Live data from Hacker News

OpenAI Charter

blog.openai.com

161–170 of 183 posts

Re: OpenAI Charter

#161
post #156

Earlier quoted context omitted.

>> What is the responsible disclosure process for ideas and research around AI? Basically, we're so far away from AGI that there's no need to worry about disclosig anything. The recent advances in machine vision and speech processing are impressive, but only in the context of the last 50 years or so. A trully intelligent agent will need much more than this and there doesn't seem to be anyone alive today who knows how…

AlphaZero is existing general AI that learns boardgames on its own and reaches superhuman levels in less than a day. It seems to me quite plausible something like that could be generalised to understanding the physical world rather than that of chess or go. Admittedly we can't do that today but am not sure it's far off. Musk gives it five years in the recent movie https://www.youtube.com/watch?v=rqlwUAxoxa8&feature=y…

AlphaZero is not general AI. It can play Go (and chess and shogi) but that's very literally all it can do- it can't process speech or recognise objects in images, or any other task you might choose to test it against. It is extremely powerful, but also extremely limited, in the same way that computers have been for a very long time.

For instance, computers are decidedly better than humans at arithmetic. Yet nobody (today) thinks that just because computers can do arithmetic very quickly and accurately, they are more intelligent than humans.

Is there anything special in Go or chess that makes a specific skill at them a necessary and suficient condition for general intelligence? We are impressed at the performance of AlphaZero because humans find it very hard to play Go etc well. On the other hand, we also find it extremely hard to, e.g. divide two arbitrary 100-digit numbers -but we are not impressed by a pocket calculator as much as we are by AlphaZero. Might there be an element of psychological bias in our ability to be impressed, then?

Re: OpenAI Charter

#162
post #43

Earlier quoted context omitted.

point is: AI is different than your usual game, in that the winner might appear randomly, and destroy the world if she makes a mistake. so i believe open ais points are warrented

That's not going to happen, though. 1. AI is still very rudimentary, despite the advancements in particular applications made over the last decade. 2. The market doesn't trust AI because there are no validation methods trusted across commerce. Until a Verisign for AI emerges, AI will be regarded with suspicion in all business settings. AI/ML also goes above the heads of pretty much everyone who is not directly workin…

i love your points, and agree mostly.

that is why i express support that small entities like openai are thinking about ai science fiction outcomes.

i didnt say we need a cross government, 1 trillion funded, manhatten like project to protect us from agi, which is what we would need IF there was a reasonable chance that AGI will be developed in the next years.

in the meantime let some smart minds think about it, just as a little insurance and preparation

Re: OpenAI Charter

#163
post #158

Earlier quoted context omitted.

When would be a good time to start worrying? What external indicator will we get that we need to worry? And what makes you sure that, whenever that external indicator happens, we'll have enough time to stop problems from happening? If you're not familiar with it, I'm making the argument from here and it makes good further reading: https://intelligence.org/2017/10/13/fire-alarm/

When would be a good time to start worrying? Let's say two hundred years from now, give or take a few decades. I imagine that by that time we might have made some progress on the subject of intelligence, what it is and how to reproduce it. At that point, it might be prudent to worry about an AI getting out of control. We have no way of knowing what external indicators we might get. We have no idea what AGI would look…

Ok, so what makes you say 200 years? And more importantly, what would make you change your mind about that number (in either direction)?

> [...] if there is a threat that you don't know anything about, there is also nothing you can do about it.

I disagree with this, at least somewhat. For one thing, even without knowing the specific threat, there's lots of stuff you can do which is just generally helpful - e.g., try to spread to other planets. Will it necessarily help? No. But there are a lot of threats it will protect against.

More importantly, people actually working on the problem of AI safety say they are doing things that appear, at least to them, to be useful. What makes you so sure they're wrong? From the little I know of their research, it certainly seems like stuff that will likely be pretty helpful.

Last point - considering just how big a problem AGI could be if they're right, just how many resources would you want to devote to it? Literally zero?

Re: OpenAI Charter

#164
post #156

Earlier quoted context omitted.

