You know, it isn't even that tools like this are useless. There are absolutely things they could be good at. I've personally seen Claude find stupid little bugs you'd spend an hour figuring out and hating yourself for afterwards with great efficiency. I tried the first iteration of Copilot, back when it was just an aggressive autocomplete, and while I had to stop using it because it was overconfidently trying to finish my programs for me without being asked, it was great for filling in boilerplate and maybe even a couple lines of real code for the basic stuff. We have models nowadays that are actually trained to find bugs and security issues in code rather than having the entire internets thrown at them to produce something Altman & Amodei can sell to the gullible as AGI.

But there's the problem. The technology has been around for a while, we have a good idea of what it's good for and, more importantly, what it's not. "Our revolutionary expert system for finding bugs in your code" isn't nearly as marketable to the general public, and the CEO class especially, as "our revolutionary PhD level sentient AI that will solve all the world's problems if you only give us another couple trillion dollars, and also wants to be your girlfriend." And so we get Claude and ChatGPT and RAM shortages and AI psychosis and accelerated climate change instead of smaller, focused models that are actually good at their specialist subjects. Because those don't produce as much shareholder value.

@bodil

> Can someone please explain this to me?

Sure: code with the job of managing a natural language LLM isn't going to look like procedural code you're used to.

If you have doubts whether coding assistants like https://antigravity.google are any use, download it, try it on your own code with your own choice of tasks and find out.

You can throw the changes away if you are worried about getting contaminated.

You can write about your experiment here. And, you will actually know.

@bodil

Oh no, the probability engine is producing average output.

surprisedpikachu.jpeg

@bodil I don't get it either. It completely baffles me that anyone can look at the generated output and think "this is how it should be" or look at the anthropic leak and say "this is great engineering"

And once someone has emotionally invested in LLMs being the future of their career it is really hard to get a honest conversation going.

And when I test it and it doesn't deliver It seems to always boil down to: you are doing it wrong ... you are stuck in your old ways ... pre AI thinking ...

@bodil Indeed. Its good programmers are not susceptible to hubris.

Otherwise, they'd probably react to a concept like AI by claiming that programming is a form of art that can never be managed by AI, that AI is crap and anything produced by it must also be crap - proven by the fact that AI uses the crap from crappy websites where crappy - eh - programmers have posted it, and they'd utterly fail to make any distinction as to where AI can be a useful tool and what should better be done manually.

First and foremost, they'd scream defiance about the code quality of AI itself, amusingly ignoring the fact that this code has very obviously been written by programmers. But they'd be AI programmers, and therefore obviously worthless bastards.

My, am I happy that programmers are so absolutely immune to hubris.

@bodil

> It's a good thing programmers aren't susceptible to hubris in any way, or this would have been so much worse.

I've been told hubris, as well as laziness and impatience are the three great virtues of a programmer.

https://threevirtues.dev/

The essential virtue is integrity and it's missing on that list. Without it, the original explanations would crumble.

I guess it's either one of those off-by one errors - or it was a daft claim right from the start.

@bodil The "it's a tool and you just have to check its output" argument enfuriates me like little else does.

Nobody ever fucking does that. People don't even take reviews of real code seriously, and I call bullshit on every sloperator who claims to have read and understood the output of their oversized autocomplete engine.

And don't get me started on the asshats who also let the slopthing write their docs. Or their business emails. Or anything else, really.

@bodil
It does seem to work though? A lot of users think that their product is good, or even the best, and that it keeps improving.

It may be built from bubblegum and clotheshangers, but apparently that works just as well as good engineering.

At least in short run. It may give them headaches in the long run. But a company like anthropic will be ecstatically happy if they make it to the long run at all.

@bodil I use claude for programming and it takes a couple of iterations to get good code.

This versus "lets just ship it" approach. It's the same copied-from-stackoverflow quality you previously got from some offshore teams. Passes functional requirements = done.

@bodil

It's not like we've all been cobbling stuff up in a hurry from StackExchange C&Ps for the last decade anyway.

Because there's a point where no matter what yr fkn agile velocity, the far end of the backlog is red shifting towards some management event horizon, and it now hurts too much to think properly.

So you do the thing that makes the pain go away because you are a thinking feeling human.

Jira does not care and adds another ticket with a t-shirt size attached.

@bodil First you must realize Knuth could have subtitled "the art of programming" as the lazy art ;). So lazy, in fact, we create programs just to write programs for us.

Now, artificial ignorance is spewing back from it's training, which is largely a similar pile of...

On top of that pile, as information entropy suggests, there is even some degeneracy from that, which is why you can't have AI coprophagia. Rather than being a Yogi a little smarter than the average programmer, it is condemned to being a little dumber than the average of the training input. Simple physics.

@bodil And considering "we" couldn't actually figure out to prioritise test before LLMs does not fill me with confidence that the test phase will be prioritised now.

Because LLM introduction for most is about production velocity and cost cutting.