@TedUnderwood@sigmoid.social
@TedUnderwood@sigmoid.social avatar

TedUnderwood

@TedUnderwood@sigmoid.social

I 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

This profile is from a federated server and may be incomplete. Browse more on the original instance.

TedUnderwood, to random
@TedUnderwood@sigmoid.social avatar

Jack Clark’s Import AI newsletter is always fun; this one especially because it explores the shared interest Meta and the CCP have in preventing fine-tuning of their models.

In a “Palantír”-level irony, one of the projects to prevent all further learning is called SOPHON. https://open.substack.com/pub/importai/p/import-ai-371-ccp-vs-finetuning-why?r=62esl&utm_medium=ios

TedUnderwood, to random
@TedUnderwood@sigmoid.social avatar

If you’d like to make it possible for people to follow you from Bluesky — and you’re in a domain that permits it — all you have to do is follow @bsky.brid.gy@bsky.brid.gy.

I think these connections have the potential to strengthen everyone. We need critical mass!

TedUnderwood, to random
@TedUnderwood@sigmoid.social avatar

This is a test to see if a message posted in the Fediverse can automatically be forwarded to Bluesky via Bridgyfed:
https://fed.brid.gy/

simon, (edited ) to random
@simon@simonwillison.net avatar

As a rough estimate, how often do you navigate to a website or individual page by directly typing its URL (including browser URL bar autocomplete)?

TedUnderwood,
@TedUnderwood@sigmoid.social avatar

@simon Hi Simon. On a loosely related topic, Bridgyfed is working now and you have fans over in Bluesky who would happily follow you there if you just hit follow on @bsky.brid.gy@bsky.brid.gy. I know you already have an account there, but this might be easier than cross-posting?

tedunderwoodillinois, to random

I've seen "critical AI" people saying generous things about the "carefully tailored and curated" datasets that were used to train statistical ML models five or ten years ago — in order to cast the monstrosity of GenAI into sharp relief.

This is quite different from what I was hearing about statistical ML in those quarters five years ago. Which gives me an idea for how GenAI will eventually gain broad acceptance!

TedUnderwood,
@TedUnderwood@sigmoid.social avatar

@budgibson @tedunderwoodillinois Just wait for the next round of technical advances, at which point we'll start hearing things like:

"At least ChatGPT-era AI was modest and well integrated with human needs. These new AI Agents, on the other hand, raise frightening questions of legal liability."

tedunderwoodillinois, to random

I’m skeptical of “declining standards” stories. My theory is that smartphones/soc media have given young people a direct access to the world that my generation never had. They see peers becoming influencers, boasting about internships, &c. School is not the only place to put aspiration. + https://www.harvardmagazine.com/node/85660

TedUnderwood,
@TedUnderwood@sigmoid.social avatar

@WenyiShang @tedunderwoodillinois Interesting! Yes, it was a somewhat monastic experience, especially because small colleges in the US are often in remote locations.

But I also really think students today are better informed.

TedUnderwood, to random
@TedUnderwood@sigmoid.social avatar

How can we evaluate the factual accuracy of long answers from LLMs? Researchers from DeepMind / Stanford demonstrate a strategy that uses LLMs + search to assess factuality: it's more accurate than human evaluation and 20x cheaper. h/t Marc Lanctot on Threads arxiv.org/abs/2403.18802

samuel_wade, to random
@samuel_wade@mstdn.ca avatar

@TedUnderwood Hi, I hope you’re well. I was wondering if you have any thoughts on detecting AI-generated text—is there any method for doing that that’s better than reading tea leaves? Could provide an interesting bonus angle to something else we’ve been working on.

TedUnderwood,
@TedUnderwood@sigmoid.social avatar

@samuel_wade My understanding is that there's no silver bullet here. It's likely going to be possible to train a model that can detect a particular generative model in a particular social context. But generalizing that is very difficult, and solutions are fragile to any change in the system. So in general people recommend not putting much hope in this, I think.

TedUnderwood, to random
@TedUnderwood@sigmoid.social avatar

IRL, if all particle accelerators started generating weird results and scientists saw a flashing countdown clock superimposed on their vision, 100% of them would assume it was a grant deadline.
https://www.threads.net/@tedunderwoodillinois/post/C41478vtflW

TriflingTree, to aiart
@TriflingTree@mastodon.social avatar

Peaches 🍑 on a branch
Watercolor illustration
Dall-e3 AI Art

TedUnderwood,
@TedUnderwood@sigmoid.social avatar

@TriflingTree even watercolors of peaches have lost their innocence

tedunderwoodillinois, to random

Hello Mastodon! This is the Threads alt of TedUnderwood@sigmoid.social.

TedUnderwood,
@TedUnderwood@sigmoid.social avatar

@stefan_hessbrueggen @tedunderwoodillinois I won't there, yet, but I'm checking here for now.

TedUnderwood, to random
@TedUnderwood@sigmoid.social avatar

It’s fine for the bridge to Bluesky to be opt-in. But however it’s constructed, I am looking forward to sharing some of the fine content here in a space where algorithmic discovery is easy and uncontroversial.

TedUnderwood,
@TedUnderwood@sigmoid.social avatar

It was imo unfortunate for the cause of open social media that it got paired with a finical dogmatism inclined to throw roadblocks in the way of basic operations like searching content.

