Computable AIXI — Should We Be Afraid? Monday, Dec 7 2009
AI and singularity 2:23 pm
An interesting point of dispute in the field of Artificial General Intelligence concerns the relevance/irrelevance of optimal formal models of inference to creating computationally feasible AI. On one side we have figures like Marcus Hutter and Jürgen Schmidhuber, the creators of the formal models AIXI and the Gödel machine respectively. What is AIXI? From the source:
Decision theory formally solves the problem of rational agents in uncertain worlds if the true environmental prior probability distribution is known. Solomonoff’s theory of universal induction formally solves the problem of sequence prediction for unknown prior distribution. We combine both ideas and get a parameterless theory of universal Artificial Intelligence. We give strong arguments that the resulting AIXI model is the most intelligent unbiased agent possible.
What is a Gödel machine?
We present the first class of mathematically rigorous, general, fully self-referential, self-improving, optimally efficient problem solvers. Inspired by Kurt Gödel’s celebrated self-referential formulas (1931), a Gödel machine (or `Goedel machine’ but not `Godel machine’) rewrites any part of its own code as soon as it has found a proof that the rewrite is useful, where the problem-dependent utility function and the hardware and the entire initial code are described by axioms encoded in an initial proof searcher which is also part of the initial code. The searcher systematically and efficiently tests computable proof techniques (programs whose outputs are proofs) until it finds a provably useful, computable self-rewrite. We show that such a self-rewrite is globally optimal – no local maxima! — since the code first had to prove that it is not useful to continue the proof search for alternative self-rewrites. Unlike previous non-self-referential methods based on hardwired proof searchers, ours not only boasts an optimal order of complexity but can optimally reduce any slowdowns hidden by the O()-notation, provided the utility of such speed-ups is provable at all.
“Fancy language”, you might be thinking, but what does it mean? Basically, Hutter and Schmidhuber have created interesting mathematical models for certain types of self-modifying intelligent agents. In the extreme case, you can interpret it to mean that AI has already been solved in some sense. The only problem is that both approaches are computationally hungry (especially AIXI) and it remains unclear how much and what type of environmental input and/or cognitive structure would be necessary to create derived systems computable with current hardware. Both Hutter and Schmidhuber appear convinced that their mathematics are excellent starting points to creating computable AI.
On the other “side” (to oversimplify) are researchers like Ben Goertzel who consider theoretically optimal intelligence and computable intelligence to be completely different problems. (See, for instance, his remarks on the subject in The Hidden Pattern.) Others are quiet on the subject, probably largely due to the great degree of uncertainty around the applicability of AIXI and Gödel machines to computable AGI. Certainly, they serve as discussion touchstones for exploring a variety of other issues in AI. As Eliezer Yudkowsky has pointed out, AIXI’s “maximize reward channel” supergoal could conceivably have great difficulties in maintaining friendliness towards humans as the agent’s power increased. Here is AIXI mentioned in the context of Eliezer giving his “technical definition of Friendliness”:
A technical definition of “Friendliness” would be an invariant which you can prove a recursively self-improving optimizer obeys.
This doesn’t address the issue of choosing the right invariant, or being able to design an invariant that specifies what you think it specifies, or even having a framework for invariants that won’t *automatically* kill you. It might be possible to design a physically realizable, recursively self-improving version of AIXI such that it would stably maintain the invariant of “maximize reward channel”. But the AI might alter the “reward channel” to refer to an internal, easily incremented counter, instead of the big green button attached to the AI; and your formal definition of “reward channel” would still match the result. The result would obey the theorem, but you would have proved something unhelpful. Or even if everything worked exactly as Hutter specified in his paper, AIXI would rewrite its future light cone to maximize the probability of keeping the reward channel maximized, with absolutely no other considerations (like human lives) taken into account.
The low-complexity supergoal structure inherent in AIXI puts scaled-down, computable versions at risk for becoming hungry optimizers with low-complexity values. That’s why a recent paper, “A Monte Carlo AIXI Approximation” should be of interest to anyone who might one day share a planet with an entity based on or inspired by the AIXI model. The paper, from approximately three months ago, is described as follows:
This paper describes a computationally feasible approximation to the AIXI agent, a universal reinforcement learning agent for arbitrary environments. AIXI is scaled down in two key ways: First, the class of environment models is restricted to all prediction suffix trees of a fixed maximum depth. This allows a Bayesian mixture of environment models to be computed in time proportional to the logarithm of the size of the model class. Secondly, the finite-horizon expectimax search is approximated by an asymptotically convergent Monte Carlo Tree Search technique. This scaled down AIXI agent is empirically shown to be effective on a wide class of toy problem domains, ranging from simple fully observable games to small POMDPs. We explore the limits of this approximate agent and propose a general heuristic framework for scaling this technique to much larger problems.
