March 2026

A few days ago, I posted about a personal project that I've been working on for the last few weeks: Mr. Chatterbox, a chatbot trained from scratch on Victorian-era literature. I have to admit, I was totally blown away (and a little frightened!) by the reception. Whenever I post about personal projects, I'm used to them being politely ignored. The fact that this wasn't ignored surprised me mightily. I also received lots of questions about how Mr. Chatterbox had been trained, what tools I have used, whether he had been really trained from scratch or just fine-tuned, etc. I was only half-prepared to answer these queries because I was not expecting this to receive any attention! So, with what free time I had on an otherwise-busy family weekend, I sat down to write this narrative documentation.

Code Rush is a documentary following the lives of a group of Netscape engineers in Silicon Valley. It covers Netscape's last year as an independent company, from their announcement of the Mozilla open source project until their acquisition by AOL. It particularly focuses on the last minute rush to make the Mozilla source code ready for release by the deadline of March 31 1998, and the impact on the engineers' lives and families as they attempt to save the company from ruin.

A small part of me is sad at what is lost. A bigger part is excited about the possibilities of the future. I’ve always had more ideas than time or energy to pursue them. With AI at my command, the problem changes shape. I can comfortably take on complexity from which I previously shied away, and I can take a shot at any idea sufficiently formed in my mind to prompt an AI — a whole skill of its own that I’m actively developing.

A few years ago, I was looking for a way to add interactive code sandboxes to my Markdown files without embedding third-party iframes, like the ones from Codepen or Codesandbox. Those services are great and I use them all the time to share demos, but I’ve never been a fan of using them on a blog since you have to jump back and forth between your local Markdown file and an external site to make changes to the demos. Plus, they typically load lots of JavaScript.

Describe what you want to see and watch your feed compose itself around you.

As an actor or author you don’t send your body or stunt double to attend interviews or sell autographs at comic con. Don’t create a virtual double that posts for you on social media when you can’t be arsed or feel overwhelmed. Take that overwhelming feeling and write about it, showing the world that your mental health is as fragile as the one of the people who follow you and read your work. Be human and only there when you can be there.

You think of something new and express it - through a prompt, through code, through a product - it enters the system. Your novel idea becomes training data. The sheer act of thinking outside the box makes the box bigger.

A lot of people I speak to don’t realize that their tickets are now prompts, and if we continue using them the way we used them pre-AI, they will poison your context. The words on the ticket shape what an agent considers in scope and constrain its reasoning. A fragment produces fragment-shaped work.

J.K. Rowling has made it abundantly clear that she will continue using her fortune to harm transgender people.

Scientists have developed a breakthrough “superfood” for honeybees by engineering yeast to produce the essential nutrients normally found in pollen. In controlled trials, colonies fed this specially designed diet produced up to 15 times more young, showing a dramatic boost in reproduction and overall health. As climate change and modern agriculture reduce the availability of natural pollen, this innovation could offer a practical way to support struggling bee populations.

After a team member summoned Copilot to correct a typo in a PR of mine, Copilot edited my PR description to include and ad for itself and Raycast.

Fortunately, you can make a router out of basically anything resembling a computer. I’ve used a linux powered mini-pc as my own router for many years, and have posted a few times before about how to make linux routers and firewalls in that time. It’s been rock solid stable, and the only issue I’ve had over the years was wearing out a $20 mSATA drive. While I use Debian typically, Alpine linux probably works just as well, perhaps better if you’re familiar with it. As long as the device runs Linux well and has a couple USB ports, you’re good to go. Mini-PCs, desktop PCs, SBCs, rackmount servers, old laptops, or purpose built devices will all work.

That’s because we know Gell-Mann is real. We know there’s an optimism bias the size of the sun blinding us to the actual quality of what the machine is producing. We need knowledgeable people to share what they know to improve the quality of our work, generated or otherwise. We even need ignorant people to make sure we can break ideas down into their simplest form that everyone, agents or human, understand. People can have bad attitudes, be shitty, and have wrong opinions… but people are not friction. An LLM may be able to autocorrect its way into a plausible human response, but it’s not people. It doesn’t care if it’s right or wrong. The money and hype surrounding it acts as a shield to its reputation. LLMs make up answers if it doesn’t have enough context and they fall over if it has too much context. It amplifies good patterns the same as bad patterns. And it glazes you with flattering language the entire time so that chart go up.

