AI & Logical Induction - Computerphile
ฝัง
- เผยแพร่เมื่อ 3 ต.ค. 2018
- Continuing to address the challenges of AI safety, Rob Miles discusses a paper from the Machine Intelligence Research Institute (MIRI).
Read the paper for yourself here: bit.ly/LogicalInduction
More from Rob Miles: bit.ly/Rob_Miles_TH-cam
/ computerphile
/ computer_phile
This video was filmed and edited by Sean Riley.
Computer Science at the University of Nottingham: bit.ly/nottscomputer
Computerphile is a sister project to Brady Haran's Numberphile. More at www.bradyharan.com
Definitely check out the paper for this one (link in the description). Even in a video this long there's so much cool stuff in there that we didn't have time for! I might make a more technical follow-up video myself if people want that
Absolutely
That would be great!
Please make a followup video, this is a really interesting subject
I am really looking forward to extra bits on this paper on your channel.
You need to eat more and do some exercise to gain some muscle.
Then people say philosphy is useless...this is pure formal epistemology, and it really shows in the references section of the paper BTW (Carnap, Priest, Hintikka, etc...)
Yes, Kant talks about the edges of computability. Except the philosophers of today can't compete or converse on the mathematical level. Which is to say philosophy isn't useless, but philosophers are... Lolz.
@@DJjakedrake how did you come to be so confident in such a false statement? Burnt out from that Williamson?
@@DJjakedrake we tend to be specialists nowadays. In the past, people could be mathematicians, philosophers and even artists all at the same time.
@@IcthiCalm down. It's a "Truthism".
@@whannabiand in the future we will be able to be none of those.
This is what’s been missing from recent computerphile videos. Rob!
He's great, and looks like a real-life Alex Kidd, too.
A young Terence McKenna...
Although I think this is the worst he's ever done, you're still right ;-)
>"spherical chickens in a vacuum"
I always heard this expressed as "spherical cows on a frictionless surface".
I heard it as an anecdote about physics.
A rich man came to biologist, statistician and physicist and asked them to predict the outcome of a horse race. Biologist looked at body structure and physical health of the horses, and named the probable winners. Statistician looked at outcomes of past races and named the probable winners. Then came physicist's turn. He was still busily writing and calculating. The rich man got impatient and asked, what is he doing. The physicist answered, "I am working on a model of spherical horses in a vacuum"... :)
I know it as "spehrical cow in a vacuum"
It turns out that physicists don't like being placed on a frictionless surface in a vacuum.
This video explained so much more than the title promised it would.
I just learned a lot. Thank you!
Great video.
One small caveat: the agents must be risk neutral and have a discount factor equal to one, for the conclusions in the video to be right (otherwise, for instance with risk-averse rational agents with a discount factor smaller than one, a 50% bet would be traded at LESS than 0.5, and vice-versa).
Really love all Rob Miles' videos!
wouldnt such agents go bankrupt in the limit, though?
i love this guy. More of him please. Computerphile used to do videos on real world stuff like cross site scripting (tom scott is dope) and like more feet on the ground real world programming things, not so etherial. This guy is more in the weeds, which i like.
The guy writes on toilet paper sheets and all.. but then he rolls the dice, and out of nowhere, the dice turns green! We can see it's trajectory in slow-mo UNDER THE CUP! Blew my mind! Direct thumbs up!
P.S.A.: There's an abridged version (from 131 pages down to 20) of the paper on logical induction. The link to it is given in the original article (see Description)
Get this genius a glass of water when you interview him next.
or make him a cup of tea.
@Tomasz
and of course make sure, if at all possible, that it's made by a safe AGI agent. And probably with no vases or children around, just for extra caution
He did, you can see it 14 seconds in
Computerphile needs infinitely more theory/math videos.
That is what Numberphile channe is for
This is a topic I've thought about for a long time. I'm excited to learn the theory.
Probably one of my favorite computerfile videos.
"we're not going to get too far into it" (looks at the video length) -rrr-right
The paper is 131 pages
@@RobertMilesAI wow! I will check out the paper. Btw, I'm shaking right now :D I read superintelligence upon your advice and watched all your videos! Thank you, you're awesome!
FYI here's a full lecture where one of the co-authors of the paper talks about it in more depth: watch?v=UOddW4cXS5Y
It's a great talk, I highly recommend watching it before trying to read the paper, which is quite technical.
