Simon Willison

@antirez I wrote that piece before anyone had got it working, I need to revisit and test your stuff out! On a trip without fast enough internet right now so won't be able to try it until Monday

@simonw Hi! Just read your post on DS4, please note that you can run my GGUF 2-bit quantized right now if you wish: github.com/antirez/llama.… And a vertical ds4 inference engine is coming soon, I'm on it. youtube.com/watch?v=todMmp…

I added a new feature to my blog (built entirely on my phone with Claude code for web) that imports my iNaturalist photos and adds them to my site's overall timeline simonwillison.net/2026/May/2/sig…

@ARKeshet Because tokens are still pretty cheap, and reviewing what someone else's LLM did is a lot more work than reviewing what yours did where you remained in complete control of how you prompted it

@simonw > as opposed to firing up their own LLM to solve the same problem Why spend tokens and attention if someone did that already? If you don't understand or don't trust it, you'll get into the same spot. That would be better spent on contracts, admissibility and evidence.

The Zig project's rationale for their blanket ban on AI-assisted contributions makes a lot of sense to me - for them, time spent reviewing PRs isn't about the code, it's about growing new contributors for the future of the project simonwillison.net/2026/Apr/30/zi…

@Indian_Bronson Read further then the headline: "Using our privacy-preserving analysis tool on a random sample of 1 million claude.ai conversations" - they wrote about that tool Clio in detail a while ago anthropic.com/research/clio

“We looked at 1M conversations” Surprise!

@_catwu Is the model behind that regular Opus 4.7?

Claude Security is now in public beta, built into Claude Code on the web. Point it at a repo, get validated vulnerability findings, and fix them in the same place you're already writing code claude.com/product/claude…

@muhamedhkimi It's backwards compatible

@simonw "Exciting updates to your LLM library, but have you considered the potential impact on existing workflows that heavily rely on the original syntax? Some users might need to adapt to the changes, which could cause temporary disruptions."

I released LLM 0.32a0 this morning, a major backwards-compatible refactor of my LLM Python library and CLI tool for working with language models - the new changes should help LLM work better with reasoning models and other new frontier capabilities simonwillison.net/2026/Apr/29/ll…

@labomen001 Where do they say no security vulnerabilities discovered with AI? That's not in their code of conduct document ziglang.org/code-of-conduc…

@simonw I think you're missing one big point - they forbid *anything* AI-related, even security vulnerabilities (if it's disclosed that they're found with LLMs). So I think it's quite a bad policy as-is.

@rjchint @sarah_edo We've seen this pattern before though: SEO had huge positive effects on accessibility because suddenly people who didn't think about accessibility at all were incentivized to improve their HTML in a way that also benefited screen readers

@simonw @sarah_edo Like 90% of websites don't support accessibility features so its like hard to see that many websites support and MAINTAIN webmcp

I gave a talk last week and someone asked why we need an agent model for the web- The web needs an established interface layer. WebMCP, though not complete, is part of a strong foundation- it exposes structured endpoints agents can use directly, giving a clearer defined course. I, for one, don’t want agents YOLO’ing around like the wild west. We can build structured well lit paths for better security, rate-limiting, and give observability and control to users.

I particularly appreciate how this rationale isn't based on the idea that LLM code is of poor quality compared to code written by hand - the quality of the code isn't the deciding factor at all here

@sarah_edo I had a fantastic conversation a few months ago about WebMCP as accessibility technology I hadn't thought about it before but it feels like a great opportunity to massively increase the utility of many web apps for screen reader users under the guise of "adding support for AI"

@sarah_edo @Clara_Gold Right Pier 39 is being visited by a giant 2000lb Steller sea lion called Chonkers right now Go see Chonkers! ktvu.com/news/massive-s…

SF Native here- the whole post is about SF tech. Go to the redwoods. Go to Santa Cruz. Make friends with dog walkers, doctors, bartenders. Get involved in the food culture (and the after parties). Go see some live music. Stare out at the bay while eating seafood in mission bay. Have a picnic with friends in Golden Gate Park. Ride a boat in stow lake. There’s so much more to the city than the tech climb and grind. You can be in tech without having it swallow you whole.

