To me the least important part about LLM code is the ability to maintain it. I don’t give a fuck about the code at my job or my job or my employer. Fuck them. They’re not going to keep me employed for decades, why should I make sure to create code that can be maintained for that time when it won’t be me doing it. I hope my company goes out of business.

Both of your coworkers’ plans are to actually not review the code. They say Claude will do it. This guarantees you will get garbage and break things. They do actually have techbro misunderstandings of the reliability of these tools and perhaps a misunderstanding of how good they personally are at planning and documentation if they don’t expect more iteration than one PR and review.

You should also expect the cost of tokens to skyrocket eventually, just as an operational thing.

But this brings up the real point: why is tech lead giving a shit about, “I think we need to lean on Al to speed up PRs”? Aside from that not actually being the usual outcome, it’s also not their job to care about it. They are acting like C suite or management or a suck-up. “Speed up PRs” isn’t a meaningful thing to anyone except the bean counters trying to create metrics of success to use to discipline labor and climb the ladder. Is team lead spending all their time on PRs? If their new job is reviewing Claude PRs (which they imply and then contradict), they’ll be spending more time on that now, not less. They might do more PR reviews in less time but they’ll either be reviewing for common inevitable mistakes (you cannot rely on Claude to do the review lmao) or just let garbage through and have to fix it later. Either way they’re sacrificing dev time to do more review time, telling themselves it’s faster but likely not tracking whether milestones are being met faster

But most importantly why are they doing managements’ job. Y’alls job is to promise less to manager so you over deliver but make it sound like you’re technical experts that are irreplaceable. Not sell the idea that writing and reviewing code has been automated and you gotta get your metrics up.

I think it is possible to have ethical “LLMs”, but the technology does not exist. An ethical LLM would require us to re-imagine the entire process, from data gathering, to tagging and organization, to ML training, to applications. There is no ethical application for the existing technology. (I put LLM in scare quotes, because with the required changes, you might not categorize the end result as an LLM any more, though there are broad applications for the use of machine-learning techniques on large bodies of information)

First and foremost, ethical LLM technology must be environmentally sustainable. We are in the middle of a run-away climate crisis and there is no excuse to be wasting energy like this. Everyone responsible for LLMs and proof-of-work cryptocurrency should be executed right alongside the oil execs. It must preserve attribution. It must not be a tool for crime-prediction, surveillance, and social manipulation (e.g. profile-based advertising). It must be accessible and democratic.

In my mind, instead of having software which synthesizes language (human or programming languages), it would function as a highly sophisticated search engine. You might prompt “I want to make a REST client which can speak the Matrix protocol” and based on the corpus of training material it can identify the required aspects of an implementation and source material for each required component. This would be a very useful tool, but it is not useful in the way LLMs are being deployed today, which is to steal dead labor (the corpus of free software, documentation, internet discussions, and literature in general), in order to make living labor redundant.

That said, I would make the argument that LLMs are not the core problem in software engineering. They are deeply exacerbating the actual problem which is a crisis of complexity (which is why I think using them makes you an automatic dickhead). We’re sprinting into a world where we rely on essential components (the kernel or the web browser, for instance), which have completely outrun the capacity for responsible public maintenance. The deficit in maintenance is papered over by using LLM’s to “accelerate” development and review duties, but this only results in even more complexity. I believe this crisis is going to reach a tipping point fairly soon, and it is completely unrelated to the financial situation of the looming AI bubble.

The problem is the vast majority of software engineers are charlatans who would spend 99% of their time making annoying abstractions to piss me off personally instead of solving the problem. Now expand this to ai they can make those annoying abstractions so vast and so quick that you basically need ai to delete their code as well lol. The entire thing is a solution in search if a problem that only really exacerbates the weird dogmas software devs have put in place for years. So for the most part I use this robot shit to delete all their robot shit.

If you have to use it never get it to start the work for you only use it to tidy up work you have already done yourself, write the plan and MVP and some thin docs and get it to tidy up after using abstractions you approve.

This way you know what the code is doing, keep your skills and it really only does the boring work for you and a higher tendency to get a reasonable output.

The other use is as a searcher tool for unfamiliar codebases you can get them to map it out for you or locate something you’re looking for. Since im forced to use it at work this is how I keep my delivery speed high without completely removing my brain from my skull, it gives me direct control and autonomy where ai is a searching or formatting tool at best. It also has the main target of my already working code to go on which means it spends less time doing random shit. If you can basically target it at the stuff you don’t like doing and try to keep doing what you enjoy.

This is basically the only way i have managed to “keep up” without completely losing my soul or passion for programming. I still need to work on personal projects outside of work to keep sane though and that will be relegated to human written docs, books and typing

But don’t get me wrong im the biggest ai hater here i detest the technology and what its doing. This is only if you HAVE to use it how not to hurt yourself in the process coming from someone who actually likes solving real problems instead of inventing problems to solve. And you’re not wrong for rejecting the tool because the idea a human just needs to review slop code is ridiculous but also not even what happens in practice and blindly trusting it is kicking the can down the road but also harming your own career position and typically only promoted by people who were shit a writing code.

In theory; yes, under capitalism; no.

I think we need to lean on Al to speed up PRs

Same old story. Good, Fast, Cheap: which two are gonna get picked?

