Full Width 
Shortcuts 
Sign Up 
Log In 
Articles 
Reading List More from the jsomers.net blog
Now that we have actually good AI, I have this vision of a form of computing that doesn’t involve me using a computer so much. Imagine you had the day’s emails to go through. It would be nice if the ones that required a simple decision could be dispatched with a few pen-strokes: I could write down a date that would work for that meeting; check a box to accept that invitation; etc. If an email required me to review a draft, I'd love to mark up a print version on my couch, sans screen, and have those notes scanned and sent off as if I'd done the whole thing on Google Docs. The point is not to give up on virtuality, but just to save the end user from having to interact with it. It's great to be able to send information to anyone in the world instantly; but let me do it without the glaring screen and the thousand distractions. Is such a thing even possible? I'm not sure, but I just now uploaded a first draft of this post, handwritten in a small notebook, into ChatGPT. Its transcription was nearly perfect. A first draft of this blog post Here's another example: when I really want to play with re-arrangements of a complicated outline, or when I want to collaborate with someone else on one, I find myself laying physical note cards on a table. Having real objects to work with allows for more flexibility than software. If the problem demands it, you can stack cards, draw on them, cut them, tape things to them, etc.—and none of these improvised ways of organizing has to be coded up in advance. Space turns out to be a good way to organize information. I gather that in the paper days if you had a big project you were working on, say a book, it would spill out into the room you were working in: chapter outlines pinned up on the walls, stacks of books in meaningful piles on the floors, folders with drafts and clippings. Certain sections of the work in progress would become associated in your mind with certain parts of the room. Chapter 3 would be over there. I rarely use paper or physical space this way because it lacks critical conveniences. A huge wall-mounted paper calendar is maybe the best way to plan and visualize a large coordinated effort. (In "making-of" documentaries you find that movie shoots are often planned this way.) Everyone can see it and point to it because it's at human-scale; you can express many dimensions of information simultaneously using shape, color, position, size, and any other physical attribute. But I have a hard time understanding how I'd keep such a calendar up to date. In practice, like almost everyone else, I use a virtual calendar that automatically adds events as I'm invited to them, syncs with other people's calendars, reconciles time zones effortlessly, and interoperates with other programs like email. Could we get the best of both worlds? In other words, shouldn't one goal of rapid technical advancement be some melding of the physical and virtual worlds such that I can sit quietly in an easy chair with pen and pad; or lay cards out on a table to organize my thoughts; or turn a room into the embodiment of a project; and yet have the same flexibility, portability, persistence, and remixability as in the digital versions of these things? I spend a lot of my day on screens. There are many problems with these things, articulated well by Bret Victor in the context of his Dynamicland project. Screens are small, antisocial, and they have a tiny vocabulary of affordances compared to physical objects. Plus they have the problem that they make it difficult to just use your calendar, todo list, or map—or even just respond to a friend's message—without encountering something else along the way, like a social network, short-form video, Slack, the news, or some other notification. To state the obvious: your phone is the best place to keep your calendar and inbox and todo list because you always have it with you, but of course that makes it ripe for other intruders. Bundling makes your phone indispensable, but also a menace. If nothing else I'd like it if operating systems and web browsers helped me be less distracted and frenetic, instead of encouraging exactly that multi-tasking freneticism. When I opened my phone or computer, it'd be nice if it was constrained to operate in a mode purpose-built for whatever task I intended to use it for. If I want to look something up, for instance, my phone should be a look-up machine (ie no texts, no apps, no ads, just a place for my question and the answer). If I want to compose a word of the day entry, I should launch straight into a browser with the tabs I regularly use for that, including the CRM, Webster's dictionary, and the OED; if I want to work on an article, my computer should assume the form of a typewriter, word processor, or McPhee mode note-processor depending on what stage I'm at. In each of these modes everything else should disappear, inaccessible. At least then you could mimic in software that thing you get from physical objects—which is that they are usually built to do one, and only one, thing well. My alarm clock, for instance, is just an alarm clock; and that's what I like about it! Not too long ago I spoke to a roboticist for an article I was writing, who worked with large autonomous earth-moving machines—e.g. a retrofitted excavator that could lift a boulder, scan its every edge and dimple, then model how it would settle amongst other boulders in a retaining wall before placing it there. He imagined a future in which such machines enabled a return to natural materials in the built world. He talked about old stone walls he’d seen in New England. Those walls, made of loose rocks found in situ, are lovely and sturdy, and adaptive—constantly rebuilt as farmers go about their work and notice areas that need patching up, adding stones they find lying around. But this is the very reason such walls aren't really built anymore. They're too labor-intensive. We live in a prefab world because the scarce thing now is not material or money but "the works and days of hands." I am moved by the idea that our future could feel less futuristic than pastoral. High tech could save us from high tech. We'd go back to the old interfaces without giving up the conveniences of the new ones. Read, write, communicate, create—and hardly ever see or touch a screen. I'm not sure that'll happen or be what people want. But shouldn't we be thinking of ways to use the new magic to spend less time tapping and clicking?
