CertiK and Columbia University researchers introduced TensorLean, which applies Lean proofs to the GPU kernels behind AI models, including DeepSeek's DeepGEMM and the FlashAttention library, according to Foresight News via Binance News. The team reports finding correctness issues in several of them.
I trained as a nuclear engineer, and in that world a safety-critical calculation is not trusted because it passed a few tests. You prove the bounds. AI compute is slowly arriving at the same idea: kernels are tuned for speed, and every tuning step can quietly change the math.
This matters for crypto too. If AI agents are going to move money, the arithmetic underneath them should be provable, not just benchmarked. Separate research on formalizing GPU code in Lean, published on arXiv, points in the same direction.
Would you trust an AI agent with a wallet if its core math had no proof?
#FormalVerification #AI
I trained as a nuclear engineer, and in that world a safety-critical calculation is not trusted because it passed a few tests. You prove the bounds. AI compute is slowly arriving at the same idea: kernels are tuned for speed, and every tuning step can quietly change the math.
This matters for crypto too. If AI agents are going to move money, the arithmetic underneath them should be provable, not just benchmarked. Separate research on formalizing GPU code in Lean, published on arXiv, points in the same direction.
Would you trust an AI agent with a wallet if its core math had no proof?
#FormalVerification #AI