Not sure what kind of work you do OP so this isn’t specifically aimed at you but (hot take alert) honestly LLM code is fine for most corposlop projects. That kind of code has almost always been shit that barely worked, it was rushed and even when they gave a “senior engineer” free reign to design fancy abstractions they fall flat about 80% of the time and just make the code harder to maintain in the long run.

In my experience LLM code from latest frontier models is at least as good as the code of a developer of an average skill level.

The dumb thing here isn’t AI usage it’s as others have already said that devs seem to be hell bent on having themselves replaced. Use AI as much as you want and then lie through your teeth that you’re actually super needed to supervise it properly.

Yes. You get an abliterated model and an open source coding agent and you say, “Agent, how would you cause as much damage to the military industrial complex as possible while making sure to cover your tracks?” And then you make sure nothing tracks back to you because you don’t use any services connected to you including your own internet. Piece of cake, in Minecraft.

Whenever you see anything like human-in-the-loop, read or hear that synonymous with fall-guy.

I don’t know how to push back on this while being taken at all seriously. It has me considering that maybe I’m in the wrong for rejecting a tool that is considered useful by all but a few stick in the mud devs, and that those devs will never have a shot at a job again unless they pull their heads out of their asses and join our modern “AI”-driven world.

Right now you are arguing against money, greed. The cost is still being highly subsidized and at work it’s basically impossible to argue against money and this shit literally rots peoples brains. It feels like the same kind of position as being COVID conscious coming into 2027 where some people are so damaged after repeat infections the concept can’t even register and you seem crazy for wearing a mask as they cough in your face.

Anytime I have to use Super Intelligence shit I document from prompt-to-flawed output, documenting any new vulnerabilities and risks, cost, etc. In theory any middle manager pre COVID/AI brain rot would be able to use this to stab their peers in the back when shit hits the fan and somebody hopefully somewhere in the C-Suite realizes they were ripped off. When the actual cost of running a prompt is passed to the business consumer they should balk at what they get, it should be criminal. I don’t think many executives will actually take any accountability, even admitting that they were duped.

We “invested” in some hardware to experiment with LLMs before agents took off, so we’re thankfully running open models on infra we own. That removes a lot of the moral problems imo

As a former 100% “not using that ai shit” guy, I still can’t give up reviewing every single line of code in the stuff I’m responsible for. Add enforcing my opinions on it, it probably would’ve been faster for me to just do it.

But what I have found immensely useful is sicing it on code I absolutely do not give a single fuck about. Vscode extension doesn’t work how I want? Clone it, tell qwen to fix it, go back to doing valuable work. If it fixes the extension, great, if not, oh well. And if it’s slop, it’s not my problem.

Code you wrote yourself and make less dogass in the first place needs to be changed less after reviewing it no? And doesn’t manually reviewing code help you understand what you’re working on and how to write better in general?

I got a hunch this is me not understanding the industry, but I feel like LLMs only make sense in a world where productivity is measured based on code output.

Need more mindless fluff? Great news. Need to spend more time writing fluff and you wish something else could do review? Does your team not write high volumes of code? Train the LLM on a high volume dataset.

Like iunno. Maybe that’s part of it. Just a broken industry practice getting more brokener.

Also what another user said with the tokens. Once the bubble pops and/or the financial crunch on LLM companies start the cost is going to jump to the point where none of this won’t matter. (I hope anyways)

The people who think these tools work only do so because they don’t seems to know what software is or how it’s made.

This idea of “human in the loop” needs to go away (as in the problem is using an LLM to make code. That cannot be compensated for, and people need to go back to taking this job seriously). It doesn’t work and it only leads to ever increasingly complicated process fever dreams to try and compensate with the dissonance in their head that they so strongly believe the hype that they refuse to accept they’re being lied to by the same kinds of people that made social media addictive.

also, these tools are used by business people who have even less understanding of software to give more work to less people. And the way people respond to that with “I can’t move that fast, so my new aim is to move faster” is so incredibly self defeating. It’s so unserious.

My eyes roll out of head whenever these kinds of people realise that the way to make the LLMs reliable enough (though often their concept of enough is broken and what they’re producing is not good) is by doing software engineering practises with an unrealistic level of discipline and somehow still sell this as them doing less work

Does roleplaying with Mpreg Fred Flintstone count?

It isn’t possible for a human being to write code responsibly.

Deepseek (the not very smart but very cheap one from China) is running a 2-3x profit on their gpu hours. That means that even with last gen tech running not generally considered very “good” models it’s still profitable to sell inference at 15 cents per million tokens in(.003 cents cache hit!) and 60 cents per million tokens out.

While the deepseek model loops and makes mistakes more than a Claude for example, when used with a harness it is equally effective if a little slower but at over ten times lower cost.

The point of this isn’t to bamboozle you with facts and jargon but to show in a clear way that what has happened in programming with llms will not go away with some bubble pop.

If you want to stay in your industry, get okay with it. I like deepseek and glm.

From the computer security side of things it has been worth it to recognize that we are currently in a triage process industry wide, that the best process will have to be fixing bugs as fast as possible no matter how and we can deal with the problems introduced by those patches later.

Maybe but I think that leads to getting fired for not producing enough output.

Possible? Yes.

Do many devs even bother? Lol.