When I first started writing professionally, for the Atlantic’s website, I taught myself “reporting” with a simple self-made curriculum unfolding over six or seven articles. The first two pieces I wrote from my head, with reference to things I already knew or to books I’d read. For the third, I layered in firsthand experience, when […]
On podcasts it's pretty common to hear something like this: So Alexander Hamilton has just finished law school, and he's trying to make a name for himself. He's only been in New York a few years. So he takes on this case... The problem with the past tense ("Hamilton had just finished law school, and […]
In 1963, the philosopher Edmund Gettier published a three-page paper in the journal Analysis that quickly became a classic in the field. Epistemologists going back to the Greeks had debated what it meant to know something, and in the Enlightenment, a definition was settled upon: to know something is to have a justified true belief […]
More in programming
Details of how the White House tried to suppress the vote by messing with mail-in voting. An early proposal was that all ballots would have to be sent via certified mail.
Today we are releasing the version 1.0 of Lexxy. Lexxy is a rich text editor for Rails built on Lexical. It already powers Basecamp, Fizzy and many others, and it will become the default editor in Rails. I recently presented it in Rails World (slides, video coming soon). This is the article version of my talk. Trix hit a wall Trix has been our editor since 2015, and every Rails app’s editor since Action Text shipped in Rails 6. It’s small and reliable, and it has served millions of people for a decade. But in the last few years our customers kept asking for features like tables or code highlighting, and we kept struggling to deliver them. The reason is the Trix document model. A Trix document is a flat list of blocks. A block is a line of text with some labels attached, like quote, bullet list, bullet. There is no tree, and a block can never contain another block. Nesting is an illusion: at render time, adjacent blocks with the same labels get wrapped together. A flat list of blocks with very limited extensibility options That design bought a lot of simplicity, but you can’t express something like a table with it. Two cells next to each other would carry exactly the same labels, so Trix would merge them into one. The model can say “this bullet is one level deeper”. It cannot say “this cell is different from the cell next to it”. A tables issue has been opened since 2015! The model just can’t do it. The second problem was maintenance. An editor built on contenteditable behaves differently in every browser and even changes from time with operating system releases. In 2024, three iOS releases in a row broke typing, dictation or the caret in Trix, and each one cost us real effort to work around. Check this one as an example. Why Lexical This was a conversation we had at 37signals for years: 2022. We started a project to add tables to Trix. We gave up after a week. The document model can’t represent two-dimensional things. 2023. I built a proof of concept with Tiptap inside HEY. We liked it, but we never started a serious project with it. 2024. We built House, our own Markdown editor, for Writebook. Not WYSIWYG, but WYSIWYM: what you see is what you mean. A wonderful editor for long-form writing like books or technical documentation. 2025. We tried House in another product, and it didn’t fit. For most apps, WYSIWYG was just the right answer. 2025. We had the discussion again, and this time we looked at the whole field. Four years of the same conversation David ruled out Tiptap, CKEditor and the other commercial editors: an open source core with features kept proprietary, and a sales team behind them. We didn’t want our editor to depend on somebody else’s licensing decisions. Then we found Lexical: MIT, from Meta, and very powerful. I spent two weeks evaluating it, and in May we made the call to go with it. A tiny core. Pick the rest. Lexical’s core has zero dependencies and weighs forty-two kilobytes. In a way, it validates the approach that Trix pioneered. The document is an immutable state you never mutate directly, you just get new snapshots when performing updates; contenteditable is an input device and a rendering surface, never the source of truth. It has a DOM reconciler to update the actual DOM very efficiently, and other primitives to deal with handling commands and node transformations. Everything else, from lists to tables to markdown, is a package in the orbit. Lexxy uses thirteen of them. Lexical solved the maintenance problem too. Meta’s products like Facebook or Instagram use Lexical and their user count is in the hundreds of millions. This means that even small issues with new keyboards and devices are fixed fast by the dedicated Meta team that maintains it. Furthermore, Meta’s business is the products, not the editor. We much rather liked this structure of incentives for the long-term investment an editor represents. The iceberg The plan was simple. Pick Lexical, wire it up to Action Text, add a toolbar and ship it. Well, it didn’t go exactly like that. What you envision, and what's under the water Lexxy today is thirteen thousand lines of vanilla JavaScript on top of Lexical. A great editing experience is very hard to get right. An editor is a machine where the user can change the state in a thousand different ways. For example. you have two images one after the other and want to put the cursor between them, but there is nothing there to put a cursor in. Or somebody pastes from Google Docs, and you have to turn a pile of inline styles and empty spans into clean markup. And then Safari, and Android keyboards, and the clipboard, and undo, and… From the first pull request to Basecamp 5 The first pull request landed in May 2025. Fizzy launched with Lexxy in December, and Basecamp 5 in May this year. Basecamp