Hacker News Reader

Comment by kypro | Hacker News Reader

I guess I have a fairly cynical take on this. I think this just exposes that AI is fundamentally going to undermine the value of education in certain fields. The fact some tests can be "cheated" by AI really just suggests that some skills have little to no value in our future AI-prevalent world. Stopping people cheating is going to change that.

In the same way that in the past being able to do fast, accurate calculations in your head or on paper might have landed you job as a "computer", today technology makes the idea of hiring someone to be a computer absurd. And therefore any tests that test someone's ability to multiply and divide large numbers is basically a worthless test in today's world.

Fields where there's value humans can add won't be cheatable by AI (at least today's generations). Tests in these fields will hold value. In some cases it will mean the education and tests will evolve to educate and test areas where humans can still add value - arguably this has already happened in maths. But many fields could be replaced entirely.

I don't mean to sound unsympathetic to those who's jobs and professions will replaced by AI in the coming years and decades. It makes me sad that someone could spend years obtaining and perfecting skills which might soon become redundant, but for better or worse AI systems like ChatGPT will fundamentally change the value of some skills and educational fields. The sooner we accept this the better we can adapt society for it.

View on HN

Yeah, no. Training people doesn't work like that. Education is about training people, we don't utilize the output of the kids and reward them with grades that they can exchange for food and toys.

The whole point of testing is to put them in hypothetical situations and grade their progress with purpose of being aware of their development so we can improve it. Another thing we do is selecting the particularly good ones for further advanced training.

The problem with cheating is that it provides wrong data about their progress, you don't want to end up with a generation that cheated their way up without learning anything.

The selection for further training is probably not that big of a problem, its mostly about supply and demand and ChatGPT won't change that.

Do you think Trump and Sam Bankman-Fried took your advice? I need a pool of people to flip burgers and take orders and those people better not take cookies from the jar. For everyone else, it's called creating shareholder value. Your advice makes for great burger flippers.

You’re either a burger flipper or a con man? That’s your binary state on mankind?

>The whole point of testing is to put them in hypothetical situations and grade their progress with purpose of being aware of their development so we can improve it. Another thing we do is selecting the particularly good ones for further advanced training.

You're missing the point. Why are we making them achieve progress on something that is being trivialized by AI? Let them use AI as a complementary tool and test them on something that requires measurable human skill. If this is difficult it's a sign that perhaps you're not teaching anything valuable.

Respectfully, I think it's you who are missing the point. There is value in teaching humans to communicate and synthesize information. It's possible that in the future the best thinkers and philosophers and historians and political analysts will be AI, but 1/ we are very very far from that, and 2/ it's important for civilisation that we don't leave thinking to the machines.

Exactly, educations is not a vocational school where we teach people useful skills so they start doing whatever they learned. Machines have long been able to do things that we teach in the schools, the purpose of teaching is developing people. Once they are developed(that means they built understanding of how nature and society works and how to interact with) and its time to expect an output, only then we can talk about teaching useful skills.

Vocational schools do exist and they are useful but that's not what we are concerned about. People who cheat not to learn useful skills simply spend their time and money.

> There is value in teaching humans to communicate and synthesize information.

Why just use the written word via essays? You could use other forms of media such as video, physical art, or maybe programming to teach this as well. The five paragraph essay isn’t the only game in town!

Similar arguments were/are made about the use of calculators.

It will go the same way. You need to understand the basics (to various degrees depending on the direction of your education), then learn how to use the new tools we are creating.

Honestly, I'd prefer we treat this as an opportunity to introduce tight feedback loops into essay writing the way we do with programming. Then evaluate the end result as a whole rather than a way to nitpick grammar rules.

Curriculum learning. It is pedagogically useful to teach simpler and easier versions of something before teaching the harder versions.

While you likely always have a calculator with you which can do division much more quickly than you can do long division, if you want to teach students more advanced math, it is nonetheless useful to have previously taught them long division of numbers, so that you can teach them long division of polynomials.

If students haven't learned how to do long division of decimals, and how that works, it is harder to teach them to do long division of polynomials.

Of course, nowadays there are computer programs which can do long division of polynomials more quickly than a person can, but this is just an example.

Suppose there are a sequence of levels, where for each level n, it is much easier to teach someone level n+1 if they already have an understanding of level n. If computers can do levels up to level k faster than humans can, it may still be useful to start teaching someone at level 1 and then proceed to level 2, and so on, even though computers can already do those levels, until finally it is time to teach them level k+1 which we have not yet managed to teach computers to do.

