Vanessa Kosoy

AI alignment researcher supported by MIRI and LTFF. Working on the learning-theoretic agenda. Based in Israel. See also LinkedIn.

E-mail: vanessa DOT kosoy AT {the thing reverse stupidity is not} DOT org

30

Sorry, that footnote is just flat wrong, the order actually *doesn't* matter here. Good catch!

There is a related thing which might work, namely taking the downwards closure of the affine subspace w.r.t. some cone which is somewhat *larger* than the cone of measures. For example, if your underlying space has a metric, you might consider the cone of signed measures which have non-negative integral with all positive functions whose logarithm is 1-Lipschitz.

20

Sort of obvious but good to keep in mind: Metacognitive regret bounds are *not* easily reducible to "plain" IBRL regret bounds when we consider the core *and* the envelope as the "inside" of the agent.

Assume that the action and observation sets factor as and , where is the interface with the external environment and is the interface with the envelope.

Let be a metalaw. Then, there are two natural ways to reduce it to an ordinary law:

- Marginalizing over . That is, let and be the projections. Then, we have the law .
- Assuming "logical omniscience". That is, let be the ground truth. Then, we have the law . Here, we use the conditional defined by . It's easy to see this indeed defines a law.

However, requiring low regret w.r.t. neither of these is equivalent to low regret w.r.t :

- Learning is typically no less feasible than learning , however it is a much weaker condition. This is because the metacognitive agents can use policies that
*query the envelope*to get higher guaranteed expected utility. - Learning is a much stronger condition than learning , however it is typically infeasible. Requiring it leads to AIXI-like agents.

Therefore, metacognitive regret bounds hit a "sweep spot" of stength vs. feasibility which produces a genuinely more powerful agents than IBRL^{[1]}.

^{^}More precisely, more powerful than IBRL with the usual sort of hypothesis classes (e.g. nicely structured crisp infra-RDP). In principle, we can reduce metacognitive regret bounds to IBRL regret bounds using non-crsip laws, since there's a very general theorem for representing desiderata as laws. But, these laws would have a very peculiar form that seems impossible to guess without starting with metacognitive agents.

50

Is it possible to replace the maximin decision rule in infra-Bayesianism with a different decision rule? One surprisingly strong desideratum for such decision rules is the learnability of some natural hypothesis classes.

In the following, all infradistributions are crisp.

Fix finite action set and finite observation set . For any and , let

be defined by

In other words, this kernel samples a time step out of the geometric distribution with parameter , and then produces the sequence of length that appears in the destiny starting at .

For any continuous^{[1]} function , we get a decision rule. Namely, this rule says that, given infra-Bayesian law and discount parameter , the optimal policy is

The usual maximin is recovered when we have some reward function and corresponding to it is

Given a set of laws, it is said to be learnable w.r.t. when there is a family of policies such that for any

For we know that e.g. the set of all communicating^{[2]} finite infra-RDPs is learnable. More generally, for any we have the learnable decision rule

This is the "mesomism" I taked about before.

Also, any monotonically increasing seems to be learnable, i.e. any s.t. for we have . For such decision rules, you can essentially assume that "nature" (i.e. whatever resolves the ambiguity of the infradistributions) is collaborative with the agent. These rules are not very interesting.

On the other hand, decision rules of the form are not learnable in general, and so are decision rules of the form for monotonically increasing.

**Open Problem:** Are there any learnable decision rules that are not mesomism or monotonically increasing?

A positive answer to the above would provide interesting generalizations of infra-Bayesianism. A negative answer to the above would provide an interesting novel justification of the maximin. Indeed, learnability is not a criterion that was ever used in axiomatic constructions of decision theory^{[3]}, AFAIK.

^{^}We can try considering discontinuous functions as well, but it seems natural to start with continuous. If we want the optimal policy to exist, we usually need to be at least upper semicontinuous.

^{^}There are weaker conditions than "communicating" that are sufficient, e.g. "resettable" (meaning that the agent can always force returning to the initial state), and some even weaker conditions that I will not spell out here.

^{^}I mean theorems like VNM, Savage etc.

20

Intuitively, it feels that there is something special about mathematical knowledge from a learning-theoretic perspective. Mathematics seems *infinitely rich*: no matter how much we learn, there is always more interesting structure to be discovered. Impossibility results like the halting problem and Godel incompleteness lend some credence to this intuition, but are insufficient to fully formalize it.

Here is my proposal for how to formulate a theorem that would make this idea rigorous.

Fix some natural hypothesis class for mathematical knowledge, such as some variety of tree automata. Each such hypothesis represents an infradistribution over : the "space of counterpossible computational universes". We can say that is a "true hypothesis" when there is some in the credal set (a distribution over ) s.t. the ground truth "looks" as if it's sampled from . The latter should be formalizable via something like a computationally bounded version of Marin-Lof randomness.

We can now try to say that is "rich" if for any true hypothesis , there is a *refinemen*t which is also a true hypothesis and "knows" at least one bit of information that doesn't, in some sense. This is clearly true, since there can be no automaton or even any computable hypothesis which fully describes . But, it's also completely boring: the required can be constructed by "hardcoding" an additional fact into . This doesn't look like "discovering interesting structure", but rather just like brute-force memorization.

What if instead we require that knows *infinitely many* bits of information that doesn't? This is already more interesting. Imagine that instead of metacognition / mathematics, we would be talking about ordinary sequence prediction. In this case it is indeed an interesting non-trivial condition that the sequence contains infinitely many regularities, s.t. each of them can be expressed by a finite automaton but their conjunction cannot. For example, maybe the -th bit in the sequence depends only the largest s.t. divides , but the dependence on is already uncomputable (or at least inexpressible by a finite automaton).

