kottke.org posts about programming
Charity Majors, writing about how high-performing engineering teams are dealing with the transition from pre-AI to AI-native development: AI enthusiasts are in a race against time, AI skeptics are in a race against entropy.
This is not a situation where one side is right and the other is huffing paint. (O, that it were!) Each side is grappling with a real, alarming, escalating threat to the company’s existence, and the closer they look the more (again: real, alarming) evidence they find.
The enthusiasts are not wrong. We are starting to see real, non-imaginary, discontinuous leaps in capabilities from teams that lean in hard to working with AI. And this does not feel like a normal technology cycle where you can wait for the dust to settle; teams that sit this out while competitors are hustling could be out of business before the dust settles. That’s a real, existential threat.
The skeptics are also not wrong. When you ship code faster than engineers can read it, in domains where nobody has full context, you are making withdrawals from a trust account that took years to build. Reliability degrades, institutional knowledge evaporates. You end up with systems nobody understands, products burbling into incoherence, and on-call rotations that grind people up and spit them out. That is ALSO a real existential threat.
She goes on to say that “the wins and costs are happening to two different groups of people. There is no natural feedback loop.” Interesting read.
Interesting observation by Mitchell Hashimoto (creator of Vagrant and Ghostty) on how a company’s or product’s choice of programming language matters less in the age of agentic programming:
On the interesting side is how fungible programming languages are nowadays. Programming languages used to be LOCK IN, and they’re increasingly not so. You think the Bun rewrite in Rust is good for Rust? Bun has shown they can be in probably any language they want in roughly a week or two. Rust is expendable. It’s useful until it’s not then it can be thrown out. That’s interesting!
Hashimoto is talking about this complete rewrite of Bun (a Javascript/Typescript toolkit that’s owned by Anthropic and includes “a fast JavaScript runtime designed as a drop-in replacement for Node.js”) in a completely different programming language (Rust) in just 6 days.
6,755 commits, branch name claude/phase-a-port, PR opened May 8th, merged May 14th.
Six days. A full rewrite of a production-grade JS runtime, merged in six days.
Let that number sit in your mind for a second.
Whether or not you think that taking this six-day-old code completely rewritten & tested mostly by LLMs and deploying it in production is a good idea, it’s something that many more companies are comfortable doing. Simon Willison riffing on Hashimoto’s thoughts:
I was talking to someone who worked for a medium sized technology company with a pair of legacy/legendary iPhone and Android apps.
They told me they had just completed a coding-agent driven rewrite of both apps to React Native.
I asked why they chose that, given that coding agents presumably drive down the cost of maintaining separate iPhone and Android apps.
They said that React Native has improved a lot over the past few years and covered everything their apps needed to do.
And… if it turned out to be the wrong decision, they could just port back to native in the future.
This also applies to other layers of the tech stack (database, etc.) to various extents as well as to some other types of software, e.g. it’s trivial to export your bookmarks from one bookmark manager to another if they both have APIs or import/export capabilities — or, with a bit more effort, you can write your own.
BTW, this also goes for the big AI companies — it’s pretty easy to switch between different flagship models or to the increasingly powerful local models.






I had a lot of fun playing around with this collection of generative design tools, especially the textual ones. I wore out the “randomize” button on each of these. (via sidebar)
Lauren Goode convinced her editors at Wired to let her spend a couple of days at a tech company called Notion learning how to vibe-code (i.e. AI-assisted computer programming): Why Did a $10 Billion Startup Let Me Vibe-Code for Them — and Why Did I Love It?
Expanding a mermaid diagram or alphabetizing a list of dog breeds hardly seemed like sticking it to the coding man. But during my time at Notion I did feel as though a trapdoor in my brain had opened. I had gotten a shimmery glimpse of what it’s like to be an anonymous logical god, pulling levers. I also felt capable of learning something new — and had the freedom to be bad at something new — in a semi-private space.
Both vibe coding and journalism are an exercise in prodding, and in procurement: Can you say more about this? Can you elaborate on that? Can you show me the documents? With our fellow humans, we can tolerate a bit of imprecision in our conversations. If my stint as a vibe coder underscored anything, it’s that the AIs coding for us demand that we articulate exactly what we want.
During lunch on one of my days at Notion, an engineer asked me if I ever use ChatGPT to write my articles for me. It’s a question I’ve heard more than once this summer. “Never,” I told her, and her eyes widened. I tried to explain why — that it’s a matter of principle and not a statement on whether an AI can cobble together passable writing. I decided not to get into how changes to search engines, and those little AI summaries dotting the information landscape, have tanked the web traffic going to news sites. Almost everyone I know is worried about their jobs.
One engineer at Notion compared the economic panic of this AI era to when the compiler was first introduced. The idea that one person will suddenly do the work of 100 programmers should be inverted, he said; instead, every programmer will be 100 times as productive. His manager agreed: “Yeah, as a manager I would say, like— everybody’s just doing more,” she said. Another engineer told me that solving huge problems still demands collaboration, interrogation, and planning. Vibe coding, he asserted, mostly comes in handy when people are rapidly prototyping new features.
