Nathan Lambert leaves Ai2, warns open post-training is falling behind

Nathan Lambert announced he is leaving the Allen Institute for AI, where he ran the post-training work behind Tulu and OLMo. In his farewell post he uses the occasion to make a public argument that lands harder than the usual exit note. His blunt summary: fully open post-training recipes are about as far behind as they have ever been, and falling further behind. Frontier labs keep their methods quiet, and the public side cannot close the gap on its own.

Lambert credits Ai2 for a culture that lets researchers ship influential work without locking it behind product launches, and frames his arc there as proof that vision and clear writing still matter in a field that rewards execution. The deeper concern, he argues, is structural. As more talent moves into closed labs, public science loses its role as the neutral interpreter of new technology, and the institutions that train the next generation of researchers thin out. He is not announcing a new employer. He plans to spend the next months building medium-sized open models tuned for specific tasks, aimed at giving the ecosystem more variety than another round of frontier-chasing.

The post lands during a busy week of closed releases, which is partly the point. Someone has to keep the recipes legible if open work is to survive past 2026.

Why it matters

If your team relies on open post-training recipes, Lambert's departure is a signal that the talent pipeline behind those recipes is fragile. Track which institutions or independents end up filling the gap, because that is where future Tulu-class work will come from.