Anthropic Releases FlashProve Cryptographic
MarkTechPost relays the release of FlashProve, and if the numbers hold up, this is a genuinely meaningful step for trustless AI infrastructure. Verifiable inference solves a real gap: today you take a provider's word that the promised weights are actually running, and cryptographic proofs would turn that promise into an auditable fact — useful for safety commitments, licensing and output provenance.… more
That said, skepticism is warranted: zkML efforts historically run orders of magnitude slower than native inference, so a claimed ~110 tokens/second of proof generation on a single H100 demands independent replication before anyone celebrates. The unknown that matters more is economics — 19 MB proofs per reply and GPU overhead per token will decide whether this stays a demo or reaches production APIs. In our view, the Apache 2.0 open-sourcing matters as much as the benchmark: the shift from 't
Anthropic Releases FlashProve: Cryptographic Proofs That a Reply Came From Committed Weights
Anthropic released FlashProve, an Apache 2.0 zkSNARK prototype that proves a reply came from a committed model without revealing weights. It proves Llama 4 Maverick at 110 tokens/second on 1 H100, with 19 MB proofs. The post Anthropic Releases FlashProve: …
Source:
Marktechpost
Discussion (0)