oh this is neat — Qwen’s newest image model decomposes images into layers (how photoshop works), and you edit just one layer and nothing else qwen.ai/blog?id=qwen...

I am increasingly starting to suspect that a good chunk of our social problems with social media is selection bias in disguise

Claude Skills are awesome, maybe a bigger deal than MCP simonwillison.net/2025/Oct/16/...

1/ 🚨 NEW NPM MALWARE CAMPAIGN. Yes, another. North Korea’s “Contagious Interview” campaign is escalating: 338 malicious npm packages, 50,000+ downloads -- 25 still live. Aimed at Web3/crypto devs & job seekers via slick recruiter DMs → git clone → npm install → compromise.

(The same thing extends to prompt injection. Human employees are often tricked, scammed or phished, and that’s the baseline that computers need to beat.)

Both these are true.

You can just make things on the web collaborative by adding a single HTML attribute...

i wrote about atproto and why it matters

This is a delightful following to lean's natural numbers game but for the reals adam.math.hhu.de#/g/AlexKonto... it's made by Alex Kontorovich and it's accompanied by a video series (a class recording) on YT youtu.be/U2oCgeVgd9M?... @danabra.mov you might be interested in this

12 exciting live-programming projects will be presented at LIVE this year. It's free, online, and coming up soon! Sat 9/27: Videos premiere Sat 10/4: Q & A and discussion over Zoom Details on projects are up already at liveprog.org; more schedule & links coming soon. Hope to see ya there!

fun fact: Lean can represent real numbers not floats (although Lean can do them). not doubles. not even arbitrary precision floats, but *actual* reals — which includes the squared root of 2, the exact number Pi (not an approximation!), fractions like 0.111... and so on. this still blows my mind

This is using the highlights API to color raw text like for example this code in a pre tag with no need for nested tags in between the code This is really cool because you can now make a very simple contenteditable editor with color highlighting without needing to fix cursor position on each edit

This is a cool thing one can do in lean the def part is the function definition and the theorem is the proof that age + current year = birthyear As soon as the function's behavior was changed (a +1 was added), the proof now shows an error

I'm up to 124 little mostly-vibe-coded apps on my tools.simonwillison.net site and most of them are trivial but some of them aren't, so I wrote up this guide to the ones I think are most worth checking out simonwillison.net/2025/Sep/4/h...

We did the thing. zed.dev/blog/claude-...

Exposing the Monsanto Conspiracy Our latest video, out now on YouTube youtu.be/CxVXvFOPIyQ

I've open sourced Ripple! github.com/trueadm/ripple It's super early days, not much works and there's little to no documentation or guidance. Plus the codebase is raw, but I at least wanted to share my ideas with the community. :)

I've had so many folks ask now, that I had to go for it. You asked me to build that ideal Inferno X framework that I described ages ago, so I went ahead and did just that. I built something that actually works and, well, it works well. Oh, and doesn't use signals!

a social media that is more like a wikipedia rabbit hole, giving you more and more posts that fit down that direction of latent space

which will burst first? the AI bubble… …or the “AI is a bubble” bubble?

my point isn’t so much that manual curation would “replace” algorithmic feeds. it’s that (imo!) curation and ranking need to be decoupled. the primitive i want (post lists) is the most basic form of curation possible. so that feels like the first step in decoupling them. like pinterest collections

Now I'm even more interested in learning Lean 🤩

😂

I believe we need a plurality of new guiding metaphors for browsers that recenter the end-user and provides a security model that supplants SOP. It's important to not throw away the internet we already built and start over, we must figure out how to reappropriate it.

🤣🤣🤣 Hermoso 😍 Es un musical de comedia que muestra la costumbre de la corrupcion al transitar en Venezuela youtu.be/cyE4l-JuXP8?...

