Mastodawn

I don’t really like supportive tooling for writing code. I work in vim and prefer bazel (with occasional support from a language server client for things like large scale renames).

I’m a reasonable person I swear to god. I know that’s just a preference, my coworkers who prefer IDEs or whatever are equally valid, and I try to help them where I can. If they say they can’t use generated files because they don’t autocomplete with their IDE, or if they want 4 repos instead of 1 because they can’t have multiple languages at once, I’m cool with that, really, I try to be accommodating, I try to help where I can.

that said, my coworker who writes his code with chatgpt is just not working out for me. I mean for one thing the code is just a mess: meaningless parameters passed everywhere, error checks that can never trigger because the implementations of the interfaces don’t throw the errors that are being checked.

but the main problem is that it’s not *designed* at all. every single one of these things is like Alton Brown used to say about unitaskers. this code passes the (generated) unit tests and that’s it. trying to use it for anything that it wasn’t explicitly and specifically designed and tested for is like trying to use a pork puller bear claw when what you want is a fork.

it’s not extensible, it’s not maintainable, if you want to add a parameter or tweak an implementation detail the only reasonable call is to rewrite it from scratch.

And what’s so frustrating about that is that it LOOKS fine at a glance. It even looks fine in code review, one chunk at a time!

I’m three hours into trying to edit in a ListFoo call to a client which already does GetFoo, CreateFoo, and DeleteFoo. all it does is delegate to a different, existing client which has all the CRUDL methods. Even so, the code to do this is seven hundred lines long. the tests are over a thousand lines.

it should have been a five minute task but nothing that I would need to be able to do this has been done. there’s no “convertFooToBar()”, there’s three slightly different implementations for the three existing methods - each reasonable but none possible to trivially consolidate, because the tests fail if you use the version in Get for the Create call. and then you have to figure out why, and there’s no *reason* it’s that way, it just *is*. there’s no *why*, just *how*.

I could leave it, I could write a fourth convertFooToBar(), but - come on, really? no I can’t. everyone knows that will be a huge problem someday even if we don’t know why or when. So now I’m 3.25 hours in because I needed a coffee break to bitch about this and it’s gonna take another 2. I coulda redone the whole thing from scratch in 5 and I feel like a chump for not noticing that an hour or two ago.

Deborah Hartmann Preuss, pcc 6d ago

@gnat so, your colleague is getting "more" work done and your work is slowing down ... 🙁 That's not going to play well, at all. #condolences

@deborahh ah, thanks for the concern! You’re very kind. It’s not gonna cause me a problem, though. I’m a senior swe, no one’s holding a stopwatch to my work. And I got this done yesterday; a one-day problem.

Deborah Hartmann Preuss, pcc 5d ago

@gnat I coach software folks, though, and suddenly, a year ago, everyone started worrying about their "performance".
I don't think it's unrelated 😢

("Performance" so often is a euphemism for random personnel reduction).

a wandering happenstance 3d ago

@gnat first of all, my condolences, it sounds excruciating.

I never have to work with this kind of generated code (and tests) so I have no way of guessing the answer to this question: how is the code coverage on the unit tests? It must be abysmal if there are a bunch of unused parameters and checks for error conditions that cannot occur …

@joXn funny point. I never bothered to check! It’s probably not much worse than my usual - 60ish % branch coverage, 75ish % line coverage. I usually consider that sufficient because when I try to go higher I wind up writing many tests proving the obvious.

It is meticulously commented but the comments just say what the code says

//GetFoo gets the Foo
func (c *client) GetFoo() (Foo, error) {

or

// Process the result
baz = processResult(bar)

It has perfect, idiomatic test coverage but the tests test that 2+2=4.

// assemble
MockResultProcessor.EXPECT().ProcessResult(gomock.Any()).Return(foo)

// act
res := processResult(bar)

// assert
assert.Equals(t, res, foo)

If you see a hundred lines at a time it hits the part of your code review brain that goes “ah yeah boilerplate ok sure”. when you see a thousand lines you go “ahhh fuck. this is nothing. this is trash.”

Flamekebab 6d ago

@gnat ugh, this irks me no end. I love comments but I don't need a comment to tell me that .ManyToMany() establishes a many to many relationship!