I think it's the same reason why we are teaching assembly even if there are compilers. Or teaching Latin even though it's a dead language

This is an interesting point of view.

I always viewed grades as a way of establishing intellectual hierarchy, which probably wasn't the most healthy way to look at them; a way of deciding where amongst my peers I stacked up. In University, we had class ranks.

Those with the best grades and thus highest rank would be given the best opportunities for wage-slave class work. Eventually people hit University and split up according to passion, goals as well as grades. Of course, there were exceptions; entrepreneurs were the folks that buck the system and don't need grades to succeed. In a sense, the grades were exchanged for food stamps and toys, only later in life. The winners drove the lambos and the losers are still living in their parents' basements.

> Those with the best grades and thus highest rank would be given the best opportunities for wage-slave class work

I don't believe that this observation is true. Those who got into tech, bitcoins etc. at the right time got the lambos - not the most brilliant ones. Tech is full people with inert brains who make much much more mone than some of the most brilliant scientists out there.

It has been good time to be into plumbing instead of astrophysics if you are optimising for money BTW. Just like knowing ReactJS has provided better income opportunities than being rocket scientist.

> Just like knowing ReactJS has provided better income opportunities than being rocket scientist.

that's because rocket science requires massive capital investment for said scientist to achieve anything - this aint no more backyard rocketry in the modern day.

ReactJS provides better income because it does not require massive capital investments on the dev to produce usable output.

The idea that societal benefit should match financial benefit is just a story made up because of the inclination that people who are moral should be rewarded etc. It has no economic basis.

Which means testing student under supervision and letting them us AI for assignments is probably going to work fine.

Sure, I also find AI and all other devices that we already had(calculators, programming languages, simulators etc) very useful for learning. Incorporating the latest and greatest tools into education works well.

> you don't want to end up with a generation that cheated their way up without learning anything.

This sounds like kind of a good summary of what we are doing to ourselves as we speak. Case in point: the crypto scandal, the housing bubble etc. A ton of people made out like thieves.

Training people (or anything) using a flawed objective function is not useful.

Writing essays, specifically expanding a list of well-reasoned bullet points into communicative written prose will be a worthless operation in the future, and likely already is.

Have you not noticed that when you write something down with the purpose of summarising it in a structured way you tend to understand it better?

There's no market for kids essays, we never made them write these things with the purpose of being useful output. The market for essays for grown ups is also very small, making kids write things down was not with the purpose of covering the case of them becoming journalists or something. It's just that makes them learn to think and structure their thought process. That's why people take notes and write down things even if its not going to be graded or read by anyone else.

Agreed. Yet many students, parents, and even teachers lost sight of this purpose.

School essays and actual essays aren't even the same thing.[0] It sounds like a more didactic use of these kids time would be arguing on forums about their favorite anime.

[0]: http://paulgraham.com/essay.html

> School essays and actual essays aren't even the same thing.

Yes. And one interesting thing about ChatGPT is that, when asked to write an essay about a topic, it defaults to the standard introduction-body-conclusion organization of a school essay assignment [1]. Since its training data must have included texts organized in many other ways, I wonder if it was tweaked to default to the school-essay format.

[1] https://www.gally.net/temp/202212chatgpt/defaultessay.html

Long form essay isn't great for that, especially as taught in school.

Writing an idea down as concisely as possible is important, the rest is just embellishment and is easily handled by AI.

There is a progression though.

For students, the objective is to write what to an adult would be mediocre essays, because they are on a path to writing adult level essays.

The AI can write mediocre essays, but not adult level essays.

Assuming that a student must write a bunch of mediocre essays in order to progress past the AI to write adult level essays, then they need to be evaluated on writing a bunch of mediocre essays.

If the teacher cannot distinguish between AI output and the student's output, and the result is that we give up on evaluating mediocre student essays, then the students are ultimately inhibited from progression to adult level essays, which, currently, are superior to the AI's output.

Automatability may make an output ubiquitous to the point where marginal value for any given output is small, but if it was really "worthless" no one would bother automating it in the first place.

We automated a lot of arithmetic and then some higher math decades ago, but that hasn't meant that knowing how to do it is worthless.

Effective communicative prose is an inherently high-value part of coordinating human activity. The number of roles exclusively dedicated to it might shrink but it's part of nearly every activity in a specialized society. What's most likely is that people will use it as a time-saving tool but everyone will be evaluating how good the output is and tweaking it. And that means until we're at a point where people trust AI to actually evaluate the value of AI output, many of these roles (and certainly anyone building AIs that do this) will need to know how to recognize good output -- and the training involved in that probably involves practice in personally doing it.

