On Thu, Jul 6, 2023 at 7:58 PM Matt Mahoney <[email protected]> wrote:
> ...
> The LTCB and Hutter prize entries model grammar and semantics to some extent 
> but never developed to the point of constructing world models enabling them 
> to reason about physics or psychology or solve novel math and coding 
> problems. We now know this is possible in larger models without grounding in 
> nonverbal sensory data, even though we don't understand how it happened.

We don't understand how it happened. No. We don't really understand
much at all. It's all been a process of hacking at some very old ideas
about training to fixed categories. Speeded up by GPUs developed for
the game market. And most recently enhanced by the accidental
discovery that context, through "attention", seems to be central.

That distributional analysis of language might result in categories
useful for broader reasoning was always plausible to me, so I find no
shock in it with LLMs. I just think LLMs are limited by not being able
to find novelty. Their categories are fixed at the time of training.
In reality I think the categories can shift. And be novel. That they
are chaotic attractors.

Anyway, I encourage those who are inclined to think about theory to
focus on the fact that simply allowing the size of the models to
increase seems to be of central importance.

------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T42db51de471cbcb9-M156f326bfc2335867341f308
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to