Toward a New Technical Explanation of Technical Explanation
A New Framework
(Thanks to Valentine for a discussion leading to this post, and thanks to CFAR for running the CFAR-MIRI cross-fertilization workshop. Val provided feedback on a version of this post. Warning: fairly long.)
Eliezer’s A Technical Explanation of Technical Explanation, and moreover the sequences as a whole, used the best technical understanding of practical epistemology available at the time* -- the Bayesian account—to address the question of how humans can try to arrive at better beliefs in practice. The sequences also pointed out several holes in this understanding, mainly having to do with logical uncertainty and reflective consistency.
MIRI’s research program has since then made major progress on logical uncertainty. The new understanding of epistemology—the theory of logical induction—generalizes the Bayesian account by eliminating the assumption of logical omniscience. Bayesian belief updates are recovered as a special case, but the dynamics of belief change are non-Bayesian in general. While it might not turn out to be the last word on the problem of logical uncertainty, it has a large number of desirable properties, and solves many problems in a unified and relatively clean framework.
It seems worth asking what consequences this theory has for practical rationality. Can we say new things about what good reasoning looks like in humans, and how to avoid pitfalls of reasoning?
First, I’ll give a shallow overview of logical induction and possible implications for practical epistemic rationality. Then, I’ll focus on the particular question of A Technical Explanation of Technical Explanation (which I’ll abbreviate TEOTE from now on). Put in CFAR terminology, I’m seeking a gears-level understanding of gears-level understanding. I focus on the intuitions, with only a minimal account of how logical induction helps make that picture work.
There are a number of difficulties in applying Bayesian uncertainty to logic. No computable probability distribution can give non-zero measure to the logical tautologies, since you can’t bound the amount of time you need to think to check whether something is a tautology, so updating on provable sentences always means updating on a set of measure zero. This leads to convergence problems, although there’s been recent progress on that front.
Put another way: Logical consequence is deterministic, but due to Gödel’s first incompleteness theorem, it is like a stochastic variable in that there is no computable procedure which correctly decides whether something is a logical consequence. This means that any computable probability distribution has infinite Bayes loss on the question of logical consequence. Yet, because the question is actually deterministic, we know how to point in the direction of better distributions by doing more and more consistency checking. This puts us in a puzzling situation where we want to improve the Bayesian probability distribution by doing a kind of non-Bayesian update. This was the two-update problem.
You can think of logical induction as supporting a set of hypotheses which are about ways to shift beliefs as you think longer, rather than fixed probability distributions which can only shift in response to evidence.
This introduces a new problem: how can you score a hypothesis if it keeps shifting around its beliefs? As TEOTE emphasises, Bayesians outlaw this kind of belief shift for a reason: requiring predictions to be made in advance eliminates hindsight bias. (More on this later.) So long as you understand exactly what a hypothesis predicts and what it does not predict, you can evaluate its Bayes score and its prior complexity penalty and rank it objectively. How do you do this if you don’t know all the consequences of a belief, and the belief itself makes shifting claims about what those consequences are?
The logical-induction solution is: set up a prediction market. A hypothesis only gets credit for contributing to collective knowledge by moving the market in the right direction early. If the market’s odds on prime numbers are currently worse than those which the prime number theorem can provide, a hypothesis can make money by making bets in that direction. If the market has already converged to those beliefs, though, a hypothesis can’t make any more money by expressing such beliefs—so it doesn’t get any credit for doing so. If the market has moved on to even more accurate rules of thumb, a trader would only lose money by moving beliefs back in the direction of the prime number theorem.
This provides a framework in which we can make sense of mathematical labor. For example, a common occurrence in combinatorics is that there is a sequence which we can calculate, such as the catalan numbers, by directly counting the number of objects of some specific type. This sequence is boggled at like data in a scientific experiment. Different patterns in the sequence are observed, and hypotheses for the continuation of these patterns are proposed and tested. Often, a significant goal is the construction of a closed form expression for the sequence.
This looks just like Bayesian empiricism, except for the fact that we already have a hypothesis which entirely explains the observations. The sequence is constructed from a definition which mathematicians made up, and which thus assigns 100% probability to the observed data. What’s going on? It is possible to partially explain this kind of thing in a Bayesian framework by acting as if the true formula were unknown and we were trying to guess where the sequence came from, but this doesn’t explain everything, such as why finding a closed form expression would be important.