was the real test: twenty years of content written with Trix, and people who use the editor all day, every day. Zoltán Hosszú and Samuel Péchèr were the key people who made this happen. The took a very green version of Lexxy, added a ton of features (including Tables) and polish, and they fixed countless bugs. They also pulled off a remarkable milestone: seamlessly switching millions of Basecamp users from Trix to Lexxy. What’s included? We didn’t want a to build Trix with tables. We had Lexical and we had agents to help, so we wanted to be ambitious here. We went for the whole package. Features In terms of major features: A color highlighter, built in instead of this being a custom Basecamp extension, as it was with Trix. Tables, with an interface we worked hard to keep simple and accessible. Markdown. You type it, you get rich text. Code blocks with syntax highlighting as you type, in more than twenty languages. Image galleries you can navigate and reorder with the keyboard. Prompts. Type a character, get a menu: mentions, emoji, or whatever your app needs. Links by pasting a URL over selected text. Previews of attachments like videos and PDFs, rendered as your app renders them. Action Text Native Lexxy is also Action Text native. Action Text stores attachments in a canonical format that Trix doesn’t speak, so it translates on save and again on render. We taught Lexxy to emit exactly that markup. What you see in the editor is what gets saved, and what gets saved is what your app renders. Your existing content, attachments and views keep working. That opened another door. Action Text now talks to an editor adapter, with an implementation for Trix and one for Lexxy, so switching is one line: config.action_text.editor = :lexxy. Here the credit goes to Sean Doyle, who started that pull request before Lexxy existed and took it to the finish line with our input. It ships with Rails 8.2, and we hope other editors will use it too. Extensibility And you can extend Lexxy. Extensions are built on Lexical’s own mechanism, and this is not a second-class API: Lexxy itself is thirteen extensions, tables included, and Basecamp has nine more. class MyExtension extends Lexxy.Extension { get enabled() { … } get allowedElements() { … } get lexicalExtension() { return this.defineExtension({ name: "my-extension", nodes: [ … ], register(editor) { … } }) } initializeToolbar(toolbar) { … } dispose() { … } } Lexxy.configure({ global: { extensions: [ MyExtension ] } }) My favorite of how extensible is Lexxy are voice notes in Basecamp: you record, you see the waveform while you talk, and it becomes a player inside the document. About a thousand lines, without forking or patching anything. Performance Lexxy is fast, because Lexical is fast. In a ten thousand word document, Trix takes thirty-eight milliseconds to process a keystroke. Lexxy takes four. Above fifty milliseconds, the editor starts feeling sluggish. Compared to Trix, Lexxy brought a whole new performance regime. Milliseconds per keystroke by document size Accessibility Accessibility in Trix was not great. In general, building accessible experiences for rich text editors built on top of contenteditable is quite hard. We had a dream team to help with Lexxy accessibility. Bruno Prieto worked with Michael Berger, our accessibility champion at 37signals, to bring the bar to where we wanted it to be. Bruno is an outstanding programmer who happens to be blind, so he knows one thing or two about accessibility, and he delivered. As a result, in Lexxy everything is reachable with the keyboard. The editor announces itself properly to screen readers, and it gets a thousand details right so that the editing experience using a screen reader is fantastic. You can learn more about accessibility in our docs. Security The latest AI models have resulted in an unprecedented explosion of vulnerabilities found, and we took this thread quite seriously. Lexxy counted with programmers of the caliber of Jeremy Daer and Mike Dalessio helping to make it more secure. We have put a lot of attention to sanitizing the editor contents, validating attachment URLs and making sure that the types of attachments and nodes the editor support are allow-listed. Lexxy also comes with preliminary Trusted Types support, to offer CSP-level control over certain DOM manipulation APIs. The trusted types policy is there, but we are not enforcing it everywhere yet. Agents We started Lexxy using Claude since day one. A main lesson was that an agent needs to drive the editor like a user does. The best decision we made in this project was moving the system tests from Capybara to Playwright: three browsers instead of one, a suite that runs in seconds, and a much more faithful clipboard, keyboard and focus. This represented a tremendous improvement in how agents could close the loop by themselves. Write a test, see it fail, fix it, see it pass. We have more than six hundred browser tests today. Moving the suite to Playwright changed how fast we could write tests With a solid testing foundation in place, we could start fixing bugs in large batches. As mentioned, getting a text editor right implies a ton of work, and the kind of backlog we got at some point would have have buried us in pre-agent times. We also used agents to validate the fixes: an agent reproduces the bug in the public Lexxy sandbox, checks that it’s gone with the branch applied, and labels the pull request. Agents were essential to get Lexxy done with the people and the deadlines we had: we are a small company, and the same people were shipping two products in parallel. 275 cards closed The new Rails default We believe Lexxy is the best rich text editor out there right now, and we are going to make it the default editor in Rails next. If you’re starting a Rails application today, use Lexxy. If you’re using Action Text with a standard configuration, switch. It’s one line, and we’ve worked hard to make it seamless.