These engineers seemed reasonably assured that humans will remain in the loop, even as they drew caricatures of the future coder (“100 times as productive”). I tend to believe this, too, and that people with incredibly specialized skills or subject-matter expertise will still be in demand in a lot of workplaces. I want it to be true, anyway.
A very interesting read. Over the past several months, I have been reading a lot about LLMs and coding, particularly pieces by experienced coders who have switched to using LLMs to code. There is a lot of silly (and perhaps dangerous) hype around AI, but over the past several months, LLMs and supporting tools have gotten unnaturally good at programming, when directed properly. Here are some of the things I’ve read recently in case you’re curious about what’s possible now:
I’m curious to know if any experienced (or inexperienced) coders among you have tried any of the recent suite of AI-assisted coding tools and what your experience has been. (Your general thoughts about AI — particularly its potential downsides, which have been amply documented elsewhere — are best left for some other time & place. Thx.)
Elon Musk has claimed that his “DOGE” team has found evidence of “massive fraud” at the Social Security Administration, alleging that 150-year-old Americans were receiving benefit checks. I saw this claim easily debunked over the weekend, but Wired has a good writeup of it. Basically, the programming language that these systems are written in (COBOL) often uses an arbitrary date as a baseline…most commonly a date from 150 years ago.
Computer programmers quickly claimed that the 150 figure was not evidence of fraud, but rather the result of a weird quirk of the Social Security Administration’s benefits system, which was largely written in COBOL, a 60-year-old programming language that undergirds SSA’s databases as well as systems from many other US government agencies.
COBOL is rarely used today, and as such, Musk’s cadre of young engineers may well be unfamiliar with it.
Because COBOL does not have a date type, some implementations rely instead on a system whereby all dates are coded to a reference point. The most commonly used is May 20, 1875, as this was the date of an international standards-setting conference held in Paris, known as the “Convention du Mètre.”
These systems default to the reference point when a birth date is missing or incomplete, meaning all of those entries in 2025 would show an age of 150.
The SSA also automatically stops benefit payments whenever someone reaches the age of 115.
Tim Carmody has a great appreciation of HTML in Wired magazine: HTML Is Actually a Programming Language. Fight Me.
HTML is somehow simultaneously paper and the printing press for the electronic age. It’s both how we write and what we read. It’s the most democratic computer language and the most global. It’s the medium we use to connect with each other and publish to the world. It makes perfect sense that it was developed to serve as a library — an archive, a directory, a set of connections — for all digital knowledge.
I love HTML!
Let’s say you’ve got a bunch of books that need to be sorted alphabetically by author. What’s the fastest way to accomplish this task? Luckily, efficient sorting is a problem that’s been studied extensively in computer science and this TED-Ed video walks us through three possible sorts: bubble sort, insertion sort, and quicksort.
For more on sorting, check out Sorting Algorithms Visualized, sorting techniques visualized through Eastern European folk dancing, and a site where you can compare many different sorting algorithms with each other. (via the kid should see this)
When you write some code and put it on a spacecraft headed into the far reaches of space, you need to it work, no matter what. Mistakes can mean loss of mission or even loss of life. In 2006, Gerard Holzmann of the NASA/JPL Laboratory for Reliable Software wrote a paper called The Power of 10: Rules for Developing Safety-Critical Code. The rules focus on testability, readability, and predictability:
- Avoid complex flow constructs, such as goto and recursion.
- All loops must have fixed bounds. This prevents runaway code.
- Avoid heap memory allocation.
- Restrict functions to a single printed page.
- Use a minimum of two runtime assertions per function.
- Restrict the scope of data to the smallest possible.
- Check the return value of all non-void functions, or cast to void to indicate the return value is useless.
- Use the preprocessor sparingly.
- Limit pointer use to a single dereference, and do not use function pointers.
- Compile with all possible warnings active; all warnings should then be addressed before release of the software.
All this might seem a little inside baseball if you’re not a software developer (I caught only about 75% of it — the video embedded above helped a lot), but the goal of the Power of 10 rules is to ensure that developers are working in such a way that their code does the same thing every time, can be tested completely, and is therefore more reliable.
Even here on Earth, perhaps more of our software should work this way. In 2011, NASA applied these rules in their analysis of unintended acceleration of Toyota vehicles and found 243 violations of 9 out of the 10 rules. Are the self-driving features found in today’s cars written with these rules in mind or can recursive, untestable code run off into infinities while it’s piloting people down the freeway at 70mph?
And what about AI? Anil Dash recently argued that today’s AI is unreasonable:
Amongst engineers, coders, technical architects, and product designers, one of the most important traits that a system can have is that one can reason about that system in a consistent and predictable way. Even “garbage in, garbage out” is an articulation of this principle — a system should be predictable enough in its operation that we can then rely on it when building other systems upon it.
This core concept of a system being reason-able is pervasive in the intellectual architecture of true technologies. Postel’s Law (“Be liberal in what you accept, and conservative in what you send.”) depends on reasonable-ness. The famous IETF keywords list, which offers a specific technical definition for terms like “MUST”, “MUST NOT”, “SHOULD”, and “SHOULD NOT”, assumes that a system will behave in a reasonable and predictable way, and the entire internet runs on specifications that sit on top of that assumption.
