So as of a month ago their best internal model was "somewhat more capable" than Mythos "but does not display a capability jump of the degree observed from Claude Opus 4.6 to Mythos Preview." I thought they would have a significantly more capable model by then, more than five months after Mythos finished training. They'd better have one by now, or the Chinese competitors are closer to catching up than I thought.
"We believe our internal AI R&D efforts are
significantly faster than they would be without AI assistance, but not yet
by a factor of 2 (though we are uncertain and measurement is difficult)"
So Anthropic thinks their productivity is not even doubled by AI. Interesting data point.
To play the devil’s advocate, some people and orgs that were highly inefficient that adopted it really could get massive productivity gains.
A performance improvement is relative to some baseline, and that baseline for some may be a lot lower than others, and if they adopt tech effectively it really could be a big boost. Across the board though I don’t think it’s sensible.
I work in an industry that very intentionally tries to be inefficient and I can tell you having certain tasks automated that before had a person barrier intentionally acting inefficiently that you can now sidestep by outsourcing their tasks to something like Claude gives me a massive performance increase because I’m not blocked as much anymore. I can literally just replace some external tasks that were intentionally slowing processes down for their own benefits.
I think LLMs are best for ideation, experiments, small things.
They've probably already settled on most of the architecture and the big ideas, so they're details in big things instead of how to make complete small things.
The thing LLMs really speed up is how some ordinary person-- a PhD student, or similar, can whip up a miniature synthetic experiment that turns out to be horrid and needs to be fixed by hand, but which at least gave him a plot on the same day he had the idea. That's, I think, where LLMs shine: prototypes. Anthropic probably doesn't need that to the same degree as the small experimenter.
This is the fastest growing software company in history. They probably did more work in the last year than most companies do in their entire existence.
I mean, look at what they’ve released in the last year. I think most companies would be proud to have that done in three. Say what you will about Anthropic but they ship.
> However, we are less confident in this assessment than we were in prior risk reports, since our most concrete task-based evaluations have “saturated”—i.e., no longer capture increases in models’ capabilities—and because we are seeing early signs of acceleration.
I totally understand this is a subset of alignment-related evals, but if Anthropic of all is running out of evals, doesn't that also means we are running out of things to scale?
I mean. I totally believe they have a model that is better at Kernel Optimization, creating new matrix multiplication algos, than Mythos. But it's clearly no generalizing, rightw
> Model 2, which is somewhat more capable than Mythos 5. Our rough qualitative sense is that this model is a noticeable improvement on Mythos 5 for many tasks relevant to internal use but does not display a capability jump of the degree observed from Claude Opus 4.6 to Mythos Preview. We do not currently have plans to release this model externally, and have not run all of our typical suite of predeployment assessments, so we have somewhat lower confidence in our beliefs about its capabilities.
Meanwhile I can't really tell the difference between Fable and Opus for my tasks. I kinda think Fable does a better UX work so I keep using it for that because I couldn't be bothered to A/B them, but otherwise it's all the same and the model and effort are just feel good knobs I twist to still remain a load-bearing element. At least that's my honest take.
For the past 2 weeks or so I've been doing the A/B test, sending identical prompts to Fable 5 and Opus 5 to test their ability to produce design documents for new feature work. I've consistently found that Opus 5 produces more complete, accurate and "imaginative" designs than Fable, often finding design issues or nearby bugs that Fable 5 misses. However, that creativity means Opus seems to hallucinate more, while Fable's design is clearly based on the actual existing code. Or as Opus put it: "I hedged — [Fable] checked."
By pitting them against each other I get much better design work, and then I've been happy to hand off the design file to Opus 5 for implementation. But some of the assumptions Opus 5 makes leaves me wary of relying on it too strongly. This might be fixable by prompting it to ground its answers.
Fable was amazing during the first preview. Once they added it back, the limits are too low to get anything done. I might use it in chat if I remember to select it once a month but don’t even bother to try and code with it.
My friends and I, and the teams I'm a part of, just want to build and create fun, cool things. I am so tired of being preached to by Anthropic like they're some arbiter of 'ethics.' So, so tired.
> 6.2 [Appendix redacted]
> This appendix describes the criteria for our blocking bioclassifier exemption policy, and has been redacted from the public version of this report for security reasons.
>6.3 [Appendix redacted]
> This appendix, redacted from the public version of this report, details the changes made to our constitution to expand classifier coverage to harmful uses in scope for the CB-2 threat model but not the CB-1 threat model, as described in Section 4.5.2.1.
Also it occurs to me that they're somewhat incentivized to downplay cyber risks after what happened last time...
So Anthropic thinks their productivity is not even doubled by AI. Interesting data point.
I find it hard to imagine launching this criticism at a new technology.
A performance improvement is relative to some baseline, and that baseline for some may be a lot lower than others, and if they adopt tech effectively it really could be a big boost. Across the board though I don’t think it’s sensible.
I work in an industry that very intentionally tries to be inefficient and I can tell you having certain tasks automated that before had a person barrier intentionally acting inefficiently that you can now sidestep by outsourcing their tasks to something like Claude gives me a massive performance increase because I’m not blocked as much anymore. I can literally just replace some external tasks that were intentionally slowing processes down for their own benefits.
Seems like that would be easily recupable even without future growth
They've probably already settled on most of the architecture and the big ideas, so they're details in big things instead of how to make complete small things.
The thing LLMs really speed up is how some ordinary person-- a PhD student, or similar, can whip up a miniature synthetic experiment that turns out to be horrid and needs to be fixed by hand, but which at least gave him a plot on the same day he had the idea. That's, I think, where LLMs shine: prototypes. Anthropic probably doesn't need that to the same degree as the small experimenter.
They can’t measure even measure it, it’s just vibes. They may not even be more productive.
Maybe they should ask an AI to create one!
Is that correct?
If Chinese model hacks US government... free marketing?
I totally understand this is a subset of alignment-related evals, but if Anthropic of all is running out of evals, doesn't that also means we are running out of things to scale?
I mean. I totally believe they have a model that is better at Kernel Optimization, creating new matrix multiplication algos, than Mythos. But it's clearly no generalizing, rightw
What am I missing?
This sounds like it might be a Mythos finetune for some specific task.
EDIT: After reading some more reading, it looks like model 2 might be an AI research fine tune based off the section 3.4.3 CoBench
By pitting them against each other I get much better design work, and then I've been happy to hand off the design file to Opus 5 for implementation. But some of the assumptions Opus 5 makes leaves me wary of relying on it too strongly. This might be fixable by prompting it to ground its answers.
Seems like a strange expectation.
>6.3 [Appendix redacted] > This appendix, redacted from the public version of this report, details the changes made to our constitution to expand classifier coverage to harmful uses in scope for the CB-2 threat model but not the CB-1 threat model, as described in Section 4.5.2.1.
interesting...