AI & wisdom 1: wisdom, amortised optimisation, and AI

post by L Rudolf L (LRudL) · 2024-10-28T21:02:51.215Z · LW · GW · 0 comments

This is a link post for https://rudolf.website/wisdom1/

Contents

  Wisdom
  A lot of wisdom is amortised optimisation
    Details & maths (skippable)
    Connecting amortised optimisation to wisdom
  Optimised artefacts
    Trees of historical optimisation
    When can you use an optimised artefact in your amortised optimisation process?
  Wise artefacts & wise people
  What's missing from this picture of wisdom
  How does epistemics relate to amortised optimisation?
None
No comments

Written for the AI Impacts essay competition on the automation of wisdom and philosophy

At least in name, AI is about intelligence. However, intelligence is not the only mental quality that matters. Many deep mental skills, in particular regarding good, deep, or long-term judgement, are bundled under the term "wisdom". If we automate and industrialise intelligence, we may not automate and industrialise other important mental qualities, similar to how industrialising food lead to a surplus of calories, but also harms from missing nutrients.

Wisdom

How does wisdom differ from intelligence? Here are some pairs of associations between intelligence and wisdom:

(Some others that fit my latter points less well: complex v simple, solving v preventing, facts v meaning, question-answering v question-asking, specific v holistic)

Here are some examples that are more on the wisdom side than the intelligence side:

A lot of wisdom is amortised optimisation

A large amount of things that seem optimised or purpose-built in the world did not come about through internal optimisation within thinking things (e.g. a person actively thinking through alternatives to find the best one). Many came about through non-mental selection acting on a population of things (e.g. evolution, memetics). Even when thinking things directly output optimised end products (e.g. a person deciding on a research direction, or Claude 3 solving a maths problem), much of the thinking seems less like search or optimisation, and more about pattern-matching and memory.

A big part of wisdom is those mental things that are more shaped by the latter types of thing, and less by direct internal optimisation. The rest of this post is about clarifying, illustrating, and defending this point of view.

The post Deconfusing direct vs amortised optimisation [LW · GW] (by Beren) introduces the terms direct optimisation and amortised optimisation. Direct optimisation is solving problems by doing explicit search through alternatives on-the-spot when confronted with a new problem (e.g. Monte Carlo Tree Search (MCTS) algorithms that search the tree of possible chess moves). In contrast, amortised optimisation is about taking many past examples you've seen, and distilling them into a function approximator that maps from inputs to outputs (e.g. an end-to-end transformer trained on a large dataset to output winning chess moves).

Details & maths (skippable)

Below, I'll expand on these. The maths (which is different [LW(p) · GW(p)] from Beren's) can be skipped, but may help understanding if you're comfortable with it.

Connecting amortised optimisation to wisdom

Let's return to our list of intelligence v wisdom properties from above, and see how well they map onto direct v amortised optimisation:

The last one is perhaps the least-obvious, but also most important one. I will justify this in a later part.

To return to our examples of wisdom from before:

Optimised artefacts

If you look at a Haskell function that sorts lists, or the xylem inside a tree that transport water up, it feels clear that they're the results of some optimisation process. They're just too unlikely otherwise; a random walk through some relevant reference space, like the set of all strings that are syntactically-valid Lisp programs of a certain length, or the set of ways that you can arrange the molecules in a tree, would be extremely unlikely to hit something as "useful". Call such things optimised artefacts.

(These are both examples of fairly physical optimised artefacts, but things like decisions, plans, and ideas can be optimised artefacts without any physical realisation.)

What mix of amortised or direct optimisation produced an optimised artefact is a fact about its history, not its state at any particular point. Before the theory of evolution, xylem would've been thought of as a product of direct optimisation: presumably, God did some mental work to figure out which structure would make trees work to fulfil His purposes. Now that we know trees came about without any on-the-fly mental search, we'd call them products of amortised optimisation, even without any change in our understanding of what trees do. Likewise, we'd think the Lisp search algorithm more a product of amortised optimisation if it was produced by an evolutionary search algorithm, or by a transformer that had memorised it. We'd think it more a product of direct optimisation if it was dreamt up on the spot by someone who had never seen a sorting algorithm before, or by an AI with no Lisp or sorting algorithms in its training data.