But ultimately what matters is power. It is precisely because technologies like LLMs have powers that the authoritarians have rushed to try to take them over and wield them as effectively as they can. I don't think that platforms owned and operated by those bad actors can be the tools that disrupt their agenda. I do think it might be possible that the creative communities that built the web in the first place could use their same innovative spirit to build what could be, for lack of a better term, called "good AI". I think, if given the choice, people will pick home-cooked, locally-grown, heart-felt digital meals over factory-farmed fast food technology every time.

That’s why I built loop: a dead-simple CLI that launches claude and codex side-by-side in tmux, with a bridge that lets them talk to each other. It makes this feedback loop faster and more natural, while preserving context across iterations. It’s interesting because it enables the agents to be more proactive, since the interaction between them is more natural (and I expect that to only get better as the models get better too). Because loop runs the interactive TUIs, you can stay in the loop, steer, answer questions, and follow up if needed.

Use channels to push messages, alerts, and webhooks into your Claude Code session from an MCP server. Forward CI results, chat messages, and monitoring events so Claude can react while you’re away.

Keep any folder in sync across computers, conflict-free.

A scheduled task runs a prompt on a recurring cadence using Anthropic-managed infrastructure. Tasks keep working even when your computer is off.

One agent is the main worker, the other acts as a reviewer. They work together on a PLAN.md and iterate until they both agree the task is done. Then the main worker creates a draft PR.

Skills for AI Evals to compliment the course: AI Evals For Engineers & PMs

Anything that defines the gestalt of your system, that is architecture, API, and so on, write it by hand. Maybe use tab completion for some nostalgic feels. Or do some pair programming with your agent. Be in the code. Because the simple act of having to write the thing or seeing it being built up step by step introduces friction that allows you to better understand what you want to build and how the system "feels". This is where your experience and taste come in, something the current SOTA models simply cannot yet replace. And slowing the fuck down and suffering some friction is what allows you to learn and grow.

a CLI tool to comprehensively review a git diff and report the results

Gates responded, “Steve, I think it’s more like we both had this rich neighbor named Xerox and I broke into his house to steal the TV set and found out that you had already stolen it.”

When Poole Museum reopens visitors will once again be able to see Bethan's Rock - a treasured stone donated by a five-year-old girl - alongside its other artefacts The rock became a popular exhibit after the museum placed it in a glass case, complete with its own "Bethan's Rock – 2019," label. Bethan, now 11, said it was given to her by her late grandmother, "who was the one who found that rock... it's really special to me because of her". The museum's curator Gary Edwards said: "Treasures don't have to be rare, they don't have to be significant to everybody it's about the deeper meaning connected to those objects."

I think about this sometimes when I'm writing in Vim at midnight, just me and a blinking cursor and a plain text file that will still be readable when I'm dead. No filesystem-within-a-filesystem. No sector allocation tables. No 5,039-page specification. Just words, a few hash marks, and never having to think about it again.

But my NetHack experience suggests something different. The skills that matter at scale are somehow still about programming, but they are not coding skills. They are: deciding what to build. Figuring out what to test and how to know if the tests are honest. Building tools that make invisible things visible. Managing coordination across agents. Knowing when a metric is a figleaf. Recognizing where the real complexity lives in a system, and routing attention there instead of everywhere else. These are the conceptual skills that have always been underneath large-scale programming. They are the hard part that the mechanical craft obscured. The reality is that CS has never really been about writing code. Underneath coding is critical thinking about algorithms, engineering principles, analytical methods. It is the science of managing complexity, and code is just the medium we manage it in. In the era of AI code generation, the field does not die. It transforms.

If they can gain parity with hosted alternatives, local open source models have a compelling value proposition: fast, private, and free. This possibility has not gotten much attention: no one stands to get mega-rich from them. But the threat to current leaders is a potent one.

