Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

DeepMind: A Generalist Agent

deepmind.com

Post reply on HN
Print view View original

201–210 of 345 posts

Re: DeepMind: A Generalist Agent

#201

Post by riversflow »

dekhn wrote: post #126
adamsmith143 wrote: post #120

Earlier quoted context omitted.

There's no evidence otherwise. You have to believe that the mind has a materialist basis or else you believe in woo woo magic.

sure, but I think it's fair to say that brains probably aren't doing lballistics calculations when a baseball player sees a pop fly and manveuvers to catch it. Rather, brains, composed mainly of neurons and other essential components, approximate partial differential equations, much like machine learning systems do.

> sure, but I think it's fair to say that brains probably aren't doing lballistics calculations when a baseball player sees a pop fly and manveuvers to catch it.

Well, I know you were talking about throwing, but there is some[1] talk/evidence in the evolutionary biology/neural darwinsm community that complex language development was a consequence of human developing the ability to hunt by throwing rocks (a very complicated and highly mathematical task). From my understanding after developing the required shoulder/arm morphology to throw at high speed brain sized tripled in early hominids.

So the brain actually might be doing something closer to math than we might think.

[1]https://www.sciencedirect.com/science/article/abs/pii/002251...

[2]https://link.springer.com/referenceworkentry/10.1007/978-3-5...

Re: DeepMind: A Generalist Agent

#202

Post by meekmind »

mach1ne wrote: post #124
kaivi wrote: post #115

Before you visualize a straight path between "a bag of cool ML tricks" and "general AI", try to imagine superintelligence but without consciousness. You might then realize that there is no obvious mechanism which requires the two to appear or evolve together. It's a curious concept, well illustrated in the novel Blindsight by Peter Watts. I won't spoil anything here but I'll highly recommend the book.

First you have to define consciousness, and especially the external difference between a conscious and non-conscious intelligence.

Likely insufficient but here is a shot at a materialist answer.

Consciousness is defined as an entity that has an ethical framework that is subordinated to it's own physical existence, maintaining that existence, and interfacing with other conscious entities as if they also have an ethical framework with similar parameters who are fundamentally no more or less important/capable than itself.

Contrast with non-conscious super-intelligence that lacks physical body (likely distributed). Without a physical/atomic body and sense data it lacks the capacity to empathize/sympathize as conscious entities (that exist within an ethical framework that is subordinated to those limitations/senses) must. It lacks the perspective of a singular, subjective being and must extrapolate our moral/ethical considerations, rather than have them ingrained as key to it's own survival.

Now that I think about it, it's probably not much different than the relationship between a human and God, except that in this case it's: a machine consciousness and a machine god.

To me, the main problem is that humans (at large) have yet to establish/apply a consistent philosophy with which to understand our own moral, ethical, and physical limitations. For the lack of that, I question whether we're capable of programming a machine consciousness (much less a machine god) with a sufficient amount of ethical/moral understanding - since we lack it ourselves (in the aggregate). We can hardly agree on basic premises, or whether humanity itself is even worth having. How can we expect a machine that we make to do what we can't do ourselves? You might say "that's the whole point of making the machine, to do something we can't" but I would argue we have to understand the problem domain first (given we are to program the machine) before we can expect our creations to apply it properly or expand on it in any meaningful way.

Re: DeepMind: A Generalist Agent

#203

Post by gurkendoktor »

idiotsecant wrote: post #183
37ef_ced3 wrote: post #175

Earlier quoted context omitted.

Let's be real. Not long from now all creative and productive work will be done by machines. Humans will be consumers. Why learn a skill when it can all be automated? This will eliminate what little meaning remains in our modern lives. Then what? I don't know, who cares?

>Then what? Growing tomatoes is less efficient than buying them, regardless of your metric. If you just want really cleanly grown tomatoes, you can buy those. If you want cheap tomatoes, you can buy those. If you want big tomatoes, you can buy those. And yet individual people still grow tomatoes. Zillions of them. Why? Because we are inherently over-evolved apes who like sweet juicy fruits. The key to being a success…

Humans are great at making up purpose where there is absolutely none, and indeed this is a helpful mechanism for dealing with post-scarcity.

