Dyna-2: A 1M-Hour Scaling Law for World-Action Models

Great to see improvements with scale. However, when improvements come with scale, will the benefits accrue to the few who can raise sufficient capital to train the most data on the biggest model?

I think this is a pretty big moment. If this holds (cautious since this company also has a vested interest) it could mean that we see similar dynamics that we saw in LLM / GPT stuff