Sounds like you need a device to throw ink at the page, to bypass friction with the delivery device. ink-jet-pen?
was going to suggest laser pointer based pen but then checked up on how did laser printers actually work again.. turns out heat to transfer powder to paper is actually produced by the drum and, if I understood correctly, laser is basically drawing negative image to counteract an electrostatic charge to prevent the powder sticking to non-printed parts of the paper. TIL. probably (I'll likely forget again soon)
Amazing channel! Keep up the great work!
Thanks Rob I enjoy listening to your explanations
Just wanna say thanks for your videos, always look forward to them!
This is so interesting, wish the video was a bit longer. I think he was about to talk about how the algorithm reacts to paradoxes
Very happy to find this video!
Very very interesting, great video! It is pure gold. Since the term is dropped multiple times, the rational choice theory has its limits. Individual entirely rational actions can lead in sum to irrational outcomes.
What's on that bookcase?
Sapiens - Yuval Noah Harari
Soonish - Zach Weinersmith
Run Program - Scott Meyer
Amazing ! Love it ! This concept is just super duper cool I'm so excited to read the paper now :3
I love it! Really fascinating.
Rob Miles! My favorite presenter.
It is now my life's goal to qualify all of my initial thoughts on solving a problem as "in a 'spherical chickens in a vacuum' sort of way"
Little correction: The price of futures does not actually depend on the expected future price. It is only a function of the current price and interest rates. That is the case because the predicted future price of the good is already reflected in the current price. If you would predict the price to go up in the future, you could also buy the good now and sell it in the future. By "no arbitrage" assumption, the expected value of doing this and selling a futures contract must be the same. As such the price in a futures contract will just be the current price plus interest for the time period.
Doesn't that assume completely durable goods? Buying strawberries could be different than strawberry futures, because that option for arbitrage wouldn't be available.
Edit: not trying to dispute you, just asking a question
@@toast_recon The future contract is durable - that you'd buy the current contract and then sell it later. Of course, real strawberries do begin to decline immediately but they were already purchased months or years ago in a futures market.
Very difficult topic to explain. Well done!
is there an implementation of explained algorithm?
What a legend! He has Soonish on his bookshelf! AI isn't my area but this paper looks like a really interesting. Thanks for bringing it to my attention CP.
This was really well explained!
The graphics were cool in this video
Great video!
Love you Rob!
22:00 Almost made me choke from laughing while eating...
I love Rob's tangents!
Super interesting from a mathematical and computer scientist's point of view.
Problem I already see is that with those contracts people blow a bubble of unrealistic prices for real goods by gambling. That leads to increasing prices for food and other real and basic resources so that people who already have hardly enough for a living (in developing countries for example) have to pay even more just to survive.
The Dutch invented shipping insurance. It's a Dutch Book because if you set the numbers right the sponsor makes a profit regardless - either the trades work or they claim it from the insurer. The original was 'I always win', not 'I always lose', but in probability that's kind of the same concept, just flipped.
25:47 I'd point out that by "loads of money" you mean infinite money (as time approaches infinity). You can in fact make loads of money for very high but finite values of "loads", because this efficient trader is very slow. It does a lot of things "in a timely manner" based on a definition in the paper, but the definition of "timely manner" is not very timely.
For example, you could make a lot of money by buying a lot of "The thousandth digit of pi is 9" and selling a lot of shares for every other digit. You couldn't get *unboundedly* high amounts of money because the inductor would eventually learn that the thousandth digit is 9, but you could probably get a lot in the meantime because until it ends up figuring out you're right it'll value all of those at $0.10.
17:50 According to Gödel there might be statements which can't be proved or disproved in a finite amount of time. An example could be Riemann's hypothesis about the roots of his zeta function. So the confidence value would be 1 - but w/o provability.
Amazing video also estimates :D
just a question why do they always write on that dot matrix paper with green lines on it...I've never seen that paper before.
Miles is so smart, it literally hurts me to listen with full attention for too long at a time.
When you say probability theory doesn't include a framework for including beliefs, that may be true for traditional frequentist probably theory, but it is absolutely a big part of Bayesian probability theory.
