I’m hearing this more and more from both students and teachers. It seems like everyone in education is frustrated and can’t figure out how to use AI for education productively yet.
TBH I’m also thinking about how to use AI at work for learning. (My work environment is 100% all-in on AI.) Everyone can see and admit that AI coding is making us lazier and that it produces too much code to do old-style line by line reviews. Some people are using agent teams to churn out millions of lines of code where there’s zero hope of reviewing it. We don’t yet know how to keep quality high and complexity low while doing agent coding, we’re just having them write lots of tests and hoping that when things break they’re easy to fix. Agents should get better at broad architecture in the future, but they’re definitely not there yet.
> The amount of work involved has also tripled
This has to be one of the big ironies of using AI. I’ve heard this from other students too. The problem is that old-style assignments where the level of difficulty was hand-crafted is way way too easy for an LLM, therefore teachers need to crank the level of difficulty so the student has to actually spend some time on the assignment. But LLMs are getting better every month and the difficulty is so high that no students can possibly do the assignment by hand and now small errors in assignment design can lead to large errors in effort and time required.
> It seems like everyone in education is frustrated and can’t figure out how to use AI for education productively yet.
Oh, it's terrible. I'm a computer science instructor at a university (I have 20 years of industry experience prior) and it was already bad enough before trying to predict where the market was going to be 4 years in the future. But a year ago, it was tough to predict where the market was going to be in one month.
So yeah, we can teach all kinds of stuff. But is it the right stuff? With one month of visibility, who can tell?
But now I feel like it's settled down a little bit. And now I feel like I have more like 7 to 8 months of visibility. One thing that's helped is that we see the hiring differential between seniors and juniors. So now we can answer the question, what does the market want? It wants the expertise between junior and senior. And I'm confident it'll want that for at least the next 7-8 months.
Good luck cramming that into 4 years, though! I figure we have to try to bridge the gap as best we can. Eventually, seniors will become rare enough that the expense isn't worth it. It'll become better to hire cheaper most-qualified juniors.
There's still some seriously open-ended questions though. How much does coding matter, for example? If you ask senior devs who don't do any coding how much they rely on their coding expertise to make decisions, you'll get a zillion answers across the spectrum. This becomes even more true for senior devs who use AI for coding. Heck, ask 100 devs if it's important that you can even read code or not and see what comes out!
And then there's the question of how do you teach those systems level thinking skills? Historically, universities haven't taught those skills. You came out of the University with general problem solving skills in technology, and then at your job you were exposed to a small corner of the codebase to fix some bug. Gradually, over the years, you learn more and more of the codebase and become more and more senior. We have to get that into our curriculum. I'm working on that as we speak and will be working on that for the next several years, it looks like; even if my position vanishes, I'll keep attacking it—this is a tough problem that's fun to try to crack.
Anyway, I hope someday my forward visibility gets back up to 18 months. Won't that be amazing. I'd also like to buy some RAM. ;)
> We don’t yet know how to keep quality high and complexity low while doing agent coding, we’re just having them write lots of tests and hoping that when things break they’re easy to fix
Implement more versions with different languages, frameworks, libraries and AI agents. Then you can do differential execution for testing, which is easier than imagining testing scenarios from pure code. You can fuzz or use production data (shadow mode execution) to get notified early of divergences.
Keeping a population of diverse replicas might be the simplest way to derisk AI code. When 2 replicas diverge an agent can decide - was it a bug in one of them, or was the spec unclear? Either way, if you fix a bug or clarify the spec you come ahead.
A separate class of problems is related to intent divergence. For that my solution is to keep a full log of all user messages during the lifetime of the project. Even 10,000 messages collected over weeks and months can be read and summarized in a few minutes. That means you can check for intent divergence with agents, but you always start from the raw source (your human messages) not from intermediate summaries and diverged code.
Combining the 2 ideas would make each replica agent derive the user intent separately from the same log.
This is a nice idea, but I’m in a big enough and low-level enough code base that using other languages or stacks isn’t very viable; we build the stack, so it’s software and hardware and compiler and driver. We do sometimes use multiple agents and adversarial review. The agents tend to come up with more complete corner-case testing than humans, the most common issue is that they don’t fully understand all design assumptions and parameters and they over-test for things that can’t happen using very complicated test setups, and they don’t test enough for simple mainline usage; they generally don’t know what the common workflows are. I had the same complaint about humans before LLMs too. :P We often try to be clever about testing and design all this logical coverage while ignoring the boring & easy stuff, we think like engineers and not users.
I taught my Claude to teach me new skills, where I wanted to understand its work deeper. Specifically, I wanted to learn coccinelle semantic patching, so I told it to find potential places to apply it but withhold the solutions from me. Instead, it walks me through producing a patch myself.
Yeah I’ve been meaning to try this more seriously, to have the agent be teacher and give me hints and force me to do the typing. It’s somewhat difficult to both find the time to do this at work, and also to have the self-control. I’ve done some half-hearted attempts, but when it gets to the repetitive and easy to understand parts, I start to slide back to letting the agent do it, and then sooner or later it’s handling trickier things again.
The other thing I want to try more at work is, rather than programmers at desks individually talking to agents, do pair or group programming with an agent in the room where the people discuss the design while the agents listens, asks questions, and writes the code live. We’ve had a few meetings where we do this, and it feels much better than individual vibe coding, because as a group we review the design as we go and prevent going down paths I might have gone down by myself. It combines the coding and review into a single session and ensures everyone present is comfortable with the design.
In that, 500 person frosh lectures are dead. 32 students per class, 6 classes a day are dead.
What remains? The Oxbridge model that has been going on for nearly a millennia.
