Andrew Ho

Former OpenAI researcher (eight months, 2025–26) who left on 29 July 2026 to found a company selling high-end reinforcement-learning datasets to frontier labs, starting with biology and statistical reasoning. A new entrant making the same bet as this atlas: that expert-built data, not scale, closes the capability gaps.

Background

Trained as a chemist (University of Washington, 2016) and spent his early career in computational biology: genomics work at BGI, then roughly five years at BioAge Labs (from 2017) building ML pipelines over genomics and proteomics data, with published biomarker research on human mortality. He then moved through fintech and startups (Ambient Finance 2022–23, Ivy Natal 2024–25) before joining OpenAI.

At OpenAI he worked on life-science evaluations — he is a co-author of GeneBench-Pro (released 30 June 2026), GeneBench and LifeSciBench, and contributed to Humanity's Last Exam. That benchmark work is what his company is built on: he saw frontier models reach only about 30% success on bioinformatics tasks.

What they run now

Building a company that produces two kinds of RL data for frontier labs: long-horizon scientific reasoning tasks with controlled ground truth (bioinformatics, statistical analysis), and routine lab work including multimodal judgement such as evaluating photos of cell cultures or Western blots. Planned expansion into chemistry, materials science, healthcare and general knowledge work. No name, funding or hires announced as of September 2026.

Career

2026 – presentFounder, New RL-dataset company (unnamed) — Biology and statistics datasets for frontier labs
2025 – 2026Researcher, OpenAI — Life-science benchmarks: GeneBench-Pro, GeneBench, LifeSciBench
2024 – 2025Ivy Natal
2022 – 2023Ambient Finance
2017 – 2022ML / computational biology, BioAge Labs — ML pipelines for genomics and proteomics
—Genomics, BGI — Per RuntimeWire
2016University of Washington, Chemistry

On the record

Why data, not scale · 2026-07

LLMs generalise poorly and have 'spiky' capabilities even in heavily funded areas like coding; most economically valuable skills are missing from existing datasets. source

Size of the market · 2026-07

Forecasts that frontier labs will spend more than $100B on targeted data acquisition (no timeframe or method given). source

Judgement-heavy work · 2026-07

Models excel at verifiable tasks like maths and code but struggle where 'research taste' and context matter; most work is not easily encoded into a gradable environment. source

Lab valuations · 2026-07

Thinks frontier labs are overvalued and stuck in a 'Red Queen's race' of spending; advised colleagues to take liquidity in tender offers. Says he holds ~$700K of OpenAI stock he cannot sell before an IPO. source

Recursive self-improvement · 2026-07

Sceptical that recursive self-improvement is a near-term breakthrough. source

Interviews & talks

‘I’m just stuck’: Meet the former OpenAI researcher sitting on $700K of equity that he says is overvalued — Fortune, 2026-07-30
Ex-OpenAI researcher bets $100 billion will flow into training data — The Decoder, 2026-07
Andrew Ho leaves OpenAI to build reinforcement-learning datasets for scientific AI — RuntimeWire, 2026-07

Connections

OpenAI life-science evals team — Former colleagues; co-authors of GeneBench-Pro

Why it matters here

He is the clearest recent example of the move Julian is making: someone who built evaluations inside a lab, saw exactly where models fail, and left to sell the missing data back. His thesis — judgement-heavy, hard-to-grade work is where the money goes — is the same argument as the taste lane, applied to science. Worth watching as (a) proof labs will buy from a two-month-old specialist, (b) a peer to compare notes with on pricing and first contracts, and (c) evidence of what an ex-lab founder's pitch sounds like.

How to reach

Public and outspoken on X (@andrewho03) and posts long threads about his thesis — replying substantively to one of those is the natural opening. He has not named co-founders or investors, so a founder-to-founder note about specialist data economics (not a sales pitch) is likely to land. His personal site lists a contact address. Speculation: he may be hiring domain experts and could be open to sharing supply-side lessons.

What we could not establish

Company name, co-founders and funding not announced as of 19 Sep 2026.
No LinkedIn profile confirmed.
Exact role at OpenAI (team, title) not public beyond benchmark authorship.
Roles at Ivy Natal and Ambient Finance not stated.