@Janet on cohost

@Janet

22/9/22

@Janet7/7/2023, 7:47 AM

eevee @lexyeevee

eevee @lexyeevee7/7/2023, 6:58 AM

mozilla keeps saying its new "AI Help" feature is "trained on MDN articles" or "limited to MDN content", which makes it sound like they trained their own model or at least fine-tuned it

but that's not what's happening — as far as i can tell, what they actually do is find a couple MDN articles relevant to your question, send the ENTIRE CONTENTS off to ChatGPT, and ask it nicely IN ENGLISH to answer the question using only what's in those articles

#LLMs

@aune7/7/2023, 7:15 AM

ah, i sure did miss that the person pushing this was the new "chief product officer"

former head of ML at twitter

good thing they picked him up presumably on the cheap :|

#mozilla

eevee @lexyeevee7/7/2023, 7:17 AM

you'd think the former head of ML at twitter would know what "training" means

@Janet7/7/2023, 7:47 AM

but... what you suggest would mean to sink time and work into your own model???!

You must log in to comment.

in reply to @lexyeevee's post:

gull *Plus!*@gull7/7/2023, 7:02 AM

absolutely miserable. i could just go to chatgpt et al. myself, ask it to roleplay as a helpful and accurate resource, ask the question, and get the same (bad) result. when will ai guys learn that These Things Cannot Actually Think Or Reason They Are Only Good At Putting Words Together Convincingly

wiggler @wiggles7/7/2023, 7:02 AM

yeah every ai project is literally just some string concatenation being fed to an openai api call

it would be funny if it wasn’t so sad

stu @stu7/7/2023, 10:24 PM

it's very funny that this was my first thought of how to do it as a lazy programmer and apparently no one came up with a next thought

@aune7/7/2023, 7:07 AM

can we get someone vaguely responsible at the helm of mozilla

please

while there's still a mozilla to salvage?

ruby @srxl7/7/2023, 7:12 AM

so basically just prompt injection. is that what training is now. why are llm bros

Raymond Neilson @deltavee7/7/2023, 12:51 PM

The sad part is, a) that's actually exactly how all these document summarization things work, and b) it's actually less inaccurate that way

in reply to @lexyeevee's post:

@aune7/7/2023, 7:21 AM

sure does show how much everyone "involved" in the LLM scene either knows or values behaving honestly

Natalie @nex37/7/2023, 7:24 AM

god this is so fucking embarrassing. someone please fork this immediately so I can update all the links on the Sass website

ruby @srxl7/7/2023, 7:41 AM

i'm hoping to help organize a forking effort, if that's of any interest

eevee @lexyeevee7/7/2023, 7:43 AM

mdn has serious backing and a lot of contributor + reader attention, is my concern. it's one thing to fork it, another thing entirely to keep it up to date

ruby @srxl7/7/2023, 7:45 AM

i'm not interested in forking the content, i'm well aware that would be a huge undertaking. i'm more interested in just forking yari, since thats the part mozilla seems intent on fucking up. the content and the site are separate, so i think a separate site hosting the same docs would be feasible

JP @vectorpoem7/7/2023, 5:45 PM

watching mozilla fuck up feels so much like watching the democrats fuck up.

eevee @lexyeevee7/7/2023, 6:06 PM

you were the chosen one.jpg