For a *really* good treatment of Bayesian probability theory I'd highly recommend Jaynes' book "Probability Theory: The Logic of Science". He makes a big point of pointing out that probabilities should be treated as degrees of belief which absolutely depend on a person's knowledge and he lays out all the mathematics needed for "updating one's belief" when you get more information or discover something by analyzing it (like in your square root example). This rule is simply Bayes theorem.
I think you misinterpreted him. He wasn’t saying that probability theory doesn’t describe how to update one’s subjective probability based on new evidence (rather, he says the opposite, that it does. He is talking about Bayesian probability.), but that it doesn’t describe how to update one’s subjective probability over time based purely on one taking more time to reason out the logical implications of the things one already knows (or already thinks likely).
He says that most probability theory assumes “logical omniscience”. E.g. if X and Y are two statements that turn out to be logically equivalent, standard probability theory requires that P(X)=P(Y), but determining if two statements are equivalent takes time and computation, possibly very large amounts of it.
And if you haven’t had time to check yet, then it seems like your probabilities for X and Y have to have the potential to be different, even though X and Y might turn out to be logically equivalent.
One way or another, I don't think the predicted jet fuel prices from last year are holding up right now ;)
Also, great video!
I can not literally imagine that the computers we use are result of countless transitions and one of those transition is just pure mathematics.
So what you're saying is now's a good time to buy Bitcoin?
No that's just the induction hypothesis
It was when you wrote that comment
Can't disprove that
always buy. trust me.
You should be doing arbitrage between all alt coins, silly.
One technical correction. "Most" futures contracts (as in most kinds, not most contracts) actually settle. If you forget to close out your contract (purchase an offsetting contract), you could find yourself the proud owner of a tank car of orange juice. Joy! Now what?* Of course, the way out is to close out your contract before the settlement date.
* That's basically the evidence the SEC (CFTC?) used in their case against the Hunt brothers in 1980 -- they held their contracts (LOTS of them) to maturity.
This is exactly why some Bayesians say that no probability is without prior. The only way to deal with probabilities on this sort of level is to be very explicit about what you know and don't know when you model your current state of knowledge.
Mathematician vs doubling cube: "Oh, powers of two".
Maybe I’ll try to peek at that paper (my old brain is hurting already). But it seems to me that formal specifications and rational choice theory are good but limited; they only work to the extent of one’s understanding of the universe of discourse (even if you were logically omniscient). But it seems like the unaccounted for possibilities are a pretty serious concern in these cases - I guess I’m just restating the basic problem of induction. Which isn’t to say that it’s not useful to try and optimize what can be done re. what we do know and the limited resources available to process that knowledge (I am a card carrying Bayesian) just that a dose of humility re. any conclusions seems useful.
If you can't use 1 or 0 probability, then it's ensuring the logic will fail because of the Halting Problem, instead of having limiting "Renormalization" boundaries, (I guess that's the logical objective of calculating probabilities in the first place). Very interesting discussion, thank you.
I guess that logical induction must be perfected before they continue developing self driving cars. There are so many variables in such a short amount of time that it will have for computations.
Can you do a video on honeypots, honeynets, and honeyfarms?
*Bringing the 'wisdom of markets' into AI research (and codifying it) intuitively seems like something that will be a game changer. It just makes sense, as it's how the world (and many natural systems within it) work.*
You're sort of looking at it backwards. Mathematics informs the structure and implementation of markets, as well as models of how they behave in the real world. Mathematics also informs computery stuff in exactly the same way. It's like saying "bringing Maxwell's equations into the structure of computers, the way it structures everything else, might just be a game changer." And what do you think people have been doing all along?
Look to the natural world instead. Things like evolutionary algorithms. Will they, in the end, be all that useful? I don't know but you'd be adding something to the algorithmist's toolbox that isn't already there.
Does having a bunch of algorithms trading their predictions on a market have any better or worse or different consequences than doing Bayesian inference on their validity?
in the part about the gathering evidence and accumulating enough beliefs to narrow your probability down, I actually sort of see it like a -- imagine a hologram to be your understanding of the thing. you see the hologram by perceiving many images of it from many different angles, each helping you to formulate a better understanding of what is being shown (for me, probabilities are not really like numbers, but more like blurry images that slowly make more and more sense with time), so even if you perceive it from all angles, you can never really "see" the hologram for what it really is. the more angles I see of it, it'll create a sort of meta-logical understanding of what it really is which I can now apply to things of similar nature.
so, how do you notice things of similar nature? well, I guess you could recognise aspects of details of the thing, but I personally "feel" it, and then the similar aspects start to show themselves.
what I'm trying to say is, while watching this video, I'm realising that I perceive things a bit backwards. like he said, the probability theory assumes logical omniscience and so therefore, if the pattern is not understood, it cannot be recognised. the brilliance to my approach is actually the assumption that never is it possible to perceive the whole thing at once (omniscience), and so therefore a gradually sharper understanding of what's happening in the image, allows for that continual revelation of what the probability really is, that logical omniscience assumes.
