I recently came across a technical report about Talus, and there's a key technical detail worth mentioning.

Many people might think that Talus simply runs large language models directly on the Sui blockchain—but that's not actually the case. What it really does is deploy Workflows on-chain, then lets AI Agents complete tasks by executing these on-chain workflows. This design approach is fundamentally different from running models directly on-chain.

As for the usual topics like transparency, verifiability, and composability in on-chain AI, I won't elaborate here. The architecture design itself is much more intriguing.