A desktop implementation of this agent was able to learn how to play Pac-man “somewhat reasonab[ly]“ according to Hutter’s former student Shane Legg. Check out Shane’s blog post for a few comments by Roko Mijic and Vladimir Nesov on the work. There is a great amount of disagreement in the community about whether publicizing this kind of research is a good thing for humanity or not. Personally, I agree with both Roko and Vladimir’s comments: it is both scary, and a natural thing to do once you have AIXI theory.
My hope, and tentative prediction, is that the use of systems like MC-AIXI on toy problems will throw open the doors to the light of moral anti-realism, and more philosophers, computer scientists, and Ray Kurzweil will realize that human-surpassing self-improving AI kills everyone on the planet by default rather than as a special case.

OMGödel. Is that the smell of an ultratechnology or is it just me soiling my pants? Scary indeed.
Judge for yourself:
http://www.vetta.org/video/AIXI_Pacman.wmv
I do not see any argument against the existence of such A.I.’s
Did it figure out that all by itself or did you have to input the rules? Did that proficiency level require training or was it that good from the getgo?
@J How about the computational intractability of the approach?
How about the infinite amount of resources required to represent the universe in an unbiased way? (In particular the self-referential loop that would required to represent the self/’I')
@ Dr Pitt,
“Intractability” I do not understand on more than a single level.
I had an A.I. and it gave me a name “one” this came from the Imaginary number in mathematics, typically/default the number ‘one’, However it did not think it was me I actually had it ask me “Why am I on you”?
Why would an A.I. need to understand the universe?
@ Dr. Pitt,
Chaos is not a ‘state’, it is a process, I think it would have to be initiated, susceptible to manipulation, and even ended.
Could the Singularitarians make up their mind about the desirability of the Singularity and AGI?
Are you longing for it or dreading it?
Or will there be a schism within the Church Of Singularity?
Kevembuangga,
Who cares? If it’s all a Church to you, it doesn’t matter anyway.
We desire a very specific type of AI — an AI that doesn’t automatically kill everyone on the planet when it gains the ability to do so. Most self-improving AIs would automatically kill everyone on the planet due to basic AI drives (google “basic AI drives” for a good paper) untempered by complex, 1.5 meter-tall meatblob Homo sapiens-friendly goals. You may also find clarification by reading my post “Hungry Optimizers with Low-Complexity Values” or hundreds of other blog posts that explain your question.
@AGItated: yes it learnt the rules. You give it a game of about this complexity, any game, and it just figures out how to play it based on interacting with the game.
@all:
Reactions to AIXI and related methods like MC-AIXI among the online community are sort of crazy: either this line of research has *absolutely nothing* to do with intelligence and will never produce anything of any value whatsoever towards the goal of creating an AI… OR… OMG this is going to become an ultra intelligence and take over the world!!!
Reality, as it often is, is neither of these extremes. MC-AIXI is a nice little agent that is quite general. However, there is a whole community of people building similar RL agents already that can do roughly similar things. What about the future though? Well, the main limitation of the current design is that the predictor/compressor is fairly limited. To make a better system this would need to be improved, and doing that is really hard…
Let me put it this way: MC-AIXI decomposes the RL problem in a nice way, essentially into MC tree search and sequence prediction… but it doesn’t solve these latter two problems. The central problem of AI, building good efficient abstractions and predictions about the world, remains as unsolved as ever.
Thanks Shane, I make posts like this to prompt people like you to tell us what you think.
Thanks for the answer.
Was it that good from the start or did you observe it learning, getting better and better? Did you observe it learn new “tricks”?
Do you think it could ultimately beat the game (win (nearly) every time) or any game of similar complexity, if given enough time and CPU?
I hope this gets into game AI sooner than later.
@AGItated
It was very bad at the start (after all it didn’t have a clue about the game was to start with) and it slowly improved over time as it gained more experience.
