Ex-OpenAI Employee Starts Company for High-Quality RL Datasets, Citing Poor LLM Generalization

rickasaurus · x · 2026-07-30

An departing OpenAI employee noted that despite billions invested, LLM generalization remains poor with "spiky" capabilities—such as coding models still requiring manual PR cleanups. He highlighted that public data is exhausted and hard to synthesize.

To solve this, he is launching a new company focused on producing high-quality reinforcement learning (RL) datasets. Industry experts echoed this, emphasizing that analytical and clickstream data is now extremely valuable as it cannot be generalized across companies, making it key to future model improvements.

Related event: Ex-OpenAI Staffer Leaves to Build High-Quality RL Data Firm(3 posts)→

Original post →

More from Companies & People

Companies & People channel →