Zoltan P Majdik

I don’t see the mate in 3, but I’m not disciplined enough to not go straight for Bf5x and then attack the rook.

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.

Incorporating the Rhetoric of Scientific Language into Sentence Embeddings using Phrase-guided Distant Supervision and Metric Learning Kaito Sugimoto, Akiko Aizawa. Proceedings of the Third Workshop on Scholarly Document Processing. 2022.

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/

“open to taking a small step back towards the warm embrace of the British monarchy” 😆

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...

It was always a tenuous argument. Given the increasing scarcity of new text for training, it never made sense that models wouldn’t become cheaper to train.

OTOH, DeepSeek has 1000s of Nvidia chips, and what we know about their model cost is at best a PR statement. Most of today is overblown.

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."

It's been interesting how the media have been so focused on intent rather than reception.

Table X — values and logarithms of haversines [a big table of values] from Logarithmic and trigonometric tables, Hedrick, E. R. (1920)

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.

Something tells me that I wouldn’t have to worry about blue shells if I did.

What if I’m too good to not be in the lead? 😁

What if Mario Kart spikes my blood pressure though? [f’ing blue shells, all-the-f’ing time!]

⬇️⬇️

Measuring the Persuasiveness of Language Models Anthropic developed a way to test how persuasive language models (LMs) are, and analyzed how persuasiveness scales across different versions of Claude.

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...

ORD? Their international arrivals are the *worst* — and sadly 90% of where I end up having to cross the border.

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. +

Who isn’t?

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.

We’re probably <2 years from in-prompt fine-tuning. “Here’s a link to the docs of the new python blackbearbruin data frame library. Can you tell me how to solve problem X with it?”