Limits of machine learning

Other Friendly
21

I think that it could be “evolution” but I don’t have a clear enough understanding of why it is or isn’t yet

I think that would need to be sorted out before discussing a more complicated case like AlphaZero. Do you agree?

I agree.

And have you read FoR?

Not as much as BoI. I was trying to reference this idea:

“two conditions must hold for an entity to be a cause of its own replication: first, that the entity is in fact replicated; and second, that most variants of it, in the same situation, would not be replicated.”

For the first condition I could say that specific chess moves for specific board states are stored in the neural network (in the form of weights and biases) and will be replicated in future playouts or “real” games.

And for the second point I gave this example:

“Any other randomly selected move would not be copied because they would all lead to an eventual game loss.”

I wanted to point out that not any random, generic chess move is being selected. It’s a very specific variant of all the possible chess moves. Random moves would simply be ignored. And most of the possible moves are random moves. So most of the possible variants would be ignored.

“For X to be a cause of Y, two conditions must hold: first, that X and Y both happen;
and second, that Y would not have happened if X had been otherwise.”

X - a move for a given board state wins the game
Y - the move is stored in the neural network of AZ

If X were otherwise (such as a randomly selected move that didn’t bring AZ any advantage) then Y would not have happened.

“In general, we may say that an event X causes an event Y in our universe if both X and
Y occur in our universe, but in most variants of our universe in which X does not happen, Y does not happen either.”

An issue I ran into here is that it seems the the chess move X would be the same in all the variants of our universe. Like there aren’t some variants where knight to e5 wins the game for board state S, and some variants where knight to e5 doesn’t win the game for the same board state S.

22

Suppose I bake a variety of cupcakes, pick my favorites, and bake more similar ones. In that scenario, do you think cupcakes are replicators?

23

@Elliot: Do you also think calculators create knowledge? Sometimes the information regarding the answer to some math problem didn’t exist before the calculator did some computations and output it. I think you’re either 1) attributing guiding and directing intelligence to tools (similar to giving a shovel significant credit for digging a whole, or maybe only an automated shovel with a motor) or 2) separating knowledge creation from intelligence or anything equivalent or similar to intelligence (so you don’t think the software does guiding or directing intelligence, but you want to give it credit anyway). But I don’t know which, and maybe it’s something else.

Yes I think calculators create knowledge - information suited to solving one or more problems(s) in one or more context(s).

I do separate knowledge creation from intelligence.

I don’t give software (or calculators) credit for stuff like having a problem in the first place, figuring out what is a good problem to solve, or what is a good way to go about solving it. I don’t think those are essential characteristics of knowledge creation.

24

What about an abacus? Pencil and paper?

25

Ideas can be replicators and a recipe for a cupcake is an idea. We start with a variety of different recipes and then we select the one that we like the best. We bake a bunch of cupcakes and share them with people who ask for the recipe, etc.

So I think the recipe for the cupcake could be a replicator. But the cupcakes themselves aren’t really replicas of each other. They all have different weights, amounts of frosting, physical structures, etc. The thing they really have in common is that they are made with the same recipe.

26

I think reaching agreement is going to be very hard. I don’t know if it’s worth attempting. Partly, I don’t know how much you care or are likely to stick around, or how willing you are to pursue tangents and do work, e.g. read books if they’re relevant. You recently left the forum for weeks. If you want to, you could speak to these things and try to sell me on continuing.

27

A good joke is a replicator because it’s funny and it causes me to want to share it with others, who then keep sharing it, etc. Is that correct?

A good cupcake recipe is a replicator because it makes delicious cupcakes and it causes me to want to bake cupcakes and share them with people, who then get the recipe and keep sharing it, etc.

I am definitely struggling with the replicator concept. I was considering reading The Selfish Gene. Or maybe I’m just misunderstanding the relevant sections in FoR.

