I would say that these results might be relevant for a university CS program setting, but I would make the distinction between this and actually learning to program.
The context of this task is definitely a very contrived "Let's learn OOP" assignment that, for example, just tires to cram in class inheritance without really justifying it's use in the software that's being built. It's a lazy kind of curriculum building that doesn't actually tell the students about OOP.
In that sense it's no wonder that AI is not that helpful in the context of the assignment and learning.
I wouldn't chalk this up to "AI doesn't help you learn". I would put this in the category of, in an overly academic assignment with contrived goals, AI doesn't help the student accomplish the goals of the course. That conclusion could be equally applied to French literature 102.
And that's very different from whether or not an AI coding assistant can help you learn to code or not. (I'm actually not sure if it can, but I think this study doesn't say anything new).
The more prevalent automation is, the worse humans do when that automation is taken away. This will be true for learning now .
Ultimately the education system is stuck in a bind. Companies want AI-native workers, students want to work with AI, parents want their kids to be employable. Even if the system wants to ensure that students are taught how to learn and not just a specific curriculum, their stakeholders have to be on board.
I think we're shifting to a world where not only will elite status markers like working at places like McKinsey and Google be more valuable but also interview processes will be significantly lengthened because companies will be doing assessments themselves and not trusting credentials from an education system that's suffering from great inflation and automation
Speak for yourself, but that's been how many companies have been operating for decades at this point.
AI to me will be valuable when it's helping humans learn and think more strategically and when they're actually teaching humans or helping humans spot contradiction and vetting for reliable information. Fact checking is extremely labor intensive work after all.
Or otherwise if AI is so good, just replace humans.
Right now, the most legible use of AI is AI slop and misinformation.
Unskilled AI users are people who use AI to do their thinking for them, rather than using it as a work partner. This is how people end up producing bad work because they fundamentally don’t understand the work themselves.
GenAI isn't a thinking machine, as much it might pretend to be. It's a theatre kid that's really motivated to help you and memorized the Internet.
Work with them. Let them fill in your ideas with extra information, sure, but they have to be your ideas. And you're going to have to put some work into it, the "hallucinations" are just your intent incompletely specified.
They're going to give you the structure, that's high probability fruit. It's the guts of it that has to be fully formed in the context before the generative phase can start. You can't just ask for a business plan and then get upset when the one it gives you is full of nonsense.
Ever heard the phrase "ask a silly question, get a silly answer"?
If I don’t exercise, will I ever become one?
Sim racing is a good example of this, there have been several drivers who went pro. Sims are used by racing teams and people like Max Verstappen, Lando Norris use sims heavily to improve their race craft.
I do a bit of sim racing and raced against pro drivers.
Athletic ability is a huge part but nearly every sport has a chess component. Additionally, ability to perform under pressure and dealing with stress/anxiety is another part of sports.
Sometimes I’m doing something in a new to me language, using an LLM to give me a head start on structure and to ask questions about conventions and syntax, and wondering to myself how much I’m missing had I started just by reading the first half of a book on the language. I think I probably would take a lot longer to do anything useful, but I’d probably also have a deeper understanding of what I know and don’t know. But then, I can just as easily discover those fundamental concepts to a language via the right prompt. So am I learning? Am I somehow fooling myself? How?
But the magic is in the “find patterns” stuff as memorization is just data storage. If you think of the machine learning algorithms as assigning items a point in a space, then it does uncover neighbors, sometimes ones we might not expect, and that’s interesting for sure.
But I’m not sure it’s analogous to what people do when they uncover patterns.
Definitely interesting to ponder though.
One of the first precepts of ML is that “memorization is not learning”.
Learning is generalization, application to new circumstances.
Schooling might not have learning as a product, but that’s a different problem.
Schools compromise their curriculum so that every student has a chance in the interests of fairness.
Understanding how concrete people navigate a domain and noting the common points between them can be illuminating.
Trying to calculate a generalisable statistical result from them… probably not so much.
My recommendation would be to encourage students to ask the LLM to quiz and tutor them, but ultimately I think most students will learn a lot less than say 5 years ago while the top 5% or so will learn a lot more
We’ll see a new class division scaffolded on the existing one around screens. (Schools in rich communities have no screens. Students turn in their phones and watches at the beginning of the day. Schools in poor ones have them everywhere, including everywhere at home.)
I specifically remember Telluride Mountain School’s banners in town advertising a low-tech approach.
