Recent searches
Search options
What's amusing to me here is it's a dev comparing the "old approach" of learning via StackOverflow to learning via LLMs. My generation of programmers disrespected SO as dumbing down learning because so many devs were just copy/pasting the top answer instead of understanding it - it was faster so of course they did. LLMs are the same: worse but faster. So the article is not wrong, but it doesn't realise they're fighting a losing battle, one that we fought & lost already. https://nmn.gl/blog/ai-and-learning
There's also the possibility that every previous gen of developers is just wrong about how bad the next gen is at development. Maybe we were wrong about SO, and maybe they (and us) are wrong about LLMs. Fortunately I have no skin in the game anymore, having banished myself to the fringes and not having any students / juniors to train anymore. They'd probably just roll their eyes at the old man anyway 
@sinbad A close friend is a distinguished hardware engineer. There's always a nice bit of grounding in a chat with him about the ridiculous overhead of C.
With certainty someone somewhere is and have been ranting about similar observations of the context of his work. And they're probably technically right as well.
@sinbad When I was a junior developer, I was criticized for being “IDE dependent”. And for having to look up functions every time instead of having them memorized.
@sinbad I think he has a solid point.
Sure, lots of code-monkeys would just copy-paste from SO, as they did from other forums before it. Or they'd pester a senior for the answer, and transcribe it with the same disinterest in actually learning from it (I've been that senior). But those with the inclination had the opportunity to learn. Not just from reading the existing responses, but from asking their own clarifying questions, and standing a reasonable chance of getting an answer from somebody who understood the topic and the issue.
A large language-mangler doesn't even have a wrong understanding of what it gave you, and it's not even Searle's Chinese Room. Sure, you can ask it a question, but the answer to that question will itself be a word-collage with nothing deterministic about it. It's text-in-a-blender, all the way down.
I remember the days of Wikipedia being useful mainly as a way to discover what questions you needed to ask about the topic at hand.
LLMs will give you material for questions, alright, but there's no guarantee those will even take you in the right direction.
My hope is that enough of the LLM-transcribers will be made to clean up their own mess and, though the debugging process, come to understand both the medium of code and the fact that LLMs don't know anything.
@KatS Yeah - you're both right, my point was mostly that railing against it is kinda futile, trends gonna trend and someone will just have to deal with the fallout
@sinbad Absolutely.
In the meantime, it's like the way 'baby on board" stickers translate to "the road is probably the last thing this driver is paying attention to."
The use of LLM-generated code can at least be useful as a first-pass filter for candidates, and for evaluating forum responses.
I mean, you're right that we're all getting splattered with the fallout one way or another, but at least you can avoid having to mop up that mess on your own shop floor.
@sinbad Where's the version for the people who learned from 'The Really Big Bookcase of Manuals'? Nobody enjoyed inserting the update pages alone but we made it into a team event and called out any interesting changes we noticed. I didn't bother reading the installation notes last time I updated Linux. I trusted that other people I don't know would protect me, which is always a silly risk to take.
@sinbad I will say in defense of SO, that a lot of documentation is shit, and that's mostly what I've used it for. I could spend an hour searching for the specific term this particular library uses for a concept, and then another hour trying to parse it; or I could look it up on SO and get an understanding there and then dig around the documentation.
@sinbad My theory is that we grew up with machines that you could actually understand completely. A Speccy fits in a brain. A single person can know everything about it.
That stopped around the Pentium era. A single person, no matter how smart, stopped being able to understand everything about a machine.
Over time as the mountain of stuff you can never sensibly understand piled up, people just embraced it. Don't know, don't care, make it work somehow, ship it. Ah well.
@TomF @sinbad having grown up mostly with machines from the pentium era, the way I learned Computer Stuff was to pick a topic, find books on it, and learn whatever I could about that chunk of the computer. I don't think it's possible to know the whole machine, but it's possible to draw a little box around part of it and learn it's contents completely, and then draw another box next to that, and so on. I think I have a pretty good breadth and depth of understanding, but I'll always be learning.