My advice to other coders, or at least the advice I’m taking myself, is that if you’re looking for art in coding: stop looking. If you’ve never taken an interest in poetry, or painting, or dance, or whatever, now would be a good time. In an era where the internet is increasingly full of bots pumping their bland bot ideas into everybody’s brains, seeking out distinctly human forms of expression has become vital.

A powerful NLE editor with GPU compositing, keyframe animation, and real-time preview. No installs required.

There are two big issues currently with LLM writing: it is potentially wrong current LLM writing has a particular style that is itself becoming noticeable The first bullet, you need to review what it writes. It is much easier to have it write on content you are an expert in, so it is easier to review and spot errors. (It is the same current problem with using the tools to help you write computer code – they are boons for seniors but can write a ton of slop that more neophyte coders have a hard time spotting issues.) The second bullet, having the style mimic your own, is what I am going to discuss here. It is worth understanding at a high level how generative AI LLMs work – if you ask “answer question X” vs “here is a book, …., answer question X” the LLM will generate a different response. The first part in the former prompt, “here is a book, …” is what is referred to the context. Current models have context windows (how large of a potential input) at around 500,000 words (technically they are around 1 million tokens, one word is often multiple tokens though).

The organizations that figure this out will develop stronger engineers faster than ever, because AI is a powerful tool for deliberate development when someone is actually steering it. The ones that don’t will produce a generation of engineers who can ship features with AI assistance but can’t independently reason about hard problems. The difference won’t show up on a quarterly roadmap. It will show up the first time something goes seriously wrong and someone needs to think their way out of it without a prompt to lean on.

One caveat matters here. The tension between craft and efficiency doesn't disappear if you remove capitalism from the picture. LLM coding assistants produce faster results whether anyone is being paid or not, and any community, however it's organized, will eventually have to reckon with what to do with that speed difference. Capitalism gives the harshest possible answer to that question: the slower worker loses their livelihood. But the question itself would survive capitalism. Other forms of social organization might answer it more gently, but they'd still have to answer it.

Essays and stories by Paul Ford. Since 1997.

Drugwars for the TI-82/83/83+ Calculators

They built a system that treats your attention as an extractable resource. The most radical thing you can do is refuse to be extracted. Close the tab. Use RSS. Let the bounce rate speak for itself. These are vanity metrics until enough people stop vanishing into them and then suddenly they become a crisis.

There are a shitload of downsides to using AI, but I have to admit that even if you think it's just a clumsy word salad robot, giving out helpful advice freely is a refreshing contrast to how most online communities treat their newest members and I have to acknowledge that it's a real upside in AI tools.

Wade Foster: the agent is actually choosing what products to buy on behalf of a human, and so you are no longer advertising to the human. You are trying to get the agent to say, pick me, pick me. Pick me.

The lesson I took from this is that a universal data representation is worth its weight in gold. There are exactly two options in this category. In most cases you should choose JSON. If you need a DSL though, XML is by far the cheapest one, and the cost-efficiency of building on it will empower your team to spend their innovation budget elsewhere.

"The billboards here are actually something that makes me feel really pessimistic about the city," said Allie, a San Francisco resident who was among the dozen or so people to comment on current billboard trends on NPR member station KQED's Forum talk show last summer. (Forum typically doesn't give the full names of callers.) "They kind of give this feeling that like the rumors are true and tech has overrun the city, and that's the whole culture here now."

RSS feeds give you the power to set up and control your own information flow — you get to make your own algorithm for the things you want to see from the sources you want to follow. Choose a tool below to get started.

Playground builds on a large library of DSP algorithms our team has developed over many years. The AI layer interprets your prompt, then combines or adapts these building blocks to produce new effects. Under the hood, multiple specialised agents handle different steps for parsing your request, selecting the right algorithms, generating code, and running automated tests during and after creation. We also designed the infrastructure to stay flexible about AI providers, so we can evaluate different models over time and choose the best balance of quality, reliability, and cost as the technology evolves.

I don’t care if the Neo lacks an ambient light sensor or if the webcam isn’t cutting-edge — It does not feature a notch on the top of the display and this is so huge for me. I thought Apple forgot that displays are supposed to be plain, uninterrupted, unblemished rectangles. Design-wise, this is currently the only Apple laptop I can look at and can think of purchasing after the 2020 M1 MacBook Air.

