Hacker News Reader

Comment by fartfeatures | Hacker News Reader

I've always wanted to find out if this would work:

If I knew most of my students were using LLMs I would encourage the rest of them to use them too. I would then change the marking criteria such that if they get a single point wrong that's 20% off the entire grade. 5 mistakes gets you a big fat zero.

LLMs are great but they make mistakes. At this point your students would have had to spend so long checking, re-checking and triple checking the LLM's output that they will have accidentally learn what they need to. They will likely need to cross-reference multiple LLMs and at least have read their output which is likely an upgrade on today.

Or some variation of the above, I'd be interested to hear your thoughts.

View on HN

This sounds like it would come at the cost of the honest students who don't want to use LLM's and train their own understanding.

Reviewing is just not the same as working through a problem yourself. You often can take shortcuts when checking answers for correctness, which generation (with mind or LLM) of the solution can't take. This isn't limited to Math but equally true for programming and many other skills.

I'd push back slightly on this: I spend more of my time reading / reviewing code than writing it yet I still find reading / reviewing code much more mentally demanding. I don't think I'm alone in that.

The people that do their own work would be at a huge advantage as the LLM(s) can work as reviewers instead of doing the work so there is more chance of catching a mistake before you lose 20%.

On top of that there will be some issues where the LLM is just plain wrong. Being able to work out when that is the case is a hugely useful skill in 2026. It is also something the people who blindly feed their work into an LLM and print its answer without reading it will be unable to discover. Much to their own detriment.

I'm not trying to suggest that reviewing (code) can't feel (more) mentally demanding. The thing with at least reviews is that if you have worked through the problem yourself, you will also hit certain walls where stuff didn't work or wasn't as beautiful as you wanted (ie trial and error), and you have to activate your brain to think of other solutions, ie really think a problem through multiple times with active feedback. As a reviewer you don't get the active feedback usually.

Now this i think applies in lesser degree to the comment I was replying to. With maths, and some comp sci problems, you often can check an answer much faster than actually solving it, eg by using simple substitution for simple math problems. That is a useful skill, but a different one from solving the problem yourself. It employs different types of skills.

I think you might be letting perfect be the enemy of good here. Right now some students are using an LLM to do all of their work. I don't suggest for a moment my method is perfect but I do claim it could be better than the status quo (or at least warrants further investigation in my opinion).

> I spend more of my time reading / reviewing code than writing it yet I still find reading / reviewing code much more mentally demanding. I don't think I'm alone in that.

That's because we write much smaller quantities of code than we read. Not that LoC is a useful metric, but I'd be surprised if I write more than a few hundred lines a code a day in a legacy codebase. Most of the time is spent planning and deliberating over what to write, and perhaps revising code a few times. Meanwhile, understanding what a snippet of code does can easily spiral into having you read thousands of lines of code across several modules.

But that's for work. writing is a stronger learning tool than reading, and we needed to write tens of thousands of lines of code before we got into the door and started reading more than we wrote. It's still preferable for students to do the same and write as much for their homework as possible.

Explaining that paradigm shift from learning to working isn't trivial, and is often poorly done. But it's needed if we ever want people to frame these tools as productivity boosters, and not shortcuts through tedium. Learning by its nature involves some tedium.

What you would grade with that approach is how well your students can operate a LLM as well as the failure rate of whatever LLM infrastructure the individual student is using. This is usually not the goal of your educational format and hence not what you would grade for - apart from some (meta-)skill courses offered by uni libraries and the like. Those are mostly ungraded, though.

Anyone that knows the subject matter well enough will spot the LLMs mistakes and correct them. If you don't know the subject matter well enough to spot a mistake then you don't know the subject matter well enough.

Reply on news.ycombinator.com