Comment by xutopia
9 days ago
I’m under the same impression. I don’t think LLMs are the path to AGI. The “intelligence” we see is mostly illusory. It’s statistical repetition of the mediocre minds who wrote content online.
The intelligence we think we recognize is simply an electronic parrot finding the right words in its model to make itself useful.
I fear that AI will be intelligent enough to negate human general intelligence before it is itself generally intelligent.
It's so attention needy, and it's transforming our culture
They already transform our language. Largely.
That's pre-training. Post training with RL can make models arbitrarily good at specific capabilities, and it's usually done via pooled human experts, so it's definitely not statistically mediocre.
The issue is that we're not modelling the problem, but a proxy for the problem. RL doesn't generalize very well as is, when you apply it to a loose proxy measure you get the abysmal data efficiency we see with LLMs. We might be able to brute-force "AGI" but we'd certainly do better with something more direct that generalizes better.
Maybe i'm misunderstanding your point, but human's have pretty abysmal data efficiency, too. We have to use tools for everything... ledgers, spreadsheets, data-bases, etc. It'll be the same for an AGI, there won't be any reason for it to remember every little detail, just be able to use the appropriate tool, as needed.
I don't see how you can come to that conclusion if you've actually used e.g. Opus 4.6 on a hard problem. Either you're not using it, or you're not using it right. And I don't mean simple web dev stuff. In a few hours Claude built me a fairly accurate physics simulation for a game I've been working on. It searched for research papers, grabbed constants for the different materials, implemented the tests and the physics and... it worked. It would have taken me weeks. Yes, I guided it here and there, especially by telling it about various weird physics behavior that I observed, but I didn't write one line of code.