Hacker News Reader

Comment by ipnon | Hacker News Reader

The adage that we overestimate technology in the short term and we underestimate it in the long term holds. There is no way we all have a personal AGI assistants that drive our cars and work jobs for us this decade. At the end of this century we will all have this technology, barring any one of the possible apocalyptic catastrophes besetting us at present. The question is whether this advance is possible without some irrational exuberance today. We need the cult of AI to inspire us to think it’s even possible in the first place.

View on HN

No the majority of researchers in the field who have been researching for decades already know deep learning isn't the correct way forward it's just the vocal minority claiming it not to be true for nothing more than purely financial interests.

The autonomous driving question is an interesting one. It's probably some combination of techno-optimism, naive extrapolation of deep learning advances over a fairly short period of time, and--yes--at least some degree of a massive game of topper because everyone else was feeding at the funding trough and saying things were right around the corner. How much of each doubtless depends on the person.

And a lot of other people listened to the vocal "experts" and assumed that they must know what they're talking about.

No the majority of researchers in the field who have been researching for decades already know deep learning isn't the correct way forward

(playing Devil's Advocate here)

That doesn't even make sense. Correct way forward to what exactly? I guess if you assume that the only valid goal of AI research is to eventually achieve AGI then you might have a sort-of valid point. Except that nobody "knows" that deep learning isn't the way forward to AGI. Nobody "knows" that it is either.

But all of that said, current deep learning research has absolutely created systems that do amazing things and create tremendous value for society. So in what sense can we say that it "isn't the correct way forward"? Do you suggest abandoning DL completely? If so, in favor of what?

(OK "Devil's Advocate" portion over)

My own position has long been that "just deep learning" might be sufficient to achieve AGI if we eventually make networks that are big enough to support the right kind of emergent behavior. But I've also long been skeptical that this is the most direct or efficient path to AGI. I'm an advocate of hybrid symbolic/sub-symbolic systems that integrate deep learning with other AI strategies.

> My own position has long been that "just deep learning" might be sufficient to achieve AGI

as soon as you say agi, you've lost the plot.

> as soon as you say agi, you've lost the plot.

What plot? Would you care to elaborate?

Hard to imagine how that term could matter. It's just short-hand for some "thing" that people care to talk about. I'm not interested in quibbling over definitions or whatever. Nor am I interested in the kind of discussions that get all wrapped in what "general" means and reduce to arguing that even humans don't have "general intelligence", etc. To me, going down those rabbit-holes is really "losing the plot."

AGI is what someone says when they want to sound like they're talking about AI, but don't know anything about AI, and still feel the need to sound as if they're wearing a lab coat.

It's like when you're in a room where a bunch of people who can't name a medical school are trying to explain to each other what r0 means in a discussion that they believe is about COVID.

It's a red flag.

You do realize that they were trying to have a rational discussion here, right? They're pretty unambiguously asking for information and trying to understand your viewpoint.

If AGI is the wrong term to use then either get over it, or explain what term they should use. Particularly when they ask you to explain why the term is wrong.

The idea of a self aware AI is a little difficult for most people to define. Even if the term AGI isn't accurate, don't you think it's important that people be able to discuss the topic? Especially when those people are quite explicitly trying to learn more about it.

"You do realize that they were trying to have a rational discussion here, right?"

Rational is another HPMoR red flag word.

All rational means is "follows a rationale," or a rules system. Astrology, anti-vaxxing, belief in Hermetic magic, and chemtrails, while all ridiculous and incorrect, are also fully rational. So is Dungeons and Dragons, or internalizing Star Trek lore.

"They're pretty unambiguously asking for information and trying to understand your viewpoint."

And I gave it, clearly and politely.

"If AGI is the wrong term to use"

No, it's not the wrong term to use. Swap it with a synonym and you have the same problem.

Look. What if I started rambling about fixing aging? Like, straight up immortality. Would you think I was a compelling medical light, or an outsider with dreams that don't make sense given today's realities?

Flying cars? Pocket fusion devices?

"bUt ThEyRe ScIeNtIfIcAlLy PoSsIbLe"

It doesn't matter if I'm "using the wrong term." I could call them aerial vehicles, floating carriages, hover-Datsuns, whatever you want.

The problem is the idea. Anyone who's plying these ideas has completely missed the boat, and doesn't recognize that they're reciting bad science fiction.

"The idea of a self aware AI is a little difficult for most people to define"

I see you're still missing the boat.

"don't you think it's important that people be able to discuss the topic?"

It is approximately as important as discussing vampire repellant strategies.