If we assume that AI will automate many/most programming jobs (which is highly debatable and I don't believe is true, but just for the sake of argument), isn't this a good outcome? If most parts of programming are automatable and only the really tricky parts need human programmers, wouldn't it be convenient if there are fewer human programmers but the ones that do exist are really skilled?
AI does not think
Ergo, AI will not take "programming jobs"
It may highlight some "fraud people" (do not know how to say it in english .. you know, people who fake the job so hard but are just clowns, do not produce anything, are basically worthless and just here to grab some money as long as the fraud is working)
Sure. So we can keep paying money to your employer, Anthropic, right?
Why?
Because LLMs are capable of sometimes working snippets of usually completely unmaintainable code?
on some ChatGPT interfaces.
>AI will improve more rapidly than the education system can adapt.
Is entirely obvious, and:
> Within a few years it won't make sense for people to learn how to write actual code, and it won't be clear until then which skills are actually useful to learn.
is not obvious, but quite clear from how things are going. I expect actual writing of code "by hand" to be the same sort of activity as doing integrals by hand - something you may do either to advance the state of the art, or recreationally, but not something you would try to do "in anger" when faced with a looming project deadline.
This doesn’t seem like a good example. People who engineer systems that rely on integrals still know what an integral is. They might not be doing it manually, but it’s still part of the tower of knowledge that supports whatever work they are doing now. Say you are modeling some physical system in Matlab - you know what an integral is, how it connects with the higher level work that you’re doing, etc.
An example from programming: you know what process isolation is, and how memory is allocated, etc. You’re not explicitly working with that when you create a new python list that ends up on the heap, but it’s part of your tower of knowledge. If there’s a need, you can shake off the cobwebs and climb back down the tower a bit to figure something out.
So here’s my contention: LLMs make it optional to have the tower of knowledge that is required today. Some people seem to be very productive with agentic coding tools today - because they already have the tower. We are in a liminal state that allows for this, since we all came up in the before time, struggling to get things to compile, scratching our heads at core dumps, etc.
What happens when you no longer need to have a mental model of what you’re doing? The hard problems in comp sci and software engineering are no less hard after the advent of LLMs.
Architects are not civil engineers and often don't know details of construction, project management, structural engineering etc. For a few years there will still be a role for a human "architect" but most of the specific low level stuff will be automated. Eventually there won't be an architect either but that may be 10 years away
Coding via prompt is simply a new form of coding.
Remember that high level programming languages are "merely" a sop for us humans to avoid low level languages. The idea is that you will be more productive with say Python than you would with ASM or twiddling electrical switches that correspond to register inputs.
A purist might note that using Python is not sufficiently close to the bare metal to be really productive.
My recommendation would be to encourage the tutor to ask the student how they use the LLM and to school them in effective use strategies - that will involve problem definition and formulation and then an iterative effort to solve the problem. It will obviously involve how to spot and deal with hallucinations. They'll need to start discovering model quality for differing tasks and all sorts of things that look like sci-fi to me 10 years ago.
I think we are at, for LLMs, the "calculator on digital wrist watch" stage that we had in the mid '80s before the really decent scientific calculators rocked up. Those calculators are largely still what you get nowadays too and I suspect that LLMs will settle into a similar role.
They will be great tools when used appropriately but they will not run the world or if they do, not for very long - bye!
High-level languages are deterministic and reliable, making it possible for developers to be confident that their high-level code is correct. LLMs are anything but deterministic and reliable.
It isn’t deterministic like a real programmer isn’t deterministic, and that’s why iteration is necessary.
Checking output can be done by testing but test code in itself can be unreliable and testing in itself is no correctness guarantee.
The only way reliable code could be produced without human touching it would be using formal specifications, having the LLM write the formal proof at the same time as the code and using some software to validate the proof. The formal specification would have to be written using some kind of programming language, and then we're somewhat back to square one (but with maybe a new higher level language where you only define the specs formally rather than how you implement them).
Therefore, I still see a need for highlevel and even higher level languages, but ones which are easy for humans to understand. AI can help of course but challenge is how can we unambiguously communicate with machines, and express our ideas concisely and understandably for both us and for the machines.
It's obviously not quite the same as programming, but my English professor assigned an essay a few weeks ago where we had to ask ChatGPT a question and then analyze its response, check its sources, and try to spot hallucinations. It was worth about 5% of our overall grade. I thought that it was a fascinating exercise in teaching responsible LLM use.
This reminds me of folks teaching their kids Java ten years ago.
You’re teaching a tool. Versus general tool use.
