Comments
An excerpt from some recent correspondence between Andrew Conner (a) and myself, which is perhaps particularly relevant given Tyler Cowen's recent assessment of o3 being AGI.
Milan:
over the last couple of years i've gotten to conviction that the term "AGI" isn’t well-specified enough to be useful
to unpack this a bit, the core issue is that AI is already superhuman at many things, and still far below human at many other things
there’s this weird thing where everyone now says "AGI isn’t well-specified as a concept" as one of their talking points, but then proceeds to talk about AGI timelines
(sorta like how there's a popular talking point of "ivy league education doesn’t really track quality" but ~everyone still gives a lot of weight to ivy league degrees in practice... a reflexivity)
there's a parallel to industrial revolution.
consider an "artificial general machine", as potentially envisioned in the 19th century.
it turns out that "machining" isn't one thing, and its deployment is massively uneven. there are some tasks that machines are so much better than humans (digging holes, moving dirt, bending metal) that it's incomparable.
others are ~equal to us but save time or automate a task (dish washer).
and others are only barely possible. most clothes manufacturing is still human-driven because it's a hard machine problem.
port to intelligence. Your pocket calculator is super intelligent at a very limited set of tasks. Very quickly, simple circuits outperform even the best humans. LLMs opened this space many orders of magnitude.
there's clearly still a gap, in ways we're only beginning to understand. there are still classes of problems the average 8 year old can do that LLMs cannot (mostly related to intuitive reasoning, world modeling, etc). sometimes, we come up with a very legible set of examples (like ARC), which can be directly optimized for (and we do).
this isn't saying that AI can't pass humans in all domains, but that what we're doing is SO fundamentally different that it'll be entirely blurry until it isn't.
just like answering "when did machines outperform humans in mechanical tasks?" is meaningless.
it is indeed meaningful to say things like "AI is better than nearly all humans at nearly all tasks", and plan for that reality. it's a huge change, but will probably stay fuzzy. we may not even notice the day it is true.
Cross-posted to my blog.
Good grief. Maybe we can retrospectively say that AGI was created in the 1990s. Maybe Babbage's Analytical Engine was AGI.
This is an absurd and discrediting statement. Sorry, Tyler Cowen, you've said some things I thought were interesting or insightful on blogs or podcasts before, and Ezra Klein respects you a lot, which goes a long way for me because I respect Ezra a lot. But this is sheer madness.
As Andrew Connor points out in the quoted text in this post, there are simple tests of fairly rudimentary intelligence like ARC-AGI-2 that, as far as we know (the full test results are still forthcoming), o3 isn't human-level on — and it's not like ARC-AGI-2 is a test of AGI, nor will ARC-AGI-3 be. Getting human-level on ARC-AGI-2 or ARC-AGI-3 is necessary, but not sufficient, for AGI. It's a very low bar to clear.
And that's just for starters. I recently posted a great talk by François Chollet here that points out many weaknesses in LLMs, and I can't imagine o3 overcomes all or even most of them.
There are all kinds of tests we could come up with for AGI, such as a deliberately challenging, adversarial Turing test with no time limits and with judges who are savvy about LLMs. Or the ability to fully and autonomously replace a highly skilled knowledge worker such as a journalist, an academic research assistant, or a paralegal with only the same level of input and supervision that a human in that role would require. Can current frontier AI systems like o3 do long-term, autonomous hierarchical planning at that level of complexity? No. Absolutely not. Not even close.
I think people should interpret Tyler Cowen making such a ridiculous statement as a signal that the AGI discourse is rotten and off the rails. This sucks.