Smart as large language models have gotten, they still need to practice — and one former OpenAI researcher thinks that’s a business opportunity.
Andrew Ho, a former OpenAI researcher and co-author of GeneBench-Pro, is leaving the lab to build a startup that makes high-quality reinforcement learning data for large models: training tasks that let a model practice a problem and then get scored on the result. The company’s first products will focus on biology and statistical reasoning. The name? Not public yet.
Ho argues that LLMs remain badly lopsided. Even in programming — an area that gets massive investment — models can work through practice problems and patch code, but real delivery still often takes a human to clean up.
And many real-world tasks don’t have suitable practice problems at all. They depend on specific context, and it’s hard to automatically grade whether the model got the answer right.
His approach is to generate data and questions that mirror real scientific research…