Eniko Fox6d ago

agentic AI in particular is so fucking funny. i run an absolutely *tiny* indie studio and i still ask people "hey could you run me through how this works?" all the time because knowing how things work is a vital part of creating a quality product

how does that work with AI? "hey could you run me through how this works?" and people just go "idk the AI did it ¯\_(ツ)_/¯"

Show threadEniko Fox6d ago
i feel like the only way businesses fall for this is when they're big enough nobody at any level really knows fully how the product gets made, because it's abundantly clear to anyone who actually knows how products are made that "nobody knows how this works" is the biggest red flag ever
Show threadEniko Fox

"you can just read the code the agent wrote"

oh fuck off. the whole idea is that agents can churn out code at way higher volumes than people can generate, and the bottleneck when people wrote the code and not "agents" was already code review, because making sense of code is harder than writing it

the only thing you've done is made the code review bottleneck so, so much worse. and this will help you be more productive... how exactly?

May 11 at 10:00amWeb
Show threadEniko Fox6d ago
so what's the alternative if you can't review the avalanche of slop code? you just don't. and that's basically akin to live coding in the production environment. anyone who attempts this for long enough will be punished for their hubris
Show threadThe Seven Voyages Of Steve6d ago
@eniko I absolutely cannot understand anyone who would be happy to ship things they don't understand, or that they can ask a person they trust about if they don't. I get the argument that as a sole coder (now) I'm an outlier and teams delegate understanding between themselves all the time, but a team is different to an LLM. A person can earn your trust, an LLM can only delude you into trusting it because it has no real memory or integrity or anything to lose by screwing you
Show threadmirek kratochvil6d ago
@sinbad @eniko
you're not an outlier
(except if like 50% of the foss world is outliers)
Show threadJames Thomson6d ago
@eniko I have seen the argument that they will just get a newer, smarter, AI to fix all the problems generated by the old one, and it's giving “I speak of none but the computer that is to come after me” from Hitchhikers.
Show threadPeter6d ago
@jamesthomson @eniko It's also the exact same bullshit people have been saying about climate collapse for the past twenty years. "Oh, we will just invent something that takes out all the CO2 afterwards. We don't have to worry about it now! Growth can continue forever!!"
Show threadDCoder 🇱🇹❤🇺🇦6d ago
@jamesthomson @eniko
My coworker already uses one agent to write code and another agent to code review it 🫠
Show threadAbyssal Rook6d ago

@eniko I'm admittedly limited in my coding experience, but I'm less worried by code generation than by error fixing.

I vaguely trust the first draft to at least sorta focus on the original intent. Every iteration where code fails to compile or gives an incorrect output, though, creates a new layer of problems to focus on, and if an LLM will do ANYTHING to look good to you (as seems to be the trend), then who knows what bullshit it'll put in there just to produce something that functions?

Show threadTony6d ago
@eniko one answer you'll get is that as long as it includes complete code coverage in tests, it should be good.
But here's the thing: the agent wrote those as well - without context of the bigger system - and bugs can and have manifested in code that was 100% covered in tests
Show thread0xC0DEC0DE07EA6d ago
@longhairmoto @eniko also perverse incentives: to increase code coverage, we removed this error handling. Number go up!
Show threadEniko Fox6d ago
@longhairmoto yeah, bugs have never existed even in codebases with good test harnesses and they definitely have never happened in codebases with bad to mediocre test harnesses
Show threadozeng6d ago
@eniko @javierg immediate rejection if it looks like vibe code. Low-effort submissions get low-effort reviews.
Show threadGabriele Svelto5d ago
@eniko also cherry on top: I am convinced that all this focus on code is because it was one of the few datasets on the internet that remained unpolluted by LLM outputs. Polluting it with slop pretty much guarantees that whatever the current models are capable of doing, the future ones will not be able to do anymore.
Show threadGabriele Svelto5d ago
@eniko what I also find absurd in this situation is that I've always used relied on the fact that if you're trying to solve a problem there's a big, big chance that someone already solved it better than how I'd solve it. And that means there's a library or a tool somewhere that does what I need. And finding it and learning it means I need to write very little code, and I don't need to maintain that code later, because somebody else is maintaining it. Deliberately writing a lot of code is dumb.
Show threadMachine Lord Zero6d ago
@eniko "I produced 2 million lines of code so I am clearly doing my job brilliantly and should be the one promoted yes yes"
Show threadDaniël Franke 6d ago
@eniko Let's not kid ourselves, this AI code will mostly be reviewed by AI tools, because who cares if it breaks, it's not their fault anymore, it's the AI's fault.
Show threadEniko Fox6d ago
@ainmosni this is akin to live coding changes in the production environment and anyone who attempts it for long enough will be punished for their hubris
Show threadRatsnake6d ago
@ainmosni @eniko this is already how some teams at my company operate
Show threadRatsnake6d ago
@ainmosni @eniko our agentic pilot projects actively brag about how no human sees the code anymore.
Show threadEniko Fox6d ago
@ratsnakegames @ainmosni they will be punished for their hubris
Show threadRatsnake6d ago
@eniko @ainmosni and i'm gonna be collateral damage, unfortunately
Show threadEniko Fox6d ago
@ratsnakegames @ainmosni at least your brain will still work? silver linings i guess
Show threadMarta6d ago

@eniko
I need a rubber stamp with this sentence to stamp it on peoples faces.

"Writing code is the easy part. Reading code is the hard part."

Show threadHP van Braam6d ago

@eniko exactly this!

I've been trying to explain this to people as well.

But it seems "we have the ai write tests too, that we also don't read" is apparently fine.

Show threadBogusMeatFactory6d ago

@eniko preach! This is the biggest talking point for me. When stuff falls apart, you can absolutely ask a person "Hey, why is your code like this?" And you can get some idea as to what the problem is.

With AI, you can't. It doesn't have context behind its decisions so you not only have to find out what's wrong, you have to figure out why and do so with zero guidance or input.

Show threadMichael Ormsby6d ago

@eniko I’m retired now, but my career was writing code for about three and a half decades. Absolutely agree that documenting code, making sense of code, and understanding the pattern of interaction between different developer’s code is the hard part. And planning for reuse.

In my experience managers of software development almost never got that. Invariably they saw code as some kind of monolithic, homogeneous product that could be mass produced by the cheapest supplier, measured and sliced.

Show threadPhil Haigh6d ago

@eniko and we all know how much developers like doing code reviews, right?

Turn my job into one never ending code review? I’d rather quit, to be honest.

Show threadJae6d ago

@eniko a lot of places have started use AI to do code reviews, and started to shift around who takes fault when the AI generated code goes bad (either the person who did the prompting, or the person who did the review).

So something goes wrong and one of those two gets fired, and the process that incentivizes rubber stamping PRs is never given scrutiny.