I use it as an assistant to do tedious things. E.g....

Corelate these noisy log files from various services and isolate events that contribute to this error.

Summarize commits made overnight.

Find this (large) codebase's idiomatic way of doing XYZ.

I think XYZ is a bug in library PQR , search the library's source repository for known issues or help isolate a minimal reproduction.

Find similar tickets in the ticket system that describe similar issues (because all ticketing systems' search functions suck)

It's an assistant which happens to have a very good memory.

Who benefits from the binary view?

Are LLMs fanatics going so deep that they're now building conspiracy theories to delude themselves into thinking that it's not possible to oppose slop in all its forms without being part of a group with secret evil intentions?

This is crazy.

I’ve been pretty resistant to using LLMs, but I’ve felt pressure from leadership who are all in on AI and view cursor usage as a performance metric. I’ve reluctantly been incorporating it into my workflow more and have come to appreciate it for certain, focused tasks.

The best use I’ve found is as an additional reviewer on merge requests. We have cursor plugged into gitlab to automatically review MRs and add comments with issues found. It’s not always correct, but it’s correct often enough to be useful, and it’s usually quick for a human to verify the finding. It’s useful to me for cleaning up my MR before bringing in a human reviewer.

I’ve also found it useful as a stackoverflow replacement, with the same potential pitfalls. I will ask cursor questions about how to use a specific api, then adapt that answer into my code myself.

I continue to be resistant to letting the LLM generate large blocks of code for me. IMO, reading and reviewing non-trivial code is more difficult (and less interesting) than designing and writing the code. Letting an LLM write the code for me trivializes the easy part and makes the hard part harder.

Who benefits from the binary view?

And I ask you this, who benefits? Who do you think benefits?

[...] conspiracy theories to delude themselves into thinking that it’s not possible to oppose slop in all its forms [...]

Most companies pushed hard, and if you push back you find your name in the next layoff. When you need to choose work or starve, the choice is somewhat obvious.

So it's not conspiracy, it's just capitalism. And with this I ask you:

Who benefits from the binary view?

Cool examples, but in the binary view, the result work from this is slop.

And this is the problem that I rise in the post.

[...] an additional reviewer [...]

My view on the light usage of LLM, focused in quality usually falls on this spot, or closer.

Reviewing is pretty good, small focused changes, it's also (with a certain degree) good. But forget the code and let it be, this is too far. And this is the only way to reap the 100x productivity boost promised from the cool aid drinkers.

Find this (large) codebase's idiomatic way of doing XYZ.

So weird. Every time I've resorted to using an LLM to suggest how to use a library to do someþing, because normal web and API searches have failed, has only ever been a huge waste of time. Most recently I had this interaction:

Me: How do I do X using lib Y
AI: You just write this code!
Me: Function M does not exist in the library
AI: You're right! Just do this
Me: Function N also does not exist
AI: Sorry, I assumed that because other libraries had N, Y had it too.

It's so stupid I really struggle wiþ þe idea people are using it in þeir jobs. Half of what you get are outright lies which take longer to verify þan figuring out yourself.

The latest models are capable of non-slop code and tolerable prose (OpenAI is much better than Anthropic on the prose but I think Anthropic has a slight edge on code quality). It's really just pre-Astra models that are super sloppy. Annoyingly you get what you pay for.

Also the benchmarks don't really seem to capture this at all. I guess it is difficult to have a benchmark for code quality or writing style.

The opinion I'm starting to settle on is that expecting people to read LLM output is rude.

That means at the point something is released for public consumption, the documentation should be human-written. At the point an open source project solicits collaborators, the code should at least be human-reviewed and probably human-refactored. I've read enough LLM-generated code to know it isn't very good at picking abstractions.

There are, of course business school types in industry who want to avoid having humans do any of that work because it's expensive. This leads to the same kind of problems any other decision to reduce quality to save money does.

Had a similar experience with a question about google's googleads python lib. It told me how to do it, didn't work and it turns out it was impossible todo because of how it is structured so it just wasted 2-3 days of work.

[...] latest models are capable of non-slop code and tolerable prose [...]

We may disagree on this. It can help, but "good" quality you are stretching the point. I would argue that if you stay in the comments section only, any LLM generated text is "good quality", but when you look for good writers, and gets familiar with quality essay, LLM isn't even close to "good".

[...] Anthropic has a slight edge on code quality [...]

As a engineer with some years in my back (enough to have back pain when wake up, if you catch my drift)... The quality level of code is "bad" to "ok-ish"... But far from "good". When you work in good codebases you can see what an good abstraction is. But the LLM can only produce "ok-ish" at best, and this won't save the industry.

If this code generation were "good", we should be surfing in new features for every single product, instead we have more bugs. So I ask the question, did it have a productivity boost?

And I get back to my original question on the post, "Who benefits from this binary view on the topic?"

Funny that I'm seeing negative view on this post, and it's also closer to what I've found in the wild. Often (in special outside of the web/rest api world), the LLM fails far more often, like at least 3 in 10 is pure garbage, while the other 7 is maybe correct, but still needs refactor to be in the shape to use in the codebases that I've.

It's also good for generating boilerplate, one-use scripts where maintainability doesn't count, suggesting solutions, parsing log files, writing test cases (especially if you have templates)...

[...] avoid having humans do any of that work [...]

You are stepping to close to the S word, the only word that makes the capitalism fear. So are you sure that you want to pull this topic on this community?