@gnat I think about this post from @ceejbot all the time when thinking about coding with AI: https://blog.ceejbot.com/posts/programming-as-theory-building/

AI cannot have a theory of the system or a point of view on the problem space, so it's actively detrimental to introduce into a codebase imho. It hinders the team's development of a theory of the system, which is the only long-term artifact that holds value.

Programming as Theory-Building – Ceejbot's notes

This internet thing seems to have taken off.

@gnat I'm not dogmatic about it. Some people use syntax-based auto complete and AI coding is basically that if you squint really hard. But God it's frustrating in a lot of cases.

GunChleoc 6d ago

@paddy @gnat What helps my productivity in an IDE is stuff like navigation help (e.g. finding all calls to a function or where a function is defined), function suggestions, autogeneration of boilerplate. Also, refactoring options like pulling out common code that I have selected into a separate function, because I like to clean things up.

All that was available and working before we had LLMs.

John Faithfull 🌍🇪🇺🏴󠁧󠁢󠁳󠁣󠁴󠁿🧡✊🏻✊🏿6d ago

@gunchleoc @paddy @gnat 👍 Indeed. Intellisense in Visual Foxpro - 1996/7??

@paddy oh man, yes! This post feels so good to read - like finding a word that’s been on the tip of your tongue for months. I really feel like you have succinctly said what I have been trying to express for months

@gnat I read it and was like "oh shit that's the thing I can't put my finger on that I'm trying to communicate here"

🐧sima🐧6d ago

@paddy @gnat @ceejbot my view is that you don't just need a theory of a system, but intuition. a gut feeling for which direction to start digging, so that you produce better than random attempts, which you can then validate against your theory of the system

as an example in the linux kernel, I judge locking designs by how they feel. it works really well, with the caveat that intuitive insights have no timeline for when they'll arrive. and sometimes that waiting is nerve-wracking

Tom Ritchford 6d ago

@sima @paddy @gnat @ceejbot

> my view is that you don't just need a theory of a system, but intuition. a gut feeling for which direction to start digging,

This is how I wrote programs in my teens. But that doesn't get you very far.

@TomSwirly @sima well apparently it gets you to some significant success writing locking designs in or around the Linux kernel, ha. not all that many higher echelons to reach!

🐧sima🐧5d ago

@gnat yeah, I've done enough locking design in the kernel that I've done a talk about locking design patterns http://blog.ffwll.ch/2023/07/eoss-prague-locking-engineering.html

links to my blog posts in there, too

or put differently, i barely chuckled seeing that toot you're replying to fly by 😏

EOSS Prague: Kernel Locking Engineering

EOSS in Prague was great, lots of hallway track, good talks, good food,excellent tea at meetea - first time I had proper teain my life, quite an experience. ...

Tom Ritchford 4d ago

@gnat @sima Having been involved in the review of a "lock-free queue" in C++, I'm extremely skeptical that intuition alone, without proving correctness, is going to get good results.

I do rely on intuition to at least write concurrent code, but that intuition is backed up by a pretty solid mental model of how the specific concurrency I'm using works.

I'll bet if I objected to some aspect of your locking, you'd have a very solid reason for me, not just intuition!

C J Silverio 2d ago

@sima @paddy @gnat The intuition is an effect of having the theory-- a thing that our human brains do for us in the background. So yes! Agreed!

Tom Ritchford 6d ago

@paddy @gnat @ceejbot

I do agree with you, but I have to quibble: we don't know for sure that some AI in the future can't have, effectively, a theory of the system.

I mean, we are essentially meat computers, so we know it's possible.

However, I am quite certain that today's LLMs have no such theory.

@TomSwirly @paddy I actually have an MS in this exact topic - my thesis was on decision theoretic natural language generation. Even got a derivative paper published in AAAI. On that basis I can tell you that in my educated opinion, we are not much more likely to have this than any other iain banks technology.

Anything can happen, but I advise everyone to be wary of thoughts like “maybe someday” when it comes to AI - the field is littered with these predictions and they come to pass so rarely!

@gnat @TomSwirly I'm very open to changing this view if the material conditions change, but I also see no harm in baking my views on this subject on the material conditions that exist today instead of trying to guess which ones might exist in the future.