Briefly: 4-10 students and a proctor. You do the reading/work, you meet a few times a week. The proctor facilitates discussion and guides students. Many schools do this, but not many students get to see it.
Why? Because it is very expensive. The proctor has to be at least a graduate in the same subject, likely having a graduate degree in it. These are small classes by design. And don;t even get started on the training that the proctors have to do to get ready and not be a lawsuit generation machine.
I feel this is really the only way forward. Which is almost maximally unfortunate.
Yes, 'Cheap' Big State University will still exist, but it's program will be so degraded by AI that it's graduates will be seen as worthless. While 'Very Expensive' Small Liberal Arts College with the Oxbridge model will guarantee that the students are at least minimally educated to bullshit well enough. Meaning that the rich and poor gap grows and grows.
> My understanding is that those who are, don’t use words as crude as “made” to describe a role in the making of a film.
Oh, no, film insiders use ‘made’ constantly, or even cruder - attribute credit without using a verb at all. Saying it’s a “Nolan film” is super common. Here’s a VFX person saying “Nolan’s CGI”: https://youtu.be/AFLcP-t0-kI (I’m ex-film industry, and still film industry tangential, and found this clip being discussed on a film industry mailing list.)
Film people are far more likely to pay attention to who the director is than non-film people, and it’s very common for film people to talk in terms of directors ‘making’ the film or otherwise casually assign ownership. I think this is especially true for famous directors and good films, and maybe a bit less true for unknown directors and mediocre films.
I’m sure exponential is what parent meant, but geometric (edit: polynomial) is what actually happens, not exponential. Code complexity increases linearly when it doesn’t interact with other features, and geometrically when it does. In the worst case, complexity increases as the square of lines of code. That is still very, very bad, but it’s not exponential.
Everything in the second group eventually grows faster than everything in the first group, no matter how small and big you make the fixed numbers, so they are two different classes.
Geometric is exponential; I goofed above. Geometric and exponential are sometimes used interchangeably, but sometimes people use geometric to refer to discrete intervals and exponential to refer to continuous functions.
Re: throughput vs latency - what is the implication for having a signed floor divide sprinkled in with other math? Is Robison likely to be the better choice unless I’m doing divide operations on long vectors/arrays of numbers?
If the goal is wide vector divides, what about GPU? How does this algorithm compare on GPU? Are GPU uses less likely to rely on and benefit from throughput behavior if the vector of divides is spread across all threads?
I don’t do image gen regularly… so I’d love more context. I can see some less good features in the images but I’m not fluent with what today’s generators would do. I agree with GP that it would be quite useful to know about the prompts & iteration effort. If generation times changed, they could be relevant & helpful.
> good luck predicting where I’m going to have my next cuppa.
Why are you assuming there isn’t someone else trying to hack you at the next cafe? Actual damage can happen one person at a time, and can occasionally lead to larger scale compromises. Randomizing your location isn’t going to reduce your risk. Now that everyone who’s paying attention knows they can MITM your Mac, you shouldn’t trust any network.
Unfortunately, the goal with OpengGL is not to teach graphics and rendering concepts. Same goes for DirectX, Vulkan, WebGL, etc. These APIs are being designed primarily for performance, not for ease of learning. For teaching and learning, it’s a great idea to start with rasterization & ray tracing in software first! That can teach you the graphics concepts first, and also provide good motivation to learn a graphics API if you want it to go faster.
Vulkan, OpenGL and DirectX make GPU programming more difficult than it has to be for multiple reasons, some of which other people here have mentioned. They are also low level, focused on high performance, and are primarily designed and written for game engines. Making them easy to learn for newcomers is pretty low on the priority list. ShaderToy makes it easy by taking care of everything except the GLSL, you might like it. You can also try CUDA, which is generally a lot easier than Vulkan. For a really easy/fun time with CUDA, try Warp - it’s CUDA in Python.
Interesting, had never heard of Kono Express. It definitely looks a little ‘Unreal’, has video game vibes. The quality is pretty good at a glance, especially the environments and effects, but it doesn’t feel all the way there under scrutiny. This reminds me of all AI generated stuff (so far) - looks great from a distance but from up close you see glaring problems. The lipsync feels automated and robotic all the way through well past ep3, for example. Very little action in the whole series.
TBH I think it’s a big mistake to compare this to hand-animated Disney material. They are just very different things, and the debatable comparison will just cause people to focus on what’s wrong with Kono. If you think it’s good, just talk about the good qualities and share it. Also don’t forget the AI is trained on Disney material, and with heavy AI, it’s highly questionable whether you can credit one person for “cranking out stuff”.
What is the ink & paint comment about? The article doesn’t mention ink & paint…
TBH I’m also thinking about how to use AI at work for learning. (My work environment is 100% all-in on AI.) Everyone can see and admit that AI coding is making us lazier and that it produces too much code to do old-style line by line reviews. Some people are using agent teams to churn out millions of lines of code where there’s zero hope of reviewing it. We don’t yet know how to keep quality high and complexity low while doing agent coding, we’re just having them write lots of tests and hoping that when things break they’re easy to fix. Agents should get better at broad architecture in the future, but they’re definitely not there yet.
> The amount of work involved has also tripled
This has to be one of the big ironies of using AI. I’ve heard this from other students too. The problem is that old-style assignments where the level of difficulty was hand-crafted is way way too easy for an LLM, therefore teachers need to crank the level of difficulty so the student has to actually spend some time on the assignment. But LLMs are getting better every month and the difficulty is so high that no students can possibly do the assignment by hand and now small errors in assignment design can lead to large errors in effort and time required.
reply