Armin was once one of the most prolific programmers in Python.

[wordshaper@weatherishappening.network]

@cwebber What's telling, I think, is that all these people go on about how much they're doing and how great AI is to help them build more *but there's no actual demonstrable stuff being done.* I mean, if AI was some kind of Nx multiplier you'd think we'd be getting N times more actual functionality out of software but mostly it seems like the N multiplier only applies to blog posts about how AI multiplies their programming.

[cwebber@social.coop]

Feeling FOMO about AI? Well here's my advice!

Stay on top of what's happening. Which doesn't really require *using* the tools. Just see what people are doing.

Whether or not you do use it, stay a practitioner. And don't fall for the FOMO.

Your career won't end because you're not making the choice to use AI. (If your employer makes you use it, that's another thing.)

If you use AI, use it for "summarize and explore" tasks. DO NOT use it for *generate* tasks. That's a different thing.

If you want to differentiate yourself, *learning skills* is the differentiation space right now.

These things are easy to pick up. You can do it whenever. But keep learning.

If you see generated examples, don't paste or accept them. Type them in by hand! The hands on imperative: actually trying things congeals core ideas.

And if it doesn't help your career... well, your consolation prize is: you'll stay interesting.

[cwebber@social.coop]

Also, I think using hosted models is strictly unethical for surveillance and energy usage reasons.

It *is* true that there are models you can run locally that are much, much more efficient, and I suspect the energy costs on training them can be dramatically reduced.

I don't use either presently, but using a local model to help you navigate a codebase (as opposed to generating code) is a very different thing, I think. But it's also not what most people are doing!

And hosted AI models, as I said, I think are fully objectionable from an ethics perspective.

Datacenters are an antipattern in the general case. AI datacenters, triply so.

[gugurumbe@mastouille.fr]

@cwebber I was quite curious LLMs, but I recently had a disappointing experience. I had a common latex problem, but with a more unusual technology stack. It went to “Do this — I get an error — OK do that —…” for a few rounds, nothing surprising. At some point I crossed a line, and it went “OK there’s no way to do what you want with this tech”. As usual, 30 seconds of grepping around in the source code gave me the solution.
Anyway, I wouldn’t trust it for summarize and explore.

[cwebber@social.coop]

@gugurumbe I'm not saying people *should* use it for summarize and explore, I'm saying that's a different category of concern, if done with a local model.

However, I'll also point out you were trying to debug LaTeX, which I would argue is a nearly impossible task no matter how many resources are thrown at it ;)

[gugurumbe@mastouille.fr]

@cwebber My XKCD-style password is so strong you can’t crack it: “undefined reference begin document”

https://xkcd.com/936/

[geichel@mastodon.ie]

@wordshaper @cwebber I don't think you appreciate just how many man years go into writing production level code. My productivity has tripled but if takes weeks to get a prototype in front of 100k+ users. Is not like we're going to release clawd and watch the world burn

[wordshaper@weatherishappening.network]

@geichel ah. Yes. You could say that neither I nor @cwebber have worked on large, complex systems, or worked on significant pieces of infrastructure, or handled production level code and thus are not well equipped to judge. That is indeed a thing you could say.

[aparrish@friend.camp]

@cwebber i don't know that i'd trust these models for summarization or navigation. even when the outputs are technically correct, they can leave out certain information or frame the information in a misleading way, papering over whatever makes the code unique and materially suited for the task at hand

[aparrish@friend.camp]

@cwebber (this is actually my main concern about llms. i think people really underestimate how much llms reproduce the values and expectations in their corpus, their reinforcement learning tasks, their explicit engineering, and their product design. and they underestimate the effects that this will have on their understanding of code and the horizon of what's possible to do with code)