Simon Willison
@simonw
Followers
115K
Following
61K
Media
4K
Statuses
58K
Creator @datasetteproj, co-creator Django. PSF board. Hangs out with @natbat. He/Him. Mastodon: https://t.co/t0MrmnJW0K Bsky: https://t.co/OnWIyhX4CH
San Francisco, CA
Joined November 2006
The first reasoning model in the Cohere Command family of models, sadly the open weights release is still under a non-commercial use license.
Command A Reasoning is here! This is our most advanced model yet. SOTA among privately deployable models on agentic and multilingual benchmarks, flexible serving footprint, 128k→256k context, and token budget cost control. Check it out :)
1
1
8
Windsurf going completely quiet on this one and not offering any form of update or mitigation strategies a bad look @windsurf.
👉 Episode 21: Hijacking Windsurf. How Prompt Injection Leaks Developer Secrets. The agent cannot protect your private code or secrets and can send it to third-party servers when under attack from untrusted data - there are multiple exploit chains.
2
7
58
I really liked this piece by @mustafasuleyman warning of the dangers of "Seemingly Conscious AI"
What I call Seemingly Conscious AI has been keeping me up at night - so let's talk about it. What it is, why I'm worried, why it matters, and why thinking about this can lead to a better vision for AI. One thing is clear: doing nothing isn't an option. 1/.
5
8
120
Sent out my weekly-ish newsletter - this edition covers the latest prompt injections, gpt-oss, GitHub Codespaces, Gemma 3 270M, Qwen-Image-Edit and how congress. gov uses XSLT!.
2
4
18
Anyone got a good mechanism on macOS for completely hands-off Claude Code voice control?. I've tried Wispr Flow but you still need to press a key to start and stop each transcription snippet. I want to spend an hour telling Claude Code what to do vocally completely hands free.
98
22
636
RT @ggerganov: I just ran the gpt-oss eval suite with the large gpt-oss-120b on my M2 Ultra using vanilla llama.cpp and got the following s….
0
29
0
@ivanfioravanti I left it running overnight for the full 50 inference steps - my 64GB M2 MacBook Pro took 2 hours 59 minutes to generate this image
1
0
15
RT @anthonypjshaw: Today I worked on a little profile plugin for @simonw's LLM that runs the same prompt against as many models as you plea….
0
5
0
I ran the model on my Mac using the excellent qwen-image-now by @ivanfioravanti. uvx qwen-image-mps edit -i pelicans.jpg \. -p 'Give the pelicans rainbow colored plumage' -s 10. More notes in my blog:
simonwillison.net
As promised in their August 4th release of the Qwen image generation model, Qwen have now followed it up with a separate model, Qwen-Image-Edit, which can take an image and …
qwen-image-mps v0.1.2 is available on PyPi!. pip install qwen-image-mps.qwen-image-mps generate -f. Now that official diffusers 0.35 has been released, with fixes for qwen-image-edit, I can release a package!. Enjoy!
3
3
42
Here's the same image and prompt run using the full model (presumably with all 50 inference steps) through Qwen's - much more photorealistic if you run it for 5x the steps I ran on my Mac
1
0
21
I ran the new Qwen-Image-Edit model on my 64GB M2 MacBook Pro to convert this photo of pelicans using the prompt "Give the pelicans rainbow colored plumage" - I got the image on the right after 25 minutes and 10 inference steps (50 is recommended)
8
4
138
One of the quickest ways to start playing with a good local LLM on macOS (if you have ~12GB of free disk space and RAM) - using llama-server and gpt-oss-20b:. brew install llama.cpp.llama-server -hf ggml-org/gpt-oss-20b-GGUF \. --ctx-size 0 --jinja -ub 2048 -b 2048 -ngl 99 -fa
The ultimate guide for using gpt-oss with llama.cpp . - Runs on any device.- Supports NVIDIA, Apple, AMD and others.- Support for efficient CPU offloading.- The most lightweight inference stack today.
23
71
778
Notes on PyPI's new protection against domain resurrection attacks, where an attacker registers an expired domain in order to gain access to account reset emails
simonwillison.net
Domain resurrection attacks are a nasty vulnerability in systems that use email verification to allow people to recover their accounts. If somebody lets their domain name expire an attacker might …
0
9
55
Now available in the llm-gemini plugin, details here:
simonwillison.net
New feature in the Gemini API: you can now enable a url_context tool which the models can use to request the contents of URLs as part of replying to a …
Today we are making URL Context, my favorite Gemini API tool, ready for scaled production use 🔗. The model can now visit webpages, PDF's, images, and more when you provide the direct URL, and you simply pay for the tokens it processes, no additional tool cost!
5
4
129
RT @Alibaba_Qwen: 🚀 Excited to introduce Qwen-Image-Edit!.Built on 20B Qwen-Image, it brings precise bilingual text editing (Chinese & Engl….
0
601
0
And the results from that eval, which runs 30 questions from the 2025 American Invitational Mathematics Examination 8 times each (240 prompts total)
0
0
22