We train AI's by giving them flawed objective functions as well. Nobody is impressed if you are able to copy an image, just use ctrl c and ctr v! All imagines the AI is trained on are worthless, we already have them, why train the AI to generate images we already have?

Or, maybe practicing things we know how to do is a good way to practice?

Why should I take your deliberations seriously? You’re saying that your writing is a worthless operation already.

Before ChatGPT, there has been criticism education as we know it today. The notion of “education” has not always been about training, and putting someone in a hypothetical situation and grading progress — indeed, the measurement of progress itself — is not present in all notions of education, or all methods of learning, or even all systems of determination of truth (epistemology). Behavioral conditioning with reward incentives are also not the only method in education.

One of the interesting voices about this John Taylor Gatto. Mainstream methods of “training” as you described them, may indeed be the most common method today, but it is by no means the most effective, and Gatto goes into detail about it in his book, _The Underground History of American Education_.

For example, an older notion of education is not so much to train people to mechanically repeat results (like you would train people in a factory), but rather, to develop informed and educated citizens that are able to reason, think critically, and participate meaningfully in a republic. There’s little point in having a voting body of citizens unable to think for themselves, and are merely regurgitating ideas they hear.

Well that's a strawman argument right there. Obviously by training I don't mean perfecting repetitive tasks or anything of that sort. If that was the case, I would have advocating on training prompts or fast typing or something of that sort.

You can check my other replies in this thread to understand better.

I said a lot more than that, namely that there are other notions of education than what you are describing:

- That there is some outside agent that observes progress and "improves" someone

- That such progress can be quantified, and comparable

- That such measurement is the only way to acertain the truth of progress.

- That there is such a thing as an universal standard by which someone needs to be educated

- That the purpose of education is to produce productive members of society

- That a person starts education as a "blank slate", ready to be filled by what the teacher is teaching

- That the capacity to learn and the potential to be educated is fixed, and therefore, some should be selected for even further education

- That learning only happens when there's an outside agent educating the person

There are so many problematic things with this frame, even if it is fairly mainstream.

Yeah, no. Have you looked at the way education is performed and assessed lately?

> The whole point of testing is to put them in hypothetical situations and grade their progress with purpose of being aware of their development so we can improve it.

Yes, that is the stated purpose, but it does not match reality. I have only seen a couple of examples of this approach to assessment in all of my years of education. Most teachers optimize for their time commitments and "coverage" of the syllabus material.

>any tests that test someone's ability to multiply and divide large numbers is basically a worthless test

Right now, teaching elementary school aged children to solve complex multiplication and division is considered an effective way to teach people critical thinking, persistence, and basic logical reasoning structure. Structured education typically relies on an ability to assess student improvement, so 3rd graders don't get a calculator when they take a multiplications table test, even though we'd all use one at our desk at work.

I would surmise that likewise there is value to students learning to read, analyze, and report insight about a text, to attempt to identify symbolism and patterns, to communicate their ideas effectively, and to be assessed about that - even if by the time they're in the workforce they'll usually just ask an AI to do it.

Learning how to multiply and divide large numbers was one of my earliest exposures to numerical algorithms.

I usually use a calculator or Python for large numerical computations, but the algorithmic understanding is something I use every day.

Someone who can use a calculator to solve a problem, still has to understand the concepts being used in order to know what buttons to press on the calculator. They will have learned what those calculations mean, and how they work.

Using an AI to crap out an essay with no understanding of the topic is not the same thing. While an essay is the way we test students, it is a stand in for measuring understanding, which is the ultimate goal.

I think that the only solution may be for the essays to be written in controlled environments (school). If AI forces schools to do more school in school and less homework, I count that as a massive win. What I would give for the countless hours of my childhood and teenage years wasted on useless homework…

> The fact some tests can be "cheated" by AI really just suggests that some skills have little to no value in our future AI-prevalent world.

That's like saying 'speech has no value'. It doesn't have value on its own, but if you can't speak then suddenly there is a whole pile of stuff you can no longer do. The ability to put your thoughts down in writing and to organize them into a hopefully coherent form is an act that has implications far beyond the writing itself. Besides the fact that it can help you to organize your thoughts in the process. Plenty of times when I write a longer piece I find out halfway that one or more of my assumptions were wrong and then I get to fix that and re-evaluate my position. It isn't rare at all to go through several such iterations for a single text and at the end of it I have a much better insight into the material that I thought I already knew well enough to write about.