Reading my recent computing retrospective, I realised there was a big section missing: the people in my life that made an impact and helped shape my career. Outside my immediate family, one person made an outsized contribution, and I’m fairly certain that without his influence my life would have taken a very different path. The fact that I’m still here in 2026, still writing code and being fortunate enough to have a career in something I love is testament to him. So I’d like to take a few moments to talk about my old secondary school teacher, George Dryden. Denied Back in 1995, I had a problem. I knew I wanted to study computing at university and build a career out of my passion, but there was a snag. For those unfamiliar with the UK schooling system, when you’re 15-16 you take a set of GCSE exams in a broad range of subjects. After that, you pick around 3 subjects to really focus on over a period of 2 years. These are called A Levels, and they are a big step up and are meant to prepare you for a degree-level course at university. Admission to university is also governed by these results - if you want to study computing, you’re going to need a computing A-Level, and most universities will only accept you (or “make an offer”) if you achieve a certain grade. And whilst I had taken computing at a GCSE level, my school did not offer a computing A-Level course. I instead had to settle on “Design & Technology”, which just didn’t inspire me. Instead of working on my portfolio and projects, I spent most of my time daydreaming and writing code on the Acorn Archimedes computers that were the staple of every 90s UK school. No disrespect to the teachers - they were all awesome - but it just wasn’t for me. I was miserable, and by the end of my first year, I was well on my way to failing outright with my entire future plans seemingly going up in smoke. Someone noticed That’s when George stepped in. He’d taught me computing right the way through my GCSEs, and with no A-Level course on offer, that was officially where his involvement was supposed to have ended. It didn’t. He had noticed my constant presence in the computing labs - before and after school, during lunch breaks, free “study” periods - working on some little pet project or digging into RISC OS internals. I remember him as warm, with a wicked, dry sense of humour, and a refreshingly spiky attitude to authority - I always got the sense he’d worked out for himself which rules were worth taking seriously and which ones weren’t. And he always had time for me. I spent years pestering him with questions that had nothing to do with anything on the syllabus, and he’d always find a way to answer them that actually made sense. He was just as supportive of my odd little obsessions. At one point I’d got deep into the BBS scene, which I thought was the coolest thing ever, and decided what the school really needed was an internal BBS running on its own network. So I wrote one. It was deeply cringeworthy, obviously - but George helped me put posters up around the school advertising it, and even gave it a mention in assembly one morning. I think about five people in total ever checked it out. It didn’t matter: a teacher had stood up in front of the entire school and treated my weird little project like it was worth something, and that was a hugely validating moment for me. Off the books He recognised the passion, and eventually he made a suggestion: What if I quit the Design and Technology course, and instead attempt the A-level course myself? Personally, I also suspect he was enjoying himself. There was some internal school politics behind why computing wasn’t offered at A-Level in the first place - I never knew the details - and I think the prospect of one of his students simply going out and getting the qualification anyway appealed to him on two separate levels. It would get me where I wanted to go, and it would wind up exactly the right people. It wouldn’t be easy, he warned. The school would be against it, plus it was a two-year course which I’d have to cram into one year. I’d have to do it all myself - studying, lesson planning, coursework - he couldn’t help me in an official capacity, but he said he’d advocate for me and help where he could. If I had assignments, he’d send them off to be graded and would give feedback in his own time. He’d enter me in for the exams and also gave me a set of keys to the computer lab so I could use it whenever I needed. It was the first time anybody outside my own family had really shown faith in my abilities and encouraged me to take a stand. It was a pivotal moment for me - I realised if I wanted something badly enough I would have to fight for it, but I could still make it happen. I didn’t have to take “NO” for an answer - a lesson I think I picked up from watching him as much as from anything he ever actually said to me. After a few weeks of dithering, I took the jump. I remember a few awkward meetings with the school administration but thanks to his behind-the-scenes work, I was soon following my dream. An intense year