You really can’t be confusing Ray Kurzweil with Ray Dalio…     See Less

He's taking credit for something he didn't invent. This test was proposed many times in the past several years     See Less

couple of thoughts on this: the moving the goal posts thing, that's a rare moment where I agree with Gary Marcus. It's a lie, they do move the goalposts but in the wrong direction. i     See More t a novel discovery that AGI can be understood in a way that if you compare it with todays LLMs or with hypothetical systems like them 20 years ago, one would have had to conclude those are not AGI. Marcus knows that, Hassabis knows that. OpenAI must know this too but had conflicting interests when tying their future to building AGI. There is some genuine confusion in the field about this, and also political game. Bengio is politically busy. He knows this too.
As you infer, the benchmarks don't tell you if a system as a capability like a human that does the same test and scores as high as the AI. Indeed a key issue is that those systems are pattern matching, though by scaffolding and inference scaling they're given a little more dynamic. I don't think LeCun is right that you can't build a system that is somehow based on what we now have with LLMs. What is missing is that these systems build world models progressively, with attempted consistency. it's possible this requires architectural breakthroughs as its likely that systems with external long term memory will always destabelize and loose coherence. it's also not elegant to try to slap AGI on top of AI instead of designing it as AGI from the ground up, that's where I am with LeCun.
The jagged intelligence is a red heering, though practically speaking you will be able to build overall very human level-ish systems despite that. I don't think invention is that hard, with not too high standards for it. LLMs have vastly more knowledge than humans, systems attuned for that will be able to surface ideas that are new to most humans but not for our collective knowledge base, in which these things perform superhuman stereotypical reasoning. I've heard some ideas that let me believe some of the limitations on world model consistency are solvable at architecture level in the short term (if it happens is another question).
I don't think AGI has an extremely sparsely populated solution space. It indeed seems most likely they will first stumble upon solutions adjacent to the LLM space, though entirely different designs can be thought of as well. Our brains will remain vastly more efficient than those LLM AGIs. and that's also interesting but once you get to Sheggs full (not minimal) AGI any other intellectual endevor looses relevance, problem solving then is just keeping it running long enough. without todays stupid mistakes.
I see some overlap between Antropics advances and Chollets ideas about AGI, maybe they get there first with some neurosymbolic approach, that gives lots of prominence to code. Using code to "reason" externally in a real python environment gives a real boost to these systems fluid intelligence.
The relevance of consciousness is that it affords just that continual learning and fluid intelligence in humans but AI may be a bastard. I think a few concepts from artificial life are gonna help with continual learning and giving AI open ended intelligence, possibly via scaffolding a slightly more adjustable and consistency seeking architecture. There are these fast in weight learning LLMs like RWKV 8 or newer archs by google like titan. maybe that's quite close to AGI. I also like neuromorphic concepts on AGI, but little money is invested there.
Only once these systems surpass human reliability in a structural sense they reach their full potential. "Jagging out" LLMs with more expertise areas to reason in within distribution is never gonna make it. Sutton is right to call out the age of experience. rethinking AI like that would eliminate all weaknesses, but realistically AI may cheat with superhuman knowledge and some more consistency and some more flexibility beyond its seen distributions. AGI doesn't have to mean ithas to think like us to surpass us in thinking. Just like planes are faster than birds without flapping wings.
still it seems useful to think in human terms as anchor. we are not pattern matchers. we are a flow like a stream of water instead, that flow is bend for pattern matching in us. that's a problem with matching distributions. we're not stochastic parrots but thermodynamic engines. for us, first comes motion, then algorithms. we're like weather systems, that afford intelligence. with neuromorphic designs, you can have that in dance AI too. but I suspect if you twist and tortue LLM alike architecture long enough, they can do that too.    See Less

Hasabis test is ridiculous. It took one of the greatest scientific minds in human history to develop relativity. AGI would be a system that can function in multiple domains as well as an ave     See More ) human. If you say every domain you would be talking about a superior intellect.    See Less

Eistein couldn't tell space and time are separate things and agreeing is your test of learning that sounds like a very low bar     See Less

I think we've already achieved AGI maybe call the Einstein level test ASI because a computer capable of that level of discovery is next level     See Less

There are many things that we humans can't do that a being with higher intelligence could. Would that intelligence also claim itself as general, and us as sub-general? The same way we hu     See More other animals as non-general intelligence?
I feel like general intelligence doesn't actually exist the way we say. A dog possesses general intelligence, but they can only generalize within their intelligence level. They can learn tricks or skills, and they can remember things, and whatnot. We humans can generalize within the boundaries our intelligence allows.
So, in my opinion, saying AGI is an AI that can do everything a human can is too much an anthropocentric definition.    See Less

They are making shyt up at this point.
Next test: The AI has to be as funny
as Dave Chappelle 😂
Gtfoh lmao     See Less

The music is so annoying     See Less

Gemini is fantastic, but the guardrails make using it almost as painful as GPT-5.x     See Less