I more so meant that to learn something new the model would have to update its own weights (I have my reasoning for this in another reply in this thread).

When I said “fundamentally unable to” I meant that current LLM architectures do not have the capability to update their own weights (although I probably should’ve worded that a bit differently)

When I mentioned prompt engineering, I more so meant that people where explaining what to do in a if/else manner to get the LLM to play tiktaktoe (not chain of thoughts or any of those techniques).

In my opinion, learning is both 1) acquiring new skills, and 2) improving upon those skills with repetition. I think it’s very debatable if an LLM could learn something truly novel (or even something like an existing game with some new rules, I.e., chess but with the game pieces in different positions) with in context learning. Secondly, no matter how much you play tiktaktoe with an LLM, it will never improve at the game.

This is just my two cents on why I don’t believe LLMs to fit the criteria of “emerging AGI” that the researchers laid out. Imo I think that to fit that criteria they would need to implement some type of online learning but I definitely could be wrong.

I doubt that any model currently is in the “emerging AGI” category (even by there own metric of “general ability and metacognitive abilities like learning new skills”).

The model(s) we currently have are fundamentally unable to update their own weights so they do not “learn new skills”. Also I don’t like how they use “wide range of tasks” as a metric. Yes, LLMs outperform many humans at things like standardized tests, but I have yet to see an LLM who can constantly play tiktaktoe at the level of a 5 year old without a paragraph of “promt engineering”

I’m not the most educated on this topic (still just a student studying machine learning) but imo I think that many researchers are overestimating the abilities of LLMs