@TomF @sinbad I think the main thing that changed is that things like stack overflow and LLMs are low friction resources that are just practical enough that curiosity, tenacity, and a lot of free time aren't prerequisites anymore for some areas of Computer Stuff (and other topics). I don't think there are any fewer tenacious curious people with the time and support to go further than that, but it might seem like that because there's a lot more programmers now.
@TomF @sinbad I like to think that having a lower barrier of entry is a good thing, because people can still be nerd sniped into becoming great programmers and may be motivated to do so when they reach the limitations of low friction learning resources, but I don't have any proof of that, and I've seen research to suggest the opposite may be true with LLMs so idk ymmv.
@aeva @TomF yeah you’re probably right - everyone has to go through the stage of doing it badly before they learn the whys and hows of how to build things well, perhaps it’s just the pathway there that is different. Although I do wonder about the erosion of what’s considered good enough, since management is always looking for a quick fix. That hasn’t changed but expectations of what acceptable quality vs speed/cost looks like may well have bad consequences
@sinbad @aeva Low barrier to entry is crucial. My generation's machines - you plug them in, you have BASIC starting at you. The next 2-3 generations - nah, you had to buy a $$$ package before you could program. Shocking difference in literacy.
But you could dig behind BASIC and get results. You could know - you could know it ALL. You could just keep digging, and actually reach the bottom.
Now... you know you can't possibly do that. There's no end. So maybe people don't even start?
@Scmbradley @sinbad @aeva Could do - I never got into that world, but yeah that's very plausible. It's also I think why writing new games on old machines (8-bit and 16bit) is so attractive. Especially with the fantastic emulators we have these days, you really can Know Everything.
@Scmbradley @TomF @sinbad i think rpi is in a similar ballpark of complexity to other modern PCs, but arduinos maybe are simple enough. or at least the old ones might be. idk about the new ones that are kitchen sink system-on-a-chips with stuff like bluetooth
@aeva @TomF @sinbad one nice difference (?) between then and now: tiny LEDs should be a lot cheaper now*, and this could potentially allow us to build circuits that give a lot more feedback about what's happening in any machine you build for this purpose.
also: you can design a custom PCB and have it fab'd and shipped here in a relatively short time now.
[*] but with the recently self-imposed trade war, who knows
@JamesWidman @aeva @sinbad This looks simple enough to learn: https://hackaday.io/project/189704-the-l7-a-very-simple-8-bit-cpu
Also, if you don't need to learn the internals of the Z80, basically the whole of the rest of the Speccy is covered down to the gate level in this excellent book:
http://www.zxdesign.info/book/
@TomF @sinbad [note for people who didn't grow up in the UK in the 1980's: in this context, "speccy" refers to this:
https://en.wikipedia.org/wiki/ZX_Spectrum
...which is *not* among the first page of results that i see when i do a web search for "speccy". (that search doesn't even return hits about the ZX Spectrum emulator named "speccy".)]
@sinbad I think I was lucky to start out before SO & a glut of open source.
It's been easy to avoid putting in the work. I'm not sure if LLMs make it worse but it's unlikely to improve it.
Pre-LLMs, I've seen devs not wanting to write an exponential back-off function because there was a library for it.
It suits businesses & as long as there's no real repercussions for companies outputting low quality then the incentives aren't there. Hopefully new devs see the goal is to replace them entirely
@sinbad Mind you it will be hard to replace most developers but I think that's what businesses are most invested in seeing. So, I'm not sure I'd just assume it'll be fine in 10 years and effectively help the AI tools get better at programming problems.
@sinbad I'm still struggling with the author's cognitive dissonance on display (which is weird because the author should struggle with it instead)
- thinks "AI" makes things worse
- builds a tool that's going to make things worse
- thinks word salad machines can be "interrogated"
as for the battle, I don't think many fought it at all - I have but it's incredibly draining and in the end you only get to fire one person
requiring devs by law to understand what the code does might be easier