Don't underestimate the value of these foundational skills.

I'm not sure I buy this argument. People still need to learn that 2+2=4, derivative of sine is cosine and similarly simple things in non-STEM fields. Sure these questions can be answered by ChatGPT trivially, but it doesn't mean they don't need to be taught.

That's not the way the system works. You spend a semester or a year going deep about Calculus, Geometry, AP Euro, or whatever. These are all things most people don't need to know unless they specialize..

Perhaps instead, we should have a semester or year classes through 8-12 + college discuss the nature of manias, like analyzing the tulip bubble, crypto bubble, 2008 mortgage bubble, etc. These are all common things that happen over and over again. Lessons which can permanently (and frequently) be called upon.

You can measure and grade people with these topics that aren't officially taught. We just decide to teach less important things

I'm not convinced calculus, geometry and European are "less important". I don't even think you can understand truly understand the tulip bubble without understanding early modern European history (or 08 and the crypto bubble and US history for that matter -- e.g. the birth of the Fed and all that). It's not enough to simply say "don't get caught in a speculative bubble" because 1. it's hard to know when you're in a bubble (housing has utility after all, part of the reason this was so disastrous was because hardly anyone defaulted on a mortgage before) 2. people will get hurt regardless (it's not like only the speculators got hurt in 08) 3. this is sorta just "no shit" advice in general, and yet people still get caught up in speculation. So if you want to convince anyone, you need to be able to formulate arguments on broad history rather than relying on pedagogy to simply convince people not to take some unclear action.

There's probably a lot more to be said about math education in general, but I'll just note that this seems to be a recurring debate in general. Personally, I wished my math education was more rigorous and more abstract. Obviously not everyone uses calculus, but also not everyone takes calculus. I'd also rather a society where a significant portion of students are trained to think of things in terms of rates of change and accumulation of change and apply that to solve physical problems, rather than the society in which even talented students are just trained to add up spare change.

I'm not sure that's entirely true. They need to know the derivative of sine exists, and how to find it/when to use it, but I'm not sure that knowing that it's cosine is actually useful.

I think you meant “stopping people cheating isn’t going to change that”.

In my friendly/complementary rebuttal, the catch lies somewhere in the phase of learning foundational material. In the same way that traditional mathematics education emphasizes doing problems by hand that could be solved computer, there’s something to be said for learning to parse the research material that ChatGPT has indexed when you’re writing a term paper. It’s worth noting here that math education may be moving away from the rote memorization and routine problem solving that today’s math researchers state are essential.

ChatGPT has made it much harder to detect cheating on untimed power tests. The biggest downside I see to this is that it makes an old problem worse. If a prof has no choice but to administer a non-computer (proctored) timed exam, students who take tests slowly or have anxiety disorders are going to suffer.

At the end of the day though, it doesn’t really matter. The person cheating is failing themselves if the foundational material you’re talking about is so important. If they can somehow manage to pass enough courses while using an AI, and then get a degree out of it, or a 4.0 GPA, or whatever then the degree wasn’t that valuable to begin with.

This already mirrors the real world pretty well. Outside of your first job or two, nobody really cares what your GPA is or what college you graduated from. In the real world if you managed to cheat with an AI you’d probably get a nice promotion. I don’t personally see a downside, other than the demise of certain flawed notions of academic success that deserve to die out anyway.

Lets say we are training ChatGPT2, would you let it use ChatGPT for solving things? Do you think it would learn better if it had to learn itself without relying on ChatGPT?

We learn by doing. Relying on smart tools puts you in a local optima, it is hard for you to improve over those tools, but people who don't rely on tools continue improving and then start using the tools later.

I was just trying to allude to the nature of the universities’ business. (Being assessing that they’ve taught what they claim to teach.)

My PhD mentor is actually leading a discussion about ChatGPT use by students, and the tone is mostly concern. But it’s not all bad; it’s an incredible tool for quickly diving into a subject at a surface level and getting your bearings regarding what’s important. It saved my ass on one of my finals too when the rest of the questions were harder than I expected and ChatGPT helped me tear through the remaining T/F questions.

With schools though you often need to learn things along the way that can easily be cheated with an AI, but the end goal not so much.

For example, writing a basic screenplay can be done with ChatGPT, but it can’t write one that will likely be turned into a movie/show.