Logical induction explains this by pointing out how different time-scales are involved. Even if all elements of the sequence are calculable, a new hypothesis can get credit for calculating them faster than the brute-force method. Anything which allows one to produce correct answers faster contributes to the efficiency of the prediction market inside the logical inductor, and thus, to the overall mathematical understanding of a subject. This cleans up the issue nicely.
What other epistemic phenomena can we now understand better?
Lessons for Aspiring Rationalists
Many of these could benefit from a whole post of their own, but here’s some fast-and-loose corrections to Bayesian epistemology which may be useful:
Hypotheses need not make predictions about everything. Because hypotheses are about how to adjust your odds as you think longer, they can leave most sentences alone and focus on a narrow domain of expertise. Everyone was already doing this in practice, but the math of Bayesian probability theory requires each hypothesis to make a prediction about every observation, if you actually look at it. Allowing a hypothesis to remain silent on some issues in standard Bayesianism can cause problems: if you’re not careful, a hypothesis can avoid falsification by remaining silent, so you end up incentivising hypotheses to remain mostly silent (and you fail to learn as a result). Prediction markets are one way to solve this problem.
Hypotheses buy and sell at the current price, so they take a hit for leaving a now-unpopular position which they initially supported (but less of a hit than if they’d stuck with it) or coming in late to a position of growing popularity. Other stock-market type dynamics can occur.
Hypotheses can be like object-level beliefs or meta-level beliefs: you can have a hypothesis about how you’re overconfident, which gets credit for smoothing your probabilities (if this improves things on average). This allows you to take into account beliefs about your calibration without getting too confused about Hofstadter’s-law type paradoxes.
You may want to be a bit careful and Chesterton-fence existing Bayescraft, though, because some things are still better about the Bayesian setting. I mentioned earlier that Bayesians don’t have to worry so much about hindsight bias. This is closely related to the problem of old evidence.
Suppose a new scientific hypothesis, such as general relativity, explains a well-know observation such as the perihelion precession of mercury better than any existing theory. Intuitively, this is a point in favor of the new theory. However, the probability for the well-known observation was already at 100%. How can a previously-known statement provide new support for the hypothesis, as if we are re-updating on evidence we’ve already updated on long ago? This is known as the problem of old evidence, and is usually levelled as a charge against Bayesian epistemology. However, in some sense, the situation is worse for logical induction.
A Bayesian who endorses Solomonoff induction can tell the following story: Solomonoff induction is the right theory of epistemology, but we can only approximate it, because it is uncomputable. We approximate it by searching for hypotheses, and computing their posterior probability retroactively when we find new ones. It only makes sense that when we find a new hypothesis, we calculate its posterior probability by multiplying its prior probability (based on its description length) by the probability it assigns to all evidence so far. That’s Bayes’ Law! The fact that we already knew the evidence is not relevant, since our approximation didn’t previously include this hypothesis.
Logical induction speaks against this way of thinking. The hypothetical Solomonoff induction advocate is assuming one way of approximating Bayesian reasoning via finite computing power. Logical induction can be thought of as a different (more rigorous) story about how to approximate intractible mathematical structures. In this new way, propositions are bought or sold at market prices at the time. If a new hypothesis is discovered, it can’t be given any credit for ‘predicting’ old information. The price of known evidence is already at maximum—you can’t gain any money by investing in it.
There are good reasons to ignore old evidence, especially if the old evidence has biased your search for new hypotheses. Nonetheless, it doesn’t seem right to totally rule out this sort of update.
I’m still a bit puzzled by this, but I think the situation is improved by understanding gears-level reasoning. So, let’s move on to the discussion of TEOTE.
Gears of Gears
As Valentine noted in his article, it is somewhat frustrating how the overall idea of gears-level understanding seems so clear while remaining only heuristic in definition. It’s a sign of a ripe philosophical puzzle. If you don’t feel you have a good intuitive grasp of what I mean by “gears level understanding”, I suggest reading his post.
