LLM won't lead to AGI

[Scholar]

 

Obvious stuff but for those overhyped people in here.

[Leo Gura]

I just finished listening to that and was gonna share it. It's a surprising deep discussion.

[Shodburrito]

Of course LLM's won't lead to AGI. Most of the best AI models now are multimodal anyways.

[Ulax]

LLM? AGI?

[Scholar]

33 minutes ago, Leo Gura said:

I just finished listening to that and was gonna share it. It's a surprising deep discussion.

It's not that deep, you could have figured out 90% of the principles discussed here just by contemplating and investigating your own intelligence and intuition, whilst knowing the basic way of how LLMs work.

It's nice to see the experts parrot what I have been saying in regardsto LLMs basically simulating intuition, and that intuition is used by humans in all forms of intelligent problem solving but is not actually intelligent problem solving. I had a discussion about this precise thing with someone on the forum a year back or so.

 

They are still missing the connection to individuated consciousness though, but that's another matter.

Any easy way to gain insight into this is to inspect specifically the nature of understanding. Note that intuition is not grounded in consciousness, or individuated consciousness to be precise. It's unconscious or subconscious, so is knowledge, and the way it is cued is conscious, but the way it is prompted from your memory is not conscious.

You know this Leo because you made this insight when you realized your mind is like google and you can just ask it something and it will give you an answer. That is basically engaging the "LLM" unconscious" part of the mind or brain.

 

Understanding however is precisely consciousness. It is cannot be simulated or imitated, because it is it's own form of existence. The brain is basically an interface between that form of existence (which is a unified form of existence, something unique and impossible to simulate via mere algorithmic approaches), so the neural networks feed into individuated consciousness, which is sustained through a particular relationship within the physicality of the universe.

Edited June 12, 2024 by Scholar

[Leo Gura]

It's not at all clear how all this works.

[Scholar]

44 minutes ago, Leo Gura said:

It's not at all clear how all this works.

Sure thing buddy.

[Seth]

The title of this post is somewhat misleading. As Dwarkesh Patel notes, Francois Chollet does not seem to disagree with the popular view that LLM intuition may be what drives a logical planning algorithm in future AGI. The primary disagreement appears to be about timelines rather than the required structure. Some have hope that reasoning will be solved this decade, while Mr. Chollet considers it a significant challenge that won't be resolved soon.

Robert Miles has a great video explaining why a system 2 algorithm is so exciting for AGI. (I also recommend his other videos. He is a good resource for learning AI safety concepts.)

With system 2, planning might not just amplify intuition; the amplified intuition could be distilled back into the neural network, improving the base intuition. Of course, the real world is more complex than a finite board game, and researchers are  still seeking ways to make this process compatible with LLMs.