The philosophical problem that I see with the "AI overlord age" (although not directly related to AI) is that we'll then have the technology to change the inherent human desires you speak of, and at that point growing tomatoes just seems like a very inefficient way of satisfying a reward function that we can change to something simpler.

Maybe we wouldn't do it precisely because it'd dissolve the very notion of purpose? But it does feel to me like destroying (beating?) the game we're playing when there is no other game out there.

(Anyway, this is obviously a much better problem to face than weaponized use of a superintelligence!)

Re: DeepMind: A Generalist Agent

#204

Post by Jack000 »

mxkopy wrote: post #195
sinenomine wrote: post #185

Earlier quoted context omitted.

It remains to be asked, just why this causal, counterfactual, logical reasoning cannot emerge in a sufficiently scaled-up model trained on a sufficiently diverse real world data? As far as we see, the https://www.gwern.net/Scaling-hypothesis continues to hold, and critics have to move their goalposts every year or two.

Neural networks, at the end of the day, are still advanced forms of data compression. Since they are Turing-complete it is true that given enough data they can learn anything, but only if there is data for it. We haven't solved the problem of reasoning without data, i.e. without learning. The neural network can't, given some new problem that has never appeared in the dataset, in a deterministic way, solve that proble…

data isn't necessarily a problem for training agents. A sufficiently complex, stochastic environment is effectively a data generator - eg. alphago zero

Re: DeepMind: A Generalist Agent

#205

Post by ehsankia »

viksit wrote: post #182

(Former AI researcher / founder here) It always surprises me at the ease at which people jump on a) imminent AGI and b) human extinction in the face of AGI. Would love for someone to correct me / add information here to the contrary. Generalist here just refers to a "multi-faceted agent" vs "General" like AGI. For a) - I see 2 main blockers, 1) A way to build second/third order reasoning systems that rely on intuitio…

For me at least, the fear is not so much about the specifics, but more around the fact of what exponential curves look like. At any point, everything before looks basically horizontal and anything after looks vertical. In that sense, the fear is that while things seem quite behind right now, it could in an instant zoom past us before we even have the time to realize it. It is partly rooted in science fiction.

Re: DeepMind: A Generalist Agent

#206

Post by sinenomine »

A(G)I has become a question of compute economics, for better or for worse. Those with more tightly integrated computational capacity or a good enough logistically sound plan to acquire just enough of it soon enough win, hard.

Should we, the people, watch in awe as our best and brightest financiers chase towards the ultimate prize, the key to all that future entails?

Are those respectable people worthy of the key, and what happens to us in this wild scenario?

Re: DeepMind: A Generalist Agent

#207

Post by drcode »

Imnimo wrote: post #76
habitue wrote: post #26

Is today the day? Date Weakly General AI is Publicly Known: https://www.metaculus.com/questions/3479/date-weakly-general... (I really like the framing of "weakly general AI" since it puts the emphasis on the generality and not whether it's a superintelligence) Edit: Probably not today, but mostly because 1.2B parameters isn't enough to get it the high winograd scores that PaLM etc have. But it seems pretty clear you…

I think this is a step in the right direction, but the performance on most tasks is only mediocre. The conversation and image captioning examples in the paper are pretty bad, and even on some relatively simple control tasks it performs surprisingly poorly. That's not to say it's not an important step. Showing that you can train one model on all of these disparate tasks at once and not have the system completely colla…

Yeah the thing that was so freaky about AlphaZero is that it was more powerful than AlphaGo, despite being more general.

This system lacks that feature.