And yes, it was bloody hard work. I had to condense an entire two year course into under a year, be disciplined enough to produce my own study plan, and be critical enough of my own shortcomings that I could focus my study where it was needed. I pretty much lived and breathed it for months straight and was more-or-less a permanent fixture in the labs or school library poring over my course books. I’d make lists of questions and chat to George over lunch, and he’d provide guidance and encouragement. It was a lonely way to learn with no classmates to compare notes with, no lessons to turn up to, and right up until the end I had no real idea whether any of it was good enough - but bit by bit, it started to feel like something I could actually pull off. And sure enough, in the late spring of 1996, I sat down in an exam hall with my fellow students, the only one with an A-Level computing question paper in front of me. The final exam went by in a blur - I can only remember a few of the questions now (and a peculiar obsession with the Pascal language) - but I do remember the euphoria as the invigilator called “time’s up, pens down, close your papers NOW”. I had done it. A few nerve-wracking months later, my Mum drove me into school to pick up my results. I ripped open the envelope and saw it - I’d passed with an A grade! I literally ran up the stairs to George’s office next to the computing labs to thank him personally. I’d taken my camera into school to take a few last photos for memory’s sake and snapped this photo of him before I walked out the school gates for the last time: A different path Because of him, I managed to get into my university of choice, studying computing with a focus on networks. Because of that, I landed my first job working as a “webmaster”, and my career since has been one of the highlights of my life. All these years later, it’s a real privilege to be able to get up each morning and actively look forward to working in an industry I love. Without George stepping up for me and encouraging me to believe in myself, none of that would have happened. I wouldn’t have had the career I have, and I wouldn’t be where I am now. I met my wife when we both worked at a software company - she sat at the desk behind me - so even my home and family life can be traced back to that spring of 1996. And the A-Level itself was only half of what I took away from that year. The qualification opened the door to university, but the lesson that came with it was every bit as important: that a “no” isn’t always the end of it, and that sometimes the answer can be argued with. I’m so proud of what I managed to achieve all those years ago, and even more thankful to have had someone like George in my life to put me on the right track. Mr. Dryden I did see him again after I left. He drank in one of my local pubs - a pub I’d been going to for a good while before I was technically old enough to be in it - and I’d say hello if I spotted him in there, mostly in the months before I moved away to university. After that it was only a handful of times. For years I’d find myself scanning the room whenever I was back home and in for a pint, half expecting him to be at the bar. At some point I stopped seeing him altogether and eventually moved across the country. The trouble was I never really knew how to talk to him outside of school. He was always Mr. Dryden, or just “Sir”, I don’t think I ever once called him George to his face! I was (and still am if I’m honest) fairly socially awkward, and I never worked out how to phrase the thing I actually wanted to say: that he had changed the entire direction of my life, and I wasn’t sure he knew it. So instead I’d say hello, and ask how he was, talk about nothing much, and go back to my friends. Epilogue Sadly, 3 years ago now, I opened the latest issue of my old school alumni newsletter to read that he’d passed away. The photo at the start of this article was taken from his obituary article and I read that he’d had a long illness and had suffered from dementia at the end. I did write to him years ago by email - I don’t know if he ever got it, or was in any capacity to understand what he’d done for me, but I hope so. There’s an old saying by one of my favourite authors (Terry Pratchett) that “no one is finally dead until the ripples they cause in the world die away”. In one of his books, a character keeps the memory of his son alive by passing his name along a series of telegraph towers. It’s in that spirit that I’m writing this post - I debated it for many years as it’s very personal to me and I also have no contact with any of George’s family so I have no idea what they would make of it all. But even though it’s 30+ years ago now, I will never forget him or what he did for me - and at least now, if someone searches his name it’ll be recorded here for as long as I’m alive and running this site. Thank you, Sir. George Dryden 1942-2023
Let’s step inside the kernel and understand how it implements copy-on-write and what are its implications for the performance of user-space systems