AlphaZero is existing general AI that learns boardgames on its own and reaches superhuman levels in less than a day. It seems to me quite plausible something like that could be generalised to understanding the physical world rather than that of chess or go. Admittedly we can't do that today but am not sure it's far off. Musk gives it five years in the recent movie https://www.youtube.com/watch?v=rqlwUAxoxa8&feature=y…

AlphaZero is not general AI. It can play Go (and chess and shogi) but that's very literally all it can do- it can't process speech or recognise objects in images, or any other task you might choose to test it against. It is extremely powerful, but also extremely limited, in the same way that computers have been for a very long time. For instance, computers are decidedly better than humans at arithmetic. Yet nobody (t…

Go and chess are different in that they are designed as a competition for human thinking in a similar way that say marathons are designed to show off human running. Dividing 100 digit numbers less so. Also there was something human like in the way alphazero learns unlike say stockfish which is more calculator like.

Re: OpenAI Charter

#165
post #163

Earlier quoted context omitted.

When would be a good time to start worrying? Let's say two hundred years from now, give or take a few decades. I imagine that by that time we might have made some progress on the subject of intelligence, what it is and how to reproduce it. At that point, it might be prudent to worry about an AI getting out of control. We have no way of knowing what external indicators we might get. We have no idea what AGI would look…

Ok, so what makes you say 200 years? And more importantly, what would make you change your mind about that number (in either direction)? > [...] if there is a threat that you don't know anything about, there is also nothing you can do about it. I disagree with this, at least somewhat. For one thing, even without knowing the specific threat, there's lots of stuff you can do which is just generally helpful - e.g., try…

200 year is an entirely arbitrary period of time. The idea is that there's no way to predict when it will be possible to develop AGI but, if it happens at all, it will happen at some point in the future distant enough that nobody alive today will be around to say "I told you so".

You can ask how I know this. Obviously, I don't because I can't see into the future. But I can see the state of the art in the present and it's been baby steps for the last 70 years or so - and our capabilities have remained entirely primitive, industry hype nonwithstanding.

>> More importantly, people actually working on the problem of AI safety say they are doing things that appear, at least to them, to be useful. What makes you so sure they're wrong?

There's all sorts of opinions about how AI performance is accelerating ("exponentially"). In truth however, what has actually accelerated and in fact, plateaued in recent years is the performance in very specific tasks -object and speech recognition- and not in the general intelligence of AI systems. In fact, if you want to be more precise, we can only talk about advances in the context of very specific benchmarks, which is to say, specific datasets (like ImageNet) and according to specific metrics (say, F-score).

The problem is that all those benchmarks are arbitrarily chosen (iish; see below) and research teams spend a great deal of time tuning their systems to beat them. Which means, good performance in a benchmark tells us nothing about the extrinsic quality of a system: how it does in the real world, outside the lab and when the central assumption of PAC learning, that training and unseen data can be expected to have the same distribution (so that a system's training performance on the former predicts that on the latter) is not guaranteed. And then, performance in one type of task (e.g. classification) tells us nothing about the general capabilities of the system (i.e. general intelligence).

Singularitarians, like the people at MIRI (which your previous comment linked to) have focused on the performance of AI systems on modern benchmarks and the increase in compute, but modern benchmarks were essentially invented to allow some progress in machine learning, when mathematical results showed that progress was impossible. These previous results include Gold's famous result about learning in the limit (from the literature on Inductive Inference). The relaxation of assumptions was suggested in Leslie Valiant's paper "A theory of the learnable", which introduced PAC learning, the paradigm under which modern machine learning operates.

And to clarify my point above- modern machine learning benchmarks are not exactly chosen arbitrarily, rather they are justified by PAC learning assumptions about the learnability of propositional functions (and those, only; in fact, modern machine learning systems are propositional in nature, which severely restricts the expressive power of their models, making it much harder to realise the promise of Turing-complete learning of many of them).

... that probably got a bit too technical. My point is that just because we see imrpovement in performance today, in the field of research that we call machine learning, that doesn't mean that there is actual progress in the understanding of what intelligence is, or our ability to reproduce it. In a way, we might have changed our metrics, but we haven't necessarily improved our performance.

>> (...) try to spread to other planets.

