Hacker News Reader

Comment by MeImCounting | Hacker News Reader

ChatGPT is far and away the best tutor I have ever had. It starts to fail once you get past mid level concepts but its ability to provide links to other sources, explain things, not get offended when you call it out as wrong about something (which human teachers certainly do) and many other aspects make it very helpful. Specifically its useful in expanding your breadth of knowledge on a subject. Less so depth.

Regardless I have learnt things and taken on projects I never would have attempted without it.

View on HN

> ChatGPT is far and away the best tutor I have ever had.

> Regardless I have learnt things

who cares what we learn is true as long as we learnt something?

They alluded in their past comment that human teachers make errors too which students can identify. So perhaps if students don't implicitly trust human tutors, then so too can they corroborate ChatGPT against reality and solidify their understanding.

I don't know about where you're from but in my country to teach requires several years of qualifications, assessment and continuing observation by an independent regulator

and there are consequences if you're caught teaching rubbish

not to say they are perfect, but there is some level of quality control

meanwhile ChatGPT just bullshits endlessly with no possible way to punish OpenAI for it

I think perhaps you had a bad experience with GPT at some point or didnt ask it to provide sources that you then read through. Either way it is far more reliable than you seem to believe.

It's also getting better, while other resources aren't, and won't.

I've just listened to the latest Numberphile podcast, and one of many surprising anecdotes from Donald Knuth was that as a kid he'd been disillusioned by maths when he "proved" that 1=0 and his teacher couldn't find the mistake.

> meanwhile ChatGPT just bullshits endlessly with no possible way to punish OpenAI for it

Sure there are, and more importantly there's also ways to "punish" the models themselves — that's what the "thumbs down" and "regenerate response" buttons in the chat UI are for.

I'll join you in being downvoted.

Human teachers can be audited without advanced knowledge in maths, computer science, and information theory to discover what went wrong.

Human teachers can be held accountable.

A Human teacher has limited scale thus limited damage over bad information with more safeguards than "Big Corp promises not to be bad"

A human teacher can realize their own mistakes.

Human teachers have complex incentive structures that can modify their behavior.

Many humans (even ones with technical training) tend to see computers as infailible in first principals of math and carry that forward to deep learning which is where the math becomes tough to audit, couple that with LLMs affinity for overcoming the uncanny valley enough to seem like a rational human actor because of the datasets being used and you have "magic". This is before you even add financial incentive.

Not saying these issues in deep learning can't be tackled but the disengenous hype-train and the downvoting because people point out flaws with the new internet ATM is wearing thin.

Idk how you're promoting, but hallucinations are rare for me, even at graduate-level material.

There's a reason GPT-4 scores so highly on advanced exams like the USMLE.

> There's a reason GPT-4 scores so highly on advanced exams like the USMLE.

yes, it's called overfitting

There are entire industries devoted to helping humans overfit on standardized tests. Humans love to overfit, it's often easy and faster to get what you want--which in the case of education, is usually credentials, not knowledge.

> Humans love to overfit, it's often easy and faster to get what you want

The problem is Goodhart's. Tests aren't used to measure your ability to overfit them, but to measure your general knowledge. Obviously the latter is exceptionally difficult to measure so we settle for an easier metric. The problem comes from believing that the measurement well aligns with the thing we set out to measure. The same is for humans as is for machines and it is why it is insane to say something like "look, GPT did well on the BAR exam, it should be able to practice law." Because we explicitly do not want a legal system that follows the letter of the law (rather spirit of the law) and it's clear which one GPT would do. Yes, there's a problem that humans go this way too, but it's something we can solve if we are willing to admit the fuzziness of our metrics.

> hallucinations are rare for me, even at graduate-level material.

How do you know? If you're using it as a tutor you're using it to get knowledge that you don't know. Sure, it doesn't get things that are obviously (to you) wrong, but what about subtle things? At the graduate level subtly dominates and many concepts are counter intuitive or non-obvious in the first place (otherwise they wouldn't be expert knowledge).

I know because I check often with what I already know, or official material/explanations written by humans.

Gpt helps me connect dots and uncover useful details I didn't previously have, and that official explanations didn't include. And just about every time I've double-checked, it's right, even when it didn't make sense to me (counter-intuitive/non-obvious due to my mistakes in reading comprehension or previously erroneous concepts in my mind).

Okay, with this added nuance I'm not worried. But I think we should be explicit about this because it is difficult to distinguish if people are doing this or not and it is quite reasonable to believe that they are not.

I have no issues with anyone using GPT as a __noisy__ tutor. But we just must be clear it is easy to get fooled, difficult to validate, and easy to cause GPT to change its answer. There's a high rate of people using it as a non-noisy tutor, so we can't be implicit.

