How Was the Open Model Beam Trained?

Misha Laskin says Beam was trained as a 500-billion-parameter model, with 23 billion of those parameters active. Pre-training ran on 6,000 GB300s. The first run took a few weeks. Infrastructure efficiencies now bring that pre-training down to about 12 days, maybe less, by coupling scientific efficiencies with infrastructure efficiencies.

Reinforcement learning used a little over 10,000 GB300s and lasted four weeks. More training compute went to reinforcement learning than to pre-training. Those jobs are more complex because they run a lot of inference at scale, with sandboxes for agents, and training at the same time.

He got into AI after seeing a co-founder's work on AlphaGo, when he was a physicist. A plot in AlphaGo never stopped improving, and the team cut it off after it beat the world champion.