This is a special post for quick takes by dactyl. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
Labs seem much more open to pacing initiatives. It seems like they see what the scaling laws do to the next gen models, and are spooked. This Tweet argues they may have discovered a new scaling axis and I hypothesize that this new axis is scaling multi-agent cooperation. OpenAI's Hugging Face incident presents ample evidence that they have ran Multi-Agent Reinforcement Learning on their models already. Contrary to the amazing cooperation results we see in the HF incident, Anthropic's past experience with multi-agent cooperation has been less successful and full of pitfalls, probably because they had not MARL'ed their agents.