chat

You could argue in the same way that car safety technologies are "bandaid after bandaid to try and compensate for the fact that cars are fundamentally huge hunks of metal going really fast".

The parallel doesn't really work (there is nothing fundamentally incompatible between the purpose of a car and the basic design idea); but to try and answer, this would be a viable answer if the huge "AI labs" admitted the tech doesn't work and were collectively working on entirely new architectures (likely ones that include actual symbolic representation of the world, not just statistical links between words).

But they're not, they're using the same fundamental approach and doing ridiculous shit around that approach - ultimately a genuine, and interesting, NLP breakthrough - to try and sell it as what it's not and will never be.

you still need to do quite a bit of handholding and double checking it gets stuff right but it definitely can get medium-high complexity stuff right and overall it is a productivity boost.

See, that's one part I disagree about. First as a direct productivity boost; I realize you probably think it's making you more productive; but that's something you have to demonstrate, at scale. And the few genuine studies I've seen suggest the effect, if it exists, is marginal at best; at least two I remember suggest it decreases productivity (and at least one, possibly both, can't remember, said self-reporting from devs was that it increased it despite the effect).

More importantly, though: very little of the code you produce this way will be maintainable. Very little will also be secure, too (both "direct" vulns - they're not actually good at finding them, FYI - and more fundamental logic vulnerabilities - they'll never find these at all outside of trivial cases). If you want a concrete example of that, just look at the leaked Claude harness code: it's genuinely pathetic. They're vibe coding it all and it shows.

Also, I notice you say "you still need to..."; do you think this will eventually get fixed ? it's been nine years. There's a point where "it'll get better" is starting to become more of a meme than an answer. It was getting better, then a lot less, then lately almost not at all, despite greater and greater cost increases.

Shit, lately I've even started to wonder if they're even training new models anymore; I've been wondering if they're just not shipping the same with a different harness / different bullshit surrounding it. It's not like they can improve on the models themselves anyway (not without at least one, likely several, genuine fundamental breakthrough) - they're all out of human-produced data by now.

Yes it is, just yesterday I was having trouble with a script, so I googled it, the Google ai told me to use a command that doesn't exist. It just made it all up. Completely worthless.