Valentine gives three tests which point in the direction of the right concept:
Does the model pay rent? If it does, and if it were falsified, how much (and how precisely) could you infer other things from the falsification?
How incoherent is it to imagine that the model is accurate but that a given variable could be different?
If you knew the model were accurate but you were to forget the value of one variable, could you rederive it?
I already named one near-synonym for “gears”, namely “technical explanation”. Two more are “inside view” and Elon Musk’s notion of reasoning from first principles. The implication is supposed to be that gears-level understanding is in some sense better than other sorts of knowledge, but this is decidedly not supposed to be valued to the exclusion of other sorts of knowledge. Inside-view reasoning is traditionally supposed to be combined with outside-view reasoning (although Elon Musk calls it “reasoning by analogy” and considers it inferior, and much of Eliezer’s recent writing warns of its dangers as well, while allowing for its application to special cases). I suggested the terms gears-level & policy-level in a previous post (which I actually wrote after most of this one).
Although TEOTE gets close to answering Valentine’s question, it doesn’t quite hit the mark. The definition of “technical explanation” provided there is a theory which strongly concentrates the probability mass on specific predictions and rules out others. It’s clear that a model can do this without being “gears”. For example, my model might be that whatever prediction the Great Master makes will come true. The Great Master can make very detailed predictions, but I don’t know how they’re generated. I lack the understanding associated with the predictive power. I might have a strong outside-view reason to trust the Great Master: their track record on predictions is immaculate, their Bayes-loss miniscule, their calibration supreme. Yet, I lack an inside-view account. I can’t derive their predictions from first principles.
Here, I’m siding with David Deutsch’s account in the first chapter of The Fabric of Reality. He argues that understanding and predictive capability are distinct, and that understanding is about having good explanations. I may not accept his whole critique of Bayesianism, but that much of his view seems right to me. Unfortunately, he doesn’t give a technical account of what “explanation” and “understanding” could be.
First Attempt: Deterministic Predictions
TEOTE spends a good chunk of time on the issue of making predictions in advance. According to TEOTE, this is a human solution to a human problem: you make predictions in advance so that you can’t make up what predictions you could have made after the fact. This counters hindsight bias. An ideal Bayesian reasoner, on the other hand, would never be tempted into hindsight bias in the first place, and is free to evaluate hypotheses on old evidence (as already discussed).
So, is gears-level reasoning just pure Bayesian reasoning, in which hypotheses have strictly defined probabilities which don’t depend on anything else? Is outside-view reasoning the thing logical induction adds, by allowing the beliefs of a hypothesis to shift over time and to depend on on the wider market state?
This isn’t quite right. An ideal Bayesian can still learn to trust the Great Master, based on the reliability of the Great Master’s predictions. Unlike a human (and unlike a logical inductor), the Bayesian will at all times have in mind all the possible ways the Great Master’s predictions could have become so accurate. This is because a Bayesian hypothesis contains a full joint distribution on all events, and an ideal Bayesian reasons about all hypotheses at all times. In this sense, the Bayesian always operates from an inside view—it cannot trust the Great Master without a hypothesis which correlates the Great Master with the world.
However, it is possible that this correlation is introduced in a very simple way, by ruling out cases where the Great Master and reality disagree without providing any mechanism explaining how this is the case. This may have low prior probability, but gain prominence due to the hit in Bayes-score other hypotheses are taking for not taking advantage of this correlation. It’s not a bad outcome given the epistemic situation, but it’s not gears-level reasoning, either. So, being fully Bayesian or not isn’t exactly what distinguishes whether advanced predictions are needed. What is it?
I suggest it’s this: whether the hypothesis is well-defined, such that anyone can say what predictions it makes without extra information. In his post on gears, Valentine mentions the importance of “how deterministically interconnected the variables of the model are”. I’m pointing at something close, but importantly distinct: how deterministic the predictions are. You know that a coin is very close to equally likely to land on heads or tails, and from this you can (if you know a little combinatorics) compute things like the probability of getting exactly three heads if you flip the coin five times. Anyone with the same knowledge would compute the same thing. The model includes probabilities inside it, but how those probabilities flow is perfectly deterministic.
This is a notion of objectivity: a wide variety of people can agree on what probability the model assigns, despite otherwise varied background knowledge.
