siderealMar 1

LLM boosters: This is trained with all of the text and code on the public internet!

People who can fucking think: So it's extremely low quality, then?

Fabian [fabsch]Mar 1
Show threadsidereal
"The average of everything on github" isn't badass code, it's unfinished student projects that never worked.
Mar 1 at 5:13pmWeb
Show threadsiderealMar 1
We hear "all the text on the internet" and we might think "well, they've sure digitized a lot of classic books, that must be good, right?" but then we realize that all of the books are maybe like 1% of the data. Most of it is like facebook messenger breakup arguments and semi-literate emails.
Show threadsiderealMar 1
Humans are relying on computers to correct their grammar when humans don't even agree on what proper English grammar is smdh
Show thread2xfoMar 1
@sidereal
I was always taught that people make grammar so i was really surprised to see people ask the machine how to do grammar
Show threadAlex@rtnVFRmedia Suffolk UKMar 1
@sidereal there's major differences within the same country (for instance between Southern and Northern England).
Show threadJohnMar 1

@sidereal

Even if it was just a bunch of literature I hate to tell the LLM-pilled folks but nobody goes to the bathroom in a novel, so 🤷‍♂️

Show threadFoolishOwlMar 1
@sidereal That's kind of the thing though. What the existentialists called "bad faith", the human social fallibility that Kafka was satirizing.
Show threadprom™️Mar 1
@sidereal Do they even tell us, what they put in our mashed potatoes? For food, that's mandatory.
Show threadcmwMar 1
@promovicz
Has anyone made analogies, yet, between large language models and Soylent Green?
@sidereal
Show threadSusan CalvinMar 1
@sidereal they've presumably pulled all the bad fanfic as well as the Epstein Files....
Show threadChip ButtyMar 1
@Susan_calvin @sidereal LLMs follow Sturgeon's Law
Show threadSusan CalvinMar 1
@otfrom @sidereal Sturgeon's law allows for the possibility of some good material amidst the dross. I don't think it applies to LLMs.
Show threadChip ButtyMar 1
@Susan_calvin @sidereal fair
Show threadSusan CalvinMar 1
@sidereal I've just realised that there must be Epstein Files fanfic. Whether LLM generated or not
Show threadCassandrichMar 1
@sidereal It's more like public comments on Facebook and birdchan and such, which even before "AI" slop were dominated by wetware-bot slop working in troll farms producing disinformation, fraudulent engagement metrics, etc. The automated slop is literally trained on manual slop.
Show threadJess👾Mar 1
@sidereal one could hope they apply different weighting factors to different sources
Show threadMalfunct (he/him)Mar 1
@JessTheUnstill @sidereal mostly they don't and it is a known issue in training.
Show threadJojo Mar 1

@sidereal

For real I have C# and Python code up there that should never be copied by anyone.

Show threadSome Kind Of GarfMar 1

@sidereal *sideways glance att the most bruteforce Chip8 emulator ever, a less-than-bare-minimum Game Boy emulator, and a barely off the ground concept for an overly ambitious MtG rule enforcement engine*

I don't want answers from anything that has learned from my code.