This whole post by @benmschmidt@vis.social is great, but this passage honestly should become *the* canonical way of thinking about #LLM and personality. It's better than the Shoggoth. https://benschmidt.org/post/2023-02-19-sydney/
Remote
Ted Underwood
@TedUnderwood@sigmoid.social
mastodon 4.7.3I use machine learning to study cultural history at the School of Information Sciences, UIUC. Also #generativeAI, #digitalhumanities, #sciencefiction, #computational #socialscience. Author of Distant Horizons (Chicago, 2019). #tfr #fedi22
3976 Followers
1230 Following
19 Posts
Joined November 08, 2022
GitHub:
Threads:
threads.net/@tedunderwoodillinois
Absolutely revelatory piece from Yoav Goldberg casting light on an overlooked puzzle about last year: why did we need *reinforcement* learning (RLHF) to unlock the potential of language models? Why wasn’t supervised learning enough? #LLM #AI https://gist.github.com/yoavg/6bff0fecd65950898eba1bb321cfbd81
A short blog post reflecting on an experience I've had lately: where students show up with a fully-formed solution to a problem I thought was still too difficult for them. What is AI doing to change the way we learn?
#machinelearning #LLMs
https://tedunderwood.com/2024/08/31/will-ai-make-us-overconfident/
Everyone knows when you meet the fay folk, you don't eat the food. Maybe we need a similar set of mores about AI? E.g., don't talk to a model about Jung and ask whether it has a "shadow self." Nothing good will come of that.
Got my first update from @quintsns@mastodon.uno. If you follow it, the bot summarizes links shared by people you follow. Might be nice if there were room for a title, but nbd. It does 80% of what I would want an algorithmic #timeline to do!
It used to be a philosophical question for science fiction writers, but it's becoming a practical problem for designers. https://tedunderwood.com/2024/06/15/should-artificial-intelligence-be-person-shaped/
"Lena." A short story in the form of a Wikipedia entry from the future. By qntm. You need to be in a mood for pretty dark humor. https://qntm.org/mmacevedo
Replying to
@compthink@sigmoid.social I’ll give away a little of Yoav’s answer, because it’s kind of fascinating philosophically. What we needed to teach the models was, when to say “I don’t know.” And that’s not something we could teach by demonstration alone, because … we don’t know how much the model knows! Had to be dialogic.
New article by Whitney Arnold & Corey Arnold topic-modeling The Monthly Review (1749-1844). @dh@a.gup.pe https://www.tandfonline.com/doi/full/10.1080/10509585.2022.2158460?needAccess=true
Hypothesis: #LLM are going to make the distinction between "distant" and "close" reading pretty much moot in 5 yrs, by a) reducing barriers to entry for large-scale analysis and b) making it possible to do fun and revealing computational things at a much smaller scale, even with slippery questions about character and plot. @dh@a.gup.pe
Replying to
@andrey@sigmoid.social Thanks for your service; this instance has been really important in a crisis. It does search well; it has a translate button; it's responsive. No complaints!
Replying to
@dh Over on the bird site, @dbamman adds BERT as a data point, and it equals or out-performs GPT-4: https://twitter.com/dbamman/status/1637515288584527872
IMO Charlie Jane Anders
should get royalties on Apple Intelligence because the model of transparent AI — running mostly on mobile, quietly steering the user via reminders without ever quite becoming perceptible as a personality — is just a Caddy from All the Birds in the Sky.
Replying to
I'm glad IHE did this round-up to explore the range of opinion. Do people see much consensus emerging yet? My initial impression is no: everyone's coming at the question differently.
Replying to
@paulg @ifvictr if you’re in the advanced web interface the icon looks like three lines with little circles on them, at the top of each column; to me they look like sliders rather than dials; i forget how it looks on the ordinary one-column interface. You turn off the noise for each category of notification.