I held out for a long time, but I finally started using LLMs in my regular coding work. I couldn't deny the utility of tools like Claude any longer, when it can do in 2 minutes what would take me an hour.
There are problems: hallucinations, verbosity, bad/unperformant/inaccessible code. I have to correct it a lot. That's OK.
The main problem I find is that it's sucked a lot of the fun out of coding for me. It's like I've been pushed into a management role when I just wanted to stay a coder.
Nolan Lawson
Als Antwort auf Nolan Lawson • • •The other issue of course is that I'm not learning as much when I use LLMs. Even if it's an iterative, back-and-forth process, the tool is doing ~80% of the thinking for me.
It feels like I either need to spend more time learning outside of coding, or just accept at some level that I'm "cashing in my chips" and relying on ~20 years of actual coding experience.
Either way, I feel less excited by these tools than defeated. They're incredible magic wands, but I kind of liked doing my own sorcery?
Nolan Lawson
Als Antwort auf Nolan Lawson • • •Of course, this is a choice: I *could* choose not to use LLMs. When I ride my bike, I don't bemoan the fact that a car could do it faster – the goal is exercise.
But nobody's paying me to ride my bike. If I were a delivery driver, it'd be pretty unprofessional to show up an hour late with a cold pizza just because I like biking.
With LLMs, it almost feels like malpractice *not* to use them at this point. I can't justify taking ~3x longer to ship a feature just because I don't enjoy using them.
Nolan Lawson
Als Antwort auf Nolan Lawson • • •I'm familiar with all the arguments against GenAI, and I'm a big fan of authors like @baldur whose excellent book, The Intelligence Illusion, I've read twice. It's highly recommended. illusion.baldurbjarnason.com/
I have a degree in computational linguistics, and that's partly why I was so skeptical of these tools for so long. I still am! But more and more I felt like the world was moving on without me. To this day I feel deeply conflicted.
The Intelligence Illusion (Second Edition): Why generative models are bad for business
illusion.baldurbjarnason.comJonathan Schofield
Als Antwort auf Nolan Lawson • • •Antonio Sarcevic
Als Antwort auf Nolan Lawson • • •Matt Campbell
Als Antwort auf Nolan Lawson • • •> With LLMs, it almost feels like malpractice *not* to use them at this point.
Are you sure that feeling isn't just a result of the constant drumbeat of high-profile people talking about using LLMs? Might it be better to just ignore them and stick to your guns? Maybe you'll take 3x longer to do certain kinds of shallow coding tasks, but maybe that will be more than compensated for in overall quality.
William Pietri
Als Antwort auf Nolan Lawson • • •Ember
Als Antwort auf Nolan Lawson • • •Jared “Indie Social Web” White
Als Antwort auf Nolan Lawson • • •But the reason the car is "faster" is because of bad infrastructure—with good urbanism cycling is faster, especially e-cargo-bike. And then it's for utility, and exercise is simply a nice by-product.
And that's actually why this is an apt analogy: people can "design" for LLMs (aka car-centric design) or they can design for humans (good urbanism, car use is severely curtailed or possibly banned).
TinyExplosions 🐾
Als Antwort auf Nolan Lawson • • •Æ.
Als Antwort auf Nolan Lawson • • •if you’re spending so much of your time actually writing code that an LLM can significantly speed things up, then that’s where I’d start looking for inefficiencies.
In my time as a developer I rarely spent more than maybe 20% of my time actually coding, the rest was design and concept work. Typing those things out into code was not the main part at all.
Torb
Als Antwort auf Nolan Lawson • • •They *are* paying you to think about what you’re actually making. I think coding things yourself is a crucial step in thinking carefully about what you’re actually doing.
Of course, this assumes people care *at all* about software quality. Maybe they don’t. In which case, being good at our jobs doens’t matter.
But if it *does* matter that we actually make good stuft then I’m farily convinced that LLM will be a hindrence.
Francis 🏴☠️ Gulotta
Als Antwort auf Nolan Lawson • • •Artemis
Als Antwort auf Nolan Lawson • • •"almost feels like malpractice"
"taking ~3x longer to ship a feature"
Sounds a bit overly dramatic.
IMO:
1. If your programming is improved "3x" by an LLM then you've got too much architectural cruft. LLMs are primarily good at generating boilerplate, so why are you repeating so much boilerplate instead of abstracting it away?
2. Most "gains" that LLMs bring you during writing code will be lost during code review, debugging, etc.
jokeyrhyme
Als Antwort auf Nolan Lawson • • •I wonder what lessons we can take from other subject areas as software goes through the process of mass production and industrialization?
Will we have second-wave and third-wave software as in coffee? en.m.wikipedia.org/wiki/Third-…
Will software engineers (and other white-collar workers) sabotage LLMs as capitalism turns on us, as in textiles? en.m.wikipedia.org/wiki/Luddit…
organization of English workers in the 19th century protesting adoption of textile machinery
Contributors to Wikimedia projects (Wikimedia Foundation, Inc.)crow
Als Antwort auf Nolan Lawson • • •Su-Shee
Als Antwort auf Nolan Lawson • • •I'm completely torn between my conscience and what I can do with LLMs if.. I hold them right
Alex R
Als Antwort auf Nolan Lawson • • •I think this is the most distressing thing for me. Sure, I’ve got the experience to be able to potentially use LLMs and understand the output if needed, but I feel like I’m watching junior engineers be robbed of the same experience.
Some argue that high-level languages will become the new assembly language, but learning a bit about assembly is really useful in developing a good mental model for how a computer actually works. What happens once you have a cohort of engineers who don’t really know what a
for
loop is or why you’d use one.We’ll probably figure this out eventually, but I think a lot of people are going to be hurt in the process.
Sébastien Arnaud
Als Antwort auf Nolan Lawson • • •Matt Campbell
Als Antwort auf Nolan Lawson • • •> There are problems: hallucinations, verbosity, bad/unperformant/inaccessible code. I have to correct it a lot. That's OK.
So does that mean you have to read all the code, or do you just make changes when you or someone else notice a problem in the results?
sabik
Als Antwort auf Nolan Lawson • • •As a programmer also in a position to potentially start using LLMs, my other questions would be (a) is it still faster once you include the time to "correct it a lot", and (b) is the final code at least as good as without using it
I know the answer for (b) is "no" for generating ideas as English text, but I'm not sure anybody's tested it for code
staringatclouds
Als Antwort auf Nolan Lawson • • •The main issues I have with LLM's is accreditation & payment
Everything the LLM outputs is a statistically probable mash up of other people's work
The LLM has no understanding of its output, each subsequent word is simply the statistically most likely next word based on it's training data
Data which was stolen wholesale from the actual authors who aren't getting any credit or recompense for their work
You could even be one of them
But the company that owns the LLM makes millions
Richard W. Woodley RNKD BLTS 🇨🇦🌹🚴♂️📷 🗺️
Als Antwort auf Nolan Lawson • • •Jan Lehnardt
Als Antwort auf Nolan Lawson • • •Torb
Als Antwort auf Nolan Lawson • • •Adnan 🦙
Als Antwort auf Nolan Lawson • • •Francis 🏴☠️ Gulotta
Als Antwort auf Nolan Lawson • • •