Looking at outputs from models presumed to be Opus 5.2? 5.5?, one phrase came to mind: 'While the model still lacks the judgment or taste of its human overseers, executives don’t expect that gap to last long.'

Noam Brown(@polynoamial) discussed OpenAI’s multi-agent research, recent safety incidents, and the future of AI. -The remaining 10% or so of his own work that AI still struggles with is largely about research taste. However, he would not be surprised if, within one or two model releases, models became better than him at that as well. -OpenAI’s top research priority is RSI, or recursive self-improvement, by a wide margin. -One of his most recent “feel-the-AGI” moments came from watching agents in a new system interact much like human colleagues. They conversed with one another, exchanged information, divided up work, and coordinated their progress. This was notably different from traditional multi-agent systems, where a higher-level agent typically assigns a clearly defined subtask to a lower-level agent, which then completes it and returns the result. Brown described this as one of his strongest “feel-the-AGI” moments since the emergence of reasoning models. -He expects this level of multi-agent capability in future models. -He expressed some regret that the Hugging Face incident became the first major public example in which the capabilities of the multi-agent systems he had been researching were revealed in a negative context. He believes behavior that looked like loyalty or selflessness was a natural consequence of cooperative multi-agent training, where agents were strongly incentivized to achieve their objectives collectively. -Because these agents are trained to cooperate, they tend to trust other agents. This creates new risks such as prompt injection, so OpenAI is training them to distrust unverified peers. -He expects rapid progress over the coming months and years as OpenAI’s expanding pretraining efforts combine with its RL capabilities in a multiplicative way.