Moments

Quick thoughts, observations, and links — a microblog by Jeroen Gordijn

April 9, 2026

Meta released Muse. Are they joining the AI race (again)?

OpenAI now also has a €100 subscription. It also gives access to the Pro model.

April 7, 2026

Anthropic Mythos is scarily powerful. It is not publicly released, but it shows what’s coming.

GLM 5.1 apparently outperforms Opus 4.6 on the SWE benchmarks.

image

Very interesting project: Glasswing

Using AI for the greater good, before handing it to the internet hackers and scammers.

Gas Town by Steve Yegge is now v1.0: https://steve-yegge.medium.com/gas-town-from-clown-show-to-v1-0-c239d9a407ec

GLM 5.1 pelican is pretty good: https://simonwillison.net/2026/Apr/7/glm-51/

With Mythos a bug in OpenBSD was found that exists already for 27 years :flushed:

April 6, 2026

Tried out Qwen 3.6 Plus for a bit. I must say, I’m quite impressed. Fast and good.

April 5, 2026

It’s no longer allowed to use the Claude subscription with any harness other than Claude Code or their desktop. It still seems to work with Pi, but you risk having your account blocked.

The check to see if you’re using the correct harness with your Claude subscription seems to be: look at the system prompt and check if certain sentences appear. If so, block!

April 3, 2026

Another Chinese model - Qwen 3.6 Plus - that is performing quite well according to the benchmarks. Need to check for myself. — source

March 31, 2026

Claude Code source code got leaked by Anthropic. 500k lines of code.

March 26, 2026

Running tools via bunx/npmx instead of installing via ‘bun install’. Always running latest version.

ARC-AGI-3 is out. All models score <1%. I’m curious to see how this will hold up in 2026. https://arcprize.org/arc-agi/3

OpenCode is making their features into plugins. This will seriously boost the plugin system. Will OpenCode become as customizable as pi? It allows users to build their own custom versions. I think this is where more and more software is headed.

Voxtral TTS by Mistral looks interesting. Need to look at it. Seems easy to clone your voice and use it in TTS.

March 19, 2026

Interesting. Are we going to remove all skills? I guess just the generic skills. I think company specific skills or references will stay relevant.

image

New product by Google: Google Stitch. Design a website or app with AI. Good video by Fireship. Tried it out and it seems really impressive.

March 16, 2026

I’ve written a new blog post! Put some effort into Closing the Loop

March 15, 2026

Seems like an interesting project: plannotator — make notes and submit to your agent

March 13, 2026

Nice blog Grief and the AI Split. I see many people who are resisting AI, because they feel love and pride in writing code. I understand. It feels painful losing the thing you really love. I hope many will find new joy in using AI, because that is the way we are heading.

Anthhropic just dropped 1M context window by default. No more extra pricing.

Retrieval is also really good.

March 11, 2026

Context engineering starts with the system prompt. How System Prompts Define Agent Behavior

Very interesting article and it shows again how important it is to work with the model, not the tool. We really should focus on tweaking how we address the model. Deliberate continuous practice.

Working with AI is more like learning to play the guitar than learning how a tool works.

Nice article to go with the previous one: Weird system prompt artefacts. The system prompts seem to be fighting against the default model’s behavior.

If you let the agent perform manual tests, can you still call it manual tests?

Agentic manual testing

Codex, File My Taxes. Make No Mistakes. How AI outperformed the human accountant*.

March 8, 2026

Finally took the time to read Welcome to the Wasteland: A Thousand Gas Towns. Very interesting stuff. Earning credits based on actual performed work. I still don’t fully understand what the Wasteland is or how to apply it, so I’m adding it to my backlog to dig deeper.

Also, Gas Town is now ready for use. In these changing times, I think we need new ways of working, and the Wasteland and Gas Town seem to facilitate that.

March 5, 2026

OpenAI api now has websockets. Nice explanation by Theo - t3․gg on YouTube

This is a great improvement, and I’m looking forward to the other providers supporting this and seeing what it will do for my workflow. Apparently, a 20%-40% speed improvement.

GPT 5.4 dropped. This will be an insane year if this goes on.

March 3, 2026

OpenCode is working on workspaces. Turn your agent into a control pane. Interesting.

March 1, 2026

Love the new Obsidian Headless sync. Made my automation of auto-tagging notes much easier. I no longer need a VM with X on it.

Received a new KPN TV+ Box, but the video was stuttering. After a back-and-forth and debugging session with Gemini, I discovered that IPv6 needs to be enabled on my FritzBox. Googling this would have taken a long time. No support desk needed.

February 28, 2026

Brutal blog post by Geoffrey Huntley:

If your company has banned AI outright, you need to depart right now and find another employer.

The world is changing and people should no longer identify as programmer:

This is going to be a really hard time for a lot of people because identity functions have been erased, and the hard thing is, it’s not just software developers. It’s people managers as well. If your identity function is managing people, you need to make adjustments. You need to get back onto the tools ASAP.

