OpenAI replaced job applications with a machine learning contest

A compression challenge launching this month is how OpenAI plans to recruit its next class of junior researchers — and the leaderboard is already getting serious.

The Forbes headline reads: "OpenAI Is Now Hiring $500,000 Jobs. No Resume Required." That framing is technically defensible and almost entirely misleading.

What OpenAI actually launched on March 18 is called Parameter Golf — a research competition with a tightly constrained problem: build the best language model that fits inside 16 megabytes, weights and training code combined, trained in under 10 minutes on eight H100 GPUs. Performance is judged by bits-per-byte compression on the FineWeb validation set — a tokenizer-agnostic metric that measures how well a model predicts real text. No resumes required because the leaderboard is the resume.

The challenge runs through April 30. OpenAI says it plans to hire a "small cohort" of early-career researchers from the results in June — current undergrads, recent grads, and Olympiad medalists. These are not $500,000-a-year jobs on day one. Research Scientists at OpenAI earn between $245,000 and $685,000 in base salary according to H-1B filing data compiled by Business Insider, but that salary band reflects senior hires, not fresh graduates coming through a competition pipeline.

The real story is what's happening on the leaderboard.

Within 48 hours of launch, five serious submissions had appeared. The current top score is 1.1428 bits per byte, held by a competitor named thwu1 using a combination of int5 quantization, a multilayer perceptron architecture, and something called BigramHash — a tokenization approach that doesn't appear to be widely documented, suggesting it's original work developed for this competition. The second and third entries use int6 quantization with quantization-aware training (QAT), SmearGate activation functions, and stochastic weight averaging. These are not toy entries from people who read the announcement and clicked submit. They reflect real engineering judgment about the efficiency frontier.

The technical framing matters. OpenAI describes Parameter Golf as solving what it calls L(N) optimization — given a fixed parameter budget, what is the best achievable loss? It's one of three related challenges in the scaling-law family: L(N) (this challenge), L(T) (minimize training time to a target loss, the focus of Keller Jordan's NanoGPT Speedrunning challenge, which OpenAI cites as direct inspiration), and L(D) — minimize training data. The three challenges together form a kind of empirical map of the trade-off space that drives every decision in large model development.

That's not incidental. OpenAI is getting early signal on which compression and efficiency techniques actually work, submitted as public pull requests on GitHub, complete with training code, logs, and explanations. Whether the company is also treating the challenge as cheap R&D is a fair question — the techniques on this leaderboard (int5/int6 quantization, BigramHash, QAT) are directly relevant to on-device inference, mobile deployment, and edge AI, all areas where model size is a hard constraint.

The hiring philosophy OpenAI is making explicit here is worth noting. The company wrote in its challenge documentation: "Many researchers at OpenAI first distinguished themselves through elite mathematics and programming competitions. The Model Craft Challenge is designed in that spirit." That's a statement of values, not just a recruitment tactic. Demonstrated ability over credentials. Move the leaderboard, get the job.

The timing isn't accidental. Meta has been pulling senior OpenAI researchers with compensation packages that have been reported at up to $300 million in some cases. The Decoder reported that researchers including Trapit Bansal and Shuchao Bi — both involved in o-series and GPT-4o voice mode work — moved to Meta. OpenAI's defensive option at the top of the market is limited. So it's expanding the bottom of the funnel instead: find talented people before anyone else knows who they are.

The $1 million in compute credits OpenAI is offering through Runpod as prize support is structured as grants that require justification and a ChatGPT account — not distributed automatically. The real prize, if OpenAI follows through, is the June cohort offer.

What to watch: the leaderboard through April 30 will show whether the leading techniques cluster or diverge. If BigramHash and SmearGate become the consensus approach, that's a signal about where quantization-aware efficiency methods are heading. If the top entries keep diverging, it suggests the search space is genuinely open. Either way, the techniques that move into the top five over the next six weeks are worth tracking — they're coming from people who are good at this specific problem and have no incentive to hold anything back.