[simianwords]

> I don't foresee AGI arising out training bigger LLMs (Though investors won't realise that for a while yet).

This is what the parent said. AGI won't rise out of AlphaZero and AlphaFold in the same way AGI won't rise out of Houdini chess engine. This is the industry consensus.

[fasterik]

>AGI won't rise out of AlphaZero and AlphaFold in the same way AGI won't rise out of Houdini chess engine.

That's a straw man. Nobody thinks AGI will rise out of domain-specific systems. The question is whether domain-specific systems are necessary for AGI.

Of course, the problem is that AGI isn't a well-defined concept. But if we define it as achieving superhuman performance across several hundred domains where there are objective measures of success, it doesn't seem far-fetched to predict that it will involve some general reasoning system paired with a bunch of specialized modules.

[simianwords]

The parent said

> I don't foresee AGI arising out training bigger LLMs

I agree that AGI will involve tool usage but not only involving domain specific AI models.

But lets try to find the discriminating point in the discussion - do you believe AGI will necessarily involve training bigger LLM's or not?

I believe they are necessary. WBU?

[fasterik]

You're still intentionally misreading the OP's statement. If you read it again in context, they're clearly saying that they think training bigger LLMs is not sufficient. I think I agree with that statement, but my confidence is pretty low.

No, I don't think LLMs are necessary for AGI at all. I think there are multiple paths to AGI, some of which involve LLMs and some which don't.

[dofm]

AGI is a macguffin (or a shaggy dog) for a story told to investors. It has never been plausible on the timescales suggested and it almost certainly will not emerge from LLMs.