Progress on provable inference for large language models (www.anthropic.com)

Recent advancements from Anthropic and collaborators have led to significant progress in provable inference (PI) for large language models, leveraging a powerful cryptographic technique called zkSNARKs (zero-knowledge succinct arguments). This development addresses a critical gap in AI model verification: users currently have no way to confirm that the responses they receive are indeed from the specified model. By implementing zkSNARKs, the model provider can commit to specific model weights and subsequently generate concise proof that each response originated from that model, thereby enhancing user trust and security. The new system, FlashProve, showcases impressive technical performance by enabling a 400 billion-parameter model (Llama 4 Maverick) to generate proofs at over 110 tokens per second on an H100 GPU, with a proof size around 19 MB per request. This configuration is significantly faster—70-95 times less computationally intensive—than prior open-source systems. The implications of this progress are far-reaching, with potential applications including enhanced security in AI service deployment, verification of compliance with regulatory standards, and the ability to detect tampering by malicious actors. The release of this system, along with accompanying executable code and formal documentation, invites further innovations by the AI/ML community to build upon these advancements.