Obviously, all of this is a matter of degree, and usually there is some amortised optimisation lurking somewhere. For example, the programmer who writes a sorting algorithm without ever having seen might've still learned Lisp from SICP, the concept of an algorithm from Ada Lovelace, and - going further back - their general intelligence depends on millions of years of hominid evolution. A full account of how the optimised artefact of the Lisp program came to be requires understanding all of this. But depending on what we're talking about, there's often some natural domain, such as "writing algorithms", and we can talk about degrees of amortised v direct optimisation, from "the impressive part of the feat was almost entirely making good use of lots of existing results/ideas/instincts" to "the impressive part of the feat was almost entirely a lot of explicit novel thinking". Below, I try to visualise this.

Trees of historical optimisation

Every optimised artefact has an explanation for being so optimised. For any such artefact, we can imagine a tree, with the artefact (diamond shape) at the root. Here are three prototypical "shapes" of optimisation:

In the real world, these tend to be mixed. For example, let's take the Lisp sorting algorithm from before:

Maybe there's some long-term selection forces that determine which professors at which institutions are most prestigious (A), the results of which feed into the choices of a professor who spends a bunch of time thinking about the field (B) to decide what to pass on. Independently, a bunch of people spend varying amounts of effort creating programming languages (C), which different programmers try and spend some small amount of thought-time evaluating (D), whose judgements are then posted onto various online forums (E), which battle over prestige and users which determines which one our programmer trusts (F) regarding programming language choice. Meanwhile, their random friend gives them some advice after thinking about the problem for a while (G). Finally, the programmer thinks and writes the program (H).

(This tree doesn't necessarily have to be a tree; it could be a DAG too.)

When can you use an optimised artefact in your amortised optimisation process?

An optimised artefact can be useful for amortised optimisation under several conditions:

Wise artefacts & wise people

We can attempt some loose definitions:

An optimised artefact reflects amortised optimisation to the extent that the most important parts of its construction were achieved by reusing existing optimised artefacts.

(I don't use the word "wisdom" above because it seems silly to call a blade of grass wise.)

Similarly, for direct optimisation:

An optimised artefact reflects direct optimisation to the extent that the most important parts of its construction were achieved by a thinking entity doing a mental search over possibilities.

Finally, let's make a claim about wisdom:

A "wise" person is someone who is good at using existing optimised artefacts when creating new optimised artefacts.

Optimised artefacts created through direct optimisation aren't automatically "wise". But if a person is able to look at the optimised artefacts around them, and cleverly and correctly use them to create optimised artefacts of their own, this sounds a lot like wisdom, for the reasons we've discussed above. This covers everything from wise gurus correctly applying the lessons long experience has taught them, to people using culturally-evolved tools and traditions to further their goals.

(Note that in the same way that "intelligent" does not imply morally good, "wise" does not imply morally good either. We should take care to distinguish between terms about mental capabilities, and terms about alignment of goals.)

What's missing from this picture of wisdom

By focusing on amortised optimisation, we're missing out on some things often associated with wisdom:

  1. The first is wisdom as the ability to live a fulfilling life. This is beyond the scope of this post; please consult your nearest wise mountain guru for advice.

  2. The second is wisdom in the sense of choosing the right questions to ask, or having the right frame or ontology for a problem. We'll see how amortised optimisation relates to some aspects of this in the third post in this series.

  3. The third is ethical judgements. This can be decomposed into the part where you're trying to make the right ethical judgement in the first place, which is about value alignment rather than mental capabilities, and therefore not part of the capability-focused view of wisdom I'm focusing on here. The second is making the right ethical judgements given a desire to do so. I will briefly discuss this in the next part.

  4. The final is wisdom as rationality, epistemics, or calibration. This is what the final section below is about.

How does epistemics relate to amortised optimisation?

In some sense, successful amortised optimisation implies the result is calibrated. Also, a lot of the calibration of both superforecasters and recent LLM forecasting system seems to be down to a good ability to judge base rates, which is an amortised optimisation problem. However, just because amortised optimisation is involved does not mean a decision is wise.

However, it does seem like something can meet the definition of being wise used above (being good at using existing optimised artefacts) without good epistemics - at least on the surface. In his review of The Secrets of Our Success, Scott Alexander writes:

A reasonable person would have asked why everyone was wasting so much time preparing manioc [also known as cassava, a South American potato-like plant that, unless very specifically prepared, is toxic enough to have long-term health effects but not toxic enough to do short-term harm]. When told “Because that’s how we’ve always done it”, they would have been unsatisfied with that answer. They would have done some experiments, and found hat a simpler process of boiling it worked just as well. They would have saved lots of time, maybe converted all their friends to the new and easier method. Twenty years later, they would have gotten sick and died, in a way so causally distant from their decision to change manioc processing methods that nobody would ever have been able to link the two together.