So, like, hindsight.
23:24 super cool if every trader would affect the system equally. Reality is that the ones who will bet more, will affect the price more, and if her/his prediction is off, then humanity's predictions is off.
Anyway, the video was great, thanks :)
Depends on your timeline length. Over a 30 year length, I'm not sure if that's true.
Isn't converging on a probability of 0 or 1 in contradiction to nondogmatism as it is essentially an approximation? To make an estimation of what an estimation is would seem to indicate an inaccuracy rather than an imprecision or a problem with applicability not application.
As a fan of MassEffect this explanation reminded me a LOT of how the Geth build a "Consensus" among them, and the more there are that communicate, the better they work.
Let's just hope we don't create something exactly similar and they exile us from the planet :D
Ngl getting exiled is the fifth best case scenario, and probably not happen
Do, would the AI become a Laplace's Demon if you let It to know more?
I love how Rob explains these papers.
7:53 Time
10:10 Seeing part of the process eliminates some wrong answers.
17:20 Well-Calibrated
19:02 Prediction Markets
Whats hillarious about this is that instantly beeing able to process the stuff around you, is the semipossible superpower of sherlock holmes.
Thanks for another great video with one of my favorite "mad scientists" (no offence ;).
And I feel I have to add that I to prefer the "long form" where the expert is allowed take their own, sometimes meandering, way to expand and explain their topic, and You only "gently steer" with Your question.
Jolly good job Sir :)
Best regards.
20:00 I now understand why futures markets are a thing.
class Task extends Goal {
// you have the robot inherit you're overall goals so that they don't obstruct those while pursuing specific task
}
Maxwell Jann // TODO: make AI a human
where is the video where he talks about - should and is as ways to describe what is and what we want - he was talking about always needing two shoulds in order to express a wish in order to state a logic construct for comeing up with that wish!?? would really like to watch that video again. got stuck with me
Going into 2021, where are we at with this?
More Rob, please.
what does this say about the nature of uncertainty, i mean is it an artifact of our ignorance, or is there something inherent in the physical processes that makes it so?
Make a video on how to think in recursion
Do I understand correctly? The minds which are thinking towards AGI safety engineering are using the paradigm of neoliberal commodity markets as the mathematical instrument for ‘value alignment’ in a formal system of reason. i.e a google adwords algorithm trading cheap reasons instead of cheap adverts. What are the other predicates of this system that aren’t philosophically grounded in Classical economic theory?
Why there isn't transcriber?
Lem "Summa Technologiae" on shelf.
The none dogmatism property is interesting, nice word to use instead of agnosticism.
I'm skeptical this approach works well in presence of randomness. Take a problem where a nondeterministic T machine is much faster than a regular one. Now consider betting on such a problem in presence of randomness. In such a case you can't have a condition about market not being beatable since someone could always get lucky and verify the answer.
There are not enough serious videos about computer intelligence
check bayesian stuff thats also a good logical induction with uncertainty
Could this algorithm apply to algorithmic stable coins?
This reminds me of TRON where the various programs compete against each other to see which one is fittest or something.
Oooh, I see that copy of Iain M Banks' "Consider Phlebas" on the shelf... nice.
I think it was a 3. Also, this video was really cool.
27 quality minutes!
So you're making the Dunning-Kruger-Effect measurable :D
I've never forgot about htis guy since the "difference between a difficult problem and a very difficult problem...".
It seems to be a sieve for AI models. Bad predictors can’t pass through the (50% accuracy) grid.
Seems a bit like Kalman filtering from feedback control theory.
Dutch was the term used for most germanic communities 100-300 years ago in the USA. This is still the case with the Pennsylvania Dutch, who in turn refer to all that is outside their community as "English".
