[mhardcastle]

This goes massively against the consensus of experts in this field. The modal AI researcher believes that "high-level machine intelligence", roughly AGI, will be achieved by 2047, per the survey below. Given the rapid pace of development in this field, it's likely that timelines would be shorter if this were asked today.

https://www.vox.com/future-perfect/2024/1/10/24032987/ai-imp...

[ein0p]

I am in the field. The consensus is made up by a few loudmouths. No serious front line researcher I know believes we’re anywhere near AGI, or will be in the foreseeable future.

[comp_throw7]

So the researchers at Deepmind, OpenAI, Anthropic, etc, are not "serious front line researchers"? Seems like a claim that is trivially falsified by just looking at what the staff at leading orgs believe.

[ein0p]

Apparently not. Or maybe they are heavily incentivized by the hype cycle. I'll repeat one more time: none of the currently known approaches are going to get us to AGI. Some may end up being useful for it, but large chunks of what we think is needed (cognition, world model, ability to learn concepts from massive amounts of multimodal, primarily visual, and almost entirely unlabeled, input) are currently either nascent or missing entirely. Yann LeCun wrote a paper about this a couple of years ago, you should read it: https://openreview.net/pdf?id=BZ5a1r-kVsf. The state of the art has not changed since then.

[comp_throw7]

I hope you have some advanced predictions about what capabilities the current paradigm would and would not successfully generate.

Separately, it's very clear that LLMs have "world models" in most useful senses of the term. Ex: https://www.lesswrong.com/posts/nmxzr2zsjNtjaHh7x/actually-o...

I don't give much credit to the claim that it's impossible for current approaches to get us to any specific type or level of capabilities. We're doing program search over a very wide space of programs; what that can result in is an empirical question about both the space of possible programs and the training procedure (including the data distribution). Unfortunately it's one where we don't have a good way of making advance predictions, rather than "try it and find out".

[ein0p]

It is in moments like these that I wish I wasn’t anonymous on here and could bet a 6 figure sum on AGI not happening in then next 10 years, which is how I define “foreseeable future”.

[ToValueFunfetti]

You disagreed that 2047 was reasonable on the basis that researchers didn't think it wouldn't happen in the foreseeable future, so your definition must be at least 23 years for consistency's sake

[fastball]

LeCun has his own interests at heart, works for one of the most soulless corporations I know of, and devotes a significant amount of every paper he writes to citing himself.

He is far from the best person to follow on this.

[ein0p]

Be that as it may, do you disagree with anything concrete from this paper?

[fastball]

Fair, ad hominems are indeed not very convincing. Though I do think everyone should read his papers through a lens of "having a very high h-index seems to be a driving force behind this man".

Moving on, my main issue is that it is mostly speculation, as all such papers will be. We do not understand how intelligence works in humans and animals, and most of this paper is an attempt to pretend otherwise. We certainly don't know where the exact divide between humans and animals is and what causes it, which I think is hugely important to developing AGI.

As a concrete example, in the first few paragraphs he makes a point about how a human can learn to drive in ~20 hours, but ML models can't drive at that level after countless hours of training. First you need to take that at face value, which I am not sure you should. From what I have seen, the latest versions of Tesla FSD are indeed better at driving than many people who have only driven for 20 hours.

Even if we give him that one though, LeCun then immediately postulates this is because humans and animals have "world models". And that's true. Humans and animals do have world models, as far as we can tell. But the example he just used is a task that only humans can do, right? So the distinguishing factor is not "having a world model", because I'm not going to let a monkey drive my car even after 10,000 hours of training.

Then he proceeds to talk about how perception in humans is very sophisticated and this in part is what gives rise to said world model. However he doesn't stop to think "hey, maybe this sophisticated perception is the difference, not the fundamental world model". e.g. maybe Tesla FSD would be pretty good if it had access to taste, touch, sight, sound, smell, incredibly high definition cameras, etc. Maybe the reason it takes FSD countless training hours is because all it has are shitty cameras (relative to human vision and all our other senses). Maybe linear improvements in perception leads to exponential improvement in learning rates.

Basically he puts forward his idea, which is hard to substantiate given we don't actually understand the source of human-level intelligence, and doesn't really want to genuinely explore (i.e. steelman) alternate ideas much.

Anyway that's how I feel about the first third of the paper, which is all I've read so far. Will read the rest on my lunch break. Hopefully he invalidates the points I just made in the latter 2/3rds.

[MacsHeadroom]

51% odds of the ARC AGI Grand Prize being claimed by the end of next year, on Manifold Markets.

https://manifold.markets/JacobPfau/will-the-arcagi-grand-pri...

[fastball]

This could also just be an indication (and I think this is the case) that many Manifold betters believe the ARC AGI Grand Prize to be not a great test of AGI and that it can be solved with something less capable than AGI.

[enragedcacti]

I don't understand how you got 2047. For the 2022 survey:

    - "How many years until you expect: - a 90% probability of HLMI existing?" 
    mode: 100 years
    median: 64 years

    - "How likely is it that HLMI exists: - in 40 years?"
    mode: 50%
    median: 45%

And from the summary of results: "The aggregate forecast time to a 50% chance of HLMI was 37 years, i.e. 2059"

[Retr0id]

Reminds me of what they've always been saying about nuclear fusion.