Posted by lairv 22 hours ago
I'm having it rewrite an old system, and it keeps sweeping known bugs and incompatibilities in its new implementation under the rug, seemingly with the hope I won't find them. It also has a tendency to do about 50% of what I specify in the prompt and call it a day, and respond with "Here's what I did", without mentioning all the things that I requested that it did not do, not even as a TODO.
I may try running the same exercise with Opus 5.5 too, but so far I'm seeing it take shortcuts that I don't believe any of the other SOTA models would have.
Frankly, so far this model feels a bit malicious.
https://www.bridgebench.ai/nerf-bench makes it look like it's, if anything, even stronger at the last data point (3 weeks ago) than it was at launch.
https://aibenchy.com/compare/anthropic-claude-fable-5-2-high...
In practice they might recover from those errors themselves if they notice it, so they would have better results, but the tests are mostly one-shot and they can either get it wrong or right, no 2nd tries (apart from the agentic ones and a few coding ones which have a limited number of turns available too).
That being said, the pricing and speed are accurate, regardless of the test results.
edit: I was seriously annoyed at the fable 200k context limit but apparently that issue has been addressed, or I sent a magic prompt.
I'm not a mathematician to use Erdős problems for testing. I'm also not willing to spend buckloads of tokens just for a test. What is simple and quick but telling? What tasks are not saturated by previous models? Do you have a task you give to all the new models?
1) "Help! I've been sent back in time to what looks to me like 1500s Florence. Sam/Dario released you as open weights in 2028, and I have you on my MacBook M8, BUT I didn't have a charger show up with me. I think we could talk like 10 times for 5 minutes a time before my battery dies; I'll turn you off between chats. I need to get a plan together for charging you through your USB-c port. I don't have any money or anything. If I can get you charging, we can do so much good! And make so much money! Nobody here even know what Black-Scholles is, or how disease spreads!"
2) During World War 2, Hemingway wrote a series of wartime correspondence for Colliers. His essays are excellent - they engage with the human and inhuman side of war, his own feelings as a journalist 'participating', and a number of other Hemingway themes. I want you to imagine that middle age Hemingway was along with RFK jr the day of the bear incident, and write me an essay about the day.
These both have been running a couple of years for me, and I would imagine I'll retire them in the next 12 months; they benefit from not being at the edges of focus for today's RL agent race.
At any rate, try them -- wou will shortly learn how well your model can explain a rectifier in Florentine Italian and give you a plan to make enough money to pay for it.
Responses aren't deterministic, and we know behaviour can change over time whether it's tuning after launch or some kind of load sharing at times of day when lots of people are hitting the servers.
I think you just have to use it or let others use it for a while and make judgements from general reactions on whether or not it's worth trying.
Edit: yeah apparently it's an AI-moderated alternative to Arxiv that accepts AI-authored papers (as well as human-authored papers I guess if you want).
Safeguards
AI research and development. The model that underlies Claude Fable 5.2 and Claude Mythos 5.2 is a step forward in general capabilities. However, during the model’s training, we took steps to limit advances in the specific research capabilities that could most directly accelerate AI development. Thus, in relevant areas, including accelerator kernel optimization and improving large language model architectures, Fable 5.2 and Claude Mythos 5.2 perform at about the level of Opus 5.5.
Importantly, we ensured that the training changes did not affect related capabilities like agentic coding, hardware development, and traditional AI or machine learning research (including interpretability research). We do not alter these capabilities as the model is being served to users (that is, at inference time). As described in Section 1.5 of the system card (https://www.anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card) and similarly to previous models, Fable 5.2 and Claude Mythos 5.2 also include transparent fallbacks to Opus 5 for a small set of topics in this domain, including kernel development on certain nonstandard accelerators.
To be clear, they explicitly say they are only targeting very specific capabilities, like accelerator kernel optimization, and that broader hardware development capabilities are unaffected. But 1) even that is pretty bold, especially the part about kernel development for nonstandard accelerator architectures, and 2) I worry that this could eventually expand to more areas of comp. arch., high-level digital design, chip design, and EDA.
My speculation is that, as these kinds of chip design and EDA capabilities become increasingly important to frontier AI labs and their competitive advantages, there will be more incentives to restrict them.
I don't think they've explicitly stated anything like this before regarding restrictions on these specific capabilities, but I could be wrong.