## The Problem with AIXI

**Followup to**: Solomonoff Cartesianism; My Kind of Reflection

**Alternate versions**: Shorter, without illustrations

AIXI is Marcus Hutter's definition of an agent that follows Solomonoff's method for constructing and assigning priors to hypotheses; updates to promote hypotheses consistent with observations and associated rewards; and outputs the action with the highest expected reward under its new probability distribution. AIXI is one of the most productive pieces of AI exploratory engineering produced in recent years, and has added quite a bit of rigor and precision to the AGI conversation. Its promising features have even led AIXI researchers to characterize it as an optimal and universal mathematical solution to the AGI problem.^{1}

Eliezer Yudkowsky has argued in response that AIXI isn't a suitable ideal to build toward, primarily because of AIXI's reliance on Solomonoff induction. Solomonoff inductors treat the world as a sort of qualia factory, a complicated mechanism that outputs experiences for the inductor.^{2} Their hypothesis space tacitly assumes a Cartesian barrier separating the inductor's cognition from the hypothesized programs generating the perceptions. Through that barrier, only sensory bits and action bits can pass.

Real agents, on the other hand, will be *in* the world they're trying to learn about. A computable approximation of AIXI, like AIXI*tl*, would be a physical object. Its environment would affect it in unseen and sometimes drastic ways; and it would have involuntary effects on its environment, and on itself. Solomonoff induction doesn't appear to be a viable conceptual foundation for artificial intelligence — not because it's an uncomputable idealization, but because it's Cartesian.

In my last post, I briefly cited three indirect indicators of AIXI's Cartesianism: immortalism, preference solipsism, and lack of self-improvement. However, I didn't do much to establish that these are *deep* problems for Solomonoff inductors, ones resistant to the most obvious patches one could construct. I'll do that here, in mock-dialogue form.

## Solomonoff Cartesianism

**Followup to**: Bridge Collapse; An Intuitive Explanation of Solomonoff Induction; Reductionism

**Summary**: If you want to predict arbitrary computable patterns of data, Solomonoff induction is the optimal way to go about it — provided that you're an eternal transcendent hypercomputer. A real-world AGI, however, won't be immortal and unchanging. It will need to form hypotheses about its own physical state, including predictions about possible upgrades or damage to its hardware; and it will need bridge hypotheses linking its hardware states to its software states. As such, the project of building an AGI demands that we come up with a new formalism for constructing (and allocating prior probabilities to) hypotheses. It will not involve just building increasingly good computable approximations of AIXI.

**Solomonoff induction** has been cited repeatedly as the theoretical gold standard for predicting computable sequences of observations.^{1} As Hutter, Legg, and Vitanyi (2007) put it:

Solomonoff's inductive inference system will learn to correctly predict any computable sequence with only the absolute minimum amount of data. It would thus, in some sense, be the perfect universal prediction algorithm, if only it were computable.

Perhaps you've been handed the beginning of a sequence like 1, 2, 4, 8… and you want to predict what the next number will be. Perhaps you've paused a movie, and are trying to guess what the next frame will look like. Or perhaps you've read the first half of an article on the Algerian Civil War, and you want to know how likely it is that the second half describes a decrease in GDP. Since all of the information in these scenarios can be represented as patterns of numbers, they can all be treated as rule-governed sequences like the 1, 2, 4, 8… case. Complicated sequences, but sequences all the same.

It's been argued that in all of these cases, one unique idealization predicts what comes next better than any computable method: Solomonoff induction. No matter how limited your knowledge is, or how wide the space of computable rules that could be responsible for your observations, the ideal answer is always the same: Solomonoff induction.

Solomonoff induction has only a few components. It has one free parameter, a choice of universal Turing machine. Once we specify a Turing machine, that gives us a fixed encoding for the set of all possible programs that print a sequence of 0s and 1s. Since every program has a specification, we call the number of bits in the program's specification its "complexity"; the shorter the program's code, the simpler we say it is.

Solomonoff induction takes this infinitely large bundle of programs and assigns each one a prior probability proportional to its simplicity. Every time the program requires one more bit, its prior probability goes down by a factor of 2, since there are then twice as many possible computer programs that complicated. This ensures the sum over all programs' prior probabilities equals 1, even though the number of programs is infinite.^{2}

## Bridge Collapse: Reductionism as Engineering Problem

**Followup to**: Building Phenomenological Bridges

**Summary**: AI theorists often use models in which agents are crisply separated from their environments. This simplifying assumption can be useful, but it leads to trouble when we build machines that presuppose it. A machine that believes it can only interact with its environment in a narrow, fixed set of ways will not understand the value, or the dangers, of self-modification. By analogy with Descartes' mind/body dualism, I refer to agent/environment dualism as *Cartesianism*. The open problem in Friendly AI (OPFAI) I'm calling naturalized induction is the project of replacing Cartesian approaches to scientific induction with reductive, physicalistic ones.

