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

171–180 of 345 posts

Re: DeepMind: A Generalist Agent

#171

Post by londons_explore »

gurkendoktor wrote: post #111
hans1729 wrote: post #5

I’m not sure how to word my excitement about the progress we see in AI research in the last years. If you haven’t read it, give Tim Urbans classic piece a slice of your attention: https://waitbutwhy.com/2015/01/artificial-intelligence-revol... It’s a very entertaining read from a couple of years ago (I think I’ve read it in 2017), and man, have things happened in the field since then. If feels like things truly start…

Not you specifically, but I honestly don't understand how positive many in this community (or really anyone at all) can be about these news. Tim Urban's article explicitly touches on the risk of human extinction, not to mention all the smaller-scale risks from weaponized AI. Have we made any progress on preventing this? Or is HN mostly happy with deprecating humanity because our replacement has more teraflops? Even t…

> Or is HN mostly happy with deprecating humanity because our replacement has more teraflops?

If we manage to make a 'better' replacement for ourselves, is it actually a bad thing? Our cousin's on the hominoid family tree are all extinct, yet we don't consider that a mistake. AI made by us could well make us extinct. Is that a bad thing?

Re: DeepMind: A Generalist Agent

#172

Post by londons_explore »

triceratops wrote: post #16
f38zf5vdt wrote: post #14

Earlier quoted context omitted.

That human intelligence might just be token prediction evolving from successive small bit-width float matrix transformations is depressing to me.

> That human intelligence might just be token prediction I mean have you heard the word salad that comes out of so many people's mouths? (Including myself, admittedly)

Eating salad is good for your health. Not only word salad, but green salad and egg salad.

Re: DeepMind: A Generalist Agent

#173

Post by fullstackchris »

habitue wrote: post #80
Imnimo wrote: post #76

Earlier quoted context omitted.

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…

Agreed, I think if they were to drop the real-time constraint for the sake of the robotics tasks, they could train a huge model with the lessons from PaLM and Chincilla and probably slam dunk the weakly general AI benchmark.

I'm in the camp that thinks we're headed in a perpendicular direction and won't ever get to human levels of AGI with current efforts based on the simple idea that the basic tooling is wrong from first principles. I mean, most of the "progress" in AI has been due to getting better and learning how to understand a single piece of technology: neural networks.

A lot of recent neuroscience findings have shown that human brains _aren't_ just giant neural networks; in fact, they are infinitely more complex. Until we start thinking from the ground up how to build and engineer systems that reflect the human brain, we're essentially wandering around in the dark with perhaps only a piece of what we _think_ is needed for intelligence. (I'm not saying the human brain is the best engineered thing for intelligence either, but I'm saying it's one of the best examples we have to model AI after and that notion has largely been ignored)

I generally think it's hubris to spit in the face of 4 billion years of evolution thinking that some crafty neural net with X number more parameters will emerge magically as a truly generally intelligent entity - it will be a strange abomination at best.

Re: DeepMind: A Generalist Agent

#174

Post by chriswarbo »

f38zf5vdt wrote: post #14
hans1729 wrote: post #5

I’m not sure how to word my excitement about the progress we see in AI research in the last years. If you haven’t read it, give Tim Urbans classic piece a slice of your attention: https://waitbutwhy.com/2015/01/artificial-intelligence-revol... It’s a very entertaining read from a couple of years ago (I think I’ve read it in 2017), and man, have things happened in the field since then. If feels like things truly start…

That human intelligence might just be token prediction evolving from successive small bit-width float matrix transformations is depressing to me.

That's a poor usage of "just": discovering that "X is just Y" doesn't diminish X; it tells us that Y is a much more complex and amazing topic than we might have previously thought.

For example: "Life is just chemistry", "Earth is just a pile of atoms", "Behaviours are just Turing Machines", etc.

Re: DeepMind: A Generalist Agent

#175

Post by 37ef_ced3 »

idiotsecant wrote: post #170
gurkendoktor wrote: post #111

Earlier quoted context omitted.