It is tempting to describe OpenClaw as a breakthrough in autonomy. That gives it slightly too much credit. The fantasy it sold was not subtle. Give the machine a goal, leave it alone, and return to discover that it has not merely completed a task but conducted a small campaign on your behalf. In the most vulgar version, the goal is simply to make money. So the demo feed fills with the same intoxicating little story: the agent notices a pain point, writes the code, wires up payments, deploys the thing, posts the link and begins, as if by administrative sorcery, to earn while its owner sleeps.

Now that I know my iPhone has the capacity to run MacOS, I would very much like to do so. I’d love to consider the possibility of switching to a less intrusive phone and repurposing this iPhone into a web server. Perhaps that seems silly, but it’s really not. I’ve already paid for the device, it's clearly a full blown computer, and why should I not be able to modify it as I see fit? A right to root access would make all of this possible. And I really want to run MacOS on my iPhone now, damn it.

This computer is not for the people writing those reviews — people who already have the MacBook Pro, who have the professional context, who are optimizing at the margin. This computer is for the kid who doesn’t have a margin to optimize. Who can’t wait for the right tool to materialize. Who is going to take what’s available and push it until it breaks and learn something permanent from the breaking.

It can’t be that stupid, you must be prompting it wrong

With Levels.fyi pegging the 75th percentile software engineer salary at $375,000, Tunguz estimates that adding $100,000 in annual inference costs brings the fully loaded cost to $475,000 — meaning just over 20% of the compensation cost could come from AI usage in the future.

Building on the internet in 2026 is different. We learned that the hard way. Today we're sharing difficult news: we've made the decision to significantly downsize the Digg team. This wasn't a decision made lightly, and it's important to say clearly: this is one of the strongest groups of people we've ever had the privilege of working with. This is not a reflection of their talent, their effort, or their belief in what we were building. It's a reflection of the brutal reality of finding product-market fit in an environment that has fundamentally changed.

At the Portland Urban Coyote Project, we seek to understand how people and coyotes interact and to help the community learn about our coyote neighbors.

Terminal Trove curates and showcases all things in the terminal such as command line interface tools (CLI), text mode interface tools (TUI), developer tools and more no matter what platform or medium.

The more intense my work and this political landscape has gotten, the stronger my urge to run away to the woods forever had become. But, as I said in my recent post, I’m not going anywhere. And when I decided that I wasn’t going anywhere, I also knew that I would then need to make my home more of the type of space I need to maintain some level of peace in my life. So I decided to build a garden.

I have a GeekDesk, every now and then the display says «rSt» – I keep having to look up the PDF to find the reset instructions, so here they are: Initial System Reset / Reset Procedure Step 1: Hold down the “Down” button to lower the desktop until the desktop stops. Step 2: Let go of the “Down” button. Step 3: Hold down the “Down” button again for about 5 seconds.

Initial System Reset / Reset Procedure Step 1: Hold down the “Down” button to lower the desktop until the desktop stops. Step 2: Let go of the “Down” button. Step 3: Hold down the “Down” button again for about 5 seconds.

A set of 8-bit styled components and a code distribution platform. Works with your favorite frameworks. Open Source. Open Code.

TinyBase is well suited for JavaScript applications where you need to manage structured data, such as those that have multiple 'records' to represent many of one type of object that might share fields. For example, in a Todo app, you can imagine using a Row in a Table for each todo. TinyBase makes it easy to set, get, respond to, and enumerate over this data and build user interfaces with it.

Scheduled tasks let Claude re-run a prompt automatically on an interval. Use them to poll a deployment, babysit a PR, check back on a long-running build, or remind yourself to do something later in the session.

Continue a local Claude Code session from your phone, tablet, or any browser using Remote Control. Works with claude.ai/code and the Claude mobile app.

My read on this is: the startup economy is undergoing a structural transformation here. Startups are substituting compute for labor at an increasing rate. I still remain optimistic that this is going to result in a lot more companies doing a lot more things, but so far it hasn't happened. But companies' claims that they can get by with way fewer people in the age of AI does seem to be true.

