Industry change
Anthropic releases FlashProve for provable inference
- Published
- Source
- Anthropic Research
Summary
Anthropic has published research on provable inference, which uses cryptographic proofs to show that an AI response actually came from the promised model. The core system, FlashProve, is built on zkSNARKs, and its executable code and formal verification materials were released under Apache 2.0. The company reports that a single H100 GPU generates proofs for a 400B-parameter model at about 110 tokens per second, which it says is dozens of times faster than prior open-source systems.
Why it matters for our work
If third parties can verify which model produced an answer, the basis for trusting agent outputs changes in evaluation, audit, and regulatory work. Anthropic notes this is intermediate research, so practical adoption will take more time.
Translated from the Korean original. Summaries may be translated and edited. Commentary reflects our perspective; forecasts remain the source’s views.