Re: DeepMind: A Generalist Agent

#208

Post by justinpombrio »

viksit wrote: post #182

(Former AI researcher / founder here) It always surprises me at the ease at which people jump on a) imminent AGI and b) human extinction in the face of AGI. Would love for someone to correct me / add information here to the contrary. Generalist here just refers to a "multi-faceted agent" vs "General" like AGI. For a) - I see 2 main blockers, 1) A way to build second/third order reasoning systems that rely on intuitio…

I am quite scared of human extinction in the face of AGI. I certainly didn't jump on it, though! I was gradually convinced by the arguments that Yudkowsky makes in "Rationality: from AI to Zombies" (https://www.readthesequences.com/). Unfortunately they don't fit easily into an internet comment. Some of the points that stood out to me, though:

- We are social animals, and take for granted that, all else being equal, it's better to be good to other creatures than bad to them, and to be truthful rather than lie, and such. However, if you select values uniformly at random from value space, "being nice" and "being truthful" are oddly specific. There's nothing universally special about deeply valuing human lives any more so than say deeply valuing regular heptagons. Our social instincts are very ingrained, though, making us systematically underestimate just how little a smart AI is likely to care whatsoever about our existence, except as a potential obstacle to its goals.

- Inner alignment failure is a thing, and AFAIK we don't really have any way to deal with that. For those that don't know the phrase, here it is explained via a meme: https://astralcodexten.substack.com/p/deceptively-aligned-me...

So here's hoping you're right about (a). The harder AGI is, the longer we have to figure out AI alignment by trial and error, before we get something that's truly dangerous or that learns deception.

Re: DeepMind: A Generalist Agent

#209

Post by uoaei »

mhitza wrote: post #50
axg11 wrote: post #22

Slowly but surely we're moving towards general AI. There is a marked split across general society and even ML/AI specialists between those who think that we can achieve AGI using current methods and those who dismiss the possibility. This has always been the case, but what is remarkable about today's environment is that researchers keep making progress contrary to the doubter's predictions. Each time this happens, th…

I'm skeptical because we are building black boxes. How do you fix something you can't reason about? These billion parameter boxes are outside the reach of your everyday developers. In terms of cost of propping up the infrastructure makes them tenable only for megacorps. Most of us aren't moving goal posts, but are very much skeptic at the things we are being oversold on. I personally think we are still far away from…

> I'm skeptical because we are building black boxes.

An article came up a couple days ago that points to some interpretable features of the so-called black boxes you refer to. It's not that they are black boxes, it's that our torches are not yet bright enough.

https://vaclavkosar.com/ml/googles-pathways-language-model-a...

> Most of us aren't moving goal posts, but are very much skeptic at the things we are being oversold on.

I think a shift in perspective is warranted here. It's becoming increasingly clear that we may have vastly overestimated our own intelligence. Human exceptionalism is crumbling before us as we see how limited the tools are that pull off such incredible stunts. Judging based on other neuroscience and psychology research coming out, it really does seem like we are no more than statistical inference machines with specialized hardware that allow us to pack a lot of processing power into a small, energy-efficient system. We need to figure out next better learning algorithms, which probably depend quite heavily on the particular physical architecture.

Re: DeepMind: A Generalist Agent

#210

Post by idiotsecant »

samdjstephens wrote: post #197
idiotsecant wrote: post #188

Earlier quoted context omitted.

> I'm very much in favor of prolonged human existence. Serious question - why?

What are your axioms on what’s important, if not the continued existence of the human race? edit: I’m genuinely intrigued

I suppose the same axioms of every ape that's ever existed (and really the only axioms that exist). My personal survival, my comfort, my safety, accumulation of resources to survive the lean times (even if there are no lean times), stimulation of my personal interests, and the same for my immediate 'tribe'. Since I have a slightly more developed cerebral cortex I can abstract that 'tribe' to include more than 10 or 12 people, which judging by your post you can too. And fortunate for us, because that little abstraction let us get past smashing each other with rocks, mostly.

I think the only difference between our outlooks is I don't think there's any reason that my 'tribe' shouldn't include non-biological intelligence. Why not shift your priorities to the expansion of general intelligence?

Post reply on HN