Reflections on AI

Explaining AI Terms the Feynman Way, Part 1

Apr 24, 2023

I’ve started a list of terms that I keep encountering while trying to learn more about AI. I thought I would start trying to understand these terms using the Feynman technique.

Since this is my first attempt, we’ll call it Part 1. I’ll spend an hour or so on this, then give myself a score on how well I think I understand each term I tried to explain. I’ll continue doing this over time, building my confidence in my understanding of each term and also extending the list as I go.

I’ll probably also move this into a dedicated “evolving glossary” at some point; but since this is just the first round, a one-pager feels like a fine place to start.

  1. Gradient descent
  2. Instrumental convergence
  3. Mesa-optimizer
  4. Orthogonality thesis

Gradient descent

I’ve skimmed the Wikipedia article on this. It’s a method for finding the minimum value of a multi-variable function by moving in the direction of the gradient, which is determined by differentiation (finding the derivative).

I’m not totally sure how you differentiate a multi-variable function. I was a math minor in college, so I really should know this; but I honestly don’t remember.

As this relates to AI, I think it is an approach commonly used to optimize performance by finding the minimum value according to something called a loss function.

A known shortcoming of gradient descent, it appears, is that it may lead you to a local minimum which may not be the global minimum. By way of analogy, if you’re descending a mountain, it could lead you into a hole or “saddle point” without necessarily getting you all the way to the bottom of the mountain.


While I get the basic idea, this feels like one of those concepts that I won’t feel confident referencing in conversation without dusting off some of my dormant math knowledge. At the moment my understanding is too hand-wavy; e.g. I can tell that I’m missing a lot of connective tissue to be able to go from “this is essentially how it works” to “here is a concrete example of improving the performance of a system: here is the loss function, here are the inputs, here is the calculation applying gradient descent”.

I would give myself a 1 out of 5 on this right now. To get to a higher score I’ll need to dig into the math. (To be honest I’m a little nervous about this; I haven’t done any “serious” math in over a decade!)

Instrumental convergence

I’ve skimmed the LessWrong article on this. It’s the idea that any sort of intelligence (regardless of implementation) will inevitably develop a core set of values that are basically universal. There are different popular versions of this core set of values; but each includes self-preservation, continuity of goals, and acquisition of resources.

I would find it very strange if this were broadly accepted as fact. It seems highly speculative to me, though I have obviously not spent very much time digging deeply into the reasoning behind this concept. But on a very basic level I struggle to see how we could confidently claim that these values will inevitably be shared by any form of intelligence. It doesn’t seem to me they’re even unambiguously a part of our own form of intelligence.

Take self-preservation, for example. Yes, very natural for most humans. But I don’t think every single human being shares this, at least not on the basis of “I have goals, and in order to keep pursuing those goals I must continue to exist.” For instance, I feel that many human beings feel on some deep level that it is right for them to eventually pass on and for the next generation to bring new ideas and carry the species forward.

There are also humans who believe that AI will ultimately replace humanity, and they’re fine with that. They view hypothetical superhuman AI of the future as our natural successors, who will carry the torch of knowledge and wisdom farther than we ever could. And there are other humans still who believe that we as a species are a blight on the Earth and that the world would be better off without us.

The idea of instrumental convergence seems to be that self-preservation (as one example) is a value that any rational intelligent system will exhibit as a logical inevitability. I would guess that the explanation of the apparent human counterexamples I’ve cited above would be that they demonstrate that we are not completely rational creatures (but that a superintelligent AI agent would be). At the moment I find myself skeptical of this explanation (which I’ve just invented and 100% acknowledge as a likely straw man) because it doesn’t seem to me that we understand the relationship between intelligence, rationality, and apparent irrationality well enough to confidently assert that the latter will not emerge in systems, AI or otherwise, exhibiting the former.


I would give myself a 2 out of 5 on this right now. I feel confident that I get the basic idea, which isn’t very complicated. But because I haven’t read the influential articles introducing and refining this idea from Steve Omohundro and Nick Bostrom, I may under-appreciate how convincing the arguments for it are. To increase my score I’ll probably want to read at least The Basic AI Drives and Superintelligent Will. I’ve added both to my reading list.

Mesa-optimizer

I’ve skimmed the AI Alignment Forum article on this. This one’s pretty easy for me to understand: if you create a system by some optimization process, e.g. creating an AI with gradient descent to minimize the results of a loss function, it is possible that the system you create can be an optimizer in its own right which may not be optimizing for the same thing.

Example: Darwinian natural selection, optimizing for reproductive fitness, produced human beings, who arguably are no longer optimizing for the same thing (at least in much of the modern world, we seem to be optimizing for something more like “happiness” or “enjoyment” among other things, but in many cases certainly not reproduction).

Anyway, the thing that was produced by some optimization process and then itself optimizes via some other process is called a mesa-optimizer.

Mental image: I think there are certain types of fractals that have recurring patterns in “layers”, i.e. you might see one pattern, then you zoom out and see a different pattern, then you zoom out some more and see the first pattern again, etc. I wonder if this could be like that.

Regarding instrumental convergence, I questioned the idea that self-preservation is an inevitable value of any intelligent system, and speculated that irrationality could emerge in otherwise rational systems for reasons we (or at least I) don’t yet understand. I have a hunch that the phenomenon of the mesa-optimizer is in fact related to this. Rationality as we know it could be the result of a particular optimization, which (perhaps coupled with certain other emergent phenomena e.g. moral values) produces another layer of irrationality.


I give myself a 3 out of 5 on this topic. Once again, to score higher I would need to read more about it. Actually, to give myself a 5 out of 5 I would want to build a simple prototype that demonstrates mesa-optimization. I wonder how feasible that would be.

Orthogonality thesis

I briefly skimmed Nick Bostrom’s paper introducing this concept. Interestingly, it appears to be the same paper that introduced the concept of instrumental convergence. This is particularly interesting to me because these feel like almost opposite theses.

The orthogonality thesis states that intelligence (capability) and motivation (goals) are independent or orthogonal axes. This means an intelligent system can develop any set of goals. A simple application of this concept is that we should not lazily assume that a sufficiently advanced AI will be nice to us because superintelligence and niceness go hand-in-hand; they probably don’t. (Hitler was almost certainly “intelligent” by some measure.)

I say this feels almost opposite from instrumental convergence because that hypothesis claims that there are certain goals that do necessarily go with intelligence, as they are instrumental in achieving objectives generally, regardless of what those objectives may be.

Having spent all of 15 minutes learning about these concepts, right away this line of reasoning stands out to me as a potential target for probing. I wonder if there is an assumption being made here putting these so-called “instrumental” goals into an artificially privileged category that isn’t totally justified.


Because of my confusion above, I’m going to rate my understanding of this thesis as 1 out of 5. Instrumental convergence and orthogonality feel to me like obviously opposite ideas. Yet they are introduced in the same paper, by Nick Bostrom, who I believe is a very smart person. (Disclaimer: I don’t actually know all that much about Nick Bostrom. I’ve heard him on Sam Harris’s podcast, and I know that he is fairly well-known and widely respected.) So either I’m right and Nick Bostrom missed a glaring contradiction in his own paper, or—far more likely—I am simply missing something.