<?xml version="1.0" encoding="utf-8"?>
<feed xml:lang="en-us" xmlns="http://www.w3.org/2005/Atom"><title>Simon Willison's Weblog: Blogmarks</title><link href="http://simonwillison.net/" rel="alternate"/><link href="http://simonwillison.net/atom/links/" rel="self"/><id>http://simonwillison.net/</id><updated>2026-11-20T02:49:57+00:00</updated><author><name>Simon Willison</name></author><entry><title>Token fraud is here and it’s BAD (YouTube)</title><link href="https://simonwillison.net/2026/Nov/20/token-fraud-is-here-and-its-bad/" rel="alternate"/><published>2026-11-20T02:49:57+00:00</published><updated>2026-11-20T02:49:57+00:00</updated><id>https://simonwillison.net/2026/Nov/20/token-fraud-is-here-and-its-bad/</id><summary type="html">
&lt;p&gt;&lt;strong&gt;&lt;a href="https://www.youtube.com/watch?v=ZDpP7JXPlhc"&gt;Token fraud is here and it’s BAD (YouTube)&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;
Claire Vo runs &lt;a href="https://www.chatprd.ai"&gt;ChatPRD&lt;/a&gt;, an AI platform that helps product managers "turn an idea into a spec". The product has been live since the GPT-4 era, and Claire has been building it in public for much of that time.&lt;/p&gt;
&lt;p&gt;In this YouTube video (see also the &lt;a href="https://x.com/clairevo/status/2119985786045227197"&gt;accompanying Twitter post&lt;/a&gt;) Claire describes the &lt;em&gt;flood&lt;/em&gt; of token fraud she has seen from groups abusing the product's free trial, sometimes just to get free AI usage but often to resell tokens or run other commercial enterprises on top of the ChatPRD agent.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;I am funding a international Roblox and Fortnite map studio. They run a loop of free trials and promo codes and then tunnel to my agent and use it to build Roblox and Fortnite maps.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;Apparently one of the common prompts Claire saw running through her agent involved SVGs of pelicans riding bicycles! Presumably the token scammers are using that to help sniff the underlying models - different model families have a definite signature, and I've used it like this in the past to try and guess the identity of stealth preview models on platforms like OpenRouter.&lt;/p&gt;
&lt;p&gt;Claire is winding down her free trials entirely, which sadly looks like the right decision in the face of this abuse.


    &lt;p&gt;Tags: &lt;a href="https://simonwillison.net/tags/youtube"&gt;youtube&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/ai"&gt;ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/generative-ai"&gt;generative-ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/llms"&gt;llms&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/pelican-riding-a-bicycle"&gt;pelican-riding-a-bicycle&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/ai-misuse"&gt;ai-misuse&lt;/a&gt;&lt;/p&gt;