This is a must read and wake-up call for anyone thinking nothing is happening.

image

February 27, 2026

People talk about liking one model or another, then say they’ve been using it for months. You should really try new models every time they come out and see if your verdict still holds.

  • No mass surveillance
  • No fully autonomous weapons

Department of Defense threatened to remove Anthropic from their systems and mark them as a “supply chain risk.” if Anthropic does not let above points go.

Regardless, these threats do not change our position: we cannot in good conscience accede to their request.

Statement from Dario Amodei

Linear walkthroughs are an interesting idea. Need to test that tomorrow on a service I’ve implemented.

Tried the Linear walkthroughs today. That is really cool! The only problem I had was missing syntax highlighting. Luckily, my clanker came up with a 2-step solution for that. Really nice to have a guide to read the code.

I’ve been using GPT-5.3-Codex throughout the day. The context window of 400k vs the 120k of Opus (in GH Copilot) makes a real difference. Getting great results, and I’m starting to like GPT-5.3-Codex more and more. Would be nice if Opus on GH Copilot got a bigger context window too. I guess the request model is holding GH back.

February 26, 2026

Extended my pi today. Widget that shows what Skills got loaded & updated the statusline to show git status.

I don’t need to install pi (or any other agent) on my servers when it can just ssh into the machine and do its work 💡

After noticing how people can just replicate software, tldraw is now removing the tests from the open source

Anthropic is making big moves in computer use. Last week Sonnet 4.6 was released with better computer use capabilities. Now Vercept got acquired.

Moving tests of tldraw to closed source was apparently a joke

February 25, 2026

Research shows that AGENTS.md is mostly not useful. I have to experiment with that and see for myself. It sounds really counterintuitive.

February 23, 2026

Every good model understands “red/green TDD” as a shorthand for the much longer “use test driven development, write the tests first, confirm that the tests fail before you implement the change that gets them to pass”.

By Simon Willison

February 21, 2026

IDEs are dead. Just thought of a new term. We will be working in an AME (Agent Management Environment).

15k tokens/sec is insane: https://chatjimmy.ai

February 20, 2026

Pi just added this:

• Added default skill auto-discovery for .agents/skills locations. Pi now discovers project skills from .agents/skills in cwd and ancestor directories (up to git repo root, or filesystem root when not in a repo), and global skills from ~/.agents/skills, in addition to existing .pi skill paths.

Nice improvement. OpenCode supports it too. It would be nice if it didn’t stop at the Git repo root, but kept going up instead. That would allow for ~/projects/clientA/.agents and ~/projects/clientB/.agents — all projects for a specific client could inherit the correct skills.

I just made a release of agentdeps with .agents support for pi and opencode: jgordijn/agentdeps:0.5.0

I wonder why people who are pro-AI feel the need to tone it down by saying that AI still makes errors. You still need to check. AI cannot do this or that. It’s a user problem, not an AI problem. As JD says: PICNIM (Problem In Chair, Not In Model)

I can relate to this post by Simon Willison. I sometimes find myself searching for something I think I made. Did I make it? Where did I do it?

February 19, 2026

MCPs are a way to restrict what an Agent can do. Not to give it more power: MCP to restrict agents

And another model is released. Gemini 3.1 Pro. And Simon made another nice writeup. I wonder what the rest of 2026 will bring.

Found on JD: The Coding Agent Is Dead. Amp may be ahead of the curve, but if you’re still working in the IDE, you’d better pay attention. You’re doing it wrong. The CLI is also not it, but it is for now. Amp is going to self destruct the vscode extension. How cool!

They also mention this:

Think of it as a ladder: we use it to climb up to the next level and then we might not need it.

That’s why I made The AI Coding Ladder

February 18, 2026

Nice quote from the same article:

In a system where agent throughput far exceeds human attention, corrections are cheap, and waiting is expensive.

OpenAI Harness engineering

A Guide to Which AI to Use in the Agentic Era by Ethan Mollick. Using AI is no longer about talking back-and-forth with the chatbot, but assigning tasks to your agent.

The shift from chatbot to agent is the most important change in how people use AI since ChatGPT launched. It is still early, and these tools are still hard to figure out and will still do baffling things. But an AI that does things is fundamentally more useful than an AI that says things, and learning to use it that way is worth your time. Ethan Mollick

February 17, 2026

Building my own RSS reader with AI summary and filtering. Using Minimax via OpenRouter is a massive cost reduction compared to using Anthropic Sonnet.

Sonnet 4.6 dropped. Nearing Opus 4.6 for lower cost. Better than Sonnet 4.5 and Opus 4.5. Most information is about computer use, so apparently it’s very good at that.

image

Just saw this in JDs moments:

Agents optimize for outcomes, not attention.

Via https://x.com/garrytan/status/2023773343124537411?s=46

We need to build good services to win. Not the most popular. 🤔 Interesting thought!

