I'm trying to remember all the past theories I heard for why AGI / intelligence explosion / etc. was far away, that were of the form "OK AIs can now do X, but there isn't evidence that they can do Y, at least not anytime soon."
Today: "OK so they are getting really good at well-specified tasks. But what about non-well-specified tasks?" In the past: Recombining or extending existing ideas vs. coming up with new ideas Interpolation vs. extrapolation shallow pattern-matching vs. true common-sense understanding of the world and/or true reasoning Benchmark-passing vs. real-world usefulness Text prediction vs. taking actions? (not sure if people actually held this line, I don't remember exactly) (with the following ones my memory is also getting fuzzy because it was almost ten years ago now) 1v1 games vs. team games perfect information games vs. imperfect information games simple games vs. complex games ... Please help me remember more?
I wasn't trying to imply that no one still believes them, or even that they aren't true. Now that you mention it, "Can competently act autonomously for year-long time horizons" is one that's still true basically.
What does your tracking tell us about the trajectory of well-specified tasks versus not?
1roha
Beyond the loud noise generated by people like Gary Marcus and Yann LeCun, I found Stuart Russell to make good arguments about internal representations of neural networks lacking expressive power and/or being brittle, which for example allowed his team to beat different "superhuman" Go playing programs by exploiting specific shortcomings in their board state representation. He talked about it in many older lectures and interviews. I'm commuting and on phone, so here's just the first and certainly not best references:
https://www.reddit.com/r/ChatGPT/comments/1hzmg1s/professor_stuart_russell_highlights_the
https://www.reddit.com/r/neuralnetworks/comments/1bycyq5/would_the_loop_technique_that_beat_alphago_work
With LeCun, the first example that comes to my mind is this Lex Fridman interview prediction about LLMs never being able to learn spatial reasoning from text corpora:
https://www.reddit.com/r/OpenAI/comments/1d5ns1z/yann_lecun_confidently_predicted_that_llms_will/