Interesting idea from a dev talk today: what if AGI comes from decentralized system instead of LLMs?

I came across a dev talk earlier today from a tech event and it got me thinking about where AGI might actually came from long term.

Most of what we see right now is centered around large language models. They're trained on huge datasets and generate outputs by predicting the next token. Super impressive, but once training is done, they don't really change much.

The talk I watched was going in a different direction. Instead of training a model and stopping there, the idea was to have something that keeps running and adapting over time. They mentioned using ternary logic (+1, 0, -1) instead of just binary, which was new to me.

Another thing was that it processes information continuously instead of in batches. And instead of improving the usual way most models do, it closer to a trial and error type of evolution.

What stood out to me is that this isn't just an idea on paper. There's already some code out there, a pretty large dataset, and a working demo from what I could tell. I think one of the papers is also being presented at a conference later this year.

I'm not deep into AI, so I'm mostly just trying to understand how realistic something like this is. But it did make me wonder it decentralised systems might play a role in how these models develop over time.

I want to know what people here think. Do you see LLMs staying dominant, or could something completely different show up?