</summary><category term="youtube"/><category term="ai"/><category term="generative-ai"/><category term="llms"/><category term="pelican-riding-a-bicycle"/><category term="ai-misuse"/></entry><entry><title>PostgreSQL 19 Released</title><link href="https://simonwillison.net/2026/Nov/19/postgresql-19/" rel="alternate"/><published>2026-11-19T20:43:57+00:00</published><updated>2026-11-19T20:43:57+00:00</updated><id>https://simonwillison.net/2026/Nov/19/postgresql-19/</id><summary type="html">
&lt;p&gt;&lt;strong&gt;&lt;a href="https://www.postgresql.org/about/news/postgresql-19-released-3449/"&gt;PostgreSQL 19 Released&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;
Most years PostgreSQL ships a new release in September, but this year's release came two months later &lt;a href="https://rhaas.blogspot.com/2026/10/postgres-19-is-late-and-its-ais-fault.html"&gt;thanks mainly&lt;/a&gt; to the flurry of issues that the PostgreSQL team needed to address as part of the ongoing AI-reported vulnerability apocalypse.&lt;/p&gt;
&lt;p&gt;This release was definitely worth the wait. My favorite new feature is this one:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;For applications that send reads to standbys, PostgreSQL 19 adds "read-your-writes" consistency with the new &lt;a href="https://www.postgresql.org/docs/19/sql-wait.html"&gt;&lt;code&gt;WAIT FOR LSN&lt;/code&gt;&lt;/a&gt; command. A session on a standby can use &lt;code&gt;WAIT FOR LSN&lt;/code&gt; to wait until the standby has replayed the write-ahead log (WAL) up to a specific &lt;a href="https://www.postgresql.org/docs/19/wal-internals.html"&gt;log sequence number (LSN)&lt;/a&gt; before running a query, so the application can read data it just wrote on the primary.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;In practice, that looks like this:&lt;/p&gt;
&lt;pre&gt;&lt;code&gt;UPDATE movie SET genre = 'Dramatic' WHERE genre = 'Drama';
SELECT pg_current_wal_insert_lsn();
&lt;/code&gt;&lt;/pre&gt;
&lt;p&gt;That returns a value like &lt;code&gt;0/0306EE20&lt;/code&gt;. Store that somewhere (a signed cookie would work) such that on a subsequent request your application can do:&lt;/p&gt;
&lt;pre&gt;&lt;code&gt;WAIT FOR LSN '0/0306EE20' WITH (TIMEOUT '0.1s');
SELECT * FROM movie WHERE genre = 'Drama';
&lt;/code&gt;&lt;/pre&gt;
&lt;p&gt;The timeout is optional - if you provide one you'll get an error if the replica you are talking to doesn't catch up within the specified time.&lt;/p&gt;
&lt;p&gt;I first encountered this pattern &lt;a href="https://www.mediawiki.org/wiki/Manual:Database_access#Lag"&gt;in Wikipedia&lt;/a&gt;, which has a hand-rolled version of this built against MySQL replication:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;In addition to this, MediaWiki attempts to ensure that the user sees events occurring on the wiki in chronological order. A few seconds of lag can be tolerated, as long as the user sees a consistent picture from subsequent requests. This is done by saving the primary binlog position in the session, and then at the start of each request, waiting for the replica to catch up to that position before doing any reads from it.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;Here's &lt;a href="https://github.com/wikimedia/mediawiki/blob/1ff78194b182ee4b5b3678f7d9fc9aefd0852cf7/includes/libs/Rdbms/Replication/MysqlReplicationReporter.php#L238-L299"&gt;the relevant MediaWiki code&lt;/a&gt;, which these days uses MySQL's &lt;a href="https://dev.mysql.com/doc/refman/9.7/en/gtid-functions.html#function_wait-for-executed-gtid-set"&gt;WAIT_FOR_EXECUTED_GTID_SET()&lt;/a&gt; function and falls back to &lt;a href="https://dev.mysql.com/doc/refman/8.4/en/replication-functions-synchronization.html#function_source-pos-wait"&gt;SOURCE_POS_WAIT()&lt;/a&gt;.

    &lt;p&gt;&lt;small&gt;&lt;/small&gt;Via &lt;a href="https://lobste.rs/s/pqyuhk/postgresql_19_released"&gt;Lobste.rs&lt;/a&gt;&lt;/small&gt;&lt;/p&gt;


    &lt;p&gt;Tags: &lt;a href="https://simonwillison.net/tags/mysql"&gt;mysql&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/postgresql"&gt;postgresql&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/replication"&gt;replication&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/scaling"&gt;scaling&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/wikipedia"&gt;wikipedia&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/ai-security-research"&gt;ai-security-research&lt;/a&gt;&lt;/p&gt;

