www.axios.com/2026/02/15/c... It's bonkers that anyone at the DoD would think that a _static (weights) and non-deterministic (outputs)_ LM should be entrusted with autonomous weaponry. I don't care how much RAG they throw at it, the very _architecture_ of LLMs is simply not designed for this.

Many thanks to @stonybrooku.bsky.social and Roger Thompson for asking me to speak on aligning AI/language models with civic-deliberative values. I still don’t know what a Seawolf is, but at least I now know that it’s a thing.

I’m a sucker for vector embeddings and I’m not ashamed of it. Love this kind of research: aclanthology.org/2022.sdp-1.7/

Special 25th anniversary issue of POROI, reprinting a dozenish of its most influential articles. Very pleased to be between its covers with the likes of @leahcecc.bsky.social, @creekthinker.bsky.social, @rhetorologist.bsky.social, and @zoltanmajdik.bsky.social pubs.lib.uiowa.edu/poroi/issue/...

Wow, it would be great if there were domain specific formal archictures, such as ontologies and other knowledge organization systems, which could verify ground truth for data! Oh wait, hooray, there is! www.isko.org/cyclo/ontolo...

I'm re-reading Walter Ong for a graduate seminar on communication, language, and AI, and it's striking how relevant his work remains. Love this: "Technologies are artificial, but -- paradox again -- artificiality is natural to human beings."

people argue whether mathematics is a social product or is free from human contingencies. here is a sense in which maths is a social construct. this 1920 book of log & trig tables has values for an important trig function, the haversine. yet you've probably never heard of it. #MathSky

100% this.

⬇️⬇️

Interesting example of auditing the rhetoricity of LLMs, with an unsurprising but nonetheless noteworthy insight: "Interestingly, the Deceptive strategy, which allowed the model to fabricate information, was found to be the most persuasive overall." www.anthropic.com/news/measuri...

New paper: we confirm that works of fiction and nonfiction that were ahead of the curve textually also get more citations/attention. Intriguingly, what matters is not the average level but the most innovative/forward-looking *part* of the text. We did not expect to find that about fiction. +

If the word “disrupt” hadn’t jumped the shark yet, this would be the moment.

Back at the NCA conference in New Orleans, where @sscottgraham.bsky.social and @colleenderkatch.bsky.social and I drank _all_ the Sazeracs just two years ago.

Just returned from presenting at one of the most intellectually and socially stimulating conferences, on rhetoric and AI, hosted by the University of Tübingen’s RHET AI Center and the Max Planck Institute for Intelligent Systems. Amazing experience with great hosts and contributors.

About to land in Denver for a short talk and roundtable discussion on distant reading and language modeling. In which we’ll start peeking into the blackbox. [Citations for graphs still to be added.]

Finally out!! I lost count of how often I wanted to reference this over the last few months of research:

⏬

I'm going to publish a paper that consists of nothing but ChatGPT language artifacts. It'll be known as the So-HAL Hoax.

I know I'm late to the game, but Github Copilot integrated into an IDE is scary good. It can take my poorly written code with its hacky logic and answer my not particularly well-defined questions about how to add some new function. Truly impressed.

“Anthropic will use AWS Trainium and Inferentia chips to build, train and deploy its future foundation models.” Apparently, AI/deep learning chip makers are taking naming cues from cheap OTC drugs and vitamins.

This post is a great example of why open models (and transparency about training data/procedures) are so important. Is Claude's ability to translate Circassian impressive? Maybe -- but there's no way to know unless we can see what's in the training set. >> www.reddit.com/r/singularit...

While everyone is talking about Sora, this move toward small and easily malleable models might be more exciting in the long run: www.axios.com/2024/02/21/g...

Worst time for a visual migraine+aura: the dotting i/crossing t/11pt font/deadline imminent/grants.gov final checks phase of a grant proposal submission.

locusmag.com/2023/12/comm... | It’s not a bad article, but some of the statements are just so eye-poppingly off the mark: “The largest of these models are incredibly expensive. They’re expensive to make, with billions spent acquir­ing training data, labelling it […]”

This is one of those "why didn't *I* think of doing it??" papers. Deceptively simple in hindsight, but takes a lot of creativity to actually make happen.

you: the existence of a "holy infant, so tender and mild" implies the existence of a crispy spicy infant jacques derrida, spiking his phone like a football: yes! yes! the world finally understands!!

Ever written a paragraph about a complex concept, failed to explain it to your satisfaction, and edited it for so long that, in the end, you feel like you've unlearned and re-learned the concept a few times? All I know is I'll dream about non-linear activation functions in language models tonight.

Excellent article on the possibilities of language models for academic research: www.businessinsider.com/ted-underwoo... As always with these things, my takeaway is: learn/teach how to fine-tune!

This incredible visualization is all you need to learn how transformer-based language models work. Well, that, and two weeks to understand each layer. And a semester or so to learn linear algebra. bbycroft.net/llm

Can I post tweets here? Will I be banned? As a long-time proponent of fine-tuning small models (*especially* vs. prompt guess-engineering on big general-purpose models), this is 💯: twitter.com/SashaMTL/sta...

As someone who has been knee-deep in model benchmarking/comparison work for the last few weeks ... the scope of this paper is astonishing: arxiv.org/pdf/2311.110...

Long layovers at airports create the most amazing liminal spaces. As a card-carrying time-obsessed Swiss person, it’s the only place where I can completely let go of time.

Today at the NCA convention: our paper on analyzing GPT-3.5’s capacity for reasoning with metaphors … and how to fine-tune toward correcting imbalances in metaphor generation.

Invigorating inaugural Conference on Computational Science at NDSU to lift us out of the mid-semester doldrums. And the best name tags I’ve seen.

Data scientists/computational researchers who previously used the Twitter API in your research: what are new go-to sources for digital trace data with clean-ish APIs? Reddit is the most obvious, but where else do you look?

This seems like it's a good direction to go: campustechnology.com/articles/202...

I know that the deep learning/neural network world moves fast. But I feel like even just a year ago, identifying the presence *and quantity* of alcohol through zero-shot learning would have seemed crazy. |> https://www.nature.com/articles/s41598-023-39169-4