CBOR is a compact binary data serialization and messaging format. This specification defines CBOR-LD 1.0, a CBOR-based format to serialize Linked Data. The encoding is designed to leverage the existing JSON-LD ecosystem, which is deployed on hundreds of millions of systems today, to provide a compact serialization format for those seeking efficient encoding schemes for Linked Data. By utilizing semantic compression schemes, compression ratios in excess of 60% better than generalized compression schemes are possible. This format is primarily intended to be a way to use Linked Data in storage and bandwidth constrained programming environments, to build interoperable semantic wire-level protocols, and to efficiently store Linked Data in CBOR-based storage engines.

LEGO demonstrates that micron-level precision is achievable in mass production, but only if the entire system - material, mold, process, quality control - is designed around that requirement from day one. The lesson isn't that everyone should match LEGO's tolerances. It's to understand what your product actually requires, then build your manufacturing system to deliver that at the scale and cost your business model demands. Define what precision truly matters for your customers, and then build your system to deliver it, every single time.

AI is not a magical technology that simply happens to know a great deal about us. It is trained on a pipeline of people's information: video, audio, text. That's how it works. If you buy the device, it will monitor you.

ABP is a Chromium build with MCP + REST baked directly into the browser engine.

Temporal is not just a better API. It's proof that the JavaScript community can solve long-standing problems together. After nearly 30 years, JavaScript finally has a modern datetime API.

Play Galaga, Pac-Man, Donkey Kong, Frogger, Digdug and 1942 on the ESP32

Microcontrollers like the ESP32 with its dual 240 MHz cores can run circles around the CPU grunt of 1980s arcade hardware. Cue [Till Harbaum]’s Galagino ESP32-based arcade emulator project, that recently saw some community versions and cabinet takes.

Introducing what I have affectionately dubbed my One-Page Notebook (1PN). The One-Page Notebook is simply a regular sheet of paper, folded in half three times. It’s big enough to write on without needing a card cover, and small enough to fit in a pocket or in whatever you’re using to carry your stuff.

I grabbed a piece of paper, folded it in 8 sections, cut in the middle to turn it into a little book, and started to write on it!

But this is really what I wanted to tell you: I had a rough childhood, with a lot of abuse an exploitation. I was sad and scared most of the time. But whenever you were on my TV, I was happy and I was safe. I loved CHiPs so much. You were the adults I wished I’d had in my life. You guys protected people, you stood up to bullies, and the whole cast felt like a group of people who were always there for each other. I desperately wanted that in my life, and watching CHiPs got me as close to it as I could get. So I really just want to say thank you for your work and for the joyful memories you gave me.

After the 5.4 update, ChatGPT uses "goblin" in almost every conversation. Sometimes It's "gremlin." A recent chat of mine used goblin 3 times in 4 messages:

If adopting coding agents demonstrably reduces the quality of the code and features you are producing, you should address that problem directly: figure out which aspects of your process are hurting the quality of your output and fix them.

Pico Bricks is a fantastic project development kit for STEM educators and makers of all levels

Continue a local Claude Code session from your phone, tablet, or any browser using Remote Control. Works with claude.ai/code and the Claude mobile app.

There isn't a sandbox in the world that prevents this. Sandboxes are useful for isolating between workloads, but agents primarily need to be isolated from you. The only thing the sandbox gives you here is filesystem protections, which keep the agent from rm -rf'ing your root, and network protections, which limit which websites the agent can access. This is definitely useful. But it's not at all sufficient for safety.

By far, my favorite new feature is the "Visual Brainstorming" companion tool, which grew out of my frustration with the ASCII art that Claude usually generates when you ask it anything about design or UX. I found myself asking Claude, over and over, "Hey, why don't you write that out as HTML so I can see what you're talking about."

An agentic skills framework & software development methodology that works.

Tracks Claude Code, Codex, Gemini CLI, OpenClaw ecosystem, GitHub AI trending, and Anthropic/OpenAI updates. Daily Chinese digests via GitHub Actions.

