provable inference

Anthropic has published new research describing progress on provable inference for large language models, an approach aimed at giving mathematical guarantees about what models compute rather than relying only on empirical testing. The work is drawing attention from AI researchers and developers who see verifiable guarantees as a potential route to more trustworthy and safer model behaviour.

Anthropic has released FlashProve as an open-source project. It is a prototype for provable inference in mixture-of-experts (MoE) models, meaning outputs could be cryptographically verified rather than simply trusted. The announcement drew coverage from security-focused outlets, with interest focused on what verifiable AI inference could mean for trust and accountability in deployed models.

Anthropic has announced progress on provable inference for large language models, an effort to give mathematical guarantees about what a model actually computes rather than relying only on empirical testing. The work aims to make model behaviour more verifiable and trustworthy. Details of the methods and results remain limited in the announcement, and outside experts have not yet weighed in publicly.

Moonshot AI released Kimi K3.1 Pro, a model with 2.8 trillion parameters, while Tencent launched Hy4, an AI tool aimed at game development. Separately, Anthropic is facing an attribution dispute concerning its provable-inference research. The three developments highlight ongoing competition among major AI labs and growing legal and technical contention around model research claims.