I was studying for an AWS exam I had to take for work and I didn’t have a lot of free time to read. And I didn’t think it was worth posting until I had read a little bit more about it. I’m kind of backed up with questions on other topics which I have to post too. And I know I’ll have more questions in the future. There isn’t any other place I know of where I could ask these questions. There’s not really anywhere for me to “leave” to lol

28

If you don’t mind my asking, despite being in the Friendly category, have you been following my CF articles much? I’d be more interested in discussing those, and also I’m wary of trying to discuss other things with people who aren’t very interested in my CF writing. Plus I think they could be helpful in a general, indirect, non-specific way to many other topics including this one.

29

I’m not sure what you mean by this. Is it because the friendly category isn’t related specifically to the CF category?

have you been following my CF articles much?

I’ve only seen a notification for one article. It was the emotions unbounded and I actually shared it with my wife. I had previously read the referenced “Emotions” article but not the “Bounded and Unbounded Emotions” one. I didn’t really know there was a new “series” of CF articles.

30

I think The Selfish Gene is a good book.

Re FoR, one of the key issues is a replicator has to significantly and casually contribute to its own replication, and do it across multiple environments that aren’t tuned specifically to replicating that object. E.g.:

Anything that is copied must have made at least some causal contribution to that copying. Junk DNA sequences, for instance, are made of DNA, which allows the cellular computer to copy them. It cannot copy molecules other than DNA. It is not usually illuminating to consider something as a replicator if its causal contribution to its own replication is small, though strictly speaking being a replicator is a matter of degree.

I agree that jokes and recipes qualify. My cupcake example was intended as a case where the cupcakes themselves play too small of a causal role. The reason they’re getting replicated is mostly about the person and his choices. Similarly, I could design a factory which replicates certain cupcakes in an automated fashion (and does so without a recipe, but rather by analyzing the contents of the cupcakes with scanners), but that doesn’t mean cupcakes should get significant credit as replicators.

Software to brute force figure out the best moves in games, like tic-tac-toe, is kinda like the factory that replicates cupcakes. The cupcakes (parallel to game moves) only replicate due to the very specific environment tuned to them. It’s the environment (like factory or software) that deserves most of the credit, not the objects being replicated.

(I might not follow up but I figured I’d try this.)

31

You can find all my CF articles at https://criticalfallibilism.com (you can get email notifications by signing up there, and there’s also an RSS feed).

There are also topics for discussing all of them in the Elliot Temple category of this forum, which are posted automatically.

The Friendly category limits tangents. For more specifics, see About the Friendly category

32

@Elliot

What about an abacus? Pencil and paper?

I’m not sure about an abacus. My guess would be yes. I think it’s possible to manipulate an abacus according to mechanical rules and generate information that didn’t exist (so like a calculator but with more manual intervention). I’ve never actually used an abacus much though, so I can’t give an example.

With pencil and paper, yes. For example, consider column addition of two or more large numbers (ex: 6539 + 2633). You write down the first number over the second and then draw a line under the second. Add the rightmost digits of both numbers (9+3), write a 1 over the next left column and a 2 below the line in the rightmost column, then add the next column (1+3+3), write a 7 below the line, and so on working right to left. As you’re doing this, you do not know the end result. Only when you’ve done all 4 columns individually does the result exist. Then you read the number written below the line left to right and you know the result. The knowledge (the result) existed on the paper before it existed in your head.

33

There are multiple issues here. One is, I don’t think doing the grunt work, using a known method like long addition, is knowledge creation. You aren’t creatively discovering something new. You’re just mechanically going through the implications of what you already know.

Another is, you seem to assign credit where I wouldn’t, e.g. to the abacus rather than just to the user who did all the thinking. It reminds me of Reisman discussing whether Columbus, the leader and decision maker, should be given the credit for discovering America: George Reisman's Blog on Economics, Politics, Society, and Culture: Why Columbus Is In Fact the Discoverer of America Reisman argues for viewing his crew as helpers or assistants. That’s something some people disagree with. But not many people try to give credit to non-automated, inhuman tools.

