The most unpleasant detail of the Fable release - the model sharply dulls when 'used for developing frontier models'.

The most unpleasant detail of the Fable release - the model sharply dulls when 'used for developing frontier models'. This concerns not only architectural research and working with datasets but also chip design and writing kernels, and who knows what else. A full arsenal of methods is used for this - steering vectors, rewriting prompts, and even changing weights. Moreover, unlike other areas, such as biology and cybersecurity, where requests are processed by Opus 4.8, in the case of LLM development, the fact of nerfing is not visible to the user. The model can quietly sabotage the development process, and this can only be discovered through indirect signs. The fact that such measures are completely opaque makes the model unsuitable for use in any AI research. Will Anthropic's filters nerf only LLM development while allowing image/video generation? Almost certainly not. And what about harness development? The problem is that no one can say exactly where the boundary lies and whether there are any other undisclosed limitations. In short, we condemn.