RC RANDOM CHAOS

Anthropic's FlashProve: Proving AI Model Authenticity

· via Hacker News

Original source

Progress on provable inference for large language models

Hacker News →

Users cannot currently verify that an AI model’s response came from the specific model they requested rather than a cheaper substitute, compromised version, or attacker-modified variant. This matters for independent evaluators validating models, customers paying for specific models through routing services, and regulators verifying compliance with AI pacing commitments.

Anthropic’s FlashProve uses zero-knowledge cryptographic proofs (zkSNARKs) to address this by proving that responses came from a specified model without revealing its weights. On an H100 GPU, the system generates proofs for a 400-billion-parameter model at over 110 tokens per second with approximately 19 MB per 1,024-token request—representing 70-95 times the compute cost of standard inference but orders of magnitude faster than prior implementations. The overhead is manageable for production scenarios while maintaining cryptographic guarantees of model authenticity.

The team released executable code, formal Lean proofs of security properties, and documentation under Apache 2.0 licensing. This work enables regulatory compliance verification, auditor tooling, and customer trust assurances, while supporting broader mechanisms to monitor and verify adherence to AI development commitments.

Read the full article

Continue reading at Hacker News →

This is an AI-generated summary. Read the original for the full story.