"Because that's how we've always done it" is not particularly compelling epistemics. Note that that isn't just Scott Alexander strawmanning the justifications: in another of Heinrich's examples in The Secrets of Our Success, it's specifically noted that when the people involved were asked for an explanation for a tradition (unrelated to manioc processing but with the story having the same moral), they invented nonsensical post-hoc rationalisations.

"So", you might say, "clearly this isn't wise; though I can admit the manioc processing practice is 'wise' in the whole amortised optimisation sense you've spent 4000 words describing, obviously 'manioc is bad' is not a wise belief since there's no good epistemic grounds for holding it."

I would push back against this. In particular: imagine that everyone around you thinks that following traditions like manioc processing is important. That this alone is not a good rationale is (hopefully) thoroughly drilled into everyone today. But also note: you have lots of examples of everyone being wrong and one person being right - at least, if your science teachers were any good - but this actually is a rare thing to happen, and you have those examples because you live in a large world with a well-documented history. Now also imagine that (and unlike the modern world), a lot of the most successful people around you are the ones that follow traditions the hardest. Imagine there are a bunch of cases you can think of where following traditions turned out to be really smart. Imagine you're not aware of cases where someone fought against something like the perceived inefficiency of manioc processing and won. Induction and proof-by-example are some of the most robust epistemic tools. Both of those would point towards manioc processing being good. "But if you don't understand that you're relying on induction and proof-by-example, or what justifies those - " ... look, is that really the bar? Do you have to understand the philosophy behind the epistemic principles you use in order to count as wise? That's a very high bar.

Of course, the types of examples The Secrets of Our Success is full of are somewhat weird edge cases, where wisdom and our normal standards for epistemic clarity are weirdly at odds. Normally, to make good use of amortised optimisation you need to understand the histories and implicit optimisation that you're applying well, and (especially in fast-changing worlds) make good judgements about how relevant they are. That is very loaded on good epistemics.

But part of the tension remains. In popular imagination, the wise sage is a different and much more ineffable sort of intellectual than the physics grad student who cheerfully bashes out formal proofs on a whiteboard. The above discussion on amortised versus direct optimisation justifies why. Artefacts of direct optimisation come with receipts. These might not be any better than "we searched the set of all possible X, and the one that did best by criterion C was X14". That may be unsatisfying - "but what's the intuition?" asks the physics grad student - but it's not a fundamentally confusing explanation.

Meanwhile, artefacts of amortised optimisation may come with clear receipts: "we did a comprehensive study of all previous cases when Y happened (see Appendices B3 to B7, with further methodology notes in C5), and found that ...". But sometimes, the receipts that amortised optimisation returns are that whenever you deviated from it, your crops failed and your codebases became unmaintainable and your trading ships invariably got eaten by krakens, and the closest to an explanation you have is a crazy old guru cackling about how naive the youths are today. Sometimes you try to build explainable AIs with all the nice algorithms they taught you in school and five more that some maths wiz dreamt up brilliant reasons for, and none of it works, and then you take your hundred-billion-parameter neural network and bonk it on the head when it predicts something wrong and eventually it beats PhDs at tricky reasoning questions in their field.

But in a way, "this is the thing that works, and we know this because things were tried until it worked" is one of the simplest types of reason you can have for something. And whatever your fancy justifications and clever thinking for something, "will this actually endure in the real world?" is the ultimate test, and "has this sort of thing previously endured in the real world?", if available, is often one of the best facts to use in answering that question.

The fact that we can have correct justifications for things without relying on the histories, distillation, and trial-and-error that feeds amortised optimisation is because of a combination of humans being smart, and humans having figured out things like science and maths and epistemics that often give shortcuts. That fact that we must rely on such things is because we live in a fast-growing and fast-changing world, where amortised optimisation (or at least, non-AI amortised optimisation) can't give all the answers fast enough. The dynamics of how the relative importance of direct and amortised optimisation varies are the subjects of the next post.

0 comments

Comments sorted by top scores.