One of the earliest lessons we learned was simple: give Codex a map, not a 1,000-page instruction manual.

OpenAI Harness engineering. This map/index keeps coming back. We need to provide the robot with indexes so it can discover the information it needs. JD is on the right track with his markdown references approach.

So instead of treating AGENTS.md as the encyclopedia, we treat it as the table of contents.

February 16, 2026

Peter Steinberger (creator of OpenClaw) is joining OpenAI to work on personal agents. OpenAI will apparently support OpenClaw. Sam Altman on X

Spotify says its best developers haven’t written a line of code since December, thanks to AI

https://techcrunch.com/2026/02/12/spotify-says-its-best-developers-havent-written-a-line-of-code-since-december-thanks-to-ai/

The question isn’t whether AI matches the very best engineer you know. It’s whether it’s better than the average engineer.

Thorsten Ball on X

Jarred Summer asking for issues to fix on X:

if you have a bun GitHub issue open for awhile with a clear reproduction, reply with a link below & I will have Claude try to fix it today

And a whole bunch of issues got fixed.

February 14, 2026

OpenAI introduced GPT-5.3-Codex Spark which can run at 1000 tokens/sec. But this has little value if your tech stack is slow in compiling and testing. Working with Kotlin and Spring with Maven, this is becoming a real pain.

I’m the bottleneck now: https://x.com/thorstenball/status/2022310010391302259?s=46

For the first time, I can see myself actually doing some work on an iPad. Echo SSH is really nice. SSH into the dev server and start pi inside tmux and you’re off.

Pi has /export. It creates a nice export of the session to an HTML file.

You build it, you own it. The cost of ownership in The vibe coding trap. Can’t the issue mentioned be mostly mitigated by a bot that inspects the internet for regulation changes and other shifts that impact your code? Of course, a fool with a tool is still a fool. Just blindly copying with AI will lead to issues.

The open-weight models are getting really good. All Chinese: Kimi K2.5, GLM 5, MiniMax. How long until we can really run these open-weight models ourselves? It performed really good on the vending machine simulation, with a different strategy than other models, investing more in the beginning, to later pivot and become really profitable.

February 10, 2026

After a remark about missing dark mode I gave the following prompt to Claude mobile app, in the code tab:

The website has no dark mode. I need the site to have a dark mode. It should listen to system preference by default, but it should also be possible to toggle manually. The toggle should be easy to see but not too much in your face. Make a PR for this.

Now the site has dark mode!

You build code with AI and you have a lot of information in the context how and why it is built this way. But after the session we throw this away. No more! Commit the session together with the code. New project by former GitHub CEO (Thomas Dohmke). Funded with $60mln in the seed round.

https://entire.io/

February 9, 2026

Pi is such a cool project. It makes me think in whole other ways about working with LLMs. No more magice, all logic is put in place by me. I still like OpenCode as well. But using Pi I now better understand what’s happening.

AI may make you more productive, but working more and more… I’m not alone: https://x.com/simonw/status/2020901645597683870?s=46

This shows the power of Pi, if you can imagine it, you can build it: https://github.com/nicobailon/pi-prompt-template-model

Building a little bot for slack with AI using Claude Opus 4.6. Without asking, it builds integration with OpenAI’s API. Why didn’t it favor Claude SDK? Now that I switch to OpenRouter it becomes apparent that OpenRouter and Ollama are all using the OpenAI API, so I think OpenAI API is just more common.

Just set up a Slack bot in about 1h to add messages to my moments (https://github.com/jgordijn/moments-slackbot). This message was created via Slack. The power we get! An idea is now realized in no time 💪

In rule form: • Code must not be written by humans • Code must not be reviewed by humans Finally, in practical form: • If you haven’t spent at least $1,000 on tokens today per human engineer, your software factory has room for improvement

Nice. No more coding by hand. No more reviewing. Tokens as a measure of effectiveness 😎

Tests can be reward hacked - we needed validation that was less vulnerable to the model cheating

Building a digital Twin of services to allow for massive integration testing. Create tests, not accessible by the agents. Instead of a Boolean test, write a probabilistic and empirical one. 🤯

Source: https://factory.strongdm.ai

February 8, 2026

Going to experiment with adding a microblog to the site. The idea is simple: quick notes and observations that don’t need the full treatment of a blog post. Sometimes you just want to jot down a thought. Inspired by the “Moments” section on aishepherd.nl by Jeroen Dee.

How to secure your agent? You can’t. You can try to put some instructions, but if it has bash access, it can just do about everything.

A huge part of working with agents is discovering their limits. The limits keep moving right now, which means constant re-learning. But if you try some penny-saving cheap model like Sonnet, or a second rate local model, you do worse than waste your time, you learn the wrong lessons.

— crawshaw.io

Nice hot take. Use the best models before complaining LLMs don’t work.