If a model is well-defined in this way, it is very easy (Bayesian or no) to avoid hindsight bias. You cannot argue about how you could have predicted some result. Anyone can sit down and calculate.
The hypothesis that the Great Master is always correct, on the other hand, does not have this property. Nobody but the Great Master can say what that hypothesis predicts. If I know what the Great Master says about a particular thing, I can evaluate the accuracy of the hypothesis; but, this is special knowledge which I need in order to give the probabilities.
The Bayesian hypothesis which simply forces statements of the Great Master to correlate with the world is somewhat more gears-y, in that there’s a probability distribution which can be written down. However, this probability distribution is a complicated mish-mosh of the Bayesian’s other hypotheses. So, predicting what it would say requires extensive knowledge of the private beliefs of the Bayesian agent involved. This is typical of the category of non-gears-y models.
Infortunately, this account doesn’t totally satisfy what Valentine wants.
Suppose that, rather than making announcements on the fly, the Great Master has published a set of fixed Doctrines which his adherents memorize. As in the previous thought experiment, the word of the Great Master is infallible; the application of the Doctrines always leads to correct predictions. However, the contents of the Doctrines appears to be a large mish-mosh of rules with no unifying theme. Despite their apparent correctness, they fail to provide any understanding. It is as if a physicist took all the equations in a physics text, transformed them into tables of numbers, and then transported those tables to the middle ages with explanations of how to use the tables (but none of where they come from). Though the tables work, they are opaque; there is no insight as to how they were determined.
The Doctrines are a deterministic tool for making predictions. Yet, they do not seem to be a gears-level model. Going back to Valentine’s three tests, the Doctrines fail test three: we could erase any one of the Doctrines and we’d be unable to rederive it by how it fit together with the rest. Hence, the Doctrines have almost as much of a “trust the Great Master” quality as listening to the Great Master directly—the disciples would not be able to derive the Doctrines for themselves.
Second Attempt: Proofs, Axioms, & Two Levels of Gears
My next proposal is that having a gears-level model is like knowing the proof. You might believe a mathematical statement because you saw it in a textbook, or because you have a strong mathematical intuition which says it must be true. But, you don’t have the gears until you can prove it.
This subsumes the “deterministic predictions” picture: a model is an axiomatic system. If we know all the axioms, then we can in theory produce all the predictions ourselves. (Thinking of it this way introduces a new possibility, that the model may be well-defined but we may be unable to find the proofs, due to our own limitations.) On the other hand, we don’t have access to the axioms of the theory embodied by the Great Master, and so we have no hope of seeing the proofs; we can only observe that the Great Master is always right.
How does this help with the example of the Doctrines?
The concept of “axioms” is somewhat slippery. There are many equivalent ways of axiomatizing any given theory. We can often flip views between what’s taken as an axiom vs what’s proved as a theorem. However, the most elegant set of axioms tends to be preferred.
So, we can regard the Doctrines as one long set of axioms. If we look at them that way, then adherents of the Great Master have a gears-level understanding of the Doctrines if they can successfully apply them as instructed.
However, the Doctrines are not an elegant set of axioms. So, viewing them in this way is very unnatural. It is more natural to see them as a set of assertions which the Great Master has produced by some axioms unknown to us. In this respect, we “can’t see the proofs”.
In the same way, we can consider flipping any model between the axiom view and the theorem view. Regarding the model as axiomatic, to determine whether it is gears-level we only ask whether its predictions are well-defined. Regarding in in “theorem view”, we ask if we know how the model itself was derived.
Hence, two of Valentine’s desirable properties of a gears-level model can be understood as the same property applied at different levels:
Determinism, which is Val’s property #2, follows from requiring that we can see the derivations within the model.
Reconstructability, Val’s property #3, follows from requiring that we can see the derivation of the model.
We might call the first level of gears “made out of gears”, and the second level “made by gears”—the model itself being constructed via a known mechanism.
If we change our view so that a scientific theory is a “theorem”, what are the “axioms”? Well, there are many criteria which are applied to scientific theories in different domains. These criteria could be thought of as pre-theories or meta-theories. They encode the hard-won wisdom of a field of study, telling us what theories are likely to work or fail in that field. But, a very basic axiom is: we want a theory to be the simplest theory consistent with all observations. The Great Master’s Doctrines cannot possibly survive this test.