Jeff Grigg 5d ago

@gnat @TomSwirly @paddy

They were describing computers as "electronic brains" and saying that they would replace us since before I was born. And I'm at an age where many sensible people would retire.

The future is unpredictable.

But LLMs are *NOT* likely to get us there.

Tom Ritchford 4d ago

I do actually make my living in this field...

My claim is that we know that thinking machines are **possible**, because we are thinking machines.

I'm almost completely certain we won't build *artificial* thinking machines, because we're speed-running the death of technological civilizations, but thinking machines exist and one is typing this response.

So what are you objecting to, exactly? Is it that you're 100% certain that it's not happening, but I'm only almost 100%?

@TomSwirly @paddy which field is that? If you’re talking philosophy, cognitive science, or even robotics or something like that, that’s pretty interesting; if all you mean is that you’re a software engineer like the rest of us it’s not much of a credential in this context, sorry.

Anyway, my objection is that it was silly for you to interject “sure, you’re correct today, and in every reasonably foreseeable future, and I’m almost completely certain you’re right, but —”.

Another objection is that any definition you’re using for “thinking machine” that is flexible enough to include both you and a hypothetical descendent of chatgpt is uselessly broad.

I probably have more but I’m going to work now; have a good morning.

@gnat @TomSwirly I mean I think my whole objection here is I'm saying "machines can't" (present tense) and you're jumping in to say "machines could" (future tense) and like sure, machines could and if they do I'll update my stance here, but nobody is talking about what machines could do in the future we're talking about what they can do in the present, so why bring it up?

@gnat @TomSwirly When you combine the tangent with the fact that the LLM industry loves to use claims about what is theoretically possible to justify things that don't make sense given demonstrated abilities in the present, the whole thing comes across like you're best case trying to pull us into a different conversation than the one we're having or worst case trying to pull the same sleight of hand the LLM industry is so fond of.

Tom Ritchford 2d ago

I didn't pull you into anything. You wrote: "AI cannot have a theory of the system or a point of view on the problem space."

This isn't a statement about today, it's an absolute statement about what AIs could conceivably do, ever.

Now, my actual beliefs are these:

* LLMs are a massive heist for the benefit of a tiny number of rich people.

* It's hard to believe that these utterance averagers will ever become even competent.

1/

Tom Ritchford 2d ago

* Intelligent machines are clearly *possible*, because humans are such machines.

* But I see no path from LLMs to actually intelligent machines. I think today's AI is a boomscam.

* And it makes no difference. What's actually going to happen is that we will devastate our climate and ecosystem, our technological civilization will collapse, and computers and integrated circuits will be history.

2/

Tom Ritchford 2d ago

@paddy @gnat I do have this issue as a mathematician with a science background writing on the Internet.

People seem to interpret "possible" as "having a significant chance of happening" - say, p > 0.1.

But when I use the word "possible" I use it to mean exactly that: p > 0.

So I can believe that today's AI bubble is a scam, a bubble, a heist, and almost certainly will go nowhere, while still believing that AGI is *possible*.

/thread

Tom Ritchford 2d ago

I'm a trained mathematician, now working on the pytorch project, so somewhere between those two.

You are attacking something I did not say. The annoyance directed at me in this thread for me saying things that are undeniably true isn't very nice.

PP made the absolute statement "AI cannot have a theory of the system or a point of view on the problem space". The word "today" does not appear anywhere in there.

The person who wrote this *cannot know this for sure*.

1/

Tom Ritchford 2d ago

Actually, I wrote a longer thread: here it is.

https://toot.community/@TomSwirly/114448829400151307

I think that all this AI stuff is almost certainly bunk, and it's instead a heist, a crime, and a fake. I strongly do not believe LLMs have any theory of mind or of systems.

But as a mathematician, I also strongly object to people being certain about future events, particularly when we have irrefutable evidence that intelligent machines, such as ourselves, exist.

Are we clear now?