I'll begin with a story about a storyteller.

Once upon a time — specifically, 1976 — there was an AI named TALE-SPIN. This AI told stories by inferring how characters would respond to problems from background knowledge about the characters' traits. One day, TALE-SPIN constructed a most peculiar tale.

Henry Ant was thirsty. He walked over to the river bank where his good friend Bill Bird was sitting. Henry slipped and fell in the river. Gravity drowned.

Since Henry fell in the river near his friend Bill, TALE-SPIN concluded that Bill rescued Henry. But for Henry to fall in the river, gravity must have pulled Henry. Which means gravity must have been in the river. TALE-SPIN had never been told that gravity knows how to swim; and TALE-SPIN had never been told that gravity has any friends. So gravity drowned.

TALE-SPIN had previously been programmed to understand involuntary motion in the case of characters being pulled or carried by other characters — like Bill rescuing Henry. So it was programmed to understand 'character X fell to place Y' as 'gravity moves X to Y', as though gravity were a character in the story.^{1}

For us, the hypothesis 'gravity drowned' has low prior probability because we know gravity isn't the *type *of thing that swims or breathes or makes friends. We want agents to seriously consider whether the law of gravity pulls down rocks; we don't want agents to seriously consider whether the law of gravity pulls down the law of electromagnetism. We may not want an AI to assign *zero *probability to 'gravity drowned', but we at least want it to neglect the possibility as Ridiculous-By-Default.

When we introduce deep type distinctions, however, we also introduce new ways our stories can fail.

## Building Phenomenological Bridges

**Naturalized induction** is an open problem in Friendly Artificial Intelligence (OPFAI). The problem, in brief: Our current leading models of induction do not allow reasoners to treat their own computations as processes in the world.

The problem's roots lie in algorithmic information theory and formal epistemology, but finding answers will require us to wade into debates on everything from theoretical physics to anthropic reasoning and self-reference. This post will lay the groundwork for a sequence of posts (titled '**Artificial Naturalism**') introducing different aspects of this OPFAI.

## AI perception and belief: A toy model

A more concrete problem: Construct an algorithm that, given a sequence of the colors cyan, magenta, and yellow, predicts the next colored field.

*Colors: CYYM CYYY CYCM CYYY ????*

This is an instance of the general problem 'From an incomplete data series, how can a reasoner best make predictions about future data?'. In practice, any agent that acquires information from its environment and makes predictions about what's coming next will need to have two map-like^{1} subprocesses:

1. Something that generates the agent's predictions, its expectations. By analogy with human scientists, we can call this prediction-generator the agent's **hypotheses **or **beliefs**.

2. Something that transmits new information to the agent's prediction-generator so that its hypotheses can be updated. Employing another anthropomorphic analogy, we can call this process the agent's **data** or **perceptions**.

## Tiling Agents for Self-Modifying AI (OPFAI #2)

An early draft of publication #2 in the Open Problems in Friendly AI series is now available: Tiling Agents for Self-Modifying AI, and the Lobian Obstacle. ~20,000 words, aimed at mathematicians or the highly mathematically literate. The research reported on was conducted by Yudkowsky and Herreshoff, substantially refined at the November 2012 MIRI Workshop with Mihaly Barasz and Paul Christiano, and refined further at the April 2013 MIRI Workshop.

**Abstract:**

We model self-modication in AI by introducing 'tiling' agents whose decision systems will approve the construction of highly similar agents, creating a repeating pattern (including similarity of the offspring's goals). Constructing a formalism in the most straightforward way produces a Godelian difficulty, the Lobian obstacle. By technical methods we demonstrate the possibility of avoiding this obstacle, but the underlying puzzles of rational coherence are thus only partially addressed. We extend the formalism to partially unknown deterministic environments, and show a very crude extension to probabilistic environments and expected utility; but the problem of finding a fundamental decision criterion for self-modifying probabilistic agents remains open.

Commenting here is the preferred venue for discussion of the paper. This is an early draft and has not been reviewed, so it may contain mathematical errors, and reporting of these will be much appreciated.