> Those calculators are largely still what you get nowadays too and I suspect that LLMs will settle into a similar role
If correct, the child will be competent in the new world. If not, they will have wasted time developing general intelligence.
This doesn’t strike me as a good strategy for anything other than time-consuming babysitting.
No, it isn't. "Write me a parser for language X" is like pressing a button on a photocopier. The LLM steals content from open source creators.
Now the desperate capital starved VC companies can downvote this one too, but be aware that no one outside of this site believes the illusion any longer.
Not according to court cases.
Courts ruled that machine learning is a transformative use, and just fine.
Pirating material to perform the training is still piracy, but open source licenses don't get that protection.
A summary of one such court case: https://www.jurist.org/news/2025/06/us-federal-judge-makes-l...
> "Write me a parser for language X" is like pressing a button on a photocopier.
What is the prompt "review this code" in your view? Because LLM-automated code review is a thing now.
To retroactively grant propriety AI training rights on all copyrighted material on the basis that it's no different from humans learning is, I think, misguided.
That's a fair position: laws are for the nation (and in a democracy, that's supposed to mean the people), and the laws we make are not divine or perfect.
But until the laws change, it is what it is.
> To retroactively grant propriety AI training rights on all copyrighted material on the basis that it's no different from humans learning is, I think, misguided.
I would say it's not retroactive, it's the default consequence of what already is. Changing the law so this kind of thing is no longer allowed in the future is one thing, but it would be retro-active to say it had always been illegal.
I am sure that the 360° performance reviews have never looked better.
Your experience is contradicted by the usually business friendly Economist:
https://www.economist.com/finance-and-economics/2025/11/26/i...
jokes aside I do trust economist’s heart is in the right place but misguided IMO. “the investors” (much like many here on HN) expected “AI” to be magic thing and are dealing with some disappointment that most of us are still employed. the next stage of “investor sentiment” just may be “shoot, not magic but productivity is through the roof”
Where are the hard numbers? Number of games on Steam, new GitHub projects, new products released, GDP growth—anything.
since you referenced a trusted Economist here’s much-more-we-know-what-we-are-talking-about MIT saying 12% of workforce is replaceable by AI (I think this is too low) - https://iceberg.mit.edu/
So far no verifiable metrics show any hint of a 3x productivity boost.
I am actually hoping someone there studies such interventions the way they did with CMU's intelligent tutor — which if I recall correctly did not have net strong evidence in its favors as far as educational outcomes per the reports in WWC — given the fall in grade level scores in math and reading since 2015/16 across multiple grades in middle school. It is vital to know if any of these things help kids succeed.
This is basically what would be expected. However n=20 is too small. This needs to be replicated with x10 the n.
I surmise that would help people learn to code better.
The world worked perfectly before 2023, there is no need to outsource information retrieval or thinking.
(…I actually kind of think this. "Kind of" being the key word.)
Speaking as someone that communicates primarily through text (high likelihood of Autism) the internet was the first chance a lot of us had to ... speak.. and be heard
People have a need to be heard and understood. That’s half of what we are doing here posting.
Many (“not disabled”) people don’t fit in with their local peer group / society. The internet gave them a way to connect with other like-minded individuals.
Do I need to give examples? Let’s say: struggling with a rare disease.
Perhaps some of that violence would have happened anyway. I don't know how it all nets out.
We often fall short, but as a society we do try to make sure we're accommodating disabled people when we make big changes in our systems.
There are far, far too many people who genuinely think disabled people should just disappear or die for it to be "safe" to be facetious about that without a clear sarcasm indicator.
I see this mentality almost exclusively in americans and/or anglo people in general, it's incredible... if you're not that, I guess you're just too young or completely isolated from reality and I wish you the best in the ongoing western collapse.
(... I actually wish you're joking and I didn't catch it, though).
It is so much harder now. There are people who are willfully ignorant now, almost proud to be; snooty about it. But it's impossible for governments and institutions to lie like they used to be able to. People are trading primary source documents online within the day.
It's why the popularity of long-ruling institutional parties is dropping everywhere, and why the measures to stop people from communicating and to monitor what they're saying are becoming more and more draconian and desperate.
and you cannot simply hand-wave away the massive acceleration of the surveillance state and characterize it as a tool of the “institutional parties”
Because arithmetic itself, by definition, is.
Human language is not. Which is why being able to talk to our computers in natural language (and have them understand us and talk back) now is nothing short of science fiction come true.
LLMs are wrong infinitely more than calculators, because calculators are never wrong (unless they're broken).
