Hacker News Reader

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.

View on HN

Maybe, 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.

Human 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.

Tangential, 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.

Reply on news.ycombinator.com