Let’s be real here: when people hear the word AI or LLM they don’t think of any of the applications of ML that you might slap the label “potentially useful” on (notwithstanding the fact that many of them also are in a all-that-glitters-is-not-gold–kinda situation). The first thing that comes to mind for almost everyone is shitty autoplag like ChatGPT which is also what the author explicitly mentions.
I’m saying ChatGPT is not useless.
I’m a senior software engineer and I make use of it several times a week either directly or via things built on top of it. Yes you can’t trust it will be perfect, but I can’t trust a junior engineer to be perfect either—code review is something I’ve done long before AI and will continue to do long into the future.
I empirically work quicker with it than without and the engineers I know who are still avoiding it work noticeably slower. If it was useless this would not be the case.
I’m a senior software engineer
ah, a señor software engineer. excusé-moi monsoir, let me back up and try once more to respect your opinion
uh, wait:
but I can’t trust a junior engineer to be perfect either
whoops no, sorry, can’t do it.
jesus fuck I hope the poor bastards that are under you find some other place real soon, you sound like a godawful leader
and the engineers I know who are still avoiding it work noticeably slower
yep yep! as we all know, velocity is all that matters! crank that handle, produce those features! the factory must flow!!
fucking christ almighty. step away from the keyboard. go become a logger instead. your opinions (and/or the shit you’re saying) is a big part of everything that’s wrong with industry.
let me back up and try once more to respect your opinion
The point of me saying that was to imply I’ve been in the industry for a couple of decades, and have a good amount of experience from before all this. It wasn’t any kind of appeal to authority, but I can see how you can read it that way.
jesus fuck I hope the poor bastards that under you find some other place real soon, you sound like a godawful leader
I’m sorry, do you trust junior engineers blindly? That’s gonna lead to a much worse outcome than if they get feedback when they do something wrong. Frankly, I don’t trust any engineer to be perfect, we’re humans and humans make mistakes, that’s why we do code review as a fundamental skill in this industry. It’s one of the primary ways for people to develop their ability.
yep yep! as we all know, velocity is all that matters! crank that handle, produce those features! the factory must flow!!
In an industry where many companies are tightening the belt, yes it’s important to perform well—I kinda want to keep my job and ideally get a good bonus. It would be pretty foolish to leave free productivity on the table when the alternative is working harder to bridge the gap, where I could spend that energy doing more productive stuff.
and the engineers I know who are still avoiding it work noticeably slower
yep yep! as we all know, velocity is all that matters! crank that handle, produce those features! the factory must flow!!
and you fucking know what? it’s not even just me being a snide motherfucker, this rant is literally fucking supported by data:
The survey found that 75.9% of respondents (of roughly 3,000* people surveyed) are relying on AI for at least part of their job responsibilities, with code writing, summarizing information, code explanation, code optimization, and documentation taking the top five types of tasks that rely on AI assistance. Furthermore, 75% of respondents reported productivity gains from using AI.
…
As we just discussed in the above findings, roughly 75% of people report using AI as part of their jobs and report that AI makes them more productive.
And yet, in this same survey we get these findings:
if AI adoption increases by 25%, time spent doing valuable work is estimated to decrease 2.6% if AI adoption increases by 25%, estimated throughput delivery is expected to decrease by 1.5% if AI adoption increases by 25%, estimated delivery stability is expected to decrease by 7.2%
and that’s a report sponsored and managed right from the fucking lying cloud company, no less. a report they sponsor, run, manage, and publish is openly admitting this shit. that is how much this shit doesn’t fucking work the way you sell it to be doing.
but no, we should trust your driveby bullshit. motherfucker.
In this and other use cases I call it a pretty effective search engine, instead of scrolling through stackexchange after clicking between google ads, you get the cleaned up example code you needed. Not a Chat with any intelligence though.
“despite the many people who have shown time and time and time again that it definitely does not do fine detail well and will often present shit that just 10000% was not in the source material, I still believe that it is right all the time and gives me perfectly clean code. it is them, not I, that are the rubes”
That ChatGPT can be more useful than a web search is really more indicative of how bad the web has got, and can only get worse as fake text invades it. It’s not actually better than a functional search engine and a functional web, but the companies making these things have no interest in the web being usable. Pretty depressing.
I’m a senior software engineer
Nice, me too, and whenever some tech-brained C-suite bozo tries to mansplain to me why LLMs will make me more efficient, I smile, nod politely, and move on, because at this point I don’t think I can make the case that pasting AI slop into prod is objectively a worse idea than pasting Stack Overflow answers into prod.
At the end of the day, if I want to insert a snippet (which I don’t have to double-check, mind you), auto-format my code, or organize my imports, which are all things I might use ChatGPT for if I didn’t mind all the other baggage that comes along with it, Emacs (or Vim, if you swing that way) does this just fine and has done so for over 20 years.
I empirically work quicker with it than without and the engineers I know who are still avoiding it work noticeably slower.
If LOC/min or a similar metric is used to measure efficiency at your company, I am genuinely sorry.
I agree with you on the examples listed, there are much better tools than an LLM for that. And I agree no one should be copy and pasting without consideration, that’s a misuse of these tools.
I’d say my main uses are kicking off a new test suite (obviously you need to go and check the assertions are what you expect, but it’s usually about 95% there) which has gone from a decent percentage of the work for a feature down to an almost negligible amount of time. This one also results in me enjoying my job a bit more now too as I’ve always found writing tests a bit of a drudgery.
The other big use for me is that my organisation is pretty big and has a hefty amount of code (a good couple of thousand repos at least), we have a tool that’s based on GPT which has processed all the code, so you can now ask queries about internal stuff that may not be well documented or particularly obvious. This one saves a load of time because I now don’t always have to do the Slack merry go round to try and find an engineer that knows about what I’m looking for—sometimes it’s still unavoidable, but they’re less frequent moments now.
If LOC/min or a similar metric is used to measure efficiency at your company, I am genuinely sorry.
It’s tied to OKR completion, which is generally based around delivery. If you deliver more feature work, it generally means your team’s scores will be higher and assuming your manager is aware of your contributions, that translates to a bigger bonus. It’s more of a carrot than a stick situation IMO, I could work less hard if I didn’t want the extra money.
Senior software engineer programmer here. I have had to tell coworkers “don’t trust anything chat-gpt tells you about text encoding” after it made something up about text encoding.
ah but did you tell them in CP437 or something fancy (like any text encoding after 1996)? 🤨🤨🥹