Games

We're expected to use it at work and ive been using Claude a lot lately. Tbh LLM coding assistants have come a very long way since the early days of copilot. Frankly, ive worked with several other (more senior even) engineers that claude could code circles around. Not many, I could count them on my fingers, but its more competitive than some other folks have let on.

The other repliers are correct that these tools can easily spit out buggy code that sneaks its way into the codebase due to lack of oversight, test coverage, and general guard rails. This is pretty easy to spot with various services you likely use or have used (amazon has had several outages recently for example). In my own workplace we have seem significantly more code being merged and a correlated increase in bug density (which is multiplicative with the increase in code being merged). There are definitely problems with relying on LLMs too much.

People are still learning what these tools are good at. Right now that seems to be boiler plate generation, following very common or explicitly defined conventions, and unit test generation. That's not a lot, but its absolutely not nothing. People seem to think their program/app/service is a special snow flake with special requirements only understandable by greybeards. That is not at all the case, most programming in industry is gluing together existing tools and solutions in various arrangements, then putting a little proprietary sprinkle on top. This has been the state of software development for decades at this point.

Like most other social issues, the underlying problem is capitalism. Like the advent of all other industrial automation, the mere existence of LLMs causes capitalists to demand an increase of production from the existing work force. Of course quality control is going to be a problem.

I dislike LLMs because of their impact on the environment and that theyre being shoved into every product. I also do enjoy programming, so LLMs were something I really avoided until the office started demanding it. I'm trying to lean into it now though. I dont care for the product my company sells (the tech is fine and even interesting, just not a product or field that seems worthy of spending so much energy on), I don't like how many hours I work, and I'd rather be spending time organizing and with my family. So , ive been offloading a lot of work to get deliverables out the door to Claude, then pivoting over to organizing work while it churns. I'm lucky in that the product I work on can't hurt anyone if a bug gets deployed. I can just log on the next day and fix it, nbd. Obv that's not the case for all software, but it is the case for most of it. Frankly, I strongly encourage other workers that have jobs that LLMs can do large swaths of to do the same. Talk to coworkers, do some work for whatever org you're a member of (you are a member or an org, right?), and let Claude churn out shit in the background.

People have written so much shit about this that writing more just feels like pissing into an ocean of piss but in brief, AI code:

• Generally looks correct but is not (the most dangerous kind of correct!)
• Reduces the share of the codebase that is held in the minds of those who maintain it, meaning further work on the codebase is either more difficult or incentivizes further giving up control to an LLM (aka hosted proprietary software blob)
• Takes the hard work put into open source software and launders it into proprietary or permissively-licensed software through the usual LLM plagiarism process
• Has a corrosive effect on the nicest part of open source, which is people voluntarily choosing to work together for a shared common good

The only okay use for AI-generated code is short one-time use scripts.

AI generated code is so much worse. AI generated/assisted art can at least hypothetically serve a space-filling role in a sort-of ok fashion for a small project with no budget, particularly if its curated and being used as a fancy photoshop tool to merge sketches and references instead of making things from whole cloth. The worst case scenario with art is that it's just not really that great. With local models, which are the only ones that can actually be controlled to any meaningful extent anyways, it's also a lightweight program that's no more energy intensive than playing a modern game is.

AI generated code is an active cognitohazard and a massive threat vector. It's taking the actual core of project, something that has to be designed cohesively and made to work with countless moving parts in an intelligent manner, and replacing it with the equivalent of massaging copy/pasted stack overflow answers until they squeak through a compiler without crashing. It can spew out boilerplate GUIs and stuff you might find in an "intro to making [whatever sort of thing]" tutorial, but in a nonsensical and impossible to follow way. The inherent, inevitable end result of using it is creating an abomination that can't be maintained and you can't fix it because it's eldritch madness spewed out by an unthinking machine and trying to follow it is physically painful. These code-generating LLMs are also part of the massively bloated and inefficient datacenter models that can't be run locally.