EEG shows brain can simultaneously encode two speech streams (plos.org): "If reading aloud a children story, you may notice you are able to maintain an independent unrelated train of thought. While doing so, I notice that occasionally extra mistakes can "leak" into the story telling - e.g. you read a single word incorrectly, maybe substituting a word from your other train of thought."

This is super interesting to me in the context of humans' ability (or lack thereof) to process things in parallel. There's a classic exercise that demonstrates how much worse we are than computers at parallel processing: try alternating between counting incrementally and reciting the alphabet aloud: "1, A, 2, B, 3, C, 4, D…". Most people can get to about 4–6 and D–F pretty easily, but then break down. Machines, on the other hand, have no issue with this.

Maybe encoding parallel speech streams is the brain's way of sneaking past its own context-switching bottleneck and doing a kind of parallel processing akin to computers.

Try the exercise yourself, first with, and then without the table and see where you break down:

I'm reading A Short Stay in Hell by Steven L. Peck, which tells the story of a man trapped in a hell of his own making: an enormous library in which every book is exactly 410 pages long. He's told there's a way out, one he shares with everyone else there: to find the book that tells the story of your life.

The premise borrows directly from Jorge Luis Borges' short story The Library of Babel, and it rhymes with the infinite monkey theorem: the ape that, given enough time at a typewriter, eventually produces Shakespeare.

The catch is that the books are randomly generated. Almost all of them are gibberish: Aj;kLJjppOjnfe7ImNBzuyS@;jHnMBVFghT/.hk, page after page. But hidden in the randomness are books that contain actual words. Somewhere on the shelves is the story of your life told from the perspective of your mother. The story of your life from the perspective of your right hand. The same books again, but told backwards. With the chapters shuffled. With a single misspelling on page 212. None of these count. Only the one true book will do.

The characters ponder whether the library is infinite. It isn't. Borges specifies the format: 410 pages, 40 lines per page, 80 characters per line, drawn from an alphabet of 25 symbols. That works out to $25^{1{,}312{,}000}$ possible books: roughly $10^{1{,}834{,}097}$. That's a number nearly two million digits long. For perspective, there are about 10^80 atoms in the observable universe. If every atom in the universe were itself an entire universe, and every atom in those universes were a book, you would still have essentially none of the library. The finitude is the torture: the exit exists, it's countable, and it will take a very, very long time to get out.

Reading it, I kept thinking about Large Language Models. The space of all possible text is the library. LLMs, during training navigate this space, converging on the paths (or corridors in the library) which read as legible. The training set used to train real models (the entire web) is already legible and they use this legibility to help predict the next word given the previous.

If you want to wander the stacks yourself, Jonathan Basile built a working version at libraryofbabel.info. Every page you'll ever write is already in there.

To put the size of possible books into perspective, I attempted to print the number $10^{1{,}834{,}097}$ below. Each number uses a pixel of different color corresponding to the number.

Regression. Whenever you build something heavy (see last note), you'll eventually run into this problem. You'll find that a particular feature worked so much better in a previous version. At a time when there was less; when it was isolated and perfected. You can write tests, you can keep notes of its state. But regression will creep in on even the most decoupled feature as things get heavier. And that creates a new and (what you should consider) interesting problem for you to solve. It's also a signal that you've arrived at something heavy.

"The modern makers’ machine does not want you to create heavy things [...] Winners of major awards almost always say the same thing as they lift the trophy: ‘Wow! It’s so heavy.’ As though the weight itself validates the achievement. Simple logic: Light achievements beget light awards. Heavy achievements beget heavy awards. We accept this in the physical world. But online, we forget [...] We create more than ever, but it weighs nothing [...] [People] want to make one really, really good thing. One truly heavy thing. A book. A manifesto. A movie. A media company. A monument. — A masterpiece. [...] You ship, but you do not build. You call yourself a creator, but what have you made that could survive a month offline? A year? A decade? If you stopped posting tomorrow, would anything remain? [...] My answer is simple but not easy: Make something heavy."

