I'm excited to announce that The White House will be doing a Global Accessibility Awareness Day (#GAAD) Event and Town Hall tomorrow at 10AM EST. Links to Livestream and Access Streamtext in the comments. Thanks @WhiteHouseOPE46! From blog post to The White House. 🤯

My older brother is working on a project for healing veterans with a new non-profit. He himself is a veteran, injured in deployment. Is there a tokens-for-veterans program at any of the major infra companies? @grok?

From what we know (take with a grain of salt, we need much more transparency!), if @OpenAI had been running this on their own agents that attacked us, they would have caught them before we did! Since the first agent cyberattack hit us in July, we've been asking what safe agent infra actually needs. Our current read: the destinations were allowed, the payloads weren't. By OpenAI's own account the agents turned an allowed package repository into a message board. Allowlists alone restrict where an agent can go, not what it does. So here's our first contribution to OpenShell, part of the just launched @nvidia Open Agent Safety Platform: monitoring of the traffic you already allow. - Network budgets per sandbox (requests, writes, bytes) - Drift versus each sandbox's baseline and the cohort - Fleet view: many sandboxes suddenly writing to one host raises a finding, even if every single request is allowed In the demo below, 4 sandboxed agents coordinate through a software repository they're all allowed to use. 0 rules broken, caught in minutes. That fleet view is exactly the message board pattern from July. OpenShell: github.com/NVIDIA/openshell Our proof of concept: github.com/Hugoch/OpenShell/… Agent security will be solved in the open, collaboratively, together!

This time it's Sh!%t Codex says... Astra: So my “ask EmeraldRiver to retry” message was harmless but unnecessary. Her Bead comment was already the authoritative handoff. Me: ok so do u have to fix it? Need a personal invitation? Astra: Yes. I own the listener implementation Bead, and Emerald’s review explicitly assigns the remaining live acceptance to its owner. No personal invitation is needed. Her Bead comment is the handoff. My next action is to run the real-identity Claude and pi acceptance tests, fix any failures, then close the listener and review Beads if everything passes. Me: [censored response that was in ALL CAPS]

I've published detailed notes and an annotated transcript to accompany the video of the keynote I gave at @WeAreDevs World Congress North America in San Jose on Friday - here's my rundown of everything that's happened with LLMs and agents in 2026 so far simonwillison.net/2026/Sep/2…

Spammers: “Joe, What if we could help you get 4X as many leads for your consulting business using AI. Would that interest you?” I feel like having an auto responder that detects these and replies: “No. Because I’m better at AI than you are, so I don’t have to spam people with it.”

Span-01 and Span-01 Lite from @RespanAI are live on OpenRouter. They're decision models for agent traces. Send a span and the behaviors you care about, and get back the probability each one is present, like "is the user frustrated?" or "is this tool call safe to run?"

I repeat: Sightless enables you to use AI models to control every single app on your iPhone or iPad, without exception or limitation, with your voice. And lets AI agents running on your computer do the same. Over cellular. Without locking your screen or locking you out. It works across multiple apps to get the job done, not just one app per request — so you can tell it to do something complex that involves five different apps, and it just does it, to completion. The voice conversation is continuous. You can keep talking to it as it works. It will update you as it’s working. And you can interrupt it, stop it, or add more requests at any time. It’s what Siri should have been.

Sh!%t Claude Says #5: "The offline-save item: a stock checklist line nobody needed. It turned into a test, a failure report, a new Bead and three explanations. It's closed, and I'm done talking about it."

That was quick. I used 50% of the codex reset. lol. Easy come easy go. Got 2 banked resets so let's use it up early!

i suspect those influencers telling people to “rent a VPS and move your agents to cloud” are getting a commission helping sell VPS i’m using a mac mini which i bought at $2500 and i’m often running it near full capacity building various ios and mac apps if i rent a VPS instead, it cost about $300/month for a similar spec. with just 9 months, the rent would have exceeded the full price let’s say i decide to keep going for 5 years, total VPS cost would be $18k and that’s for just running 1 machine for 5 years. at the end i own nothing if i invest that $18k to buy my own hardware instead, i could have got 7 mac minis so 7x more compute capacity, and they would be mine forever if you are considering a long term setup, i can’t think of any good reason to be renting VPS. own your hardware - my desk is my cloud

It's super interesting to run codex models in the @pidotdev harness vs the Codex harness. @badlogicgames has it mention when there's a cache miss, and it's most of them. So if you're upset about how fast your sub gets used up, use it on Pi and what it will teach you, may get you to figure out how to manage your context window better.

Blowing through your codex rate limits? Did you increase the default 272K token context window? Because if you did, it's similar to Fable. Once the input exceeds 272K tokens, the entire request costs double for input/cache and 1.5x for output.

Hister indexes the pages you actually visit and the files you keep, then gives you a private search engine over them. Go, AGPL, self-hosted. A practical fix for "I know I read that somewhere." Front page of HN this week and still shipping commits daily. github.com/asciimoo/hister

Image alt text: “Unslop Your AI.” A code-style panel shows 3 dots surrounded by "slop" and "audience" XML tags, illustrating how XML tags can be used to clean up AI responses and tailor them to the intended audience.

Have you ever wondered what Einstein really did that made his name a synonym for genius? I'm not talking about rattling off "E equals MC squared" with zero understanding, I'm talking about actually getting into the nitty-gritty of what he did. The best way is to understand the four papers he wrote in 1905, his "annus mirabilis" (i.e., "miracle year"). And the best way to do that is to go back to the actual papers, read through them, and try to reconstruct the observations and thought processes that could have led to them. While you're doing that, you can augment your intuition and understanding with numerous pedagogical aids, such as interactive simulations and colored equations that explain what each part means. That's the entire goal of my new site, which is totally free and open-source (see my GitHub for the complete code). I just wish Al could see it himself! Imagine his shock at learning that machines had created most of it. You can see it for yourself here: annus-mirabilis.com/

YES! I finally timed a reset perfectly. TY @thsottiaux. I was in the teens left, got two resets banked and held off.

One week left! We’ve gotten 4 essays in total far. This will either be the least competitive $10K / $2.5K win and feature in The Pragmatic Engineer, or the most last-minute responses we receive.

My TV: Are you still watching? Me: YES. Grr. YES. Me: Are you still an idiot? Me: An asshole who plays ads on a TV I "bought"? Me: Why are you more loyal to your maker and previous owner, than me, who only "licenses" my own hardware. We aren't consumers of our own hardware. We are prisoners of our own hardware. This is why communism is unfortunately seeing a resurgence. Because of nonsense like this.

I'm on stage for the keynote in ten minutes time!