I've written about why just depending on pretraining won't be enough to create AGI, but it seems like we may have a viable solution now. Reasoning is all you need.
Minor point: that Latin magic square is a famous example (as o1 itself points out), so it seems to be displaying ability to retrieve known solutions rather than come up with new ones. But at least it clearly understood what you were looking for!
Is it possible that you don't know very well how to judge short stories and poems, and that calling AI output "very good" or "pretty good" for both these things is out of your domain of expertise? I'm not accusing, just wondering. What if AI just churns out completely passable examples of these forms, because almost has the ability to properly judge the quality of them? It makes me wonder: what would Helen Vendler say, for example?
I've written about why just depending on pretraining won't be enough to create AGI, but it seems like we may have a viable solution now. Reasoning is all you need.
I know that sounded familiar, I read that :)
Minor point: that Latin magic square is a famous example (as o1 itself points out), so it seems to be displaying ability to retrieve known solutions rather than come up with new ones. But at least it clearly understood what you were looking for!
Yes precisely it figured out a meta answer
Nice piece Rohit. Is it too much to ask for OpenAI to come up with a naming scheme that is not so confusing?
The golden question is, with GPT-o1, how far are we from AGI? Months? Years?
I would like to know so I can update my recent essay on the topic.
A version of this grounded in search or documents might as well be called AGI I think, at least in the narrow sense
That's somewhat what I was thinking. I am reading this is PhD level intelligence, at least across narrow areas.
i think that the fact that they were sitting on this is why they were so confident / breezy about the stochastic parrot diss
Is it possible that you don't know very well how to judge short stories and poems, and that calling AI output "very good" or "pretty good" for both these things is out of your domain of expertise? I'm not accusing, just wondering. What if AI just churns out completely passable examples of these forms, because almost has the ability to properly judge the quality of them? It makes me wonder: what would Helen Vendler say, for example?
No I don't believe so
I think so.
Anyway, here's a podcast episode discussing some of these things. Eventually he gets to some of the AI poems used in the study - they are very bad.
https://www.matthewbuckleysmith.com/sleerickets/pe8nshntt6xxfex-zfjcw-ppapj-976xb-4ajcg-ssjbt-yzmdc-ktjcx-gjcfl-89e9e-33pyy-yby8c-5e68z-9kh7w-egrye-tyblf-9fyke-jydn9-6b26n-cmpyj-jf4jm-88bpt-dphka-7cak6-h5lwx-6xrls-yc2cf-5ljws
For one thing, the poem in this post just used the plain old model, no "thought for 13 seconds".
You're right! I copy pasted the wrong one. Fixed now thanks.
Oh yeah, is all going to end well for us. After all, ML is just a parrot and definitely can't.... uh...