Claude Fable 5.2 can do "weight-space-programming" about as well as GPT-6 Astra

An evaluation comparing Fable 5.2 to GPT-6 Astra in low-level neural network manipulation tasks finds that Fable reproduces nearly all qualitative results from a recent research paper. Fable performs comparably or better on tasks like writing 9x9 Go policies and editing Qwen's weights for math performance, though it exhibits higher false confidence and occasionally struggles with weight reading and language corpus identification. Open article →