Not you specifically, but I honestly don't understand how positive many in this community (or really anyone at all) can be about these news. Tim Urban's article explicitly touches on the risk of human extinction, not to mention all the smaller-scale risks from weaponized AI. Have we made any progress on preventing this? Or is HN mostly happy with deprecating humanity because our replacement has more teraflops? Even t…

I think the best-case scenario is that 'we' become something different than we are right now. The natural tendency of life(on the local scale) is toward greater information density. Chemical reactions beget self-replicating molecules beget simple organisms beget complex organisims beget social groups beget tribes beget city states beget nations beget world communities. Each once of these transitions looks like the de…

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?

Re: DeepMind: A Generalist Agent

#176

Post by Jensson »

VikingCoder wrote: post #167
Jensson wrote: post #159

Earlier quoted context omitted.

You can use a real persons contact details as long as the AI does all communication and work. Also it has to be the same AI, no altering the AI after you see the tasks it needs to perform after it gets the job, it has to understand that itself. For teleconferencing it could use text to speech and speech to text, they are pretty good these days so as long as the AI can parse what people say and identify when to speak…

I find it interesting that you have not put any kind of limit on how much can be spent to operate this AI. Or on what kinds of resources it would have access to. Could it, for instance, take its salary, and pay another human to do all or part of the job? [1] Or how about pay humans to answer questions for it? [2] [3] Helping it understand its assignments, by breaking them down into simpler explanations? Helping it im…

I said it has to manage all communications and do all the work, so no forwarding communications to third party humans. If it can convince other humans in the job to do all its work and coast that way it is fine though.

> Does it have to make more than its total operational expenses, or could I spend ten or hundreds as much as its salary, to afford the compute resources to implement it?

Yes, spend as much as you want on compute, the point is to show some general intelligence and not to make money. So even if this experiment succeeds it will be a ton of work left to do before the singularity, which is why I choose this kind of work as it is a nice middle ground.

> You also haven't indicated how many attempts I could make, per success. Could I, for instance, make tens of thousands of attempts, and if one holds down a job for a year, is that a success?

If the AI applies to 10 000 jobs and holds one of them for a year and gets paid that is fine. Humans do similar things. Sometimes things falls between the cracks, but that is pretty rare so I can live with that probability, if they made a bot that can apply to and get millions of jobs to get high probabilities of that happening then I'll say that it is intelligent as well, since that isn't trivial.

Re: DeepMind: A Generalist Agent

#177

Post by DavidSJ »

fullstackchris wrote: post #173
habitue wrote: post #80

Earlier quoted context omitted.

Agreed, I think if they were to drop the real-time constraint for the sake of the robotics tasks, they could train a huge model with the lessons from PaLM and Chincilla and probably slam dunk the weakly general AI benchmark.

I'm in the camp that thinks we're headed in a perpendicular direction and won't ever get to human levels of AGI with current efforts based on the simple idea that the basic tooling is wrong from first principles. I mean, most of the "progress" in AI has been due to getting better and learning how to understand a single piece of technology: neural networks. A lot of recent neuroscience findings have shown that human b…

What are one or two of the recent neuroscience findings that you feel point most strongly towards what you are saying?

Re: DeepMind: A Generalist Agent

#178

Post by edouard-harris »

hans1729 wrote: post #5

I’m not sure how to word my excitement about the progress we see in AI research in the last years. If you haven’t read it, give Tim Urbans classic piece a slice of your attention: https://waitbutwhy.com/2015/01/artificial-intelligence-revol... It’s a very entertaining read from a couple of years ago (I think I’ve read it in 2017), and man, have things happened in the field since then. If feels like things truly start…

That Tim Urban piece is great. It's also an interesting time capsule in terms of which AI problems were and were not considered hard in 2015 (when the post was written). From the post:

> Build a computer that can multiply two ten-digit numbers in a split second—incredibly easy. Build one that can look at a dog and answer whether it’s a dog or a cat—spectacularly difficult. Make AI that can beat any human in chess? Done. Make one that can read a paragraph from a six-year-old’s picture book and not just recognize the words but understand the meaning of them? Google is currently spending billions of dollars trying to do it. Hard things—like calculus, financial market strategy, and language translation—are mind-numbingly easy for a computer, while easy things—like vision, motion, movement, and perception—are insanely hard for it.