However, for our original application, this is entirely insufficient. This is because in the formal language we use to define (e.g. combinator calculus) has some "easy" equivalence relations. For example, consider the family of programs of the form "if 2+2=4 then output 0, otherwise...". All of those programs would output 0, which is obvious once you know that 2+2=4. Therefore, once your automaton is able to check some such easy equivalence relations, hardcoding a single new fact (in the example, 2+2=4) generates infinitely many "new" bits of information. Once again, we are left with brute-force memorization.

Here's the improved condition: For any true hypothesis , there is a true refinement s.t. *conditioning * *on any finite set of observations cannot produce a refinement of* .

There is a technicality here, because we're talking about infradistributions, so what is "conditioning" exactly? For credal sets, I think it is sufficient to allow two types of "conditioning":

- For any given observation and , we can form .
- For any given observation s.t. , we can form .

This rules-out the counterexample from before: the easy equivalence relation can be represented inside , and then the entire sequence of "novel" bits can be generated by a conditioning.

Alright, so does actually satisfy this condition? I think it's very probable, but I haven't proved it yet.

I wrote a review here. There, I identify the main generators of Christiano's disagreement with Yudkowsky^{[1]} and add some critical commentary. I also frame it in terms of a broader debate in the AI alignment community.

^{^}I divide those into "takeoff speeds", "attitude towards prosaic alignment" and "the metadebate" (the last one is about what kind of debate norms should we have about this or what kind of arguments should we listen to.)

30

Yes, this is an important point, of which I am well aware. This is why I expect unbounded-ADAM to only be a toy model. A more realistic ADAM would use a complexity measure that takes *computational complexity* into account instead of . For example, you can look at the measure I defined here. More realistically, this measure should be based on the frugal universal prior.

1211

Thank you for the clarification.

How do you expect augmented humanity will solve the problem? Will it be something other than "guessing it with some safe weak lesser tries / clever theory"?

52

in each of the 50 different subject areas that we tested it on, it's as good as the best expert humans in those areas

That sounds like an incredibly strong claim, but I suspect that the phrasing is very misleading. What kind of tests is Hassabis talking about here? Maybe those are tests that rely on remembering known facts much more than on making novel inferences? Surely Gemini is not (say) as good as the best mathematicians at solving open problems in mathematics?

Here is a way to construct many learnable undogmatic ontologies, including such with finite state spaces.

A *deterministic partial environment* (DPE) over action set and observation set is a pair where and s.t.

- If is a prefix of some , then .
- If , and is a prefix of , then .

DPEs are equipped with a natural partial order. Namely, when and .

Let be a strong upwards antichain in the DPE poset which doesn't contain the bottom DPE (i.e. the DPE with ). Then, it naturally induces an infra-POMDP. Specifically:

- The state space is .
- The initial infradistribution is .
- The observation mapping is , where is the empty history.
- The transition infrakernel is , where

If is non-empty for all and , this is a learnable undogmatic ontology.

Any yields an example . Namely, iff and for any it holds that:

- If then for any , .

I think that for ~~any continuous~~ some non-trivial hidden reward functions over such an ontology, the class of communicating RUMDPs is learnable. If the hidden reward function doesn't depend on the action argument, it's equivalent to some instrumental reward function.

Here is a modification of the IBP framework which removes the monotonicity principle, and seems to be more natural in other ways as well.

First, let our notion of "hypothesis" be Θ∈□c(Γ×2Γ). The previous framework can be interpreted in terms of hypotheses of this form satisfying the condition

prΓ×2ΓBr(Θ)=Θ(See Proposition 2.8 in the original article.) In the new framework, we replace it by the weaker condition

Br(Θ)⊇(idΓ×diag2Γ)∗ΘThis can be roughly interpreted as requiring that (i) whenever the output of a program P determines whether some other program Q will run, program P has to run as well (ii) whenever programs P and Q are logically equivalent, program P runs iff program Q runs.

The new condition seems to be well-justified, and is also invariant under (i) mixing hypotheses (ii) taking joins/meets of hypotheses. The latter was not the case for the old condition. Moreover, it doesn't imply that Θ is downward closed, and hence there is no longer a monotonicity principle

^{[1]}.The next question is, how do we construct hypotheses satisfying this condition? In the old framework, we could construct hypotheses of the form Ξ∈□c(Γ×Φ) and then apply the bridge transform. In particular, this allows a relatively straightforward translation of physics theories into IBP language (for example our treatment of quantum theory). Luckily, there is an analogous construction in the new framework as well.

First notice that our new condition on Θ can be reformulated as requiring that

For any Φ, we also define τΦs:Δc(elΓ×Φ)→Δc(elΓ×Φ) by

τΦsθ:=χelΓ×Φ(s×id2Γ×Φ)∗Now, for any Ξ∈□c(Γ×Φ), we define the "conservative bridge transform

^{[2]}" CBr(Ξ)∈□c(Γ×2Γ×Φ) as the closure of all τΦsθ where θ is amaximalelement of Br(Ξ). It is then possible to see that Θ∈□c(Γ×2Γ) is a valid hypothesis if and only if it is of the form prΓ×2ΓCBr(Ξ) for some Φ and Ξ∈□c(Γ×Φ).^{^}I still think the monotonicity principle is saying something about the

learning theoryof IBP which is still true in the new framework. Namely, it is possible to learn that a program is running but not possible to (confidently) learn that a program isnotrunning, and this limits the sort of frequentist guarantees we can expect.^{^}Intuitively, it can be interpreted as a version of the bridge transform where we postulate that a program doesn't run unless Ξ contains a reason while it must run.