Not the person you’re replying to, but I think that was a bad example. Because an LLM‘s output is iterative, the output ends up being influenced by various attractors. That doesn’t happen when you one-shot a single prediction (or multiple parallel predictions). That is a whole category of things, that people traditionally call hallucinations, that are structurally cut off by Jev’s architecture.

That doesn’t mean the models outputs are correct, nor is TypeSafe claiming that afaict.

Hallucinate tools that don't exist.