Hey HN, we've been trying to solve LLM Hallucinations. So we built Pinocchio.
Pinocchio is a research preview for AI outputs with provenance and verification built in.
Give it a task and source material. For numerics, the values used in the computation are carried directly from the source rather than regenerated by the model across reasoning steps. Pinocchio records the computations and derivations applied to them, and a deterministic verifier can replay and check the resulting computation.
The output retains a connection to: • the exact source inputs used • the operations performed on them • the intermediate and final results they produced • the checks those results passed
Every output points back to its sources. You can trace it backwards through its derivation and sources.
Really looking forward to what you guys think about this.
P.S. we're limiting to one run at a time. Other than there are no other restrictions. It does take a while though, so I'd let it be and check back in a while.
Comments URL: [https://news.ycombinator.com/item?id=49687558](https://news.ycombinator.com/item?id=49687558)
Points: 1