So we have a science fiction problem and we're looking for science fiction solutions to it? :)

>> Last point - considering just how big a problem AGI could be if they're right, just how many resources would you want to devote to it? Literally zero?

Well, again that depends on what we can do about the problem, which in turn depends on what we know about it. I guess, like you say, if we know nothing about the problem, we can try random things like spreading to other planets or genetically enhancing the whole human race's intelligence until we ourselves are superintelligent and our risk to be taken over by an artificial superintelligence is 0.

But, having 0 certainty about the nature of AGI, we have exactly the same chances to avert the danger from it by sitting on our hands, as we have by migrating to other planets. I believe Bostrom's superintelligence scenario involves a machine that eventually colonises Mars with secret minining robots? If you're prepared to entertain the possibility of truly-Super intelligence, there's probably nothing you can do about it anyway.

Re: OpenAI Charter

#166
post #164

Earlier quoted context omitted.

AlphaZero is not general AI. It can play Go (and chess and shogi) but that's very literally all it can do- it can't process speech or recognise objects in images, or any other task you might choose to test it against. It is extremely powerful, but also extremely limited, in the same way that computers have been for a very long time. For instance, computers are decidedly better than humans at arithmetic. Yet nobody (t…

Go and chess are different in that they are designed as a competition for human thinking in a similar way that say marathons are designed to show off human running. Dividing 100 digit numbers less so. Also there was something human like in the way alphazero learns unlike say stockfish which is more calculator like.

There are competitions where people try to do arithmetic with very large numbers [1], or memorise very long sequences. The point is not what is an actual competition, which is a matter of tradition or culture; but what is easy and hard for a human to do, which is not (entirely) culturally determined.

I guess you could say that the difference between human and computer intelligence is that humans don't have total recall, while computers do. This allows computers to add large numbers without trouble or play millions of entire games at random and choose the best, etc.

On the other hand, there's always a chance that humans' incomplete recall is a hallmark of general intelligence.

>> Also there was something human like in the way alphazero learns unlike say stockfish which is more calculator like.

That's a matter of interpretation. AlphaZero plays by searching a huge space of possible moves- humans dont' play like that. It learns by playing itself millions of times. Again, humans don't learn like that.

There's nothing human about AlphaZero, nor Stockfish far as I can tell.

____________________

[1] https://en.wikipedia.org/wiki/Mental_Calculation_World_Cup

Re: OpenAI Charter

#167
post #163

Earlier quoted context omitted.

Ok, so what makes you say 200 years? And more importantly, what would make you change your mind about that number (in either direction)? > [...] if there is a threat that you don't know anything about, there is also nothing you can do about it. I disagree with this, at least somewhat. For one thing, even without knowing the specific threat, there's lots of stuff you can do which is just generally helpful - e.g., try…

200 year is an entirely arbitrary period of time. The idea is that there's no way to predict when it will be possible to develop AGI but, if it happens at all, it will happen at some point in the future distant enough that nobody alive today will be around to say "I told you so". You can ask how I know this. Obviously, I don't because I can't see into the future. But I can see the state of the art in the present and…

To fully answer your last point. Sorry- this has grown extremelly long. I'm actually doing work right now, waiting for a long experiment to run so I got some time to burn :)

So, my pet formula for deciding whether taking a risk is worth it is to multiply the probability of some adverse event occurring, let's call this event X, by the cost associated with that event, let's call it Y; so R = p(X) * Y, where R the risk and you can then decide on some risk threshold, T, where if R Now, when it comes to the Singularity (or, indeed, an Alien Invasion) we can accept the cost of the event to be infinite, Y = ∞, under the assumption that Superintelligence means game over for the species. But our knowledge of the event is nonexistent so the probability of the event, X = Singularity is... undefined. You can't plug that in my formula. You can't guess at the probability of X because nothing like X has ever happened before. You can try to extrapolate it from the development of human intelligence, but that took billions of years to evolve in a manner completely different than what we can reasonably expect for artificial intelligence (i.e. involving computers).

Basically, there's no way to calculate the risk from Superintelligence as long as we know nothing about it- 0 information means undefined risk. And therefore, no amount of resources can be justified to be spent to mitigate this risk.