The overall agenda of the paper is introduce the conceptual notion of a self-reproducing decision pattern which includes reproduction of the goal or utility function, by exposing a particular possible problem with a tiling logical decision pattern and coming up with some partial technical solutions. This then makes it conceptually much clearer to point out the even deeper problems with "We can't yet describe a probabilistic way to do this because of non-monotonicity" and "We don't have a good bounded way to do this because maximization is impossible, satisficing is too weak and Schmidhuber's swapping criterion is underspecified." The paper uses first-order logic (FOL) because FOL has a lot of useful standard machinery for reflection which we can then invoke; in real life, FOL is of course a poor representational fit to most real-world environments outside a human-constructed computer chip with thermodynamically expensive crisp variable states.

As further background, the idea that something-like-proof might be relevant to Friendly AI is not about achieving some chimera of absolute safety-feeling, but rather about the idea that the total probability of catastrophic failure should not have a significant conditionally independent component on each self-modification, and that self-modification will (at least in initial stages) take place within the highly deterministic environment of a computer chip. This means that statistical testing methods (e.g. an evolutionary algorithm's evaluation of average fitness on a set of test problems) are not suitable for self-modifications which can potentially induce catastrophic failure (e.g. of parts of code that can affect the representation or interpretation of the goals). Mathematical proofs have the property that they are as strong as their axioms and have no significant conditionally independent per-step failure probability if their axioms are semantically true, which suggests that something like mathematical reasoning may be appropriate for certain particular types of self-modification during some developmental stages.

Thus the content of the paper is very far off from how a realistic AI would work, but conversely, if you can't even answer the kinds of simple problems posed within the paper (both those we partially solve and those we only pose) then you must be very far off from being able to build a stable self-modifying AI. Being able to say how to build a theoretical device that would play perfect chess given infinite computing power, is very far off from the ability to build Deep Blue. However, if you can't even say how to play perfect chess given infinite computing power, you are confused about the rules of the chess or the structure of chess-playing computation in a way that would make it entirely hopeless for you to figure out how to build a bounded chess-player. Thus "In real life we're always bounded" is no excuse for not being able to solve the much simpler unbounded form of the problem, and being able to describe the infinite chess-player would be substantial and useful conceptual progress compared to *not *being able to do that. We can't be absolutely certain that an analogous situation holds between solving the challenges posed in the paper, and realistic self-modifying AIs with stable goal systems, but every line of investigation has to start somewhere.

Parts of the paper will be easier to understand if you've read Highly Advanced Epistemology 101 For Beginners including the parts on correspondence theories of truth (relevant to section 6) and model-theoretic semantics of logic (relevant to 3, 4, and 6), and there are footnotes intended to make the paper somewhat more accessible than usual, but the paper is still essentially aimed at mathematically sophisticated readers.

## Pascal's Muggle: Infinitesimal Priors and Strong Evidence

**Followup to:** Pascal's Mugging: Tiny Probabilities of Vast Utilities, The Pascal's Wager Fallacy Fallacy, Being Half-Rational About Pascal's Wager Is Even Worse

**Short form: **Pascal's Muggle

*tl;dr: If you assign superexponentially infinitesimal probability to claims of large impacts, then apparently you should ignore the possibility of a large impact even after seeing huge amounts of evidence. If a poorly-dressed street person offers to save 10 ^{(10^100)} lives (googolplex lives) for $5 using their Matrix Lord powers, and you claim to assign this scenario less than 10^{-(10^100)} probability, then apparently you should continue to believe absolutely that their offer is bogus even after they snap their fingers and cause a giant silhouette of themselves to appear in the sky. For the same reason, any evidence you encounter showing that the human species could create a sufficiently large number of descendants - no matter how normal the corresponding laws of physics appear to be, or how well-designed the experiments which told you about them - must be rejected out of hand. There is a possible reply to this objection using Robin Hanson's anthropic adjustment against the probability of large impacts, and in this case you will treat a Pascal's Mugger as having decision-theoretic importance exactly proportional to the Bayesian strength of evidence they present you, without quantitative dependence on the number of lives they claim to save. This however corresponds to an odd mental state which some, such as myself, would find unsatisfactory. In the end, however, I cannot see any better candidate for a prior than having a leverage penalty plus a complexity penalty on the prior probability of scenarios.*

