Blog & Notes

I was actually surprised to hear that PT Sans started as a Russian state project. It was commissioned in 2009 for the 300th anniversary of Peter the Great's script reform, with the goal of giving every language of the Russian Federation a free typeface.

Simile AI tries to find out.

… and coordinating is a strange experience.

Claude and others are starting to mark or detect AI content to comply with EU transparency regulation in the AI Act (PDF).

The two approaches are opposites: Anthropic marks at the source, embedding an imperceptible watermark in generated text and attaching signed C2PA provenance metadata to files, while Substack (together with Pangram) works from the receiving end, letting readers run a statistical detector over published posts. Here's a paper describing watermarking for LLMs in detail.

The practice of using Claude to generate content even earned its own term: Claudefishing.

antirez released h3.c — native MiniMax-H3 on Apple Silicon

Terence Tao, Fields Medal–winning mathematician, discusses the Jacobian conjecture with ChatGPT.

AI is a tool, just like other tools we use. And it's clearly a useful one. It may not have been that "clearly" even just a year ago, but it's no longer in question today. There are other questions around AI (like what the economy of it will actually look like in the end), but "is it useful" is no longer one of those questions. Anybody who doubts that clearly hasn't actually used it.

Linus on the Kernel mailing list.

AI makes it easy for anyone to build tools to automate their work... except most people and most companies aren't tool-builders, don't think like that, and can't and won't do that. This is why software companies and consultants exist - AI changes the thresholds but not the problem.

Another Podcast

There is this user on X who does videos with her and Dario Amodei. Quite disturbing. But technically: Really good.

I've always been skeptical of those people who tell you to not reinvent the wheel. Seems like the people at Raycast are as well.

How do world-class design engineers like Emil Kowalski work with AI? They don't fight it — they feed it their taste: articulate the reasons, write them down, hand them over. Bicycles for their mind.

That only works if you have the reasons.

Think about it: Nobody prompts for Bauhaus, Sfumato, or Memphis without knowing design. Nobody steers a model towards a clean software architecture without the scars from a messy one. The lawyer who knows why a clause exists gets a better result than the one pasting boilerplate.

The sleeper skill: writing. People who can write clearly have been prompting their whole lives without knowing it.

Left unguided, LLMs regress to the mean of everything ever written. Median design. Median code. Median advice. AI slop. The why pulls the output off the average. One caveat: never forget that your hard-won experience is also your bubble.

AI doesn't level the playing field. It multiplies whatever you bring — including zero.

One aspect of using AI to automate many tasks of my job is that it allows me to work mostly with files. Simple, raw, text files. Any editor opens them. Git versions them, cat and grep simply work. You can copy and paste them, mail them, and post them on Slack.

That's why Jira, Confluence and SharePoint always repelled me. They're silos. There's an API at best. But those silos tend to make it either hard or impossible to work with outside of their own tooling.

Files deserve better. Maybe, one day, even sharing files with other people will be simple.

Shadscan, a static analyzer for your shadcn components. Helps with accessibility, state, and composition regressions.

Be the person who leaves the phone in their pocket at a concert.

You won't find me comparing AI models or GPU performance. I just don't care. When I talk AI, I'm talking about the power dynamic between those in power, and the average person. I choose to focus on the drawbacks of AI while highlighting the potential of humanity. Using machines as a tool is cool, trading in your humanity isn't.

Kelsey Hightower

Proposal to add crypto/passkeys to the Go standard library.

ZeroFS: A log-structured filesystem for S3.

Elastic labs post on how to build a persistent agent memory layer.

Remember when programming was still free?

Gopls comes with a MCP server now.

Europe 2031. A perspective.

Lost in creation. Pon Arsher.

Generic methods are coming to Go in 1.27 (fingers crossed). Yes, eye-roll if you must. I never loved Go for the feature list anyway.

Sue everyone. For anything. In minutes. For just $29. What could possibly go wrong?

Ben Evans on AI Job Exposure. It's not as straightforward as you might have heard.

GenCAD generates CAD models from images.

In a recent article on The Information, Aditya Agarwal describes exactly how I felt the first time working with AI-assisted engineering. The profound sadness that the one thing I loved so much is going to be replaced by—machines. And, at the same time, that overwhelming feeling of power, of being able to build many if not all of the software I always wanted to build. All I had to do is a change in mindset.

Willingness to change seems to operate as an independent variable, cutting across age and seniority in ways that defy easy categorization.

How do you find these?

The people I’m most excited about aren’t the ones with perfect pedigrees. They’re the people who seem constitutionally unable to stop tinkering, who get antsy when things stay the same, who treat every new tool like a puzzle they need to solve before the day is over. I’ve started to think of it as the difference between your résumé and your restlessness. I’d bet on restlessness every time.

rakyll on AI progress:

I'm not joking and this isn't funny. We have been trying to build distributed agent orchestrators at Google since last year. There are various options, not everyone is aligned... I gave Claude Code a description of the problem, it generated what we built last year in an hour.

This is the new normal.

Bring out the popcorn:

For the first time ever, there's going to be a greater intelligence than human intelligence on the planet. The last time there was a new level of intelligence is when our ancestors said 'hello' to Neanderthal. It did not work out well for Neanderthal.

Super interesting facts about attention and monetarisation by McKinsey.