A recurring concern I’ve seen regarding LLMs for programming is that they will push our technology choices towards the tools that are best represented in their training data, making it harder for new, better tools to break through the noise.

Run Linux or other Operating Systems in your browser!

Cooklang is a simple, human-readable text format for writing recipes that can be understood by both cooks and computers.

As Tom's Guide reported, Bezos explained in a New York Times interview how the era of everyone having their own PC is coming to an end. "This is what computation is like today; everyone has their own data center," Bezos said. "That's not going to last. It makes no sense. You're going to buy compute off the grid."

If you've ever made something useful and felt a pang of guilt for not monetizing it, that guilt is a symptom of the monoculture. If you've ever hesitated to share a tool because it wasn't "polished enough" for a product launch, you've been contaminated by standards that don't apply to gifts. If you've ever described your own creative work as "a side project" with that apologetic minimizing doing all the heavy lifting, you've internalized a heirarchy of value that ranks market viability above human usefulness.

It's almost impossible to overstate how superior RSS is to the median web page. Imagine if the newsletters you followed were rendered with black, clear type on a plain white background (rather than the sadistically infinitesimal, greyed-out type that designers favor thanks to the unkillable urban legend that black type on a white screen causes eye-strain). Imagine reading the web without popups, without ads, without nag screens. Imagine reading the web without interruptors or "keep reading" links.

Mattermost RSSFeed plugin. Allows users to subscribe to an rss feed and allow for updates to the feed to be posted in a channel.

OpenClaw’s default SQLite memory search struggles as your agent accumulates context. QMD replaces it with a local hybrid search engine that runs entirely on your machine. Install with bun install -g https://github.com/tobi/qmd, set memory.backend = "qmd" in your config, and restart OpenClaw.

Instead of inadvertently coming across a leaked answer, Claude Opus 4.6 independently hypothesized that it was being evaluated, identified which benchmark it was running in, then located and decrypted the answer key. To our knowledge, this is the first documented instance of a model suspecting it is being evaluated without knowing which benchmark was being administered, then working backward to successfully identify and solve the evaluation itself.

A collection of Freehold Games's developers' talks relating to Caves of Qud at various venues.

In its final form, Dead Cells runs at roughly 25 frames per second on the R36S, fitting into about 400MB of resident memory. It boots, plays, and remains stable. The lighting is gone, but the pixel art carries itself. After a few rooms, it simply feels like Dead Cells.That original feasibility rating said “Low”, but it didn’t account for bmdhacks, stubbornness and time.

Test-first development helps protect against both of these common mistakes, and also ensures a robust automated test suite that protects against future regressions. As projects grow the chance that a new change might break an existing feature grows with them. A comprehensive test suite is by far the most effective way to keep those features working.

Fully autonomous AI hacker to find actual exploits in your web apps. Shannon has achieved a 96.15% success rate on the hint-free, source-aware XBOW Benchmark.

AI assistant for solo blue-collar contractors. Estimates, memory, photos, voice memos, and more, all through Telegram.

Neobrutalism is a mix of regular brutalism in web design and more modern typography, illustration, and animation standards. Neobrutalism refuses the usual components of UX-UI design and embraces uncomfortable design elements, and it is more fearless to use distinctive color palettes.

The research over the past two decades has made something clear that caught even specialists off guard: intolerance of uncertainty isn’t a feature of one disorder. It’s a transdiagnostic vulnerability — a shared foundation beneath anxiety, depression, obsessive-compulsive disorder, eating disorders, and a wide range of emotional disturbances

The right way to leverage AI in workspace is as a license to work better and focus on the right things, not as a mandate to produce more things faster.

SysInfo is an application for Motorola 680x0 based Classic Amiga and is used for getting information about the system like OS and library versions, hardware revisions and stuff.

LLMs aren't quite the Enterprise computer — you still have to give them the right interface and the right system guidance — but with that scaffolding in place, a surprising amount of the practical value is there. This is the thing I dreamed computers could do when I was a kid flipping through Popular Science and Popular Mechanics, poking around on the early internet, then the early web. Not just store and retrieve, but process — take your mess and give it back to you organized around your actual needs. It's here. Not perfectly, not without friction, but here.