TedUnderwood,
@TedUnderwood@sigmoid.social avatar

@ideaferace Yes, faceted and local has advantages and is okay with me.

But personally, I prefer a space where algorithmic discovery (and search) are just default norms. I’m glad I’ll be able to base myself there and connect to (parts of) the multiverse.

simon, to random
@simon@simonwillison.net avatar

The new "memory" feature of ChatGPT is yet another example of OpenAI building a whole new feature on top of some relatively straight-forward prompt engineering. They've given ChatGPT a new tool called "bio" which can be used to persist little pieces of information about the user.

Here's a ChatGPT session where I reverse-engineer the new tool: https://chat.openai.com/share/bcd8ca0c-6c46-4b83-9e1b-dc688c7c3b4d

TedUnderwood,
@TedUnderwood@sigmoid.social avatar

@simon It's so eerie that they're building this core function in natural language. I would have imagined it as a separate model trained to recognize snippets that ought to be persisted.

TedUnderwood, to random
@TedUnderwood@sigmoid.social avatar

It seems like a big deal that a bridge between Bluesky and the Fediverse is set to open up in a month. https://fediversereport.com/bluesky-opens-the-network/

TedUnderwood, to random
@TedUnderwood@sigmoid.social avatar

The School of Information Sciences at IL Urbana-Champaign is seeking postdoctoral researchers for AY 2024-25, renewable for a second year. We're interested in fields from children's literature through computational social science; your doctorate doesn't have to be in IS.

Applications are due March 15. It's a good idea to reach out to potential faculty mentors in advance.

Please repost and tell friends!

https://ischool.illinois.edu/research/postdoctoral-researchers

GreatBigTable, to aiart
@GreatBigTable@mastodon.social avatar

Draw Jam, on Youtube, made an interesting counter argument to the oft given argument from AI art users that is making the creation art more accessible.

That's complete crap.

As he points out, you can get a complete free art education from the free art tutorials (some by your favorite artists) that already exist on Youtube and the web in general.

You just have to sit down and put the work and practice in.

That's the rub.

https://www.youtube.com/watch?v=x1UO80zD62o&ab_channel=DrawJam

TedUnderwood,
@TedUnderwood@sigmoid.social avatar

@GreatBigTable generally providing a shortcut that gets you somewhere with less work is what we mean by “more accessible”

eg I can get to Seattle without a car — I just have to put in the work of walking 2095 miles — but a car would definitely make it more accessible

seinecle, to random
@seinecle@ioc.exchange avatar

LLM specialists: besides a powerful computer with a GPU, what kind of tricks can I use to make a locally hosted model spit out a response faster? Thx!

TedUnderwood,
@TedUnderwood@sigmoid.social avatar

@seinecle I don’t actually have first hand experience on this yet though I want to soon.

My understanding is that memory (VRAM or Apples unified memory) matters a lot. Also people sacrifice some quality by quantizing models to make them smaller.

TedUnderwood,
@TedUnderwood@sigmoid.social avatar

@seinecle the advice I’ve been getting over on bsky is that you can do a lot locally with Apple silicon

TedUnderwood, to ai
@TedUnderwood@sigmoid.social avatar

Computer-assisted studies of fiction tend to count words and trace themes. It's harder to measure the things that really keep readers turning pages: mystery, suspense, and surprise.

But large language models may change that. In this post, I experiment with a new way to measure uncertainty at different points in a story. https://tedunderwood.com/2024/01/05/can-language-models-predict-the-next-twist-in-a-story/

TedUnderwood,
@TedUnderwood@sigmoid.social avatar

If we got this method to work, the payoff might be that we could measure the creation of mystery, and distant-read things like the history of the cliffhanger. Here's a toy example with Conan Doyle's Hound of the Baskervilles.

I find that uncertainty is higher at the end of serial installments than at other chapter breaks. (The "tune in next week" effect.)

TedUnderwood,
@TedUnderwood@sigmoid.social avatar

@Virginicus Well, you can compare it to the divergence within a book in the earlier figure. The variance between books’ mean difficulties is comparable to the variance between the hardest/easiest passages in any single book.

TedUnderwood,
@TedUnderwood@sigmoid.social avatar

@Virginicus yes, although the range between outliers in Hound may also be partly just having more data points — I had to do four passes on it versus one in the other illustration

TedUnderwood, to random
@TedUnderwood@sigmoid.social avatar

If Threads is willing to commit to follower portability—which Mosseri did suggest today—I think I see the bridge as a positive development.

TedUnderwood,
@TedUnderwood@sigmoid.social avatar

@robux4 he indicated being able to take followers out of Threads to other Fediverse platforms, which is I think the part that would be hardest for Meta to swallow

  • All
  • Subscribed
  • Moderated
  • Favorites
  • •
  • normalnudes
  • tsrsr
  • DreamBathrooms
  • thenastyranch
  • magazineikmin
  • mdbf
  • Youngstown
  • ethstaker
  • slotface
  • khanakhh
  • rosin
  • hgfsjryuu7
  • kavyap
  • PowerRangers
  • Leos
  • ngwrru68w68
  • Durango
  • modclub
  • everett
  • cubers
  • vwfavf
  • InstantRegret
  • osvaldo12
  • GTA5RPClips
  • tester
  • cisconetworking
  • tacticalgear
  • anitta
  • All magazines
    •