Hacker News Reader

Comment by DougBTX | Hacker News Reader

I don’t buy this “it is just statistics” dismissal. Practically speaking, any multiple choice question is ultimately a “predict the next character, A, B, or C” challenge. Should we dismiss all multiple choice questions because they are just “predict the next token” tasks?

There is a broader question about whether it is possible to learn from reading. Can a blind person ever really understand “blue”? If not, what can be learnt by reading? Why do we rely on reading and writing so heavily for learning?

Edit: just want to note that this is a great site, really like the author’s approach and style. And maybe if I’ve not learnt from his writing, at the very least it is a good read :-)

View on HN

I don’t think I solve multiple choice questions like that.

I consider the letters to correspond to facts about the world, consider the facts symbolically, and then map back to the letters.

It has the same output, but it’s a different process.

Yes ? We already dismiss them, because they are too easy to hack, and don't give enough insight to the teacher as to the student's learning process ?

In fact now that I think of it, almost the only fields that seem to be using them would be... foreign language teaching ! (Where they are probably appropriate, since so much of it is just pattern memorisation...)

It's funny because large parts of the world outside the US already do not routinely use multiple choice questions and the motivation behind their use in the US seems to be primarily an easy grading process.

Yes, grading free form writing tasks is a lot more work, but they are also an obviously vastly superior form of exam question.

I agree. Quantum mechanics is "just statistics", and yet it still causes chemistry to happen.

I think that the argument for qualia is pretty weak - if we can have a conversation about "blueness" with a blind person, we have to admit the possibility of "understanding" existing in some form in LLMs.

Even if it's just a statistical emulation, at what point does a statistical emulation become reality, if it's a good enough emulation down to the metaphorical planck distance.

Anybody who has ever done arithmetic knows that human logic is more than just statistical pattern prediction. Humans can learn and execute algorithms in precise, deterministic ways, which current LLM's fundamentally cannot do.

That's funny. So being intelligent is executing algorithms deterministically and being a machine is making mistakes and being non deterministic?

How the tables have turned

I’m not sure that’s true. Human make mistakes performing math all the time. Human’s ability to accurately perform multiplication is very much a statistical phenomenon. Now you probably dismiss those mistakes as different than the sort of thing that a llm does when following a rigid set of instructions, but to me they seem different in degree but not kind

Why cant LLM's fundamentally execute in a deterministic way? Its a computer running computation on some fixed data. Without randomization parameters for e.g. temperature it would be pretty deterministic?

My understanding is "tech enthousiast" level, so happy to learn.

So for LLMs like ChatGPT, one issue with doing arithmetic is that the input is tokenised, so it doesn't "see" the individual digits in numbers. That will make it harder for it to learn addition, multiplication etc. You can see what the inputs to the model might look like here: https://platform.openai.com/tokenizer

So for example, the text "123456789" is tokenised as "123", "45", "67", "89", and the actual input to the model would be the token IDs: [10163, 2231, 3134, 4531]. Whereas the text "1234" is tokenised as "12", "34" with IDs [1065, 2682]. So learning how these relate in terms of individual digits is pretty hard, as it never gets to see the individual digits.

I think to extend on the question, though, the fundamental answer is "There is nothing stopping the LLM from containing the embedding of all basic math", with the proviso that tokenization makes it vanishingly unlikely (perhaps in the current generation, or within reasonable resource limits).

I see it analogous to asking a human why they don't just "learn all the answers to simple arithmetic involving integers below 10,000" - you possibly could, it would just be a huge waste of time when you can instead learn the algorithm directly. Of course, LLMs are inherently a layer on top of an existing system which solves those problems quite well already, so it'd be somewhat silly there too.

A congenitally blind person ? For a quite weak form of "understanding" then... (and they might still have some form of built-in but unused/repurposed specialized neural circuitry for sight !)

Our consciousness is biochemical statistics. It's a weak argument to dismiss AI as "not like us".

https://clementneo.com/posts/2023/02/11/we-found-an-neuron

Reply on news.ycombinator.com