Tom Ritchford (@TomSwirly@toot.community)

@paddy@mastodon.cloud @gnat@tech.lgbt I didn't pull you into anything. You wrote: "AI cannot have a theory of the system or a point of view on the problem space." This isn't a statement about today, it's an absolute statement about what AIs could conceivably do, ever. Now, my actual beliefs are these: * LLMs are a massive heist for the benefit of a tiny number of rich people. * It's hard to believe that these utterance averagers will ever become even competent. 1/

toot.community

@TomSwirly your point is that intelligence can exist. this is trivial, obvious, and widely understood by absolutely everyone. I think it is a silly thing to say. I have lost patience with this extremely verbose nitpick.

Tom Ritchford 1d ago

> your point is that intelligence can exist.

No, it isn't. It baffles me that you can read what I wrote and say that. It reads like mockery. Indeed, none of your responses are polite.

I agree; it's not worth wasting my time when there are so many kind and respectful people on Mastodon.

C J Silverio 2d ago

@TomSwirly @paddy @gnat No feedback loop: no actual intelligence.

FFFluoride 6d ago

@paddy @gnat @ceejbot I don't know if I'm tired or what, but that essay was hard to read

Rachael L 6d ago

@paddy @vaurora @gnat @ceejbot this article is soooo good. I added "legacy code archaeology" to my resume's pile of skills section many years ago after realizing how often I had to reconstruct past authors' thinking. I routinely remind folks on teams that they are writing comments for future people who will not have the context they do and that incomplete or out of date docs are better than none because it shares their thinking.

burakemir 5d ago

@paddy @gnat @ceejbot I had posted these quotes last year (without any commentary): https://discuss.systems/@burakemir/112621313656511067

I think it stands to reason that what is called "design" should be considered part of the theory.

But there is a challenge: the various artifacts we can come up with are never enough to capture the theory.

burakemir (@burakemir@discuss.systems)

Papers to be read, again and again: Peter Naur "Programming as theory building" (1985) https://pages.cs.wisc.edu/~remzi/Naur.pdf "The present discussion is a contribution to the understanding of what programming is. It suggests that programming properly should be regarded as an activity by which the programmers form or achieve a certain kind of insights, a theory, of the matters at hand. This suggestions is in contrast to what appears to be a more common notion, that programming should be regarded as a production of a program or certain other texts."

discuss.systems

C J Silverio 2d ago

@burakemir @paddy @gnat The map is not the territory. The map helps, though.

@gnat the perfect recipe for plausibility soup 😭

steeznson 6d ago

@gnat Honestly it's just helped the devs who don't bother testing their code raise a PR faster. Those people have always existed but LLMs enable them to foster bad habits to an even greater degree.

As a senior dev I got stung trying to get an LLM write a custom strategy for an objectMapper in Java when most of my prior experience was python. It gave me a plausible looking solution that was the exact inverse in terms of functionality of the correct one. Took me a day to debug (locally)

@steeznson you remind me of my other thought: the industrial revolution was driven not by considerations of quality of life, but by desire to replace expensive and competent workers by cheap incompetent ones. key word here: incompetent. we're at industrial revolution 2.0, it will be as devastating to the workforce and quality as tge first one, ruining lives of an entire generation, or two. there will be quality of life improvements too, eventually, there will be neostandardisation, new fords with conveyors, new speeds… the world of it will never be the same :-/

@gnat

Змей 🏳️‍🌈6d ago

@gnat I am also trying to be reasonable and accomodating, but at this point the only way I am looking at anyone's LLM genarated code is if they sit next to me and explain it bit by bit - not the prompt they used, not what the code is supposed to do, but what the piece of code in front of us is *actually* doing and we take it from there; I am tired of looking at code that looks like it's supposed to be working, more often than not it IS working, but it's so far from the optimal solution that I would consider it a first draft at best.
(that probably makes me hard to work with, but they're yet to prove my work is slower or worse for refusing to use LLM-based tools)

Petrichor ᚄᚔᚅᚐᚁᚆᚃᚒᚔᚂ 6d ago

@gnat Just wait till AI starts writing recipes 🤮

Jeff Grigg 5d ago

@sinabhfuil @gnat

Oh, it already has.

And some of them can literally kill you, if you follow them. Ignorant and dangerous.

Petrichor ᚄᚔᚅᚐᚁᚆᚃᚒᚔᚂ 5d ago

@JeffGrigg @gnat Really? Example?