The very act of debugging assumes that a system is meant to work in a particular way, with repeatable outputs, and that deviations from those expectations are the manifestation of that bug, which is why being able to reproduce a bug is the very first step to debugging.
Into that world, let’s introduce bullshit. Today’s highly-hyped generative AI systems (most famously OpenAI) are designed to generate bullshit by design.
I bet NASA will be very slow and careful in deciding to run AI systems on spacecraft — after all, they know how 2001: A Space Odyssey ends just as well as the rest of us do.
Ilia Blinderman of The Pudding has written a pair of essays about how to make data-driven visual essays. Part 1 covers working with data.
It’s worth noting here that this first stage of data-work can be somewhat vexing: computers are great, but they’re also incredibly frustrating when they don’t do what you’d like them to do. That’s why it’s important to remember that you don’t need to worry — learning to program is exactly as infuriating and as dispiriting for you as it is for everyone else. I know this all too well: some people seem to be terrific at it without putting in all that much effort; then there was me, who first began writing code in 2014, and couldn’t understand the difference between a return statement and a print statement. The reason learning to code is so maddening is because it doesn’t merely involve learning a set number of commands, but a way of thinking. Remember that, and know that the little victories you amass when you finally run your loop correctly or manage to solve a particular data problem all combine to form that deeper understanding.
Part 2 is on the design process.
Before you begin visualizing your data, think through the most important points that you’re trying to communicate. Is the key message the growth of a variable over time? A disparity between quantities? The degree to which a particular value varies? A geographic pattern?
Once you have an idea of the essential takeaways you’d like your readers to understand, you can consider which type of visualization would be most effective at expressing it. During this step, I like to think of the pieces of work that I’ve got in my archive and see if any one of those is especially suitable for the task at hand.
Check out The Pudding for how they’ve applied these lessons to creating visual essays about skin tone on the cover of Vogue or how many top high school players make it to the NBA.



Sándor P. Fekete, Sebastian Morr, and Sebastian Stiller came up with these Ikea-style instructions for algorithms and data structures used in computer science. In addition to the three pictured above, there are also instructions for several other searches, trees, sorts, and scans.
From Feminist Frequency, a quick video biography of Ada Lovelace, which talks about the importance of her contribution to computing.
A mathematical genius and pioneer of computer science, Ada Lovelace was not only the created the very first computer program in the mid-1800s but also foresaw the digital future more than a hundred years to come.
This is part of Feminist Frequency’s Ordinary Women series, which also covered women like Ida B. Wells and Emma Goldman.
From the BBC, an hour-long documentary on Ada Lovelace, the world’s first computer programmer.
You might have assumed that the computer age began with some geeks out in California, or perhaps with the codebreakers of World War II. But the pioneer who first saw the true power of the computer lived way back, during the transformative age of the Industrial Revolution.
Happy Ada Lovelace Day, everyone!
Seymour Papert, a giant in the worlds of computing and education, died on Sunday aged 88.
Dr. Papert, who was born in South Africa, was one of the leading educational theorists of the last half-century and a co-director of the renowned Artificial Intelligence Laboratory at the Massachusetts Institute of Technology. In some circles he was considered the world’s foremost expert on how technology can provide new ways for children to learn.
In the pencil-and-paper world of the 1960s classroom, Dr. Papert envisioned a computing device on every desk and an internetlike environment in which vast amounts of printed material would be available to children. He put his ideas into practice, creating in the late ’60s a computer programming language, called Logo, to teach children how to use computers.
I missed out on using Logo as a kid, but I know many people for whom Logo was their introduction to computers and programming. The MIT Media Lab has a short remembrance of Papert as well.
When Grade-A nerds get together and talk about programming and math, a popular topic is P vs NP complexity. There’s a lot to P vs NP, but boiled down to its essence, according to the video:
Does being able to quickly recognize correct answers [to problems] mean there’s also a quick way to find [correct answers]?
Most people suspect that the answer to that question is “no”, but it remains famously unproven.
In fact, one of the outstanding problems in computer science is determining whether questions exist whose answer can be quickly checked, but which require an impossibly long time to solve by any direct procedure. Problems like the one listed above certainly seem to be of this kind, but so far no one has managed to prove that any of them really are so hard as they appear, i.e., that there really is no feasible way to generate an answer with the help of a computer.
Given that there’s so much mathematicians don’t know about prime numbers, you might be surprised to learn that there’s a very simple regular expression for detecting prime numbers:
/^1?$|^(11+?)\\1+$/
If you’ve got access to Perl on the command line, try it out with some of these (just replace [number] with any integer):
perl -wle 'print "Prime" if (1 x shift) !~ /^1?$|^(11+?)\\1+$/' [number]
An explanation is here which I admit I did not quite follow. A commenter at Hacker News adds a bit more context:
However while cute, it is very slow. It tries every possible factorization as a pattern match. When it succeeds, on a string of length n that means that n times it tries to match a string of length n against a specific pattern. This is O(n^2). Try it on primes like 35509, 195341, 526049 and 1030793 and you can observe the slowdown.
Older posts
Socials & More