- ...And Morris?

- Um, it actually already is a word. Tnetennba.

- Good heavens, really? Could you, uh, use it in a sentence for us?

- "Good morning. That's a nice Tnetennba" [1].

______________

[1] Moss from IT Crowd on Countdown:

https://youtu.be/g9ixvD0_CmM?t=52

Edit: to clarify, if you don't know what a word means, just seeing it used in a sentence won't necessarily tell you much about its meaning, so that a language model was able to generate a phrase with the word in it doesn't necessarily tells us it understands the word's meaning.

Maybe not, but the fact that it actually made a sentence with a word in it that it could not have possibly seen in the training data tells us that it understood something about the meaning of the instructions.

The word was used because it was in the prompt and the prompt was constructed in such a way as to force it to use the new word in the place of an old word. No "understanding" is necessary, other than from the human constructing the prompt who needs to understand how the system works.

In any case, it's a language model. It has no ability to "understand" anything. It can compute the probability of a token to follow from a sequence of tokens, and that's all. There's no "understanding" there, nobody made it to understand anything.

I'm using "understood" metaphorically similarly to how people say machine learning models "learn" something when they really just minimize a loss.

Ah, thanks for clarifying. In that case, can you explain the metaphor? What do you mean literally, when you say that GPT-3 "understands" something metaphorically?

Edit: btw, "learning" in the context of machine learning is more of a term of trade with well-established connotations. For example, we have Tom Mitchell's definition of a machine learning system as "a system that improves its performance over time", etc. We don't have similarly established definitions for the "understanding" terminology. Hence my request for clarification. I literally don't understand what you mean that GPT-3 "understands" metaphorically.

I'm using "understood" as "can apply in an unseen context". For example if you a human student understood a math concept it means they were able to use it in a new context and have not just memorized how to solve a specific problem.

I think the more interesting question is the definition of meaning. I am thinking about meaning here as the relationship between symbols. So if you can explain what a words means, you can give a definition in terms of other words. If you "understood" what a word means, you have not just memorized the definition but can apply the word in unseen contexts.

Thanks for clarifying what you meant by "undestanding". I think it's a bit too broad of a criterion, for example in the video above Morris is using "Tnetennba" in a new context but there's no way to tell whether he really knows what it means (the joke is not that he doesn't know, but exactly that using "Tnetennba" in the kind of sentence he uses it does not elucidate the meaning of the word).

>> So if you can explain what a words means, you can give a definition in terms of other words.

Suppose I give you the following mapping between symbols: a -> p, c -> r, d -> k, e -> j.

Now suppose I give you the phrase: "a a a c a d e e a c"

I gave you a definition of each symbol in the phrase in terms of other symbols. What does the phrase mean? Alternatively, what do the symbols, themselves, mean?

Obviously, you can't say. Being able to give the definition of a word in terms of other words presupposes you understand the other words, also. So, just because a language model is using a word doesn't mean it knows its meaning- only that it uses the word.

There is a bit a grey area, which is why I was careful say "the model learned/understood/knowns something" in my earlier comments.

However, You seem to be making the Chinese Room argument. If you define meaning such that either no computer program could possibly "understand" meaning or it is unverifiable if it does, I don't think it makes much sense to have a discussion if GPT-3 does. Is there a test that a model could pass that would convince you that it "knows" meaning according to your definition?

The "Chinese room argument" says that external behaviour cannot be proof of consciousness, intelligence, understanding etc. What my comment above says is that you can't explain a word A by means of another word, B, unless you already know the meaning of B.

My comment is relevant to the question of whether GPT-3 has "understanding" or not, because in order for GPT-3 to understand the meaning of a word A in terms of a meaning of a word B, it needs to already know the meaning of the word B. However, this is what we wish to know, whether GPT-3 knows the meaning of any word. Observing that GPT-3 can use a new word in the place of a different word doesn't tell us whether it knows the meaning of the original word.

As of yet, no, there is no formal test that would convince me or a majority of reserachers in AI that a model "knows", "understands" or anything like that. The reason is not that I am too stubborn, say. Rather there simply aren't such tests available yet. One reason for that is that we don't, well, understand what it means to "understand". We don't have a commonly accepted formal definition of such ability. Without that, we can't really design tests to prove that some system has it.

The take away is that it will be a long time before we can know for sure that a system is displaying intelligence, understanding, etc. This may be unsatisfying- but the alternative is to design meaningless tests that prove not what we are trying to prove and proclaim the goal proven if the tests pass. This does not go well with the purpose of scientific endeavour, which is to acquire knowledge- not pass tests and make big proclamations about winning this or that competition.

In short, I'm not saying that computers can't have understanding, or that we can't know if they do. I'm saying that right now, these things are not possible, with current technology.

Depends what you're talking about by meaning.

It correctly interprets the first part of the prompt as 'farduddle ~= jump' and the second part as an instruction to generate a sentence containing farduddle, possibly utilising a corpus of existing sentences containing jump in the context of 'really fast'. But that's also a series of instructions you could imagine as a DSL a relatively simple program could parse and generate a satisfactory response to. Which I believe the OP is classing as 'structure' since it's just performing translations based on familiar syntax. Understanding the concept of 'jumping' is a step further, before we get into the more philosophical stuff about qualia and whether things that can't jump can ever truly understand the experience of jumping...

Right, in this context I define meaning as the relationship between symbols. Form would be how to structure a sentence (but that sentence wouldn't have to make sense).

If the parent comment was making some kind of Chinese Room argument than I think that's not very helpful for the discussion. "GPT-3 learned nothing about meaning because no computer program can do that by definition".

But I don't think they were trying to make that kind of argument as you seem to be suggesting because they said they were unsure if it learned anything about meaning.