Very nice.

Formula 1 Silverstone 💨

This is super cool. It's not quite one-shot, but it's the closest thing I've experienced to agentic coding generating animations. My head is already racing with ideas.

A couple of cool artworks I recently stumbled across…

Xue Jiye — "Possible Future and Definite Future"

Wang Haiyang — "Skins 06"

It’s been awhile since this campaign launched but it really is so so good. The video starts with the doom and pessimism associated with “problems” and contrasts it with an alternate scenario that’s guided by processes, focus, and collaboration with “solutions.” The MF Doom All Caps track is the cherry on top.

The Northern line was (eerily) quiet during the Tube strikes a couple of weeks back. A good excuse to grab this clip and try out color grading with LUTs in Final Cut Pro for the first time. I muted the video by default because, well...

This is a lovely UI component registry. It uses framer-motion under the hood which bloats your bundle size (making the website slower to load). But if it can create a UX like this, it might be worth it.

I recently posted a note about value capture, and how AI could make you 10X, no... 100X, no... 1,000X more productive.

This essay from Quarter Mile brings us softly back down to earth.

I threw away my IDE some time ago (goodbye IntelliJ...) and Warp is the reason for that. If you're looking for a high-performance GPU-backed Terminal (written in Rust), Warp's got you. And now, they're open source!

Excited to see where the community takes it.

Suppose AI makes you 100X more productive. Then who captures the value?

If you work eight hours a day and produce 10-100X the output, the company captures all of it. But if you work one hour a day and manage to ship the same output, then you capture all of it.

Steve Yegge gets at this dilemma on a recent episode of The Pragmatic Engineer, arguing we need to figure out what work-life balance will look like in the age of AI. For programmers, now that time-consuming copy-paste, debugging, and boilerplate tasks can be outsourced to AI. What's left is System 2 thinking: architecture, judgement calls, weighing trade-offs, deciding what to build in the first place.

The catch is not even engineers can sustain five-plus hours of System 2 thinking a day. So if the job now is mostly System 2, the eight-hour workday is doing two things: it's either padding empty hours, or burning people out.

Delta Components is now officially part of the shadcn open-source community registry. Install any component directly into your React or Next.js project via the shadcn CLI and own the code.

Visit deltacomponents.dev to find out what the registry has to offer.

The 404 page on this site (patrickprunty.com) is a little easter egg. Find out what I mean.

This is a great research paper that evaluates Agent Skills. The SkillsBench study tested 7 model harnesses (i.e Claude Code w/Opus 4.5, Opus 4.6, Sonnet 4.5, Haiku 4.5, Gemini CLI w/Gemini 3 Pro, Gemini 3 Flash, and Codex CLI w/GPT-5.2) across 84 tasks (for example, a task which asks an agent to count coins, enemies, and turtles across frames of Super Mario gameplay video). Hand-rolled human written skills boosted pass rates by 16.2%, whilst agent self-generated skills, i.e via Claude Code's skill-creator, resulted in -1.3% pass rates. The agent actually performed better without a playbook, then with the one it wrote for itself. For engineers, the real skill is no longer writing the code: it's knowing how to write a good playbook.

切腹 / Seppuku (/ sɛˈpuːkuː /) is a term used for ritual suicide practiced by Japanese Samurai by first cutting the abdomen with a small dagger, and then having a "second", or kaishakunin, complete the suicide by decapitation. The Samurai commit seppuku to restore honour, avoid capture by their enemies, atone for an error that disrupted their sworn lord, or to loyally follow their lord in death. Watching the Shōgun TV series, I must admit it's hard watching seppuku-after-seppuku, and yet, at the same time fascinating to observe (even a fictionalised version of) a culture whose attitude towards death is so contrary to how modern society view it today.

Hello world!