In late 2007 I coined the term "Pascal's Mugging" to describe a problem which seemed to me to arise when combining conventional decision theory and conventional epistemology in the obvious way. On conventional epistemology, the prior probability of hypotheses diminishes exponentially with their complexity; if it would take 20 bits to specify a hypothesis, then its prior probability receives a 2^{-20} penalty factor and it will require evidence with a likelihood ratio of 1,048,576:1 - evidence which we are 1048576 times more likely to see if the theory is true, than if it is false - to make us assign it around 50-50 credibility. (This isn't as hard as it sounds. Flip a coin 20 times and note down the exact sequence of heads and tails. You now believe in a state of affairs you would have assigned a million-to-one probability beforehand - namely, that the coin would produce the exact sequence HTHHHHTHTTH... or whatever - after experiencing sensory data which are more than a million times more probable if that fact is true than if it is false.) The problem is that although this kind of prior probability penalty may seem very strict at first, it's easy to construct physical scenarios that grow in size vastly faster than they grow in complexity.

I originally illustrated this using Pascal's Mugger: A poorly dressed street person says "I'm actually a Matrix Lord running this world as a computer simulation, along with many others - the universe above this one has laws of physics which allow me easy access to vast amounts of computing power. Just for fun, I'll make you an offer - you give me five dollars, and I'll use my Matrix Lord powers to save 3↑↑↑↑3 people inside my simulations from dying and let them live long and happy lives" where ↑ is Knuth's up-arrow notation. This was originally posted in 2007, when I was a bit more naive about what kind of mathematical notation you can throw into a random blog post without creating a stumbling block. (E.g.: On several occasions now, I've seen someone on the Internet approximate the number of dust specks from this scenario as being a "billion", since any incomprehensibly large number equals a billion.) Let's try an easier (and *way *smaller) number instead, and suppose that Pascal's Mugger offers to save a googolplex lives, where a googol is 10^{100} (a 1 followed by a hundred zeroes) and a googolplex is 10 to the googol power, so 10^{10100} or 10^{10,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000} lives saved if you pay Pascal's Mugger five dollars, if the offer is honest.

## New report: Intelligence Explosion Microeconomics

**Summary**: Intelligence Explosion Microeconomics (pdf) is 40,000 words taking some initial steps toward tackling the key quantitative issue in the intelligence explosion, "reinvestable returns on cognitive investments": what kind of returns can you get from an investment in cognition, can you reinvest it to make yourself even smarter, and does this process die out or blow up? This can be thought of as the compact and hopefully more coherent successor to the AI Foom Debate of a few years back.

(Sample idea you haven't heard before: The increase in hominid brain size over evolutionary time should be interpreted as evidence about increasing marginal fitness returns on brain size, presumably due to improved brain wiring algorithms; not as direct evidence about an intelligence scaling factor from brain size.)

I hope that the open problems posed therein inspire further work by economists or economically literate modelers, interested specifically in the intelligence explosion *qua* cognitive intelligence rather than non-cognitive 'technological acceleration'. MIRI has an intended-to-be-small-and-technical mailing list for such discussion. In case it's not clear from context, I (Yudkowsky) am the author of the paper.

**Abstract:**

I. J. Good's thesis of the 'intelligence explosion' is that a sufficiently advanced machine intelligence could build a smarter version of itself, which could in turn build an even smarter version of itself, and that this process could continue enough to vastly exceed human intelligence. As Sandberg (2010) correctly notes, there are several attempts to lay down return-on-investment formulas intended to represent sharp speedups in economic or technological growth, but very little attempt has been made to deal formally with I. J. Good's intelligence explosion thesis as such.

I identify the key issue as

returns on cognitive reinvestment- the ability to invest more computing power, faster computers, or improved cognitive algorithms to yield cognitive labor which produces larger brains, faster brains, or better mind designs. There are many phenomena in the world which have been argued as evidentially relevant to this question, from the observed course of hominid evolution, to Moore's Law, to the competence over time of machine chess-playing systems, and many more. I go into some depth on the sort of debates which then arise on how to interpret such evidence. I propose that the next step forward in analyzing positions on the intelligence explosion would be to formalize return-on-investment curves, so that each stance can say formally which possible microfoundations they hold to be falsified by historical observations already made. More generally, I pose multiple open questions of 'returns on cognitive reinvestment' or 'intelligence explosion microeconomics'. Although such questions have received little attention thus far, they seem highly relevant to policy choices affecting the outcomes for Earth-originating intelligent life.

The **dedicated mailing list** will be small and restricted to technical discussants.