Also notice how they state just for SWE-Bench Pro: "*Anthropic reported signs of memorization on a subset of problems"
303 karma · joined December 31, 2018
Also notice how they state just for SWE-Bench Pro: "*Anthropic reported signs of memorization on a subset of problems"
I'm not a big tailwind fan, but keeping styling in a separate file feels like a net negative
My guess is that 5.5 will come out soon and be significantly better so you'd want to be using Codex then, but then when Opus 5 comes out probably back to claude code
Also 5.4 has fast mode, and higher usage limits since it's cheaper
The impressive part is multimodality, very plausible since there's less focus there by other labs (especially Anthropic)
The red team post goes over some more impressive finds, and says that there's hundreds more they can't disclose yet: https://red.anthropic.com/2026/mythos-preview/
"Most security tooling has historically benefitted defenders more than attackers. When the first software fuzzers were deployed at large scale, there were concerns they might enable attackers to identify vulnerabilities at an increased rate. And they did. But modern fuzzers like AFL are now a critical component of the security ecosystem: projects like OSS-Fuzz dedicate significant resources to help secure key open source software.
We believe the same will hold true here too—eventually. Once the security landscape has reached a new equilibrium, we believe that powerful language models will benefit defenders more than attackers, increasing the overall security of the software ecosystem. The advantage will belong to the side that can get the most out of these tools. In the short term, this could be attackers, if frontier labs aren’t careful about how they release these models. In the long term, we expect it will be defenders who will more efficiently direct resources and use these models to fix bugs before new code ever ships. "
Also can see https://marginlab.ai/trackers/claude-code/
It's very interesting to me how widespread this conception is. Maybe it's as simple as LLM productivity degrading over time within a project, as slop compounds.
Or more recently since they added a 1m context window, maybe people are more reckless with context usage
Maybe you just spend more on tokens by some factor than the attackers do combined, and end up mostly okay. Put another way, if there's 20 vulnerabilities that Mythos is capable of finding, maybe it's reasonable to find all of them?
Maybe a bad example since Nicholas works at Anthropic, but they're very accomplished and I doubt they're being misleading or even overly grandiose here
See the slide 13 minutes in, which makes it look to be quite a sudden change
I think this would be very heavily used if they released it, completely unlike GPT 4.5
I believe he didn't think that reasoning/CoT would work well or scale like it has
There's some indirect exposure and potential of being granted significant equity, but his actions don't read as being for his own wealth
As an example, he walks away with nothing in the very plausible timeline where he was fired but not then reinstated
https://www.anthropic.com/_next/image?url=https%3A%2F%2Fwww-...
basically 9/13 are very close
From his description: "Github suspended my account without warning or reason, so here for the duration"
What do you mean by sweep training here?