King Arthur's stories are set 700ish years before Robin Hood's. About the same time has elapsed between Robin Hood's setting and today. I mention this because TV shows / movies about each of them seem to take place at the same time. It's unsurprising but always weird how we flatten time behind us.

Echoing you, like a lot of text technologies (see also e.g., topic modeling), LLMs continue to outpace in practice what they can do in theory. I suspect that's because language carries more than even our great expectations, and the technology is mostly improving our ability to catch it in the act.

I would be cautious about saying anything "in principle" since they have already surpassed what I thought they would be able to do. If forced, I would be surprised if they could develop sophisticated spatial reasoning without a fundamental shift in how they are trained.

I often hear, in my line of work, that large language models have surpassed the capabilities of “mere stochastic parrots.” Yet few are willing to attribute to them anything like “understanding.” What, in your view, is the best recent appraisal of the limitations of LLMs?

Brutalism holds a special place in my heart. Pollock's paint splashes aren't my thing, but I understand and appreciate their place in art history. Wine mom art has its charms. Most aesthetic styles deserve some appreciation. Except these, the stupidest objects humankind has yet produced.

At the end of the last Bush administration, I spent a lot of time in other countries juggling on street corners. My "stupid American makes a fool of himself" act made me the most money by far. Anyway if you're a US expat trying to make some money right now, you'll go far being the butt of the joke!

General reminder that one can live their ideals of social action without ever posting about it online. In fact, that's often preferable!

This is great - it's about time someone updated the discourse on LLM energy usage to reflect that coding agents use massively more prompts than occasional questions to ChatGPT Simon estimates that a day of coding agent usage comes out close to the energy needed to run a dishwasher

Spouse: I was thinking this morning, in my time off I could use that 50% off coupon for the massage. Me: I was thinking this morning, since our main mode of information access switched from alphabetical to search & retrieval, I bet we're applying alphabetical ordering to unrelated things less. +

2) Disagreement patterns evolve significantly after AI deployment. In the first year, radiologists reject AI-detected PEs in 30% of cases. This drops sharply to 12% by year two, then stabilized.

We document several key facts: 1) when an AI tool for radiology scans was rolled out in a hospital system, radiologists showed asymmetric agreement with AI's recommendations. When an AI was negative on PE: 97% agreement from rads AI positive [~10% of the time]: 84% agreement from rads

Posted a very early stage draft with rock star collaborators. Key question: when we actually roll out AI tools, how do people use them? Do they just defer completely? Does it improve productivity and ability? We look in the medical setting of pulmonary embolisms paulgp.com/papers/Radio...

Don't talk about it much, but I wound up leaving my PhD early because I was fortunate enough to have a different career path to follow. Not finishing was a bummer, though. Anyway, if all goes well I'm getting back to it this year (while still in my day job). Wish me luck! I'm excited to try again.

Important work

An absolutely extraordinary look at how to improve rendering of ASCII art, first in static images, then in motion. As he gets into contrast enhancement for complex grayscale animations, the vector lookup works starts to be parallel to work happening with language models. Just an excellent narration.

"the ability of LLMs to recover the meaning of this [intentionally mangled passage of text] is more like recovering 'Happy Birthday' from the shape of someone’s mouth as they whistle." Interesting result, and reminds me of the pattern-based logical process we use to decode encrypted messages.

Among other things, this morning I learned world leaders just text each other, I guess? I mean it makes sense, but I'd never thought about it before.

Here are 8 successful *proposals* for academic books/monographs. I am hoping that these help someone who is starting out trying to write one and wants to see what worked for me. They are provided as examples, rather than exemplars. A lot of it was just finding my own way eve.gd/2026/01/16/s...

wait this is such a handy reference? igorcarron.github.io/welcome-to-t...

Once again: if you want to use AI in a way that will actually improve what you are doing, which *is* possible, it would take more people, more money, more time, and more care, not less. But that basically never happens so everything AI touches becomes terrible.

Me, a chief technology officer at a large state agency when my wife calls because the 'delete' key on her laptop is sticking: "Have you tried turning it off and sort of softly punching that corner of the keyboard a few times?" "You want me to... punch my computer?" "I'm a professional!"

📌Hi! I'm Scott, a historian of science. Before DOGE, I helped the US fund the humanities efficiently and impactfully, to reach the breadth of the American public. Now I help make the Library of Virginia's rich collections and services digitally accessible to all. Personal account, mostly silly.📌