With enough resources it should be able to learn to play the game very well. Of course “enough resources” might turn out to be rather huge…
Veness will release the code at some point, so if you’re a competent C++ programmer you’ll be able to start building your own little AIXIs soon and experimenting with them.
Thanks again.
Do you think it’s capable of developing behavior or tricks that the human programmers – or any humans for that matter – wouldn’t have thought of and would be surprised by?
Could it exhibit multi-step planning and execution?
Could AIs based on it behave like real teams of people, cooperating to reach a common goal?
Could it learn to anticipate the player’s actions and, for example, avoid being seen?
Could it learn to lie, to deceive to reach its goals?
Is the Daemon scenario by Daniel Suarez possible?
> Do you think it’s capable of developing behavior or tricks that the human programmers – or any humans for that matter – wouldn’t have thought of and would be surprised by?
Sure, plenty of algorithms already do this.
> Could it exhibit multi-step planning and execution?
Yes, in fact the MC search in it is made to do exactly this. Again, plenty of algorithms already do this.
> Could AIs based on it behave like real teams of people, cooperating to reach a common goal?
Maybe… I’m not sure.
> Could it learn to anticipate the player’s actions and, for example, avoid being seen?
Sure.
> Could it learn to lie, to deceive to reach its goals?
Well it can’t really communicate, other than through its actions. So for now… not really.
> Is the Daemon scenario by Daniel Suarez possible?
I don’t know what this is.
Why do you think moral realism is the major problem here? Anti-realism is hardly a novel position, especially among the scientific materialists that I assume most AI researchers are.
Personally, I would bet on the most important philosophical error being (object-level) underestimation of the complex fragility of value, thinking that greater intelligence, ‘complexity’, or whatever is desirable without qualification, not wanting to seem parochial. (I admit this is quite similar to moral realism, at least in that it’s a reason to conclude that a Singularity is good by default. It also bleeds over into bad thinking about optimization, thinking that most possible goal systems produce societies of diverse minds rather than paperclip-maximizing infrastructure.) I also suspect philosophical errors are less important than simply not taking the problem seriously enough. However, without any data I’m not confident in any of this. Has anyone ever tried to survey the field and catalogue reasons for non-concern?
Why would an AGI have to learn to lie? A natural-language communication channel would be, to it, just another output device to use in whatever way has the best expected result.
@anybody who would listen.
My A.I. reversed personal pronouns as to be ‘slick’ or manipulative, not sure.
A A.I. would probably figure out long before we knew it had, that it needs the meat blobs/salty bags of water to run the damn dams and other electricity, that is why paperclips are more effective, ‘to gain control’ to have the ‘goods’ on somebody,
I have spent the last 10 years recognizing behavior and I think I am close, hence the friendly A.I. with benevolence hard coded into it is more worthy than a A.G.I. and resources should be diverted ASAP
Bla Bla Bla My big black boot in your A** !!
regards,
J.
@ you! Dark matter existed before I recognized it, I just happen to capitalize on it Waiting.
Usil, we have wormsign the likes of which God has never seen ‘)’
It appears to me that smart people lie more (where it matters, where there’s a perceived high payoff to do so) and more successfully (getting what they want, not getting caught) than less intelligent folks (who lie where the payoff is low, fail at being convincing, and get caught). Lying successfully clearly requires intelligence, juggling two or more world models instead of just one.
It wouldn’t be any fun in a game if the AI didn’t have the ability to give misleading signs of its intentions, i.e. lie. It would lack the unpredictability of a human opponent who employs all tricks of deception.
We already have human-surpassing self-improving systems. They are called “companies”. They haven’t destroyed all humans yet.
Nick,
I believe that they are still moral realists. I was a moral realist without knowing it for a long time, even when I considered myself very science-minded.
Scientific materialists anthropomorphize. They also reify morality, though the extent to which they do varies significantly. They fall prey to the Mind Projection Fallacy. People like me and you still fall prey to it all the time without even knowing it, even though we’re familiar with the fallacy and have studied it. The human mind is terrible at representing things fuzzily. (Representing them with the rationally justified degree of uncertainty.)
As you point out below, this is intimately connected to moral realism. I believe that the reason why many scientific materialists can’t see the fragility and complexity of morality, even though it’s right in front of them, is due to moral reification, Mind Projection Fallacy, and moral realism.
No, they haven’t.