# Andrew Ho leaves OpenAI to build reinforcement-learning datasets for scientific AI > Source: > Published: 2026-07-30 00:35:28+00:00 [Andrew Ho (@andrewho03)](https://x.com/andrewho03) left OpenAI on July 29th after eight months and said in [an announcement on X](https://x.com/andrewho03/status/2082615798011744270?s=46) that he is starting a company to produce reinforcement-learning datasets for AI labs, beginning with biology and statistical reasoning. Ho is carrying a specific piece of his OpenAI work into the new venture. He co-authored [GeneBench-Pro](https://openai.com/index/introducing-genebench-pro/), a benchmark released on June 30th that tests whether AI agents can navigate messy biological data, select appropriate analyses and revise their approach when initial assumptions fail. The company will initially build two types of datasets, according to Ho. The first will target long-horizon scientific reasoning, giving agents complex datasets and answers that can be graded against known ground truths. The second will focus on routine scientific work, including multimodal questions involving images of cell-culture plates and Western blots. Ho said he eventually plans to move into chemistry, materials science, healthcare and office work. ### Turning an evaluation into training data GeneBench-Pro provides a working example of what Ho wants to sell. The [preprint](https://www.biorxiv.org/content/10.64898/2026.06.29.735386v1) contains 129 problems across 10 scientific domains and 21 subdomains. OpenAI said 82 problems received external expert review, 10 representative questions were released publicly and 50 were allocated for independent benchmarking by Artificial Analysis. OpenAI reported that GPT-5.6 Sol passed 28.7% of the problems at its highest reasoning setting, rising to 31.5% in Pro mode. Each problem uses simulated data, allowing its designers to control the underlying causal structure and grade an agent against a known answer while still requiring several rounds of analysis and judgment. That construction is central to Ho's pitch. Real scientific work often permits several defensible methods, which makes it difficult to turn into reinforcement-learning data with a clean reward signal. By generating realistic data and controlling how it was produced, a vendor can create tasks that preserve the ambiguity of research while retaining a deterministic grading target. Ho said the new company can generate thousands of such problems. That is his claim, rather than a demonstrated production figure, but GeneBench-Pro gives prospective customers a detailed sample of the methodology. The announcement also doubled as a sales pitch. Ho invited labs with data requirements to contact him and said the venture would offer "industry-standard pricing and terms." His proposed advantage is proximity to the buyer: Ho said his work at OpenAI covered the process from data procurement through model training, giving him direct knowledge of which datasets labs can incorporate into large training runs. ### A founder with biotech data experience Ho's background predates his work at OpenAI. He earned a bachelor's degree in chemistry from the University of Washington in 2016, according to [his professional profile](https://www.toptal.com/developers/resume/andrew-ho), and worked at genomics company BGI before joining BioAge Labs in 2017. At BioAge, Ho says he built machine-learning pipelines for genomics and proteomics data, managed more than 100 gigabytes of biological information and worked on drug-discovery and clinical-trial programs. His published work includes [research on biomarkers associated with human mortality](https://pubmed.ncbi.nlm.nih.gov/28858850/). His [personal site](https://andrewho.xyz/) lists five years at BioAge before subsequent work in finance and at OpenAI. That combination shapes the company's initial market. Ho understands both the scientific workflows being encoded and the technical requirements imposed by a frontier model lab. The immediate product is data, but the harder work sits in translating expert judgment into environments where a model can explore, fail and receive a reliable score. Ho argues that model capabilities remain "spiky," performing well in closed and heavily trained domains while still producing weak results on ordinary, contextual work. He expects frontier labs to spend more than $100 billion on targeted data acquisition as scaling alone produces fewer capabilities for free. He supplied no timeframe or methodology for that figure, making it a statement of market conviction rather than a testable forecast. The narrower business case is already visible. GeneBench-Pro showed that a frontier model could solve fewer than one-third of a carefully constructed scientific reasoning suite. Ho is betting labs will pay an outside supplier to create the training material needed to close that gap, and that his eight months inside OpenAI gave him enough knowledge to build data those labs will actually use.