Cresta rebuilt its stack every 2 years: scaling beats patching, says Tim Shi

agihouse_org · x · 2026-11-21

Part 3 of AGI House's thread: at early OpenAI, Tim Shi worked with Andrej Karpathy on keyboard-and-mouse web agents nearly a decade before computer use shipped — it failed for lack of foundation models. "The model is the cake and RL is the icing," he now says. At Cresta the team rebuilt its stack every two years (LSTMs→BERT→GPT-2/3), teaching him scaling beats patching — hence Recursive spends its first 18 months purely on research.

Related event: Early OpenAI Member Tim Shi on AI-Driven Autonomous Research(3 posts)→

Original post →

More from AGI Musings

AGI Musings channel →