</summary><category term="mysql"/><category term="postgresql"/><category term="replication"/><category term="scaling"/><category term="wikipedia"/><category term="ai-security-research"/></entry><entry><title>The PSF end-of-year fundraiser</title><link href="https://simonwillison.net/2026/Nov/18/psf-end-of-year-fundraiser/" rel="alternate"/><published>2026-11-18T14:44:48+00:00</published><updated>2026-11-18T14:44:48+00:00</updated><id>https://simonwillison.net/2026/Nov/18/psf-end-of-year-fundraiser/</id><summary type="html">
&lt;p&gt;&lt;strong&gt;&lt;a href="https://donate.python.org"&gt;The PSF end-of-year fundraiser&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;
With my Python Software Foundation board director hat on: we've launched our end-of-year fundraiser for the non-profit that supports Python development and the Python community.&lt;/p&gt;
&lt;p&gt;If your company or your work benefits from Python, please consider donating this year!&lt;/p&gt;
&lt;p&gt;The PSF looks after the IP behind Python, operates PyPI, sponsors core developer sprints, supports community events (including running PyCon US every year), and employs a small and excellent staff team to keep everything running.&lt;/p&gt;
&lt;p&gt;We also operate the &lt;a href="https://www.python.org/psf/developersinresidence/"&gt;Developers in Residence program&lt;/a&gt;, which now includes three security developers (welcome, Brett Smith) focused on helping secure Python, PyPI, and the wider ecosystem.


    &lt;p&gt;Tags: &lt;a href="https://simonwillison.net/tags/python"&gt;python&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/psf"&gt;psf&lt;/a&gt;&lt;/p&gt;

</summary><category term="python"/><category term="psf"/></entry><entry><title>External sandboxes for coding agents</title><link href="https://simonwillison.net/2026/Nov/17/external-sandboxes-for-coding-agents/" rel="alternate"/><published>2026-11-17T22:52:29+00:00</published><updated>2026-11-17T22:52:29+00:00</updated><id>https://simonwillison.net/2026/Nov/17/external-sandboxes-for-coding-agents/</id><summary type="html">
&lt;p&gt;&lt;strong&gt;&lt;a href="https://blog.primeradiant.com/external-sandboxes-for-coding-agents/"&gt;External sandboxes for coding agents&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;
I've been exploring new ways to sandbox code that is executed by agents, with a focus on making it safer to try out local models without worrying that a mistake or prompt injection attack might cause them to take damaging actions on a machine.&lt;/p&gt;
&lt;p&gt;I worked on this project with &lt;a href="https://primeradiant.com"&gt;Prime Radiant&lt;/a&gt;, the AI research lab behind Superpowers and &lt;a href="https://github.com/orgs/prime-radiant-inc/repositories"&gt;other projects&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;&lt;a href="https://blog.primeradiant.com/external-sandboxes-for-coding-agents/"&gt;My post&lt;/a&gt; on the Prime Radiant blog introduces the &lt;strong&gt;external sandbox&lt;/strong&gt; pattern, where code executed by the agent runs inside a dedicated sandbox while the agent itself lives outside of that sandbox.&lt;/p&gt;
&lt;p&gt;To demonstrate this pattern we've released three sandbox server implementations:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href="https://github.com/prime-radiant-inc/external-sandbox-v86"&gt;external-sandbox-v86&lt;/a&gt; runs a Node.js server which then uses &lt;a href="https://github.com/copy/v86"&gt;v86&lt;/a&gt; to run Alpine Linux inside WebAssembly.&lt;/li&gt;
&lt;li&gt;&lt;a href="https://github.com/prime-radiant-inc/external-sandbox-container"&gt;external-sandbox-container&lt;/a&gt; is a Python web server which runs the sandbox in Alpine Linux inside an Apple Containers instance.&lt;/li&gt;
&lt;li&gt;&lt;a href="https://github.com/prime-radiant-inc/external-sandbox-quicksand"&gt;external-sandbox-quicksand&lt;/a&gt; is another Python web server, this time using Microsoft's &lt;a href="https://github.com/microsoft/quicksand"&gt;Quicksand&lt;/a&gt; sandboxing library, which bundles QEMU &lt;em&gt;and&lt;/em&gt; Alpine Linux in a Python package!&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Each of these implementations offers the same localhost API. You can then configure Pi, Claude Code, or Codex to send API calls to that sandbox any time they want to execute code.&lt;/p&gt;
&lt;p&gt;The sandboxes also include an optional HTTP/HTTPS allow-list proxy, so you can control exactly which hosts your agent's code is allowed to communicate with - plus a credential replacement mechanism to make authenticated API calls without sharing a secret with the agent. Here's the UI you can use to reconfigure that proxy while the sandbox is still running:&lt;/p&gt;
&lt;p&gt;&lt;img alt="The network policy page with allowed destinations, the policy JSON editor and recent proxy activity" src="https://static.simonwillison.net/static/2026/sandbox-proxy.png" /&gt;&lt;/p&gt;
&lt;p&gt;I continue to be fascinated by the challenge of safely running untrusted code. This external sandbox pattern provides a neat way to experiment with different approaches.


