I’ve just been Googling to see what became of EURISKO. The results are baffling. Despite its success in its time, there has been essentially no followup, and it has hardly been cited in the last ten years. Ken Haase claims improvements on EURISKO, but Eliezer disagrees; at any rate, the paper is vague and I cannot find Haase’s thesis online. But if EURISKO is a dead end, I haven’t found anything arguing that either.
Perhaps in a future where Friendly AI was achieved, emissaries are being/will be sent back in time to prevent any premature discovery of the key insights necessary for strong AI.
a close coupling of representation syntax and semantics is neccessary for a discovery program to prosper in a given domain
This is a really interesting point; it seems related to the idea that to be an expert in something, you need a vocabulary close to the domain in question.
It also immediately raises the question of what the expert vocabulary of vocabulary formation/acquisition is, i.e. the domain of learning.
a close coupling of representation syntax and semantics is neccessary for a discovery program to prosper in a given domain
This is a really interesting point; it seems related to the idea that to be an expert in something, you need a vocabulary close to the domain in question.
It doesn’t seem that interesting to me: it’s just a restatement that “data compression = data prediction”. When you have a vocabulary “close to the domain” that simply means that common concepts are compactly expressed. Once you’ve maximally compressed a domain, you have discovered all regularities, and simply outputting a short random string will decompress into something useful.
How do you find which concepts are common and how do you represent them? Aye, there’s the rub.
It also immediately raises the question of what the expert vocabulary of vocabulary formation/acquisition is, i.e. the domain of learning.
So my guess would be that the expert vocabulary of vocabulary formation is the vocabulary of data compression. I don’t know how to make any use of that, though, because the No Free Lunch Theorems seem to say that there’s no general algorithm that is the best across all domains And so there’s no algorithmic way to find which is the best compressor for this universe.
This is a really interesting point; it seems related to the idea that to be an expert in something, you need a vocabulary close to the domain in question.
I’m not so sure about this. I am pretty good at understanding visual reality, and I have some words to describe various objects, but my vocabulary is nowhere near as rich as my understanding is (of course, I’m only claiming to be an average member of a race of fantastically powerful interpreters of visual reality).
Let me give you an example. Say you had two pictures of faces of two different people, but the people look alike and the pictures were taken under similar conditions. Now a blind person, who happens to be a Matlab hacker, asks you to explain how you know the pictures are of different people, presumably by making reference to the pixel statistics of certain image regions (which the blind person can verify with Matlab). Is your face recognition vocabulary up to this challenge?
I think “vocabulary” in this sense refers to the vocabulary of the bits doing the actual processing. Humans don’t have access to the “vocabulary” of their fusiform gyruses, only the result of its computations.
The most sensible explanation has, I think been mentioned previously: that EURISKO was both overhyped and a dead end. Perhaps the techniques it used fell apart rapidly in less rigid domains than rule-based wargaming, and perhaps its successes were very heavily guided by Lenat. It’s somewhat telling that Lenat, the only one who really knows how it worked, went off to do something completely different from EURISKO.
In this regard, one could consider something like EURISKO not as a successful AI, but as a successful cognitive assistant for someone working in a mostly unexplored rule-based system. Recall the results that AM, EURISKO’s predecessor, got—if memory serves me, it rediscovered a lot of mathematical principles, none of them novel, but duplicating mostly from scratch results that took many years and many mathematicians to find originally.
Not that I’m certain this is the case by a long shot, but it seems the most superficially plausible explanation.
From what I remember of the papers, it was pretty clear (though perhaps not stated explicitly) that AM “happened across” many interesting factoids about math, but it was Lenat’s intervention that declared them important and worth further study. I think your second paragraph implies this, but I wanted it to be explicit.
A reasonable interpretation of AM’s success was that Lenat was able to recognize many important mathematical truths in AM’s meanderings. Lenat never claimed any new discoveries on behalf of AM.
Lenat was also careful to note that AM’s success, such as it was, was very much due to the fact that LISP’s “vocabulary” started with a strong relation to mathematics. EURISKO didn’t show anything like reasonable performance until he realized that the vocabulary it was manipulating needed to be “close” to the modeled domain, in the sense that interesting (to Lenat) statements about the domain needed to be short, and therefore easy for EURISKO to come across.
