Hacker News Reader

Comment by TonyAlicea10 | Hacker News Reader

> 2024 called and wants its talking points back.

It's a classic example of the kinds of questions LLMs get wrong. There's plenty of others. Not sure your point here. We can easily find things a 4-9 year old will talk about that an LLM will get wrong, hallucinate, etc.

> a child tutored by a good quality state-of-the-art LLM with a good teaching-focused harness could have better learning outcomes than a child without it.

There's a lot of work being done in 'could'. And it's entirely ignoring the dangers.

I'm not saying "no LLMs in education". I'm a technical educator and I give students LLM prompts and agent skills that I've built to help them learn.

This isn't that. We're talking about giving an LLM to a 4-9 year old and saying "this is your teacher".

View on HN

> It's a classic example of the kinds of questions LLMs get wrong. There's plenty of others. Not sure your point here. We can easily find things a 4-9 year old will talk about that an LLM will get wrong, hallucinate, etc.

My point is that the hallucination rate of modern LLMs, while not zero, is so low that this is no longer an issue in practice.

> There's a lot of work being done in 'could'. And it's entirely ignoring the dangers.

I agree that there are some risks. But for many children, the alternative is not human teacher, but no teacher at all. This considerably changes the risk-benefit calculus, IMHO.

> My point is that the hallucination rate of modern LLMs, while not zero, is so low that this is no longer an issue in practice.

That's just not true. Talk to an LLM about a subject you know deeply. They make up things all the time.

With stuff that a 4-9 year old is going to be learning? Color me skeptical.

Teaching children is not just about delivering facts to them. The challenge when teaching children is not "do you know the topics that you are teaching them." They challenge is in the pedagogical methods used to get them to learn.

You’re applying human cognition expectations to an LLM. Just because the topic is simple to a human doesn’t mean it improves the likelihood the LLM will behave better. It’s all inference to the machine.

I’m applying my experience with these models, as well as common sense.

What are you applying? Would you even admit LLMs are useful for this if they’re perfect at topics that kids would be learning? Or are your objections actually ideological rather than pragmatic?

I'm not sure how common sense says that LLMs will hallucinate less on topics for children. That's not how they work.

And no, LLMs are not perfect. They're non-deterministic.

And no, even if they were perfect my statement was that it is irresponsible to teach a child to trust an LLM. Even if it was perfect on these topics, as they move forward in education that trust would later fail them.

That is a problem both ideologically and pragmatically. You have to think long-term.

On a side note: ask an LLM to write a story at a 4-9 year old level. Work with it for awhile doing iterations and it will hallucinate a previous detail or change a character name as context rots. Lots of other examples of "basic" topics that still produce hallucinations.

The common sense part is not holding LLMs to a standard of absolute perfection that literally no other medium, or humans, can reach.

You’re welcome to your views, but they don’t make a lot of sense, or fit with reality. Continue angrily tilting against windmills I guess.

You were skeptical that LLMs would hallucinate on topics for children and you said it was common sense.

But now you’re saying the common sense is that they will hallucinate.

Man, it must be exhausting to have reading comprehension this bad.

Let’s try again: your claim is that LLMs “make up stuff all the time”, so they can’t be used to educate children.

I’m saying that modern SOTA models do not make up stuff all the time as it pertains to the topics a child of 4-9 would learn. That doesn’t mean they’re perfect or will never make mistakes, but there’s zero reason to think such issues would be widespread. And common sense would tell you that perfection is not required for LLMS to be a net positive in education.

I’m sure you’ll have another vapid rejoinder, but this has been boring for awhile and you seem to be enjoying being obtuse, so I’m moving on.

> My point is that the hallucination rate of modern LLMs, while not zero, is so low that this is no longer an issue in practice.

Ahahahahahahahahahahahahahahahahaha

Do you actually use these things?

> Do you actually use these things?

I use them every day for SWE. I cannot remember any hallucination from the last 6 months. It pretty much stopped being an issue from the beginning of this year, if you use top paid models. Free models or lower-tier paid models can still hallucinate, though.

I can use Claude as much as I want at work, and I get hallucinations every day. I'm not even doing anything all that fancy with it, but I get frequent hallucinatory behavior regardless of the model or how I go about things.

Every single time I have asked Claude or another "top tier" offering a detailed question about something I'm an expert in, there's something wrong in the answer. Usually many things.

Do you actually look at what comes out?

I would suggest that your ability to discern when the model is wrong might be the thing that has changed.

This is what it seems to be. I've had to review the LLM output from coworkers who have insisted that the "old" models are terrible but the new ones are very good. They are always discussing the latest models and their tweaks and things.

Their output is just.... not very good. For example, it may not fail tests, but it does so for wrong reasons and will fail under different conditions. Or it checks or guards for things in trivially redundant ways, like doing something like `if (x > 5 && x > 3)` that betray its lack of "knowing" what it's doing. And I can't even get answers from them anymore because they just feed anything I say into the LLM and copy/paste the response.

I'm basically being forced to code with LLMs via review through a person proxy. Or maybe they just have it hooked up to read it directly. It's maddening. It's like having a realistic enough chatbot just shuts people's brains off.

It would be extremely valuable to be able to reliably identify this kind of behavioral tendency in an interview setting. Some kind of reliable test that screens for likelihood to go AI psychotic.

I'm not participating in the debate, but fyi I have just asked haiku, sonnet, opus and fable that question and they all got it right. Tried with the letter s in portuguese word "sussuarana" (mountain lion) too - same thing.

> lot of work being done in 'could'

if my aunt had wheels, she'd be a bicycle.

Reply on news.ycombinator.com