34

Knowledge is created by evolution and evolution does not necessarily require creative discovery.

I don’t think that a calculator is creating knowledge when it does computations. There is no variation or selection process involved when a calculator does computations. Without variation and selection there can’t be evolution and therefore there can’t be knowledge creation. I think calculators work using logic gates in a specific configuration. I think the designer of the calculator embedded all of the necessary knowledge into the specific configuration of logic gates.

An abacus or a piece of paper is not like a calculator. Calculators do computations while pencil and paper do not (not sure about abacus but I don’t think so). Knowledge is not embedded into blank paper or unorganized beads on an abacus. They are more like tools or algorithms we can use to help us do the correct computations (in our heads) in the correct order.

35

Perhaps we disagree about what knowledge creation is? I assumed:

I don’t see anything in there that requires thinking or discovery. Knowledge creation and some other stuff, not all of which we understand, is a necessary component thinking and discovery but not the whole story.

Do you disagree with how I’m thinking about knowledge and knowledge creation?

36
  1. I think you’re trying to give credit for creating knowledge to every atom involved, rather than to the ones responsible – like crediting a shovel along with a human digger.
  2. Applying a formula you already know is a different thing than inventing a formula, and you seem to want to treat them the same. Implications and consequences of existing knowledge are different than new knowledge. It’s kind of like if you said closing a door was creating knowledge because it creates a physical state, which includes information that keeps the room quieter, and that physical state didn’t exist before when the door was open. But the door designer had already figured that out in advance.
  3. Knowledge is information that’s adapted to a purpose. A calculator or abacus doesn’t do any adapting.
37

vs.

One difference between our concepts is in “adapted” vs. “suited”.

Adapted implies the information came to be in its current form as the result of a specific process like evolution.

Suited doesn’t imply anything about how the information came to be in its current form, but does imply a positive judgment by an external entity as to the information’s suitability.

I think a concept that doesn’t rely on an external entity would generally be easier to analyze than one that does rely on an external entity. However, I think the second part of both of our concepts (which I view as similar) rely on an external entity to define the “purpose” or “problem(s)”. I don’t think either concept can stand alone. Do you agree?

38

I think discussing credit is hard without additional description.

A human without a shovel could still dig a small hole. A shovel without a human will never dig any size hole. The human is essential; the shovel is not. In that sense, the human gets 100% of the credit for the hole in a human+shovel combination.

A specific human without a shovel could dig a hole of size X in 1 day, whereas the same human with a shovel could dig a hole of size 10X in 1 day. In that sense, the shovel gets 90% of the credit for the hole’s size in a human+shovel combination.

There are other senses in which credit could be allocated.

I’m not trying to give credit for stuff like intelligence, invention, goal selection, etc. to devices like pencils and shovels and electronic circuits that aren’t capable of those things.

39

Do you think the set of all real numbers are knowledge?

For example:

10 000.01 is suited to solving the problem 10 000 + 0.01.

10 000.02 is suited to solving the problem 10 000 + 0.02.

etc.

What about the string ‘axeasapsodifhiopaowihefpaoisehfl’? That string is well suited to solving the problem “What do we get from concatenating the strings ‘axeasapsodifhiop’ and ‘paowihefpaoisehfl’?”

Wouldn’t any data or information be suited to solving some hypothetical problem? So does that imply that all data and information is knowledge?

I think that if a piece of information is “adapted” it implies that it’s adapted to some purpose. If there was no purpose it wouldn’t be adapted to anything.

40

By problem, I mean a human problem - something a human cares about. If a human cares about the answer to 10 000 + 0.01 for some reason, then 10 000.001 is knowledge. Otherwise it is merely information.

I think that’s similar to purpose, in that both rely on an external entity to specify some problem / purpose for which the information is either suited or adapted.