Hacker News Reader

Comment by jqpabc123 | Hacker News Reader

Q: Count the letters in 'Hippopotamus'.

A: 9 letters

Wrong! The "Large Language Model" obviously lacks a coherent understanding of language.

"Language" is supposedly it's bread and butter but the "language" it produces is "correct" only from a statistical perspective.

In other words, it simply can't be trusted --- because it lacks any coherent understanding of what you or it are saying/asking.

LLMs fundamentally turn computing on it's head. Instead of accurate results at low cost, they produce inaccurate results at high cost.

View on HN

> Wrong! The "Large Language Model" obviously lacks a coherent understanding of language.

It's not all-or-nothing for the entirety of language, but yeah it's known that LLMs are imprecise at mapping from tokens to characters (as a human who only saw tokens may be, to be fair).

For an unseen question, even just to follow the Q/A format and give a right-category answer already demonstrates some model of language, I'd argue.

> "Language" is supposedly it's bread and butter but the "language" it produces is "correct" only from a statistical perspective. In other words, it simply can't be trusted.

Nothing can be trusted 100%, and current LLMs are far below humans, but both the structure and the meaning of their language is correct significantly above a random-guessing baseline.

A pretrained foundation model will have representations of fiction and non-fiction content, and can be guided towards either one with RL or a system prompt.

> LLMs fundamentally turn computing on it's head. Instead of accurate results at low cost, they produce inaccurate results at high cost.

Compared to what? LLMs often do things that traditional algorithms couldn't, faster and cheaper than hiring a human.

... demonstrates some model of language, I'd argue.

Yes, it's model is all based on statistics. It lacks any real "coherent understanding".

It produces nice prose --- that simply can't be trusted for even rudimentary results.

What do you think "real" understanding is?

To me, if something can map some modality (the word "horse", a picture of a horse, ...) into corresponding representation in its internal semantic space then I'm fine to say it "understood" that. There are varying degrees to how useful an understanding is (does the semantic space link related concepts closely together? can it be used to reason, extract information, and make predictions?).

...can it be used to reason, extract information, and make predictions?

Reason? Clearly no. It doesn't understand what you're asking well enough to "reason". But it will gladly offer an attempt based on some statistical mishmash of patterns it has been trained with.

extract information, and make predictions? Yes --- just not reliably or accurately. Anything it produces needs verification --- which defeats a huge chunk of it's utility.

Would you trust a calculator that fails a simple math test? Of course not. So why trust an LLM?

Q: What is the square root of 2? A: approximately 1.41421356

Q: What is the square root of one plus one times forty five minus eighty eight? A: i * sqrt(42)

> Reason? Clearly no. It doesn't understand what you're asking well enough to "reason".

If you ask something a question ("Could a car fit in a bread-box?"), it maps from words into concepts in its internal semantic space ("car" to representation of car, associated with large size), then applies rules/patterns in a way that reaches a correct conclusion more often than an RNG baseline - what concretely would you say is missing from the process?

> Would you trust a calculator that fails a simple addition test? Of course not. So why trust an LLM?

You can weigh something in proportion to its accuracy, opposed to giving total trust or nothing at all. An 80%-accurate weather prediction is still useful, as are calculators despite potential for hardware issues or (more likely) fat-thumbing the wrong button.

> Q: What is the square root of one plus one times forty five minus eighty eight? A: i * sqrt(42)

I'm unclear on what you intend to be showing here. Is this an LLM response?

Is this an LLM response?

Yes. It shows how easily an LLM can misinterpret subtlety and ambiguity in human language.

> Yes. It shows how easily an LLM can misinterpret subtlety and ambiguity in human language.

Without parentheses or context, -42 is a perfectly reasonable answer to "one plus one times forty five minus eighty eight", to the extent that it's probably what I'd go with if asked it on a quiz.

I imagine when you read it in you're head you're putting emphasis/pauses in places that lean towards the `((1+1) * 45) - 88` interpretation, but neither I nor the LLM have that information.

If asked, GPT-4 can give reasonable responses to this issue: https://i.imgur.com/ptMB7Ci.png

It is applying out of order symbolic precedence rules to prose language. These rules are not universal and certainly not when writing in prose form.

LLMs frequently misinterpret and proceed without hesitation. They lack the necessary understanding to ask for clarification.

> It is applying symbolic precedence rules to prose language. These rules are not universal and certainly not when writing in prose form.

You have to apply some form of precedence in order to evaluate it at all. Question is whether to use PEMDAS, or left-to-right. I don't see anything wrong with the LLM's initial choice for the former, particuarly since it can explain the alternative and that it may be the more likely intention if spoken casually out loud as opposed to written as text.

To potentially get back more to the core of the issue (rather than going on a tangent about whether an answer many humans would give is a misinterpretation): How do you define understanding? What do you think is currently missing?

-- (edit, to respond to your edit:) --

> LLMs frequently misinterpret and proceed without hesitation. They lack the necessary understanding to ask for clarification.

I'd conjecture that always giving an answer and not asking for clarification first may be an artifact of the system prompt/RLHF of the LLM you're using. Do you believe that asking for clarification on ambigious questions requires understanding?

Do you believe that asking for clarification on ambigious questions requires understanding?

Absolutely! Recognizing ambiguity requires a deeper level of cognition and analysis than simply making assumptions and plowing ahead to the wrong conclusion.

LLMs clearly have no understanding of what "the letters in 'Hippopotamus'" means --- and they aren't smart enough to ask for or to really grasp an explanation.

> > Do you believe that asking for clarification on ambigious questions requires understanding?

> Absolutely! Recognizing ambiguity requires a deeper level of cognition and analysis [...]

Then if asking for clarification requires understanding, what do you make of cases where the LLM does ask for clarification (e.g: https://i.imgur.com/OXQF2nx.png)? If it's doing something that requires understanding, does that not imply understanding?

It's a step in the right direction.

Unfortunately, this not that common and typically only occurs in very simple cases. In my experience, an LLM will typically just plow ahead in most cases and let the user figure it out --- or not.

> typically only occurs in very simple cases

Would you agree then that it has understanding of at least these simple cases?

Personally I think the fact ChatGPT/etc. don't ask for clarification more often is largely just down to system prompt and RLHF encouraging it to always give an answer, not necessarily any deeper limitation of LLMs. May be worth testing how well it can recognize ambiguity when specifically tasked to do so, or standard Q&A but with a system prompt that clarifies that it can ask for information before answering.

Not necessarily.

It would be much more convincing if they were able to do so *consistently*. In my experience, they do not.

Everything they produce requires verification --- unless the results don't really matter.

> Not necessarily.

You claimed asking for clarification requires understanding, and then agreed my example showed it asking for clarification. That does logically necessitate understanding (if ¬U ⇒ ¬C, and C, then U).

> It would be much more convincing if they were able to do so consistently. In my experience, they do not.

I have a suspicion that your experience as such may largely be due to using chatbots conditioned to directly give an answer - opposed to being representative of fundamental lack of ability. Not to say that LLMs are perfect or at human-level on the task yet, though.

What rate of correctly identifying ambiguous sentences would you set? I could collect a dataset and give it a shot.

Your point aside, counting letters is not a good task for LLMs since they aren't trained on letters. They are trained on tokens which represent words or parts of words.

In other words, they don't really understand "language" and it's constituent parts.

Reply on news.ycombinator.com