Yeah, that was basically what I meant. My hypothesis was that if you gave AM to someone with good mathematical aptitude but little prior knowledge, they would discover a lot more interesting mathematical statements than they would have without AM’s help, by analogy to Lenat discovering more interesting logical consequences of the wargaming rules with EURISKO’s help than any of the experienced players discovered themselves.
Perhaps in a future where Friendly AI was achieved, emissaries are being/will be sent back in time to prevent any premature discovery of the key insights necessary for strong AI.
As silly explanations go, I prefer the anthropic explanation: In worlds where AI didn’t stagnate, you’re dead and hence not reading this.
Or in non-anthropic terms, strong AI could be done on present-day hardware, if we only knew how, and our survival so far is down to blind luck in not yet discovering the right ideas?
For how long, in your estimate, has the hardware been powerful enough for this to be so?
If Eurisko was a non-zero step towards strong AI, would it have been any bigger a step if Lenat had been using present-day hardware? Or did it fizzle because it didn’t have sufficiently rich self-improvement capabilities, regardless of how fast it might have been implemented?
Not all worlds in which you continue to exist are pleasant ones. I think Michael Vassar once called quantum immortality the most horrifying hypothesis he had ever taken seriously, or something along those lines.
Sure, but the idea that we should ignore futures where we are dead will still have some bizarre implications. For example, it would strongly contradict Nick Bostrom’s MaxiPOK principle (maximize the probability of an OK outcome). In particular, if you thought that the development of AGI would lead to utopia with probability p u, near instant human extinction with probability p e and torture of humans with probability p _ t, where
p t << p u
then one would have a strong motive to accelerate the development of AGI as much as possible, because the total probability of mediocre outcomes due to non-extinction global catastrophes like resource depletion or nuclear war increases every year that AGI doesn’t get developed. Your actions would be dominated by trying to increase the strength of the inequality p t << p u whilst getting the job done quickly enough that p u was still bigger than the probability of ordinary global problems such as global warming happening in your development window. You would do this even at the expense of increasing the probability p e—potentially until it was > 0.5. You’d better be damn sure that anthropic reasoning is correct if you’re going to do this!
If there’s quantum immortality, what proportion of your lives would be likely to be acutely painful?
I don’t have an intuition on that one. It seems as though worlds in which something causes good health would predominate over just barely hanging on, but I’m unsure of this.
I’ve just been Googling to see what became of EURISKO. The results are baffling. Despite its success in its time, there has been essentially no followup, and it has hardly been cited in the last ten years. Ken Haase claims improvements on EURISKO, but Eliezer disagrees; at any rate, the paper is vague and I cannot find Haase’s thesis online. But if EURISKO is a dead end, I haven’t found anything arguing that either.
Perhaps in a future where Friendly AI was achieved, emissaries are being/will be sent back in time to prevent any premature discovery of the key insights necessary for strong AI.
Hm, the abstract for that paper mentions that:
This is a really interesting point; it seems related to the idea that to be an expert in something, you need a vocabulary close to the domain in question.
It also immediately raises the question of what the expert vocabulary of vocabulary formation/acquisition is, i.e. the domain of learning.
It doesn’t seem that interesting to me: it’s just a restatement that “data compression = data prediction”. When you have a vocabulary “close to the domain” that simply means that common concepts are compactly expressed. Once you’ve maximally compressed a domain, you have discovered all regularities, and simply outputting a short random string will decompress into something useful.
How do you find which concepts are common and how do you represent them? Aye, there’s the rub.
So my guess would be that the expert vocabulary of vocabulary formation is the vocabulary of data compression. I don’t know how to make any use of that, though, because the No Free Lunch Theorems seem to say that there’s no general algorithm that is the best across all domains And so there’s no algorithmic way to find which is the best compressor for this universe.
(ETA: multiple quick edits)
I’m not so sure about this. I am pretty good at understanding visual reality, and I have some words to describe various objects, but my vocabulary is nowhere near as rich as my understanding is (of course, I’m only claiming to be an average member of a race of fantastically powerful interpreters of visual reality).
Let me give you an example. Say you had two pictures of faces of two different people, but the people look alike and the pictures were taken under similar conditions. Now a blind person, who happens to be a Matlab hacker, asks you to explain how you know the pictures are of different people, presumably by making reference to the pixel statistics of certain image regions (which the blind person can verify with Matlab). Is your face recognition vocabulary up to this challenge?