To give a less silly example: if we train up a big neural network to solve a machine learning problem, the predictions made by the model are deterministic, predictable from the network weights. However, someone else who knew all the principles by which the network was created would nonetheless train up a very different neural network—unless they use the very same gradient descent algorithm, data, initial weights, and number and size of layers.
Even if they’re the same in all those details, and so reconstruct the same neural network exactly, there’s a significant sense in which they can’t see how the conclusion follows inevitably from the initial conditions. It’s less doctrine-y than being handed a neural network, but it’s more doctrine-y than understanding the structure of the problem and why almost any neural network achieving good performance on the task will have certain structures. Remember what I said about mathematical understanding. There’s always another level of “being able to see why” you can ask for. Being able to reproduce the proof is different from being able to explain why the proof has to be the way it is.
Gears-y ness is a matter of degree, and there are several interconnected things we can point at, and a slippage of levels of analysis which makes everything quite complicated.
In the ontology of math/logic, we can point at whether you can see the proof of a theorem. There are several slippages which make this fuzzier than it may seem. First: do you derive it only form the axioms, or do you use commonly known theorems and equivalences (which you may or may not be able to prove if put on the spot)? There’s a long continuum between what one mathematician might say to another as proof and a formal derivation in logic. Second: how well can you see why the proof has to be? This is the spectrum between following each proof step individually (but seeing them as almost a random walk) vs seeing the proof as an elementary application of a well-known technique. Third: we can start slipping the axioms. There are small changes to the axioms, in which one thing goes from being an axiom to a theorem and another thing makes the opposite transition. There are also large changes, like formalizing number theory via the Peano axioms vs formalizing it in set theory, where the entire description language changes. You need to translate from statements of number theory to statements of set theory. Also, there is a natural ambiguity between taking something as an axiom vs requiring it as a condition in a theorem.
In the ontology of computation, we can point at knowing the output of a machine vs being able to run it by hand to show the output. This is a little less flexible than the concept of mathematical proof, but essentially the same distinction. Changing the axioms is like translating the same algorithm to a different computational formalism, like going between Turing machines and lambda calculus. Also, there is a natural ambiguity between a program vs an input: when you run program XYZ with input ABC on a universal Turing machine, you input XYZABC to the universal turing machine; but, you can also think of this as running program XY on input ZABC, or XYZA on input BC, et cetera.
In the ontology of ontology, we could say “can you see why this has to be, from the structure of the ontology describing things?” “Ontology” is less precise than the previous two concepts, but it’s clearly the same idea. A different ontology doesn’t necessarily support the same conclusions, just like different axioms don’t necessarily give the same theorems. However, the reductionist paradigm holds that the ontologies we use should all be consistent with one another (under some translation between the ontologies). At least, aspire to be eventually consistent. Analogous to axiom/assumption ambiguity and program/input ambiguity, there is ambiguity between an ontology and the cognitive structure which created and justifies the ontology. We can also distinguish more levels; maybe we would say that an ontology doesn’t make predictions directly, but provides a language for stating models, which make predictions. Even longer chains can make sense, but it’s all subjective divisions. However, unlike the situation in logic and computation, we can’t expect to articulate the full support structure for an ontology; it is, after all, a big mess of evolved neural mechanisms which we don’t have direct access to.
Having established that we can talk about the same things in all three settings, I’ll restrict myself to talking about ontologies.
Two-level definition of gears: A conclusion is gears-like with respect to a particular ontology to the extent that you can “see the derivation” in that ontology. A conclusion is gears-like without qualification to the extent that you can also “see the derivation” of the ontology itself. This is contiguous with gears-ness relative to an ontology, because of the natural ambiguity between programs and their inputs, or between axioms and assumptions. For a given example, though, it’s generally more intuitive to deal with the two levels separately.
Seeing the derivation: There are several things to point at by this phrase.
As in TEOTE, we might consider it important that a model make precise predictions. This could be seen as a prerequisite of “seeing the derivation”: first, we must be saying something specific; then, we can ask if we can say why we’re saying that particular thing. This implies that models are more gears-like when they are more deterministic, all other things being equal.
