Excuse me for being a bit snarky here...but if I could sum up the first speakers point it would "haha...I get to define intelligence...I don't have to respect any formal method of quantification...instead I can just move the goal post indefinitely to shield me from any uncomfortable truths" Either what we mean by terms like consciousness and intelligence is decidable or not. If it is decidable then machines can be applied to that endeavor... If it is not some people will get to walk around with a sense of unearned superiority bc there is no way to measure if their interactions with the rest of us invilve any thought it is similar to the pegion chess board analogy level of discourse
If deep learning was not abstract it would not be useful and it would not have issues of hallucination and alignment I would suggest that limits of deeplearning are over abstraction that creates a reachability issue for mechanistic interpetation.
"We have to build hybrid systems" The royal we are building hybrid systems Case in point deepmind built a silver medalist math olympiad level deeplearning solver They did so by combining classic deterministic algos with a countinious training model that became descreate after training. It seems trivial to me to insist that a model "must be on an analog continuous cycle like biological agents" Rather than intelligence can be treated like a tabula rasa brain in a vat In fact I would go so far as to say that insisting that an intelligent agent must be embedded in the world with an experience of time as dynamical as biological agents...is a more dangerous approach to agi than the current real time unembedded approach.
oh and btw...suggesting that real time embedding in the world "as it is" as if it is necessary for "real" progress is also a scale problem in terms of compute...so that point was a bit self refuting imo
"Why is it that we don't have AGI" Bc the term is loaded with philosophical baggage. So the question is not "why don't we have AGI" The question is "what do we have" and even if you want to call what we have now a hype bubble...in the tech space...if history repeats...that means we have a new technology that will revolutionize the world in ways that most people, even industry insiders, could not have imagined before that revolution took place.
"We need cognitive models" No the issue is we need world models that filter out noise humans don't deem as relevant unless those models can justify that the noise should be relevant bc it will align with a solution for some problem space humans are concerned about. Hand waving to theories about why this is not happening now won't change the intended goal of AI
"It may be compute" There is evidence to support this, see groking "It may be data" Scraping data from the surface indexed web is a lot more limited than the backend darkweb data pools still lurking out there, also creating synthetic data is trivial in some domains. That is why we are still seing gains in scaling in the space of robotics and AI...where synthetic data is just a simulated environment that produces useful real world robotics improvements. Hate to be this guy...but if either or both these things are valid...then they should both still fall under the scaling umbrella to my mind. No offense to the speaker...he has obviously thought about his perspectives a great deal and has raised valid points, I just don't agree they are as definitive as he seems to believe they are.
Disclaimer...I saying this in a context where I believe AGI is a cybersuper organism with human AI coberation. And in that the lag in the exponential is the result of a lag in adoption assuming that premise is true. That would mean ASI will have a lag in adoption that realizes the promises of AGI as they are marketed now. But of course all these projections are scale based so I concede if other issues besides scale exist than those circumstances, be they exigent, will be unaddressed by this conjecture.
