Breakthroughs in provable inference for large language models
anthropic inference
| Source: HN | Original article
Researchers report advances in applying zkSNARK cryptographic proofs to enable provable inference for large language models.
Anthropic unveiled a research prototype called **FlashProve** on 2 November 2026, marking a concrete step toward “provable inference” for large language models. The system leverages zkSNARKs—a class of succinct, non‑interactive cryptographic proofs—to generate a verifiable attestation that a specific model produced a given response. In Anthropic’s own measurements, FlashProve delivered 53–77 tokens per second on a dense 31‑billion‑parameter model, with an overhead of roughly 70–95 times native computation. The team notes that this two‑order‑of‑magnitude slowdown is far lower than earlier estimates, suggesting the approach could be practical for a broader range of architectures.
The development matters because it tackles a growing demand for trustworthy AI outputs. Cryptographic proof of provenance could simplify compliance with emerging regulations, enable third‑party auditing, and reduce reliance on opaque service‑level agreements. By tying a response to a particular model instance, FlashProve also opens the door to new business models where verification becomes a sellable feature.
Looking ahead, the AI community will watch whether the prototype can scale to larger models and maintain acceptable latency, and whether integration into production APIs becomes feasible. Parallel efforts from other firms could spur standards for provable inference, while further optimisations may shrink the performance gap. If Anthropic’s early results hold up, cryptographic verification could become a routine part of deploying generative AI, reshaping how developers and regulators ensure model accountability.
Sources
Back to AIPULSEN