And automation will continue to make some jobs obsolete and create new ones. For example in youth sports we don’t get great coverage of youth/school events. I’m hopeful in the not too distant future we can get game summaries and stats for all games — and editors will probably be needed to add commentary. That’s something I’d love to see.

> I guess I have a fairly cynical take on this. I think this just exposes that AI is fundamentally going to undermine the value of education in certain fields. The fact some tests can be "cheated" by AI really just suggests that some skills have little to no value in our future AI-prevalent world.

If you think about it, most of the stuff you learn at school (beyond reading and writing) has not much value apart from certifying that you are not stupid and compliant to do stupid tasks.

I had to rearrange and equation at work yesterday. I let everyone know that 9th grade was not a waste.

Really what we need to do is take a good hard look at how education system works and whether or not it is aligned with our goals. Sadly, that's a pretty massive undertaking and no one can be arsed to support doing difficult things on a societal level any more.

The value of knowledge in school is to have a "liberal arts" baseline, for future selection and pursuit of an area of interest, and a basis for comparison for the river of bullshit you spend the rest of your life wading upstream through, so you don't get fooled by age old scams and lies about how this time is different.

But isn't there value in a student researching and putting their thoughts to (virtual) paper? A lot of this testing is more about the journey than destination, and students are losing these valuable skills by cheating.

I give you another anti-cheating solution: test all exams on chatGPT, by giving it first the tasks, then the answers for self grading. If the model fails, use the problem, if not, change it until chatGPT fails. The task should always require human intervention to pass.

I think the purpose of exams is deeply misunderstood. They are supposed to be a feedback system for the teacher and the student so the student can improve. They’re not a competition or an intelligence measuring stick. ChatGPT is kind of the perfect tool for that; if a student doesn’t understand maybe AI can help then get to the right answer and give them personalized attention. Something that IMO teachers fail at due to the nature of large class sizes where teachers can’t help tutor each student.

How does GPT help that at all? It can't diagnose and correct misconceptions, only proffer its own.

And sadly, colleges and employers deeply misunderstand the purpose of exams, which forces students to misuse them.

I think this will optimize for unclear questions and tasks. The ai will fail because it can't figure out what it is you were looking for, and the student will too

You'll still be able to test writing, but going forward it will need to be done in a proctored setting. Similar how you would test grade school math given the existence of calculators.

Honest question: how valuable is the current mode of testing (based rote memorization) if much of that can be done by AI anyway? Maybe it’s time for new testing methodologies?

Where are people testing just rote memorization? At least 20 years ago when I was in lower school, most of what we did had only a very small memorization component. The SAT, ACT, SAT subject exams, etc only barely rely on memorization, and many questions have no memorization component at all. It’s why 4th graders can get top-percentile scores on some of these.

You need both memorization and applied reasoning to be taught in school. Without something to apply reasoning to you can’t really reason about anything. And without applied reasoning, memorization is pretty useless.

I think it's fine for ai to be able to answer test questions?

If the ai can get a good answer, the questions must be fairly well constructed, and it's reasonable to expect a student exposed to the same info could also come up with a good answer.

The problem is with areas like reading comprehension. The underlying skill is still human-value-added today - asking ChatGPT to produce a summary of your comment isn't anything like a replacement for reading it and understanding it myself. But educators can't read minds to determine if real comprehension is taking place, and the only way they currently know how to measure it is through tests which ChatGPT can cheat.

Some years back I heard several mathemeticians advocating that the kind of math we teach should be more proof and theory based and less computation and memorization based because computers can do the computation bits better, and so knowing more than the basics of it was like putting our effort into the wrong part of the subject, and our current education system emphasises that portion of the subject the most. I imagine many avenues of education and careers will have to start shifting like that, leave the parts the computer is good at for the comupter, specialize in the parts the computer is less good at then work together with the machine.

Right. Why run and build muscle, if your avatar in the shitverse can run the whole day? Take it easy. Lie down. Put the googles on. Dont worry about the muscles.

This is understating the problem.

In months or years, it's not clear if any human will be able to provide economic value, relative to state of the art AI.

I don't see where any fundamental barriers lie between current AI and surpassing humans across the board. What can humans do that is unachievable by AI in the foreseeable future, given the current rate of advancement?

We are still leaps and bounds away from robotics that can match the precision and versatility of human hands.

I also think in terms of knowledge work, we’ll have to work through at least a few more plateaus before we automate away most humans. Transformer models were a huge leap, but you can see us pushing against their limits already.

AI is just an enhancer to those who already have creativity and ideas. Expect a Golden Age.

Reply on news.ycombinator.com