An Ex-OpenAI Researcher Enters AI Drug Discovery at a $2B Valuation — a Talent Migration Takes Shape

A $2B valuation for a company with almost nothing yet

In March 2024, Miles Wang dropped out of Harvard's CS undergrad program to join OpenAI's reinforcement-learning team; a year and a half later he left with several OpenAI colleagues to start an AI drug-discovery company. It has no website, no public pipeline, and its founder holds no biology degree — yet it is in talks for ~$200M in funding led by Lightspeed at a $2B valuation. The deal isn't final, but the number alone reframes a question: when the very best AI talent leaves the big model labs, where do they go next?

A migration chain taking shape

Wang has stepped into a red-hot field. In the past three months ~$2.7B has poured into AI pharma: Chai Discovery announced a fresh $400M round at a $3.8B post-money; Isomorphic Labs closed a $2.1B Series B in May; and Anthropic entered drug R&D this month after acquiring Coefficient Bio for hundreds of millions and recruiting AlphaFold's John Jumper. A new chain is forming: top talent from the big model labs is migrating systematically into science.

The logic: bring general capability "down" to scientific discovery

Wang's papers — e.g. "FrontierScience: Evaluating AI's Ability to Execute Scientific Research" and "Measuring AI's Ability to Accelerate Biology in the Wet Lab" — reveal his bent: gauging how far general LLMs get in real scientific settings. His startup logic follows: transfer already-validated general cognition into the discovery workflow, rather than building a vertical small model from scratch for drugs alone. The same is happening institutionally: Anthropic's Coefficient Bio buy, Claude Science, and Jumper hire push model capability from the tool layer into the operations of research. As the "low-hanging fruit" in text, code and multimodality is picked, science is becoming the next arena to test AGI generality with high compute and high data.

A clever entry point, but a real gap remains

Wang's entry point — drug repurposing, finding new indications for approved or well-characterized drugs — is shrewd: no need to design a molecule from scratch or face lengthy Phase-I safety first; instead, AI integrates vast literature, expression profiles and real-world clinical data to surface hidden "drug–target–disease" links. In this information-dense, reasoning-over-simulation step, an LLM's evaluation-and-reasoning edge more readily converts to commercial value. But the risk is clear: between "evaluating AI for science" and "actually making a drug with AI" lies the gulf from dry lab to wet lab. Wang's strengths are RL and alignment evaluation, whereas today's AI-pharma mainstream is diffusion models, graph neural networks, protein-structure prediction and molecular dynamics — RL may not be the most fitting path. That is both the cleverness of his repurposing bet and the question this $2B valuation must answer.