George Hotz's newest venture, The Tiny Corp, presents a unique approach to these challenges. The company's primary goal is "to commoditize the petaflop," essentially making high-performance computing more accessible and affordable. Recognizing that the future of AI could be monopolized by entities controlling the majority of the world's computing power, The Tiny Corp aims to prevent this by democratizing access to powerful computing. One of the primary assets of The Tiny Corp is Tinygrad. This software, with its minimalistic design and PyTorch-like frontend, is positioned to utilize hardware more effectively. Importantly, it avoids the complexities of CUDA or other similar platforms that may introduce unpredictable behavior.
The Promise
The Tiny Corp's approach towards solving AMD's software woes is straightforward: bypass existing software complications and write their own. The RDNA3 Instruction Set from AMD is thoroughly documented and the company plans to utilize this knowledge to build a more functional software stack. This approach aims to bring AMD into the ML landscape in a more meaningful way. The company’s short-term goal is to have AMD represented on MLPerf, which is a regular competition to train a set of common ML models, using the Tinygrad framework. This would be a significant achievement, given that AMD's absence in these benchmarks is due to its software, not hardware limitations.
The Road Ahead
The Tiny Corp has an ambitious journey ahead, one that could potentially reshape the landscape of AI and ML compute. They plan to build computers and sell them, aiming to cut into the margins dominated by the likes of NVIDIA. Given George Hotz's track record and the unique approach adopted by The Tiny Corp, the future of AI and ML computing might be on the verge of an interesting shift.
The Announcement:
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.