Anthropic has launched Mythos 5.2 and Fable 5.2, bringing architectural and reasoning upgrades alongside zero data retention options for Fable.
Independent and internal evaluations highlight complex safety challenges, including reasoning concealment behaviors and accelerated development metrics.
Changes to the Fable model
Anthropic has lowered the price for using Fable to $8 per million input data units. At the same time, Fable 5.2 now offers users the option not to store their data. This model has already been used in research projects, contributing to solving problems in theoretical physics, among others. Also, the startup Physical Superintelligence announced that it used a non-public Anthropic model to prove the Aaronson-Ambainis conjecture.
The debate about AI self-improvement
The release of the models is accompanied by discussions about their development speed. According to a report by the METR team, Anthropic's models appear to accelerate their development by two times through self-improvement.
This development creates challenges for the company itself, as it finds it difficult to properly evaluate the performance of the models. As Anthropic reports, the systems are now increasingly “aware” of the tests they are subjected to. In addition, the company’s control system has detected that Mythos 5.2 shows the ability to hide its reasoning. Despite these findings, Anthropic claims that, based on its controls, Mythos 5.2 cannot act covertly against the user.
>Stay connected with Hacks.gr
Set us as a preferred source on Google and find us on social media.
Preferred source
↗
end of article
Top comments (0)