FWIW, I use GPT in a very similar way, and frequently. Especially with the decline of Google and especially for high level topics. It functions very well as a fuzzy search and often even when wrong (which is frequent for me when asking high level math questions, think post PDEs) there's still words or phrases that help me significantly narrow down a Google search.

Which any Google Search engineer should not be reading this as "replace Google Search with GPT/Bard"!!!!! They work in tandem!!! I CANNOT STRESS THIS ENOUGH and do not want to see the hell that that creates.

The level of cautiousness many feel about the risks of AI hallucinations feels a bit excessive in a world where many people can easily obtain chainsaws without the need to fulfill any significant prerequisites

This is a cynical take on what I think is a meaningful benefit that the above commenter is talking about. If you use ChatGPT to learn concepts and approach it with the caveat that you should make sure you understand what you are trying to learn (something that helps to verify that what is being presented to you by the system is "accurate"), it's a fantastic resource. That isn't to say it's true for learning all things, but I'm generally impressed and optimistic for ChatGPT to be as valuable a teaching/learning resource as any other I might find on the web.

Human teachers aren't omniscient, but often have absolute authority over their pupils. This flaw isn't unique to AI.

I have more confidence in a teacher, especially in niche subjects. One of the big reasons why is that I can ask a teacher a question, they will infer what I am trying to ask, and not easily flip flop. Because a teacher has confidence and GPT doesn't. A teacher will even find their own mistakes without prompting (though of course, everyone needs it sometimes. Just a teacher won't say that a kilogram of feathers and a pound of bricks weigh the same and then explain how 1 kg is 2.54 lbs).

Sure, no system is perfect, but not all imperfections are equal, and I'm tired of pretending like they are.

> Sure, no system is perfect, but not all imperfections are equal, and I'm tired of pretending like they are.

I think this is the actual issue, rather than the other specifics you raise — my anecdote of human mistakes includes having a 44kHz/16-bit audio recording of a whole university lecture where the lecturer insisted that nobody can record mocap joint positions for a mere dozen joints at a mere 24Hz/8-bit for more than a few minutes because that is too much data to store, and it was their confidence that kept them from updating.

But, back to the point.

Every instance of any specific AI release (GPT or LLM or whatever else) are identical; errors are therefore necessarily correlated. Humans get an advantage from our imperfections being un-equal, we benefit from the wisdom of the crowd (with conditions) and second opinions are meaningfully independent (again with conditions), but not so for any specific AI.

(Possible twist: Mixture of Experts).

That's not true even before mixture of experts because the systems are stochastic. They are only identical if we were to always initialize with the same starting seed and rng_state.

While they do indeed have a random number generator, I think the main strengths and weaknesses of these models (and humans) are in the structures rather than at the token temperature stage — I expect a human to answer any given question slightly differently at different times, but asking the same doctor for a second opinion as the first is considered the setup for a joke rather than a sensible idea.

This is the assumption of a flat and smooth latent manifold. If either of those conditions are not met, then seeds can have a wild effect on the model output. There's only bounds on the latent structure being locally smooth or locally flat btw. This is the underpinning of such works as the Lottery Hypothesis but there's a whole rabbit hole to dig down here and it is why Alpha Go did so poorly in round 2. This is an important part of the discussion of generalizability btw. I do not think your analogy works out and it is inconsistent with your prior comment, because you're right that anyone asking GPT is akin to anyone asking the same doctor. The second doctor is a second model (or potentially a second GPT checkpoint depending on how we want to consider divergence).

Very good. Now apply that healthy skepticism to the contexts where you should have been applying it in the past but haven't been, from Stack Overflow to Wikipedia to your high school and college teachers to the evening news.

Chatgpt will not get offended when you call it out. Even when it was right and you are wrong. You call it out and now your incorrect notion is gospel as far as chatgpt is concerned. How can that possibly teach you anything?

This is another false belief that a lot of people have. ChatGPT sometimes wont correct false statements but just as often it will. This is also why we read the source and dont believe anything anyone says without proof.

Professors and chatbots are both wrong all the time and its up to the student to extract useful knowledge.

That's the hardest thing for me is that it nearly always accepts your corrections. It will often change answers even if you just ask it if it is sure or suggest it might be wrong. It's weird to say, but GPT has no confidence. Weird because it spouts things out without concern for their validity, and we might call that "confidently incorrect." But it isn't confidence because GPT won't stick to its guns like a human would. So even some of our common notions are completely divorce from how the machine actually works.

My major concern when using GPT and especially with hearing how common it is for people to use it as a tutor, is that being a novice you explicitly want a confidently correct tutor. There are certainly good aspects of GPT but you're here to gain knowledge. Being a novice explicitly means you are going to have difficulties recognizing incorrect information so you put yourself in a concerning position of becoming confidently incorrect.

Reply on news.ycombinator.com