@jrconlin I have yet to see any useful application of LLMs. (I've seen plenty of useful applications for neural networks in general though.)
Conversation
Notices
-
Embed this notice
Thomas 🔭🕹️ (thomasfuchs@hachyderm.io)'s status on Thursday, 30-Mar-2023 06:54:35 JST Thomas 🔭🕹️ -
Embed this notice
Thomas 🔭🕹️ (thomasfuchs@hachyderm.io)'s status on Thursday, 30-Mar-2023 07:04:34 JST Thomas 🔭🕹️ @jrconlin IMO if you need a lot of boilerplate code, that's more of a sign that your programming framework isn't great and less of a sign that you need to send in the robots :)
-
Embed this notice
jr conlin (jrconlin@soc.jrconlin.com)'s status on Thursday, 30-Mar-2023 07:04:35 JST jr conlin Someone pointed out that they basically used an LLM to build the boilerplate code for a project they wanted.
I'll admit that's a reasonable thing since it's literally automating some boring parts and gets you past that blank page hurdle.
Could you swap that LLM for a cookbook? Sure. It just saved you from looking through it.
-
Embed this notice
Thomas 🔭🕹️ (thomasfuchs@hachyderm.io)'s status on Thursday, 30-Mar-2023 07:05:22 JST Thomas 🔭🕹️ @jrconlin Btw, I've seen some of the code generated, and... well... I wouldn't recommend using it (e.g. it was saving passwords in plaintext).
-
Embed this notice
Thomas 🔭🕹️ (thomasfuchs@hachyderm.io)'s status on Thursday, 30-Mar-2023 07:14:27 JST Thomas 🔭🕹️ @jrconlin I'd be highly careful with using it, because you don't know how these LLMs were trained (in fact, big tech doesn't want to disclose this at all, presumably because they could be sued for copyright infringement), and what the quality of the input code was like.
You'll need to thoroughly understand what it came up with, which at that point means that you need to learn everything about it anyway—in the end probably rewriting most of it and not saving any time.
-
Embed this notice
jr conlin (jrconlin@soc.jrconlin.com)'s status on Thursday, 30-Mar-2023 07:14:28 JST jr conlin Totally. When I say "boilerplate" I mean the sort of general skeleton code that you'd find in a cookbook. (e.g. I need a rust app that uses Cadence for metrics, Actix for web handling and Slog for logging. Basically this: https://github.com/mozilla-services/skeleton )
I'd still review the hell out of it, but at least it's better than digging through manuals for stuff you forgot last time.
-
Embed this notice