zkSNARKs Prove a 400B Model Wrote a Reply Without Revealing Weights, at 70-95x Compute
rickasaurus · x · 2026-11-04
A thread on 'HTTPS for AI models': FlashProofs (associated with Anthropic) uses zkSNARKs — the proofs behind Starknet — to prove a response came from an exact copy of a 400B-parameter model without revealing its weights.
Numbers: 110 tokens/sec generation on a single H100, with 19 MB proofs that verify in seconds on a CPU. The cost: roughly 70-95x extra compute. The takeaway: verifiable AI is closer than we think.
Related event: Anthropic open-sources FlashProve for zkSNARK-verifiable LLM inference(24 posts)→
More from Research
- SemiAnalysis launches GDPx benchmark: best frontier model scores just 12% — scaling01 · 2026-11-04
- TLAPS-Bench: 265 real-world TLA+ proof targets; top model GPT-6.1 Sol solves only 44.3% — tianyin_xu · 2026-11-04
- StreamDuo pairs a commentator with a wiki research agent to commentate unseen games — Zixian Gao · 2026-11-04
- Redwood researcher: AI research automation and runaway recursive self-improvement can come apart — deanwball · 2026-11-04
- UMAP 0.6 adds Adam and momentum layout optimizers, becoming defaults in 0.7 — leland_mcinnes · 2026-11-04
- UMAP 0.6 preview release candidate lands on PyPI with major changes — leland_mcinnes · 2026-11-04