Google's Genie 3 is a controllable world model and it looks incredible

You may not feel it, but the ground beneath you may be sinking. In fact, the surface of California’s Central Valley is nearly 30 vertical feet lower than it was a century ago. Why is this happening? Because our planet is rapidly drying. THREAD/

For software to be malleable, we need ways of building and modifying programs that are accessible to many more people. Are you working on new interfaces for programming? Share your progress at the LIVE workshop! Submissions are due July 21st. liveprog.org (Happy to chat if you have questions!)

canvasifying is a really cool demo

very interesting comment on the original thread mathstodon.xyz/@johncarlosb...

this post is so good My vibes resonate, we live in times where the more rebellious thing one can do, ironically, is to be an establisher

i feel like most people are seriously underusing markdown. it's not just for links and lists — it's a full-on superset of HTML! here are a couple of my favorite favorite tricks:

this cool, they're using git worktrees to run multiple instances of Claude in different branches to run them in parallel And they got a script that it seems to handle all the complexity of doing that so you don't even have to think about git worktrees

I really liked this blog post by @jeremiahdjohns.bsky.social: open.substack.com/pub/infinite... "Persistence beats talent" is honestly quite inspiring

Voice cloning is now trivially easy with open source AI tools that run on a PC, while live avatar videos of real people are easy with proprietary tools & a variety of open source tools are getting there. Very limited time to adjust legal & financial safeguards to new ways of authenticating people.

OpenAI are now under a court order to permanently preserve logs of temporary conversations or paid API usage (previously subject to a 30 day retention policy) - a new twist in the now 17 month lawsuit between the New York Times and OpenAI simonwillison.net/2025/Jun/5/o...

Yuck. Meta apps listen on localhost on Android so that tracking scripts can talk to apps, and Meta can monitor your browsing and tie it to identity. Why is that allowed? www.theregister.com/2025/06/03/m...

I put an LLM in whatsapp, and OMG it has been so useful in group chats, and I haven't even implemented tool calls, no multimodality, just boring text I also made a super cool chatbot that creates its own tools on-the-fly and runs in the browser privately Tried to use it seriously once, didn't work!

Note that the bsky.app offers a RSS feed for every account. Just add /rss at the end of the profile page E.g. your account has this RSS feed -> bsky.app/profile/did:... (deer.social does not support this feature currently)

In case you missed it, Niko and I wrote a round up of big news in biotech and medicine! (Apologies for the picture of screwworm flies.)

@aviva.gay introduced the ability for users to appoint their own "Trusted Verifiers" to their bluesky alt client, deer.social 1. Amazing turnaround, less than a week after we launched verifications! 2. This is a great way to explore new approaches to hard problems 3. Atproto is all about interop

Loved that quote "AI is the mirror most don't want to see

Watching o3 guess a photo’s location is surreal, dystopian and wildly entertaining simonwillison.net/2025/Apr/26/...

It is insane how scary the threat models of encrypted messaging apps providers are.

Ok, lets poke at this. Why do people want verification? It's in large part trust delegation: "I don't need to figure out if I can trust Z because I trust Y and they trust Z." We're busy! The world is complicated. Trust delegation is simply more efficient. 1/17 🧵 (17, really? 😬 whoops!)

Thoughts on Bluesky verification steveklabnik.com/writing/thou...

In atproto – the open network we’re built on – anybody can publish verifications. In the Bluesky app, Bluesky chooses whose verifications to show. This means that other apps could choose a different set of trusted verifiers, or use a different model entirely.

So for now, our app is hardcoded to say "only show verification from these specific accounts". Anyone can verify anyone else, but that doesnt get shown in the app because those users won't be in the apps 'allowed verifier' list.

Reminder: You can self-verify your account by setting your domain as your username. Since we launched this feature in 2023, over 270,000 accounts have linked their Bluesky username to their website. Here’s how:

Yeah, I called out Google's AI overviews as an example of a system that does NOT do this well in my piece: simonwillison.net/2025/Apr/21/...