Comment by foldr | Hacker News Reader

Comment by foldr | Hacker News Reader

Parent Full thread

>The other day I saw this Twitter thread. Briefly: GPT knows many human languages, InstructGPT is GPT plus some finetuning in English. Then they fed InstructGPT requests in some other human language, and it carries them out, following the English-language finetuning.

>And I thought: so what? Isn’t this expected behaviour? Then a friend pointed out that this is only confusing if you think InstructGPT doesn’t understand concepts.

> [conclusion that ChatGPT must understand concepts]

I think this argument is a bit mixed up.

Good quality machine translation has been possible for longer than ChatGPT has been around. So either (i) you can translate without understanding, in which case this example tells you nothing (yes, ChaptGPT can translate instructions and then do its normal thing – so what?), or (ii) you can't translate without understanding, in which case you could just use machine translation as your example to show that some computational model is capable of understanding, and leave ChatGPT out of it.

The point isn't that it can translate between languages. It's not translating the instructions, at least not explicitly. Here's what they did:

- They found a task that GPT wasn't very good at, because examples of that task weren't in the training set (in any language).

- They trained a fine-tuned variant of GPT where examples of the task were in an appended training set, but only in English.

- They told the variant to do that task again, in other languages.

- Its performance improved on all of them, not just English

If the training set has no Mandarin examples of the task, how did it get better when you ask it in Mandarin? Sure, you could fake this by having an API call to Google Translate where it takes the Mandarin request, translates to English, solves the task in English, then Google Translates back to Mandarin. But it's not doing that, it's just doing the same "predict the next token" operation on the Mandarin prompt.

I don't see how it can do that unless it really has some kind of understanding.

>I don't see how it can do that unless it really has some kind of understanding.

One possibility is that the model itself has learned that tokens are related across languages based on translation examples. If the appended training changes the model's treatment of tokens in one language, that could have a statistical knock-on effect on the weights between similar tokens in different languages.

Similarly, if you train the model that "blue" is a "colour", you'd expect it to pick up that "navy" is a "shade".

At the scale/complexity GPT operates at, how is that different from "understanding"? It seems like you've just rephrased it but with more words.

Yeah, but then it goes back to GP's original argument. If relations between translated tokens are classified as "understanding", that would mean that translation AIs are already capable of understanding:

> in which case you could just use machine translation as your example to show that some computational model is capable of understanding, and leave ChatGPT out of it.

Show me a pre-GPT machine translation model that can do what I've described.

The fact that something is hard to define does not mean it isn't meaningful. I cannot tell you what consciousness is, but I know that I have it and my sock does not.

It is trivial to tie these machines up in knots, blatantly contradicting themselves from one sentence to the next. If that is understanding then it is an understanding utterly foreign to any human, and a form of understanding that looks a whole lot like bullshit.

Because it's a purely statistical explanation that doesn't require understanding. Put differently, it's possible that GPT doesn't "understand" language itself as a concept, and instead tokens in the same language are just highly-correlated when it comes to prediction. When affecting weights between tokens, it wouldn't be surprising that those weights have effects across languages, much in the same way they work within languages - after all, it's all just probabilities to GPT.

Google Translate works in a comparable way, and nobody suggests that it is sentient.

Frankly, the argument that "GPT operates at a scale that means it must be sentient" is begging the question.

No, you're the one who's begging the question. Why can't a "purely statistical" process have an understanding? If you a priori assume it can't, then nothing could ever persuade you GPT understood anything, no matter how it performed.

And again, this magical word "just". "Just highly correlated", "just probabilities". Putting the word "just" in front of something doesn't mean you've explained it.

>No, you're the one who's begging the question. Why can't a "purely statistical" process have an understanding?

This is a related, but fundamentally different thing to the point I replied to in your original comment. You asked:

> I don't see how it can [apply training across languages] unless it really has some kind of understanding.

I provided a potential explanation that is in line with how we think GPT works, and which doesn't require it to have understanding. You may feel that GPT is complex enough that this process itself models understanding - but I disagree, and I think that's begging the question because it falls back on a fact (GPT is highly complex) that is independent of the above problem (how GPT applied training across languages.)

I am not compelled by the translation example to believe beyond doubt that GPT actually models and understands abstract concepts. I don't think the fact that its training works across languages is any proof that it parses that training in an abstract way, or that it forms abstract links between the same ideas in different languages, or indeed that it has any notions of language at all.

What would convince you? Give an example of something that would show beyond a shadow of a doubt that it understands.

What would convince you that it didn’t?

> Why can't a "purely statistical" process have an understanding?

Once you understand long division, you can do it on infinite numbers without ever having seen the specific numbers. You can get this full understanding just from a handful of examples, no need for terabytes of them.

No matter how many examples of long division examples you fed to statistical model like GPT, there will always be infinite amount of numbers you can tell it where it will give the wrong answer*, unless you cheated and actually hard coded the understanding into the model.

If it cannot understand long division just from few examples it cannot ever understand it. The very reason it needs ridiculous amounts of data is precisely because it cannot understand. If you think it understands you simply aren't trying very hard to confirm otherwise.

* in a way that reveals there is no understanding of long division, obviously a human would also give wrong answer after being awake 100 hours writing numbers on paper

Reply on news.ycombinator.com