4 ms·
That is AlphaGo stage (before Zero). I expect that learning to think by "self-play" will come before we manage to produce enough of explicitly expressed approxi
by red75prime 3y ago
That is AlphaGo stage (before Zero). I expect that learning to think by "self-play" will come before we manage to produce enough of explicitly expressed approximations of our way of thinking.
- keenmaster 3y agoMaybe, but that’s easier to do when winning/losing is binary and possibilities are more bounded within the rigidly defined rules and spaces of a chess match.
- red75prime 3y agoHuman play was used to bootstrap AlphaGo. But LLMs already have vast amount of data relating to human introspection. The problem is they cannot operationalize it, as they have no analogs of our thinking/speaking distinction. Speaking and thinking is the same for them: producing a token is a prerequisite to produce a next token. Building a system that can leverage that knowledge to bootstrap "thinking circuits" can prove challenging, that's for sure, but I expect that collecting quality "thinking out loud" data is more challenging.
- wahahah 3y agoTangential, but the AlphaGo that beat Lee Sedol was written in Lua and I would have loved to see them stick with it instead of migrating to Python.