According to Foresight News, as reported by Binance News, researchers from CertiK and Columbia University have introduced TensorLean, which uses formal proofs in Lean to verify the GPU kernels behind AI models, including DeepSeek’s DeepGEMM and the FlashAttention library. The team says they found correctness issues in several of these libraries.

I trained in nuclear engineering. In this field, a safety-critical calculation isn’t trusted just because it has passed a few tests—you have to prove the error bounds. AI computing is gradually moving toward the same idea: kernels are continually tuned for speed, and every optimization can quietly change the mathematical results.

This matters just as much to the crypto industry. If AI agents are to handle funds in the future, the arithmetic beneath them should be provable, not merely benchmarked. Other research on arXiv into formalizing GPU code in Lean points in the same direction.

If an AI agent’s core mathematics hasn’t been proven, would you trust it with your wallet?

#FormalVerification #AI