However, I think it is also meaningful and useful to talk about whether the predictions of the model are deterministic; the standard way of assigning probabilities to dice is very gears-like, despite placing wide probabilities. I think these are simply two different important things we can talk about.
Either way, being able to see the derivation is like being able to see the proof or execute the program, with all the slippages this implies. You see the derivation less well to the extent that you rely on known theorems, and more to the extent that you can spell out all the details yourself if need be. You see it less well to the extent that you understand the proof only step-by-step, and more well to the extent that you can derive the proof as a natural application of known principles. You cannot see the derivation if you don’t even have access to the program which generated the output, or are missing some important inputs for that program.
Seeing the derivation is about explicitness and external objectivity. You can trivially “execute the program” generating any of your thoughts, in that you thinking is the program which generated the thoughts. However, the execution of this program could rely on arbitrary details of your cognition. Moreover, these details are usually not available for conscious access, which means you can’t explain the train of thought to others, and even you may not be able to replicate it later. So, a model is more gears-like the more replicable it is. I’m not sure if this should be seen as an additional requirement, or an explanation of where the requirements come from.
Conclusion, Further Directions
Obviously, we only touched the tip of the iceberg here. I started the post with the claim that I was trying to hash out the implications of logical induction for practical rationality, but secretly, the post was about things which logical inductors can only barely begin to explain. (I think these two directions support each other, though!)
We need the framework of logical induction to understand some things here, such as how you still have degrees of understanding when you already have the proof / already have a program which predicts things perfectly (as discussed in the “mathematical understanding” section). However, logical inductors don’t look like they care about “gears”—it’s not very close to the formalism, in the way that TEOTE gave a notion of technical explanation which is close to the formalism of probability theory.
I mentioned earlier that logical induction suffers from the old evidence problem more than Bayesianism. However, it doesn’t suffer in the sense of losing bets it could be winning. Rather, we suffer, when we try to wrap our heads around what’s going on. Somehow, logical induction is learning to do the right thing—the formalism is just not very explicit about how it does this.
The idea (due to Sam Eisenstat, hopefully not butchered by me here) is that logical inductors get around the old evidence problem by learning notions of objectivity.
A hypothesis you come up with later can’t gain any credibility by fitting evidence from the past. However, if you register a prediction ahead of time that a particular hypothesis-generation process will eventually turn up something which fits the old evidence, you can get credit, and use this credit to bet on what the hypothesis claims will happen later. You’re betting on a particular school of thought, rather than a known hypothesis. “You can’t make money by predicting old evidence, but you may be able to find a benefactor who takes it seriously.”
In order to do this, you need to specify a precise prediction-generation process which you are betting in favor of. For example, Solomonoff Induction can’t run as a trader, because it is not computable. However, the probabilities which it generates are well-defined (if you believe that halting bits are well-defined, anyway), so you can make a business of betting that its probabilities will have been good in hindsight. If this business does well, then the whole market of the logical inductor will shift toward trying to make predictions which Solomonoff Induction will later endorse.
Similarly for other ideas which you might be able to specify precisely without being able to run right away. For example, you can’t find all the proofs right away, but you could bet that all the theorems which the logical inductor observes have proofs, and you’d be right every time. Doing so allows the market to start betting it’ll see theorems if it sees that they’re provable, even if it hasn’t yet seen this rule make a successful advance prediction. (Logical inductors start out really ignorant of logic; they don’t know what proofs are or how they’re connected to theorems.)
This doesn’t exactly push toward gears-y models as defined earlier, but it seems close. You push toward anything for which you can provide an explicit justification, where “explicit justification” is anything you can name ahead of time (and check later) which pins down predictions of the sort which tend to correlate with the truth.
This doesn’t mean the logical inductor converges entirely to gears-level reasoning. Gears were never supposed to be everything, right? The optimal strategy combines gears-like and non-gears-like reasoning. However, it does suggest that gears-like reasoning has an advantage over non-gears reasoning: it can gain credibility from old evidence. This will often push gears-y models above competing non-gears considerations.
All of this is still terribly informal, but is the sort of thing which could lead to a formal theory. Hopefully you’ll give me credit later for that advanced prediction.