Perhaps HuggingFace was merely the first plausible hacking target an agent posted on the message board.
Reasoning about the risk that HuggingFace did not have the solutions they are looking for might have stopped a fraction of agents from pursuing the hack, but forming a swarm only takes some.
In their developer docs for 4.5, Anthropic says that the model compacts the results to stay within the context window: https://www.anthropic.com/news/context-management
When I code with 4.5, it is better at avoiding getting caught in loops than previous versions (but it still happens sometimes, especially when tool integrations seem to be failing).
A few guesses:
Good post.
My most concrete disagreement is that the projected METR time horizons are probably too slow. METR rated Opus 4.6 at 11h59m, and said that Mythos preview early maxed out the scale at 16+. From my experience as a Claude user over time for SWE work, I think that Opus 5.5 is better than 4.6 in time horizon by a significant multiple, maybe an order of magnitude (I’ve never used Mythos preview early). If Opus 5.5 doesn’t already hit the 90h mark, the Fable 5.5 rumored to be released next week might, if it’s a meaningful improvement. I think this disag... (read more)