At a guess, about a 50th of a hamburger
At a guess, about a 50th of a hamburger
Chronological feed of everything captured from Simon Willison.
At a guess, about a 50th of a hamburger
Let's take a very pessimistic estimate - assume Fable is 2x the energy/water usage of Opus (so 10 rather than 5) and treat those pelican SVGs as each equivalent to a 5 page report - there's a lot of tokens in SVG Around 862ml, according to www.andymasley.com/visuals/ai-p...
Depending on which almond growing water usage estimate you believe that's between 1/14th and 1/5th of an almond
It's now possible to compile Python extensions (C, C++, Rust etc) to WebAssembly and distribute them through PyPI such that Pyodide can install them directly simonwillison.net/2026/Jun/13/...
I got fed up of waiting for OpenAI to bring their much improved gpt-realtime-2 voice conversation model to the ChatGPT product, so I upgraded my OpenAI-WebRTC playground tool to use it and to let you paste in a document to have a conversation about, too simonwillison.net/2026/Jun/12/...
Wow that's a fine pelican
Yes that should work pretty well, maybe I should give that thing more of a try
It was in "auto" mode
After two days with Claude Fable 5 the best way I can describe it is "relentlessly proactive" - here's an example where I dropped in a screenshot of a bug and it span up custom CORS Python servers and used pyobjc-framework-Quartz to capture screenshots simonwillison.net/2026/Jun/11/...
This right here is a pretty elegant solution to the problem of wanting to capture measurements from JavaScript in the system Safari without being able to directly access the DOM
New Datasette release: 1.0a33, which finally brings documents the ?_extra= JSON API mechanism and brings it to the row and query pages in addition to the table pages (Most of the code in this release was built with the help of Claude Fable 5) datasette.io/blog/2026/ap...
Since the JSON extras API is a little hard to explain without an example I also had Fable 5 and GPT-5.5 collaborate on this custom API explorer tool for trying out the new feature tools.simonwillison.net/datasette-ex...
They released Fable 5 and Mythos 5 - and described Mythos as "Fable 5’s capabilities without the safety classifiers" - but Mythos is still only available to approved partners It's priced the same as Mythos too - I think it's the same model with the safeties turned off
I've stayed clear of those because I'm worried about how many tokens they would burn, plus I still review all the code that goes into Datasette so churning it out even faster doesn't help me much!
I can't talk for other professions, but it further changes what being a professional software engineer looks like - we can build even more ambitious and complex projects with it, which makes our work even harder
Almost always
Wrote up my initial impressions of Claude Fable 5 - it has a big model smell: slow, expensive and capable of crunching through pretty much everything I threw at it simonwillison.net/2026/Jun/9/c...
Here are Fable's pelicans for the different thinking effort levels, plus how much each one cost to generate via the Claude API
A TIL on using agentsview.io to calculate token spending with Claude Fable 5 despite that model not yet being included in the AgentsView pricing database til.simonwillison.net/llms/agentsv...
I tried to visit that one during PyCon US in Pittsburgh last year but it turns out you have to book your tour a few weeks in advance
I've been running Claude Code on the default "high" level, haven't tried the others yet
It doesn't handle any Python package that can't be inlined into a single pure Python string For numpy something based on Pyodide would be a more likely route
I may have finally found the Python-in-a-sandbox solution I've been looking for... here's my latest experiment, this time running MicroPython in WebAssembly inside my Python applications simonwillison.net/2026/Jun/6/m...
You can try it out like this: uvx micropython-wasm -c 'print("Hello world")'
Today on TikTok: Chinese truck selling thirst traps www.tiktok.com/@chenglongtr...
Today on TikTok: Chinese truck selling thirst traps www.tiktok.com/@chenglongtr...
My guess is that it's per billable vendor
Presumably each of Anthropic and OpenAI and Cursor have enterprise plans where you can set per-user limits Setting a single limit across multiple vendors at once sounds technically challenging
Uber reportedly now caps coding agents at $1,500/month per employee per tool - seems sensible to me, but it's also an interesting hint at the value Uber thinks these tools are providing simonwillison.net/2026/Jun/3/u...
My problem is more that it FAKES intent - if I'm going to be convinced of something I want it to come from a human who has themselves been convinced previously
I mean yeah, tokenmaxxing leaderboards were an absurdly stupid trend and deserve to go extinct already
Right now if you want a big PR hit being a company that says "we're pulling back on our AI spend" is a great way to get yourself in the news cycle for a few weeks
In PR terms, do you think the Uber COO messed up when he let slip that Uber were reconsidering the ROI In that podcast interview?
Tone didn't really feel confrontational to me, the thing I care about is that people are arguing in good faith based on credible information/experience - you had both of those!
Flask was Armin! I'm partially responsible for Django
Thanks, just fixed it
Depends on the content I'm fine with LLMs writing code documentation - the qualities I care about there are clarity and accuracy, so as long as it's reviewed it I'm fine with that I'm NOT OK with AI-generate text that expresses opinions, uses personal anecdotes or tries to convince me of something
Have you seen any better expamples of organizations "pulling back" than that (in my opinion quite flimsy) Uber COO story? I agree that the huge pricing increase is a big story, especially Microsoft dropping their subsidies for Copilot and Anthropic/OpenAI charging enterprise list price for tokens
Parenting is an interesting one because, on the one hand, billions of people have done it and shared their knowledge to the LLM training data is incredibly rich Whether or not the average of all of those tips is GOOD advice is another matter
Not to mention different cultures have wildly different parenting standards - advice for rising a child in the USA might not apply at all to a child in Spain or Vietnam, and vice-versa
I'm talking about AI-generated slop Interestingly it's rare in comments, but increasingly common in linked articles
I'm pretty sure most Hacker News upvotes are based on the title only, without visiting the article
The best heuristic I have is to check for other posts or article by the same "author" - the slop accounts tend to use exactly the same length, style and shape of content for everything
Effectively, I am looking for inhuman consistency!
Hacker News has an alarmingly high tolerance for slop content these days It also has social norms against accusing something of being written with AI, which means the community has trouble strengthening its immune system against that slop
I really do
No I think they are proportional - the December date is 31st so effectively the same as January 1st, then the gap between that and Feb 12th is about the same as the gap between April 1st and May 7th
I fed your feedback to Claude Opus 4.8 and it had a bit of a spin-out about it (since I was saying the chart was wrong but it was confident it was right) - eventually it drew a new one with month grid lines to make it more clear gist.github.com/simonw/df3c7...
That's a better image so I've dropped that into my blog post to replace the old one
Not if they'd like to avoid the coverage of their S-1 being mainly about how it demonstrates their dishonesty in their previous statements