OpenAI Researcher Highlights How Open-Source Contributions Lead to Top AI Lab Roles

[Noam Brown, a prominent OpenAI researcher known for his work on poker AI, recently detailed how individuals without advanced degrees or extensive publication records can secure positions at leading AI laboratories. His observations, shared publicly, emphasize the value of initiative and open-source projects.]

Brown cited the case of Keller Jordan, who joined OpenAI as a machine learning researcher based on a single open-source blog post. Jordan's approach involved fully open-sourcing his research process, code, and experimental results on GitHub, rather than publishing traditional academic papers. Brown concluded that while opportunities for open research may be more limited now, improving existing papers remains an effective method for demonstrating capability to researchers and securing interview opportunities.

Modern, clean office environment with natural light.

Early Career and Breakthroughs

Keller Jordan graduated from the University of California San Diego in 2020 with dual bachelor's degrees in Mathematics and Computer Science, having no published papers at the time. His initial employment was with an AI content moderation startup.

Jordan's trajectory shifted when he identified an improvement to a paper by Google researcher Behnam. He contacted Behnam, who subsequently mentored him. This collaboration led to a publication at the International Conference on Learning Representations (ICLR).

Later, Jordan's work on "NanoGPT speed run" garnered attention from Tesla AI head Karpathy and OpenAI. This project, which was not a traditional paper, documented a reproducible and quantifiable approach to improving training efficiency.

Developer focused on coding on a laptop, illustrating open-source work.

NanoGPT Speed Run and Its Impact

The NanoGPT project, an open-source framework by Karpathy for GPT training, became a focus for Jordan. He consistently sought to break training speed records, leading to a 3.8-times improvement in the token efficiency of Transformer model training by October 2024. This achievement earned him recognition from Karpathy.

Jordan's objective was to minimize tokens and training time for a fixed model size (124M Transformer) and validation loss target (3.28 val loss). He transformed Karpathy's nanoGPT/llm.c PyTorch training code into a benchmark, reducing the tokens required to achieve the target loss from 10 billion to approximately 2.7 billion. This demonstrated a rigorously verifiable improvement.

A key aspect of Jordan's speed run design was to minimize the cost of experimentation. He compressed the code to 537 lines, enabling installation and execution in 20 minutes on an 8xH100 environment, with a single attempt costing around $8. This design choice aimed to make AI research more accessible, allowing individual researchers, students, and independent engineers to test ideas without significant computational barriers.

Intricate circuit board with glowing data pathways, symbolizing efficient AI training.

OpenAI Recognition and Muon Optimizer

The NanoGPT speed run's solid methodology, reproducible code, logs, and community engagement contributed to its credibility. The verification process included embedding a complete copy of the code within each speed run log file, ensuring transparency and ease of reproduction.

In late 2024, Jordan introduced Muon, an optimizer for neural network hidden layers. Muon set new records for NanoGPT and CIFAR-10 training speeds. It operates by orthogonalizing the update matrix generated by SGD-momentum through Newton-Schulz iteration, producing nearly semi-orthogonal updates and enhancing training efficiency. Its implementation is concise and supports bf16 precision, reducing computational overhead.

Muon demonstrated strong performance across various tasks compared to the AdamW optimizer. While AdamW is effective for models like GPT, LLaMA, and Qwen, its limitations become apparent with increasing model parameters and extended training times. The emergence of Muon suggests a potential fundamental innovation in AI model training.

Jordan officially joined OpenAI in December 2024. Despite Muon's role in his recruitment, Jordan stated in February that he would not publish a paper on it, believing that continued research on the optimizer was more valuable than publishing in what he considered a saturated academic field.

Multiple glowing pathways converging, symbolizing diverse career routes in AI.

Other Paths to Top AI Labs

Noam Brown also highlighted other successful cases of individuals joining major AI companies through unconventional routes.

Sholto Douglas, now at Google DeepMind, was noted for his contributions to the Gemini project despite a low public profile and limited first-author publications. While working at McKinsey, Douglas developed personal AI projects and engaged in insightful discussions on Jax's GitHub, which led to an interview invitation from James Bradbury at Google DeepMind.

Andy Jones, a quantitative analyst, joined Anthropic. Before the widespread adoption of test-time computation, Jones authored a paper comparing the effects of scaling pre-training versus test-time computation. His work was recognized for its clever design choices, custom GPU-accelerated environment, and rigorous ablation experiments.