6 months ago, I moved to San Francisco. It’s the best place in the world to build, and one of the worst places to stay human. My unfiltered take: 1. SF is both overhyped and underrated The overhyped part: there are a lot of people with incredible resumes who are deeply unimpressive in real life. They were at the right company, at the right time, in the right market, and got carried by the wave. They made money, got comfortable, and now spend their time “exploring opportunities” over coffee, wasting your time. The underrated part: the top 1% here is insane. But almost impossible to get. Hiring in SF feels like being a guy on a dating app: everyone you want is out of your league, and everyone in your league wants someone out of theirs. The best people have unmatchable packages, endless options, and are optimizing for maximum impact: labs, frontier companies, or startups raising $100M pre-seed rounds. If you raised $10M from Tier 1 investors, you’re not hot shit here. You’re a B-player. It’s humbling. 2. There are fewer mission-driven people than I expected Especially on the application layer. A lot of people are in “secure the bag before it’s too late” mode. And honestly, it gives me the ick. The real religious builders I’ve met are often in labs, hardware, biotech, deeptech, defense — places where the work is hard enough that you can’t fake obsession. 3. The status game favors builders This is what SF does better than anywhere else. It rewards obsession. It rewards weirdness. It rewards people who make building their entire personality. Europe punishes that. SF gives it status. If you’ve felt like an outsider your whole life because you care too much, work too much, think too radically, or refuse to be chill about things that matter, this city will make you feel less insane. 4. The market liquidity is absurd Even if you don’t build a billion-dollar company, if you manage to build a strong product with a great team, someone smart might still acquire you for $ 100M. Yeah I know, it’s not your dream outcome as a founder, but on the days you feel desperate, it helps to keep going. 5. SF does not care about the meaning crisis that’s coming Anyone paying attention here can feel that something massive is happening with AI. But I’m shocked by how little people talk about the meaning crisis coming next. Everyone wants to talk about AI liberating humanity. Almost no one wants to talk about what happens when work — the thing that gives most people identity, structure, dignity, status, and purpose — starts disappearing. The vacuum will not be peaceful. People are underestimating the chaos that comes from humans suddenly having no idea why they matter. And I really feel like no one cares. 6. Personally, I’ve never been more unhappy I moved to SF and entered the matrix. I’ve always been intense. I’ve always worked crazy hours. But here, I lost the last parts of myself that were not about building. I don’t go to events. Most networking events feel like theater for people pretending to be important. The only events worth going to are small, curated dinners with people who are actually alive. I’ve made 0 real friends. I don’t do well with transactionality. I don’t do well with people constantly performing greatness. I don’t do well with rooms where everyone is optimizing and no one is being honest. So yes, SF is lonely, transactional, delusional, addictive, inspiring, boring, extraordinary, and completely insane. But it is still the only place to be right now if you’re a founder trying to build the next wave of humanity. And for now, that’s enough.

@phoe6 Best explanation I've seen is still this Hugging Face article huggingface.co/blog/vlms-2025

Ask HN: What happens when you paste a screenshot, and ask questions in LLM? news.ycombinator.com/item?id=479407… Hello AI Twitter, Can you share your insights on this #askhn question. cc: @simonw

@jeremyphoward I wonder if Google internally are still hyper-resistant to offering an API for their core search engine

I'm so confused…

@idarbek @SpongeOfTruth Because if you don't offer backups your customers will chose someone else who does

@SpongeOfTruth @simonw Why market as one then? docs.railway.com/volumes/backups

The conclusions here feel wrong to me. The two lessons I see are: 1. Don't run agents anywhere they might be able to access production environment credentials - it's on you to know which credentials those are 2. Keep tested backups that are independent from your production host

@ArtyomAstafurov Where the heck did "Live Nude Chickens" come from?

@simonw Same prompt followed by More, Harder, Faster, and Ghibli.

Oh no

@pranesh I don't think so, this mode is a lot heavier

@simonw Whisper can run reasonably on CPU. Can this?

Microsoft's MIT licensed VibeVoice speech-to-text model (think Whisper with speaker diarization) is really good - my notes on running the 5.71GB 4bit MLX conversion on an M5 MacBook, using about 60GB of RAM at peak and transcribing 1hr of audio in ~9 mins simonwillison.net/2026/Apr/27/vi…

@GregKamradt @CharlesDardaman I can't imagine any way of running a coding agent for 8 hours without skip perms - the trick is to run it in a sandbox where mistakes don't matter

@CharlesDardaman Do I need to skip perms?

When OAI employees say, “I let codex run all night…” What framework they use? How do you set up the task so it has enough work for 8 hours?