Simon Willison
@simonw
Creator @datasetteproj, co-creator Django. PSF board. Hangs out with @natbat. He/Him. Mastodon: Bsky:
5.7K Following 203.4K Followers
I built Moonlight & Mayhem with my Codex monthly subscription, but if I had been paying API prices it would have cost $23.28 (according to AgentsView)
More details on my blog:
There was one bug I had to fix before shipping though (unlike Fable which did it all from a single prompt) - Codex initially gave the raccoons eyeballs four times the size of their bodies!
Show more
Useful tip for the ChatGPT mobile app: long press the send button to adjust the effort
@simonw i personally leave it on instant all the time and go up to high sometimes when i need something more comprehensive. pro tip: if you long press the send button on mobile you can change the slider for just this prompt. i call it the slingshot!
Show more
Waaaay back in January, @simonw was kind enough to share his thoughts on technical blogging with me. ICYMI...
"Felony humble-bragging" is a great line
At final Black Hat keynote (called a locknote, ha ha) panelists say they are surprised at how the OpenAI - Hugging Face incident debrief, as well as other reporting on AI agent escapees (Anthropic/Meta) - has turned into a marketing/PR play. "Felony humble-bragging," said one.
Show more
Our Black Hat talk on the OpenAI-Hugging Face incident is now live on youtube.
This is a watershed moment for the industry. I encourage all defenders to watch, consider how attack dynamics will imminently change, and plan for accelerating defense.
Show more
UPDATE: a challenger emerges
One of the things the pelican benchmark is still useful for is visually representing (to a tiny extent) the improvements in a single model family
Here's Meta AI's Spark (8th April), Spark 1.1 (9th July), and Spark 1.2 (today, 5th August)
Show more
I wrote about the project in more detail on my blog
Got a disappointing pelican from DeepSeek-V4-Flash-0731 at default reasoning mode - on the left - but then I bumped reasoning up to high (via OpenRouter) and got the much better one on the right
Show more
The new stateless MCP specification has rekindled my interest in MCP, and inspired some new projects, including mcp-explorer and datasette-mcp
On of the hardest parts of the project was figuring out the vocabulary! Here's what I settled on
I'm on Oxide and Friends podcast this week!
We talked about accidental cyberattacks, Kimi K3, Golden Gate Claude, the Zizians, Alameda wild turkey attacks, Soviet Marburg virus research, the Lead-crime hypothesis, and a bunch of other worthy digressions
Show more
Here's an example of the kind of report it can produce having run and graded an evaluation suite against different models
I switched my instance to Luna (it was previously on Gemini 3.1 Flash-Lite - Luna is cheaper now) - you can sign into that with GitHub to try it out
OK, GPT-5.6 Luna is a bit of a beast. Given the 80% price drop today I decided to try it in Datasette Agent, and it's furiously quick and generates all the SQL, HTML and JavaScript (for Datasette Apps) I could possibly want
Show more
I hope there are QA testing experts out there who are thinking "finally, we don't need software developers any more!", rolling up their sleeves and building really great software right now
GPT-5.6 found optimizations that "reduced end-to-end serving costs by 20%" for OpenAI to serve that model
Presumably that's billions of dollars a month in savings at this point?
Codex analysed production traffic, improved load balancing, rewrote production GPU kernels and ran hundreds of experiments on its own speculative-decoding model.
The kernel improvements reduced end-to-end serving costs by 20%, while speculative decoding improved token-generation efficiency by more than 15%.
Show more
This is detailed, fascinating and answers all sorts of open questions
I'd love to know more about the "unsecured public code-evaluation sandbox hosted on a third-party provider's infrastructure" that the agent used to stage its attack against HF after it broke out of OpenAI
Show more
ChatGPT Sites means ChatGPT in "Work" mode can build and deploy public websites running on Cloudflare Workers, including with persistence on top of SQLite
(OpenAI do not make it easy to figure out that's how the platform works, though)
Show more
