This is further illustration that raw model capability isn’t really the most interesting thing at this point. The model and the harness should be viewed as a single system. I expect that we’ll be moving towards neurosymbolic systems where LLMs act as a heuristic engine within the boundaries of a symbolic logic system.
yeah i always wondered at what point we move on from the fascination with language and into a system where language is just part of what the system works with… at some point real live physical data is going to be far more important to hoover up and wire up than just language
And also the fact that doing symbolic logic is desirable where possible because we want determinism. The model is a great heuristic engine, but it needs to be grounded in a deterministic state machine.
yeah exactly!