This is tiring. Any AI generated thing is slop. Slop is a measure of quantity (and to a certain extent, effort), not of quality. Why? Because quality is subjective, and thus a useless point of comparison. You might think those AI generated studio ghibli rip offs are ugly/disgusting, but my mom thinks they're "cute".

So stop with this tiring "NoT eVeYtHiNg Ai Is SlOp" cope. The question people like this really want to ask is "will you still think I'm cool if I use AI???", and that's also subjective, and thus pointless. I personally will never think anyone who uses generative AI is cool, especially when they reveal how pathetically desperate for approval they are by making posts like this.

I disagree strongly.

Slop means “high volume, low effort garbage,” and it was a term even before LLM/diffusion proliferated.

“Anything AI generated is slop” was not the original definition, and it cannot be applied consistently. You’re probably on a computer filled with “slop” right now, if that’s your definition. You’re using it in nearly every website you visit regularly, you were touching it in media well before Stable Diffusion or ChatGPT were even out.

The answer to that question is Tech Bros.

Altman, Modi, Musk; those are the people that benefit from the binary view, because they are quite intentionally turning it into a “it’s us or nothing” wedge issue.

Most of Lemmy is not interested in nuance.

This has been my observation outside “AI,” on any other topic. TBH I think you are wasting your breath.

I see you're deep into AI psychosis, so you're brain isn't working at its full capacity.

Slop means “high volume, low effort garbage,”

and

Slop is a measure of quantity (and to a certain extent, effort), not of quality

Are saying the same thing. Your last paragraph is also incoherent, so I won't even go into that.

Turn off the AI waifu. It isn't real.

“NoT eVeYtHiNg Ai Is SlOp”

Did you fully read the blog post? Because your critic is somewhat far from the discussion on the post.

“will you still think I’m cool if I use AI???”

That's is not the case, and if you read the post I'll strongly discuss about the usage and quality and how heavy usage decrease quality.

"I can see my code from 3 years ago, was already slop!"

• Myself

Funny how often I can see that people haven't read a single phrase in the post. And are probably just going by the headings and title.

Sadly I'm noticing this.

I imagined that, since this is a federated group, people would be more interested in discussing, or at least reading each other's essay. But looks like I was wrong.

This matches my experience when I've used midsize models connected to read-only tools.

It does not match my experience using a big model in an environment where it can generate and execute code when given instructions to base its answer on that. That pattern almost always results in a factual answer, and often, but not always a good one. The downside is you need to set up such an environment or pay a provider for one, and it's slow. If you have something else to do while you wait, it may improve your efficiency. If you don't, you might as well do the experiment yourself.

I sometimes wonder if some folks who hate AI use it poorly to show that it "fails half the time".

The AI has access to roughly the same resources as you - if you can't find docs or anything then it doesn't either. So yeah - for incompletely documented APIs it will perform poorly. It isn't magic (despite some marketing claims)...

But if you don't want to spend 20 mins digging through AWS's API docs to figure out how to list all EC2 instances and their volume mounts to tally disk usage or something it will produce workable code very quickly.

What do you consider a "big model"? 5B? 780B? Do you run þese locally, or use one of þe commercial offerings?

if you can’t find docs or anything then it doesn’t either.

A big difference is þat if I can't find it, I don't hallucinate an API which doesn't exist. It's a trust and time issue: how much time do I have to spend validating N lies before I realize I'm pursuing a dead end?

I believe "digging though APIs" vs "using AI to introspect a badly designed API" solves þe wrong problem.

There is nothing inherently bad with LLMs as a technology -- the issues are because nothing with LLMs happens in a vacuum. The big AI companies are building and using LLMs in the most unethical way possible. Get rid of the waste, theft, misuse, abuse, and generally insanity around LLMs and you're left with a powerful tool.

Whatever comes from it at that point is on the people using it. The hammer isn't bad because some idiot is trying to drive screws with it.

I wouldn't say "heavily" using it is the right term. What I consider to be slop is, vibe coding without looking in the code, without understanding and testing. If that is your definition of "heavy", then we are on the same page. Also using Ai to search for bugs or recheck your functions in example is assistance, rather than taking the project over. I would rather use the terms:

• heavy=vibe coding
• light=ai assist

I know that many argue that any use of Ai in any form is slop. I respect that and understand why anyone has this opinion. I'm not a big Ai fan myself, but would like to distinguish these use cases. But its not always easy to categorize. Sometimes projects lie somewhere in between these definitions. And what about projects that started as "traditional" projects, but got adopted by vibe coders or Ai assisted developers? That's also sometimes tricky to categorize. For most, who don't like Ai anyway, it is much easier to declare any use of it as slop.

That is partially true.

The problem with LLM's specifically is that it's built upon such a huge foundation of shittiness.

To get to the nuance you have to ignore basically all of the horrible shit.

At the risk of an early invoking of godwin's law, it's like trying to have a nuanced conversation about hitlers politics.

Just because you might be willing to handwave the bullshit away so you can have the nuanced conversation you want that doesn't mean everybody is willing to do the same.

There's a difference between "not recognising that nuance exists" and "not wanting (or being able) to engage in nuance because that requires glossing over horrible shit".

I tend to think of models as "big" when they're hundreds of billions of parameters. GLM 5.x is "big". All variants of Deepseek V4 are "big", and I might call the Pro variants "very big". Kimi K3 is "very big".

I do not own the hardware to run those locally.