Anthropic's FlashProve uses zkSNARKs to prove model responses, at 70-95x GPU overhead
winer666 · reddit · 2026-11-04
Anthropic, with collaborators from EPFL, MIT and others, published FlashProve (Nov 2): a provable-inference system using zkSNARKs to cryptographically prove a response came from a specific, committed set of model weights — without revealing the weights — and to catch tampering with the token stream in between.
Key numbers and limits
- Tested on Llama 4 Maverick (400B-param MoE, 4-bit weights): proving costs roughly 70-95x the GPU time of vLLM inference
- Contexts up to 1,024 tokens only; proof covers an integerized model under greedy decoding; architecture itself not hidden
- Code and Lean formal proofs open-sourced under Apache 2.0; positioned as a research prototype
The open question: if proving gets cheap, does it change how we trust hosted models — or does the overhead confine it to audits and regulated settings?
Related event: Anthropic Open-Sources FlashProve for Provable LLM Inference(20 posts)→
More from Safety
- AI welfare and AI safety are allies, not rivals — a researcher's four arguments — RosieCampbell · 2026-11-04
- Meta's letter to Newsom opposes mandatory outside safety checkers, diverging from OpenAI and Anthropic — ShakeelHashim · 2026-11-04
- New book "Treacherous Tech" cites Irving's call to halt frontier AI and Pachocki's warnings — DavidSKrueger · 2026-11-04
- NBC Exit Poll: US Voters Worried About AI, Overwhelmingly Oppose Data Centers — DavidSKrueger · 2026-11-04
- Binance dev-ops infected after downloading Homebrew from malicious Google search ad — HankYeomans · 2026-11-04
- FAI Seeks Plaintiffs to Sue Vendors of Unlocked 'Abliterated' Models Over Cyber Harms — michael_nielsen · 2026-11-04