Robs hair is back!!!!!!!!!!!!1 YEAH!!!!!!!!!!!
I have a problem with the probability relationships as presented, compare 13:30:
P(A), P(B), and P(A and B)
Thank you for scaring me with a superabstract paper. If you are all fancy for AI, you should probably consider to learn about didactics in the first place.
No wonder paper is "superabstract". The problem they are tackling is superabstract.
As a framework for producing safe AI (and I do not know how to code this) how about making the primary goal of the AI "minimize the impact of actions on the world in the process of achieving a task." This might make its utility more difficult to realize, but it should be safer. By "impact" I mean an increase of entropy. A human body is more organized than a smear on the floor (stepping on a baby to make tea). Converting a human body into crystalline forms of its constituents produces more entropy in the form of heat than leaving it alone.
Yeah, people have looked at these kind of information-theoretic impact metrics for AI Safety. I made a video about it a while ago. I think links don't work well in comments but just go to my channel and/or search for "Avoiding Negative Side Effects: Concrete Problems in AI Safety part 1" and "Empowerment: Concrete Problems in AI Safety part 2"
Thank you.
and that why u need to remember.......remember as much as u can anyway
reasoning.... that thing that would be really nice if more people did do...
Perhaps the last criteria that things that cannot be proven do not take values of 0 or 1 is not that obvious. It's my understanding that Goedel's incompleteness theorem implies that there exist some statements that cannot be proven to be true/false.
It seems to me like this supports the criteria because if something is not provable then you don't know whether it is true or false and therefore cannot assign definite 0/1 values to the statement.
It is controversial, because there are some meta-statements about logic itself that are unprovable. For example, you can't prove memory is valid and past exists. Or that logical deduction is actually logically valid. So even statements like "1=1" should have probability
"Goedel's incompleteness theorem implies that there exist some statements that cannot be proven to be true/false."
true
"this supports the criteria because if something is not provable then you don't know"
false,
because anything that is not-false is true, but that does not mean it is useful or meaningful.
But which number was under the cup?
The paper uses the qualifier "efficiently computable" so often that it doesn't really prove whole lot if the class of efficiently computable trading strategies is too small. Also the enumeration technique it uses is completely impractical.
I thought it just used "efficiently computable" to mean in polynomial time?
But yes, it isn't meant to be a practically useful to run, it is meant to be *an* algorithm which performs the desired task, to show that there is any such algorithm.
Knowing that there is such an algorithm (even if the only one we know of is very slow), helps make us less confused about how uncertainty about logical statements works, and lets us reason about that idea better.
@@drdca8263 I don't remember entirely the context, but I think that wasn't my point. I am aware of the fact that efficiently computable generally means polynomial time. I probably meant that for a problem class where it is not known in general what polynomial time algorithms are capable of(In the context this hasn't really been established), it is of limited value to prove results about them. It is actually a fairly common thing in math and computer science to prove things about trivial or empty sets without realising it at the time. Theoretically this is a fine paper, from what I remember of it. I think for further work on the subject it has a lot of useful insight. My complaint was in end effect that none of it is practically directly applicable with out considerable additional theoretical effort. Again not bad. It just makes me less than excited about the paper.
An explanation so complex, it is complex.
Had a brain freeze watching this. At one point my brain slammed the door and shouted through the letter box, come back Tomorrow.
What about undecidability? I assume an AGI would be Turing Complete. There is in general no decision procedure for a Turing Complete system's operation, so all the talk about proving theorems about an AGI's behavior seems vacuous.
Just because something is Turing complete doesn't mean you can't prove theorems about it. (If it did, then we wouldn't be able to prove that anything was Turing complete, now would we? :P)
There is no general decision procedure which will always tell use whether a program halts. This is true.
However, this does not mean that it is impossible to make a program which takes as input a program, and either correctly says that the input program halts, correctly says that it doesn't halt, or says "I 'unno" . It is possible to make programs that do this.
Also, it is a little unclear what you mean when you describe the AI as "being turing complete". the AI is not a programming language. It doesn't take as input a source code and run it. Now, yes, the AI could simulate a turing machine, just as you or I could with pencil and paper.
Basically, that the halting problem is undecidable implies "you can't get an answer to *all* questions of this particular form", it doesn't mean "you can't get an answer to *any* question of this particular form" (for some particular form).