Hacker News Reader

Comment by riskable | Hacker News Reader

I don't think it's possible to say how far away or "never" it is. All we know is that LLMs cannot become AGI.

My prediction: AGI will come from a strange place. An interesting algorithm everyone already knew about that gets applied in a new way. Probably discovered by accident because someone—who has no idea what they're doing—tried to force an LLM to do something stupid in their code and yet somehow, it worked.

What wouldn't surprise me: A novel application of the Archimedes principle or the Brazil nut effect. You might be thinking, "What TF to those have to do with AI‽ LOL!" and you're probably right... Or are you?

View on HN

> All we know is that LLMs cannot become AGI.

A part of it, perhaps: I think of it like 'computer vision'; LLMs offer 'computer speech/language' as it were. But not a 'general intelligence' of motives and reasoning, 'just' an output. At the moment we have that output hooked up to a model that has data from the internet and books etc. in excess of what's required for convincing language, so it itself is what drives the content.

I think the future will be some other system for data and reasoning and 'general intelligence', that then uses a smaller language model for output in a human-understood form.

I guess AGI will come from identifying and grinding away limitations of transformers (and/or diffusion networks). As it has been with almost all technologies. Then someone (or someAI) will probably find something unexpected (but less unexpected at this stage) and more suitable for general intelligence.

Why not?

What's the fundamental, absolutely insurmountable capability gap between the two? What is it that can't be bridged with architectural tweaks, better scaffolding or better training?

I see a lot of people make this "LLMs ABSOLUTELY CANNOT hit AGI" assumption, and it never seems to be backed by anything at all.

The view that LLMs alone are insufficient for AGI is based on their fundamental mathematical architecture

What is the exact limitation imposed by "their fundamental mathematical architecture"? I'm not aware of any such thing.

They don't update themselves as they work on a problem. Humans solve hard problems by forming new neural connections, LLM cannot do that. If you make a model that does that then its no longer the LLM architecture we use today and it would be called something else.

Even if we assume that "update themselves as they work on a problem" is vital for AGI (not proven, it's just an assumption), then what stops us from just giving an LLM an "update itself" tool call?

As the most naive implementation: the LLM just collects the relevant training data, and then runs PEFT on itself. Then it tests the tuned version of itself to see whether the tune was any good and is worth retaining.

Sure, that naive approach would require an LLM to be good at assembling training sets, and validating performance gains, and it would be very computationally expensive. But none of those things would somehow make it not-an-LLM-anymore.

Not the current ones but that doesn't mean you can't build something that uses the ideas from LLMs but adds to them.

Reply on news.ycombinator.com