What makes you say that? There's a vocal faction I've heard who essentially say that statistics can never be intelligent. I don't understand the argument. I think it's grounded on feeling like there should be something somewhere in the complicated system that you can point at and say it understands, but just like the old chinese room thought experiment I think the question of what it means to understand isn't the important one. Practically it's hard to deny that a system of approximation run statistically can do more or less the same things we can, because that appears to be what we are.

I think there's the much more defensible position that current techniques can't scale up to general artificial intelligence. That's pretty clear because we've already had to change details of model architecture a lot to get the models to where they are now

> majority of researchers in the field who have been researching for decades already know deep learning isn't the correct way forward

Do you have a reference for this? It’s not my experience.

There's also the AI adage that you don't get to the moon by building successively larger ladders... all this deep learning stuff is great and will unlock amazing value (esp if we fund open data+models), but we run the risk of exhausting consumer and government patience if we keep promising the world with this single family of techniques.

Not sure if the analogy holds. Human brains work on the same principle as lizard brais, no change in strategy was needed to go from lizard intelligence to human intelligence. I'm not saying that DL holds this promise though.

https://en.wikipedia.org/wiki/Triune_brain

There are distinct architectures and our brains have additional architectural features that lizards' lack.

https://en.wikipedia.org/wiki/Basal_ganglia

https://en.wikipedia.org/wiki/Limbic_system

https://en.wikipedia.org/wiki/Neocortex

Yes, but unlike a ladder and a rocket, they are not fundamentally different. This is why I don't think the analogy holds. Evolution iterated over the same model, never starting again from scratch.

That's interesting and something I hadn't heard before.

Any pointers to further reading on why deep learning might be a dead end?

They didn't say it was a dead end. I think their observation is rooted in gradient descent and how it improves but gets stuck in local maximas, when that solution might never be good enough. Hence the ladder. No matter how good you get at ladder making it with take you to the moon, but it will get you closer the entire time you improve at it.

For a contrasting viewpoint, there's always Sutton's Bitter Lesson - http://incompleteideas.net/IncIdeas/BitterLesson.html

How about a simple explanation of why it seems palpably absurd to some people that it will not be a dead end?

Consider a teratoma or lab-grown organ. Is grafting or engineering another attached sensory mechanism going to bring it closer to being an ordinary organism?

I think whether capabilities are super- or sub- human is a red herring.

Even a really primitive organism is still taking all of its computational capabilities and outputting, implicitly, decisions in one context that is its perceived reality.

A collection of computation and perception modules does not do this, without something else.

I don't think developing "something else" is obviously impossible or would require magic. But I'm not sure anyone sane would want to create it when it inherently creates unlimited risk of running amok. This is what the LessWrong people are afraid of, aren't they?

I think it's important to distinguish between something that's a "dead end" in that it will take you nowhere interesting, and something that's a "dead end" in that it will, after a certain point, cease advancing.

The former would imply that there is no point using deep learning and other similar techniques at all, and is the common implication when people say something is a "dead end".

The latter is what I believe the current generation of machine learning to be: I do not believe it will lead to AGI, and I am skeptical that it can do a great deal more than what it has already done (it can continue to refine the types of things it already does, and I expect it to do so, but I don't think it will open up new categories of things it can do many more times). But despite that, it does do some very cool things now, and as they are refined, I think they can be commercially successful and generally beneficial tools.

There's some real selection bias in that observation: if we overestimate a technology's impact in the long term, nobody's going to remember the technology or the estimates by the time the long term rolls around.

self driving cars would be easy if all the cars on the road were self driving, what makes it difficult is getting them to operate in an environment with illogical human drivers.

I'm certainly unconvinced that the first part of that statement is true. We've seen self-driving cars run into static obstacles. We still need to deal with emergent behaviour of groups of self-driving cars. There are still unpredictable environmental factors that we rely on humans to work out now.

No, pedestrians and cyclists still exist.

And animals and weather and road damage...

Self driving cars would be easy if all cars were self driving, all roads were designed for self driving cars, roads did not degrade, there are no pedestrians, no cyclists, trash or other obstructions could appear on the road, weather did not exist, and the sensory hardware could not get damaged or obstructed.

And even then... Any sufficiently complex system is chaotic (behaves in surprising, non-linear ways). And traffic networks pretty certainly constitute a sufficiently complex system.

And they already exist under these conditions: automated subway trains.

Very true Haha.

Very true but lets pretend we are in the real world where making all these changes is exponentially harder than only changing the cars. Listening to Geo-shot the main problem is to get them to co-exist with human drivers.

You are as overconfident as they are to say that there is "no way" that those things happen.

Let's try to not make false promise one way or another, just wait and see.

Reply on news.ycombinator.com