If you input "1 + 3" into your calculator and get "4", but you actually wanted to know the answer to "1 + 2", the calculator wasn't "wrong". It gave you the answer to the question you asked.
Now you might say "but that's what's happening with LLMs too! It gave you the wrong answer because you didn't ask the question right!" But an LLM isn't an all-seeing oracle. It can only interpolate between points in its training data. And if the correct answer isn't in its training data, then no amount of "using it with care" will produce the correct answer.
They also don't use it at all anymore, they barely even care about your search query.
Google is successful, however, because they innovated once, and got enough money together as a result to buy Doubleclick. Combining their one innovation with the ad company they bought enabled them to buy other companies.
I was allowed to use a calculator from middle school onward, when we were being tested on algebra and beyond and not arithmetic.
Some schools have ridiculous policies. Some don’t. Ymmv. I don’t think that’s changed from when I was in school.
But unless I practically apply what I learned, my retention is quite low.
As someone studying CS/ML this is dead on but I don't think the side-effects of this are discussed enough. Frankly, cheating has never been more incentivized and it's breaking the higher education system (at least that's my experience, things might be different at the top tier schools).
Just about every STEM class I've taken has had some kind of curve. Sometimes individual assignments are curved, sometimes the final grade, sometimes the curve isn't a curve but some sort of extra credit. Ideally it should be feasible to score 100% in a class but I think this actually takes a shocking amount of resources. In reality, professors have research or jobs to attend to and same with the students. Ideally there are sections and office hours and the professor is deeply conscious of giving out assignments that faithfully represent what students might be tested on. But often this isn't the case. The school can only afford two hours of TA time a week, the professors have obligations to research and work, the students have the same. And so historically the curve has been there to make up for the discrepancy between ideals and reality. It's there to make sure that great students get the grades that they deserve.
LLMs have turned the curve on its head.
When cheating was hard the curve was largely successful. The great students got great grades, the good students got good grades, those that were struggling usually managed a C+/B-, and those that were checked out or not putting in the time failed. The folks who cheated tended to be the struggling students but, because cheating wasn't that effective, maybe they went from a failing grade to just passing the class. A classic example is sneaking identities into a calculus test. Sure it helps if you don't know the identities but not knowing the identities is a great sign that you didn't practice enough. Without that practice they still tend to do poorly on the test.
But now cheating is easy and, I think it should change the way we look at grades. This semester, not one of my classes is curved because there is always someone who gets a 100%. Coincidentally, that person is never who you would expect. The students who attend every class, ask questions, go to office hours, and do their assignments without LLMs tend to score in B+/A- range on tests and quizzes. The folks who set the curve on those assignments tend to only show up for tests and quizzes and then sit in the far back corners when they do. Just about every test I take now, there's a mad competition for those back desks. Some classes people just dispense with the desk and take a chair to the back of the room.
Every one of the great students I know is murdering themselves to try to stay in the B+/A- range.
A common refrain when people talk about this is "cheaters only cheat themselves" and while I think has historically been mostly true, I think it's bullshit now. Cheating is just too easy, the folks who care are losing the arms race. My most impressive peers are struggling to get past the first round of interviews. Meanwhile, the folks who don't show up to class and casually get perfect scores are also getting perfect scores on the online assessments. Almost all the competent people I know are getting squeezed out of the pipeline before they can compete on level-footing.
We've created a system that massively incentivizes cheating and then invented the ultimate cheating tool. A 4.0 and a good score on an online assessment used to be a great signal that someone was competent. I think these next few years, until universities and hiring teams adapt to LLMs, we're going to start seeing perfect scores as a red flag.
Individual instructors should do something about it, even.
The fact that there is no feedback loop causing instructors to do this is a real problem.
If there were ever a stats page showing results in your compilers course were uncorrelated with understanding of compilers on a proctored exit exam you bet people would change or be fired.
So in a way, I blame the poor response on the systematic factors.
They had opted out of the lectures, believing that they were inefficient or ineffective (or just poorly scheduled). Not everyone learns best in a lecture format. And not everyone is starting with the same level of knowledge of the topic.
Also:
> A 4.0 and a good score on an online assessment used to be a great signal that someone was competent
... this has never been true in my experience, as a student or hiring manager.
For many classes this is still the case, and I lump these folks in with the great students. They still care about learning the material.
My experience has been that these students are super common in required undergrad classes and not at all common in the graduate-level electives that I’ve seen this happening in.
> ... this has never been true in my experience, as a student or hiring manager.
Good to know. What’ve you focused on when you’re hiring?