I think “vocabulary” in this sense refers to the vocabulary of the bits doing the actual processing. Humans don’t have access to the “vocabulary” of their fusiform gyruses, only the result of its computations.
The most sensible explanation has, I think been mentioned previously: that EURISKO was both overhyped and a dead end. Perhaps the techniques it used fell apart rapidly in less rigid domains than rule-based wargaming, and perhaps its successes were very heavily guided by Lenat. It’s somewhat telling that Lenat, the only one who really knows how it worked, went off to do something completely different from EURISKO.
In this regard, one could consider something like EURISKO not as a successful AI, but as a successful cognitive assistant for someone working in a mostly unexplored rule-based system. Recall the results that AM, EURISKO’s predecessor, got—if memory serves me, it rediscovered a lot of mathematical principles, none of them novel, but duplicating mostly from scratch results that took many years and many mathematicians to find originally.
Not that I’m certain this is the case by a long shot, but it seems the most superficially plausible explanation.
From what I remember of the papers, it was pretty clear (though perhaps not stated explicitly) that AM “happened across” many interesting factoids about math, but it was Lenat’s intervention that declared them important and worth further study. I think your second paragraph implies this, but I wanted it to be explicit.
A reasonable interpretation of AM’s success was that Lenat was able to recognize many important mathematical truths in AM’s meanderings. Lenat never claimed any new discoveries on behalf of AM.
Lenat was also careful to note that AM’s success, such as it was, was very much due to the fact that LISP’s “vocabulary” started with a strong relation to mathematics. EURISKO didn’t show anything like reasonable performance until he realized that the vocabulary it was manipulating needed to be “close” to the modeled domain, in the sense that interesting (to Lenat) statements about the domain needed to be short, and therefore easy for EURISKO to come across.
Yeah, that was basically what I meant. My hypothesis was that if you gave AM to someone with good mathematical aptitude but little prior knowledge, they would discover a lot more interesting mathematical statements than they would have without AM’s help, by analogy to Lenat discovering more interesting logical consequences of the wargaming rules with EURISKO’s help than any of the experienced players discovered themselves.
As silly explanations go, I prefer the anthropic explanation: In worlds where AI didn’t stagnate, you’re dead and hence not reading this.
Or in non-anthropic terms, strong AI could be done on present-day hardware, if we only knew how, and our survival so far is down to blind luck in not yet discovering the right ideas?
For how long, in your estimate, has the hardware been powerful enough for this to be so?
If Eurisko was a non-zero step towards strong AI, would it have been any bigger a step if Lenat had been using present-day hardware? Or did it fizzle because it didn’t have sufficiently rich self-improvement capabilities, regardless of how fast it might have been implemented?
That is silly. In the same vein, why worry about any risks? You’ll continue to exist in whatever worlds they didn’t develop into catastrophe.
This is a very serious point and has been worrying me for some time. This problem connects to continuity of consciousness and reference classes.
Not all worlds in which you continue to exist are pleasant ones. I think Michael Vassar once called quantum immortality the most horrifying hypothesis he had ever taken seriously, or something along those lines.
Indeed. In particular, “dying of old age” is pretty damn horrifying if you think quantum immortality holds.
Sure, but the idea that we should ignore futures where we are dead will still have some bizarre implications. For example, it would strongly contradict Nick Bostrom’s MaxiPOK principle (maximize the probability of an OK outcome). In particular, if you thought that the development of AGI would lead to utopia with probability p u, near instant human extinction with probability p e and torture of humans with probability p _ t, where
p t << p u
then one would have a strong motive to accelerate the development of AGI as much as possible, because the total probability of mediocre outcomes due to non-extinction global catastrophes like resource depletion or nuclear war increases every year that AGI doesn’t get developed. Your actions would be dominated by trying to increase the strength of the inequality p t << p u whilst getting the job done quickly enough that p u was still bigger than the probability of ordinary global problems such as global warming happening in your development window. You would do this even at the expense of increasing the probability p e—potentially until it was > 0.5. You’d better be damn sure that anthropic reasoning is correct if you’re going to do this!
If there’s quantum immortality, what proportion of your lives would be likely to be acutely painful?
I don’t have an intuition on that one. It seems as though worlds in which something causes good health would predominate over just barely hanging on, but I’m unsure of this.
Hunh. I’m glad I’m not the only person who has always found quantum immortality far more horrifying than nonexistence.