In other words, I really couldn't answer your question- 0 resources, an infinite amount of resources, they're essentially the same.

Bottom line: to make decisions you need information even more than you need the resources to implement them.

Or in other words: you can't prepare for the unknown.

___________

OK, experiment's done cooking :)

Re: OpenAI Charter

#168
post #159
post #130

Earlier quoted context omitted.

> Are you saying that if it were discovered to be both cheap and easy to create a super-pandemic using nothing more than household chemicals and a swab of your own snot, then you'd publish the results freely and in their entirety? No. > If not, how catastrophic a level of harm are you willing to risk before you start advocating concealing the results? I'm not sure. Probably quite catastrophic, because it stretches th…

> Give me a credible scenario that brings us from strong AI to the annihilation of our species without handwaving about recursive self-improvement and selective idiocy and I'll reconsider my position. What kind of credible scenario are you looking for? You're obviously knowledgeable about the subject, I'm sure you've read various stories, but it's quite easy to dismiss them all as "not credible" or "just so stories".…

A credible scenario is one in which none of the parts of the story have a giant leap between them. For example, with your scenario:

AGI gets built by a farming company

Okay, sounds fine, keep going.

gets programmed to farm more corn

Yep, this makes sense to me.

but there is no stop condition programmed in

Definitely sounds like something that could happen, with you so far.

so converts ever available land into corn fields

...and you lost me. I'm not sure how we ended up here.

The AI you're postulating is 1) intelligent and capable enough to be called "AGI" and arbitrarily terraform land into a useful field for corn, yet 2) dumb and inept enough to interpret its instructions absolutely literally, like a magic genie; meanwhile 3) the collective capability of the human species is apparently insufficient to stop this from happening.

I guess that could happen. But I take it about as seriously as alien life trying to invade us, a random meteorite killing us all, a solar flare destroying the Earth or malicious time travelers. Each of these what if scenarios also have a sensible build up followed by a giant leap in suspension of disbelief, and concluding in disaster. That's not a framework for intelligent discussion and productive rationality, it's a self-indulgent and conceptually incoherent exercise in mental gymnastics. While we're at it we could argue about how many angels will fit on the head of a pin. If I start a news cycle about how quantum computers will allow us build nanotechnology that can hack human brains, does my idea have any credibility? It could happen, right?

Our current capability with respect to artificial intelligence is so far removed from any reasonable form of strong AI that there isn't even a recognizable path forward. Established leaders in the field like Yann LeCun have publicly stated deep learning will not take us there. We don't even know how to consistently and rigorously define strong AI, let alone conjecture about how it would work or what its theoretical danger would be. That means we're trying to extrapolate the properties of hypothetical nuclear weapons from gunpowder. We are effectively children grasping in the dark, getting hysterical about a monster that may or may not be lurking under our collective beds. We can't agree on what the monster looks like, we don't have a rational explanation for why we believe it exists, but we heard the floor creak and we've seen plenty of depictions of monsters in fiction that sound sort of logical.

Re: OpenAI Charter

#169

Earlier quoted context omitted.

200 year is an entirely arbitrary period of time. The idea is that there's no way to predict when it will be possible to develop AGI but, if it happens at all, it will happen at some point in the future distant enough that nobody alive today will be around to say "I told you so". You can ask how I know this. Obviously, I don't because I can't see into the future. But I can see the state of the art in the present and…

To fully answer your last point. Sorry- this has grown extremelly long. I'm actually doing work right now, waiting for a long experiment to run so I got some time to burn :) So, my pet formula for deciding whether taking a risk is worth it is to multiply the probability of some adverse event occurring, let's call this event X, by the cost associated with that event, let's call it Y; so R = p(X) * Y, where R the risk…

Do you sincerely think that we are totally unable to predict when--or if ever--we might create general intelligence, something which stupid evolution has done through trial and error, but humans can attempt with striking brilliance and worldwide motivated research efforts? You can't make any rough prediction on whether we will be able to reach that point, or when it will happen?

Re: OpenAI Charter

#170
There must be an AI quality index before anything. NOw a days anything and everything is being decorated with AI while the real use-case, technology and maturity is found only in few places.
Post reply on HN