The children's picture book problem is solved; those billions of dollars were well-spent after all. (See, e.g., DeepMind's recent Flamingo model [1].) We can do whatever we want in vision, more or less [2]. Motion and movement might be the least developed area, but it's still made major progress; we have robotic parkour [3] and physical Rubik's cube solvers [4], and we can tell a robot to follow simple domestic instructions [5]. And Perceiver (again from DeepMind [6]) took a big chunk out of the perception problem.

Getting a computer to carry on a conversation [7], let alone draw art on par with human professionals [8], weren't even mentioned as examples, so laughably out of reach they seemed in the heathen dark ages of... 2015.

And as for recognizing a cat or a dog — that's a problem so trivial today that it isn't even worth using as the very first example in an introductory AI course. [9]

If someone re-wrote this post today, I wonder what sorts of things would go into the "hard for a computer" bucket? And how many of those would be left standing in 2029?

[1] https://arxiv.org/abs/2204.14198

[2] https://arxiv.org/abs/2004.10934

[3] https://www.youtube.com/watch?v=tF4DML7FIWk

[4] https://openai.com/blog/solving-rubiks-cube/

[5] https://say-can.github.io/

[6] https://www.deepmind.com/open-source/perceiver-io

[7] https://arxiv.org/abs/2201.08239v2

[8] https://openai.com/dall-e-2/

[9] https://www.fast.ai/

Re: DeepMind: A Generalist Agent

#179

Post by gurkendoktor »

londons_explore wrote: post #171
gurkendoktor wrote: post #111

Earlier quoted context omitted.

Not you specifically, but I honestly don't understand how positive many in this community (or really anyone at all) can be about these news. Tim Urban's article explicitly touches on the risk of human extinction, not to mention all the smaller-scale risks from weaponized AI. Have we made any progress on preventing this? Or is HN mostly happy with deprecating humanity because our replacement has more teraflops? Even t…

> Or is HN mostly happy with deprecating humanity because our replacement has more teraflops? If we manage to make a 'better' replacement for ourselves, is it actually a bad thing? Our cousin's on the hominoid family tree are all extinct, yet we don't consider that a mistake. AI made by us could well make us extinct. Is that a bad thing?

Your comment summarizes what I worry might be a more widespread opinion than I expected. If you think that human extinction is a fair price to pay for creating a supercomputer, then our value systems are so incompatible that I really don't know what to say.

I guess I wouldn't have been so angry about any of this before I had children, but now I'm very much in favor of prolonged human existence.

Re: DeepMind: A Generalist Agent

#180

Post by sinenomine »

abeppu wrote: post #130
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 think a key problem is our understanding of the quality of an ML system is tied to a task. Our mechanism of training is tied to a loss, or some optimization problem. The design, training, and evaluation of these systems is dependent on an externally provided definition of "correct". But this seems structurally different from how we or even less intelligent animals operate. DALL-E may make "better" art than most hum…

> Methodologically, how do we create agents that aren't just good at several tasks, but make up their own tasks

It's a good question, it has been asked a few times, and there are some answers[1][2] already, with the most general being to endow the agent with intrinsic motivation defined as an information-theoretic objective to maximize some definition of surprise. Then the agent in question will develop a general curious exploration policy, if trained long enough.

> Further, how do we create agents that can learn without ever seriously failing?

Another good question. One of the good enough answers here is that you should design a sequence of value functions[3] for your agent, in such a way, as to enforce some invariants over its future, possibly open-ended, lifetime. For this specific concern you should ensure that your agent develops some approximation of fear, leading to aversion of catastrophic failure regions in its state space. It's pretty self-evident that we develop such a fear in the young age ourselves, and where we don't, evolution gives us a hand and makes us preemptively fear heights, or snakes, even before we ever see one.

The other answer being, of course, to prove[4] a mathematical theorem around some hard definition of safe exploration in reinforcement learning.

1. https://people.idsia.ch/~juergen/interest.html

2. https://www.deepmind.com/publications/is-curiosity-all-you-n...

3. https://www.frontiersin.org/articles/10.3389/fncom.2016.0009...

4. https://arxiv.org/abs/2006.03357

Post reply on HN