Louis Ingenthron @LouisIngenthron@social.cleverlibre.org

@helenczerski How exactly does nuclear not cut emissions as well as wind/solar? Neither link addresses that.

@Pat Yeah, but this isn't "recall", it's "repeat". I'm working on tuning requests to summarize large blocks of text, so I was using Joe Biden's wikipedia page as test data. Meaning, the AI had the correct spelling in the input data and still messed it up in the output, lol.

I'm noticing that #LLMs have a *lot* of trouble repeating Ketanji Brown Jackson's name correctly.

They almost always replace her first name. The latest example in the output was "Tanisha Brown Jackson".

Seems... uh, problematic.

@skrishna Oh come on. Just because you can learn from a failure doesn't make it any less of a failure.

@barney @VannaBlack So tempted to go over there and start an anonymous account and post a bunch of progressive news articles and pro-diversity content to put their "no censorship" policy to the test.

@micahflee I'm the exact opposite. I resist the urge to be polite or rude to it to avoid anthropomorphizing it. It's a tool and should be instructed as such.

@shoq Can't that be roughly approximated with hashtag follows?

@ethandoescode @andrewt It's better at helping experts expand their area of expertise.

The problem for beginners is that they can't tell the difference between the 90% very useful answers and 10% absolute crap made-up answers.

Plus, AI tools like CoPilot learn from their "style", which for new coders, essentially means it just helps lock in bad habits.

But for someone with enough generalized experience in the field to be able to spot and avoid that stuff, they're incredible for learning new languages and APIs.

@andrewt @feld @phryk Can confirm. I've been using two AI assistants for a little while now (GitHub CoPilot and CodeGPT). They're absolutely fantastic as learning assistants, and 9 times out of 10 they can pull documentation and customize it to match my context faster than I could pull the docs alone.

But, for the day-to-day work of operating in large pre-existing codebases, they're only a fraction as useful. I don't use CodeGPT at all in those situations and CoPilot only provides like 10-20% of the productivity boost it does in fresh codebases.

That said, though, they still do have moments where they can shine, even in big codebases, mostly in really common (but repetitive) patterns like serialization or response handler scaffolding.