Jeff Grigg 4d ago

@sinabhfuil @gnat

If you intentionally (or ignorantly) mislead it, you can get deadly chlorine gas.

https://www.opb.org/article/2024/09/23/ai-can-generate-recipes-that-can-be-deadly-food-bloggers-are-not-happy/

But what I really had in mind is the more subtle problems of insufficient cooking time for proper food safety. I've read about that happening.

AI can generate recipes that can be deadly. Food bloggers are not happy

Apple, Google and Samsung are all working to integrate AI that can generate recipes into their virtual assistants, like Siri. People who work to develop recipes respond.

OPB

Petrichor ᚄᚔᚅᚐᚁᚆᚃᚒᚔᚂ 4d ago

@JeffGrigg @gnat Jesus, yes - undercooked pork…

PatternChaser 6d ago

@gnat we have a global problem with AIs, I think. They are NOT 'intelligent'. They can't learn. ... And they CAN'T design #software, or anything else.

If you don't understand your own (copy-and-pasted) code, how can you release it to your customers? 😱

Total agreement! 👍 😀

So I code with ChatGpt/Claude.

First, it's not like ordinary coding.
If you expect to vibe code, you are going to have a very bad time.

Second. The more definitions you give the #AI, the better.
Give parameters what you want to expect.

Third, spec it. Give as much specifications as you can. You want that text window to scroll?
Propose an array or a list structure.
Leave as little to imagination as possible, the thing has very little of it and it will try to please you hard, it will make shit up.

Fourth. Give overall instructions. I usually say something along the lines of "Do not code unless clear instructions are given". Else the thing will launch into code at the first prompt.

Fifth, I used to get it to Pseudocode. Now I just usually say "Restate the problem". Just to make sure the machine understands what it's doing.

Checkpoint. When you have code that works, designated it as "Version X.1" because inevitably the machine will fuck it, esp if you're introducing a notable change.

Seventh, learn #promptengineering, most people have NFI how to use the #LLM esp. if they are naturally hostile towards the tech.
E.g. If I really want the model to pay attention, I will say something like: DIRECTIVE: Blah blah.

Lastly, this should go without saying, the free models suck, pay the broligarch tax for the smarter engine.

It helps if you understand a little how LLMs work, today for example I gave a prompt to just keep latest checkpoint and specs and flush everything else from the session context as it tied itself into knots

There are other tips.

P.S. If this is not your sport, just mute and move on, don't be rude

@n_dimension I don’t really understand why anyone would choose this over doing it yourself. It sounds very annoying - much more annoying than managing a very junior SWE, and with worse results. Managing is already something that I have chosen to avoid in favor of being an IC, so this sounds kind of hellish to me.

To that end, I’m not gonna start using these techniques. My colleague does, and his work is causing me problems. Do you feel sure your development practices here are not causing your colleagues problems? Are you sure you’d know if they were?

If it’s not causing problems, then, sure, all the best. Your workflow is not my workflow but your workflow is okay.

You are totally right.
I would not inflict #AI programming on my peers.
I note that many of the solutions are showy, artisan code to show off how clever you are (what makes it to GitHub)

I code for myself. In languages I don't have a skill.
I am an old dog. I'm also not a very good programmer, though a few decades back, I earned a crust coding

For me, #AIprogramming is a choice between not coding at all
Or
Building systems that work and having the satisfaction of doing work I thought I was no longer capable of.

Ben Aveling 6d ago

@gnat " It even looks fine in code review, one chunk at a time!"
That's because each chunk is fine. It's plagiarising working programs, one chunk at a time, and mashing them together, pretty much at random, like some sort of Frankenstein's monster.

Jerry 🇸🇪6d ago

@gnat
Seeing the bigger picture or long term, that's not everybody's forte, unfortunately.

Joana de Castro Arnaud 6d ago

@gnat Exactly as "designed" for generating prose text: looks correct, looks plausibly human, but has no sense of anything.

@gnat This is exactly what I expect from LLM-generated code: no design, because there's no actual thought.

GrooveMD, Queer Time Machine 3d ago

@gnat Double plus boosts for tying this to Alton Brown's philosophy of unitaskers.

AI coding is the turkey stuffing of computing: just plain evil.