Recursive SI's Flint: from zero pre-training to shipped model in under 4 months
thursdai_pod · x · 2026-10-27
On the ThursdAI podcast, Tim Shi of Recursive SI explained how the Flint model came together in under 4 months from zero pre-training effort. The key was applying automated research at almost every stage of the pipeline — pre-training, post-training, RL, and inference — drastically compressing the model-building cycle.
More from Companies & People
- "Use every chip, every piece of land": the aggressive push to make tokens dirt cheap — GregKamradt · 2026-10-28
- General Catalyst launches fellowship: 20 undergrads get up to $25K compute each — Daisy4ai · 2026-10-28
- Y Combinator Releases Winter 2027 RFS: AI Bottlenecks Are Moving Beyond Models into the Real Economy — ycombinator · 2026-10-28
- Identity Labs hiring "thinker-writers" at $80k to ghostwrite CEOs' ideas — bigaiguy · 2026-10-28
- UC Berkeley math prof stops grading for correctness, gives credit just for turning in homework — Polymarket · 2026-10-28
- Ashoka University launches one-year Master's in Applied Digital Health and Responsible AI — DrDatta_AIIMS · 2026-10-28