    &lt;p&gt;Tags: &lt;a href="https://simonwillison.net/tags/sandboxing"&gt;sandboxing&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/security"&gt;security&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/ai"&gt;ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/webassembly"&gt;webassembly&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/generative-ai"&gt;generative-ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/llms"&gt;llms&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/coding-agents"&gt;coding-agents&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/claude-code"&gt;claude-code&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/codex"&gt;codex&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/pi"&gt;pi&lt;/a&gt;&lt;/p&gt;

</summary><category term="sandboxing"/><category term="security"/><category term="ai"/><category term="webassembly"/><category term="generative-ai"/><category term="llms"/><category term="coding-agents"/><category term="claude-code"/><category term="codex"/><category term="pi"/></entry><entry><title>GPT‑6.2 Astra and GPT‑6.2 Sol</title><link href="https://simonwillison.net/2026/Nov/17/gpt62-astra-and-sol/" rel="alternate"/><published>2026-11-17T22:38:52+00:00</published><updated>2026-11-17T22:38:52+00:00</updated><id>https://simonwillison.net/2026/Nov/17/gpt62-astra-and-sol/</id><summary type="html">
&lt;p&gt;&lt;strong&gt;&lt;a href="https://openai.com/index/introducing-gpt-6-2-astra-sol/"&gt;GPT‑6.2 Astra and GPT‑6.2 Sol&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;
OpenAI:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;We’re introducing GPT‑6.2 Astra, our new frontier model for your hardest projects, and GPT‑6.2 Sol, our efficient, versatile model for everyday work.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;They're available in the API as &lt;code&gt;gpt-6.2-astra&lt;/code&gt; and &lt;code&gt;gpt-6.2-sol&lt;/code&gt;. I shipped &lt;a href="https://llm.datasette.io/en/stable/changelog.html#v0-38"&gt;LLM 0.38&lt;/a&gt; with support for the new models.&lt;/p&gt;
&lt;p&gt;6.2 Astra is 80% of the price of 6 Astra - $8/MTok input and $40/MTok output, the same price as Claude Fable 5.2 (which was also 80% of its predecessor). 6 Astra and Fable 5 were both $10/$50. The cached price is an even bigger discount - $0.40/million compared to $1/million for the older model. This is great news for longer running agentic sessions.&lt;/p&gt;
&lt;p&gt;GPT-6.2 Sol is the same $2/$10 price as 6.1 Sol and 6 Sol - which are half the price of GPT-5.6 Sol ($4/$20).&lt;/p&gt;
&lt;p&gt;Here are the pelicans &lt;a href="https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2Fa940cc2750162ac9edf0ddf881067b83"&gt;for GPT-6.2 Sol&lt;/a&gt; and &lt;a href="https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2Ffe836e13c90c00233c9e9e499a941805"&gt;for GPT-6.2 Astra&lt;/a&gt;. The 6.2 Astra max pelican looks like this:&lt;/p&gt;
&lt;p&gt;&lt;img alt="Good bicycle, good pelican, one wing is on the handlebars with fingers, the other wing is behind. Feet on the correct side of the frame. Nice gradiants. Good pouch, hooked bill." src="https://static.simonwillison.net/static/2026/astra-6-2-max.webp" /&gt;&lt;/p&gt;
&lt;p&gt;I have a &lt;a href="https://static.simonwillison.net/static/2026/openai-pelicans-november-2026.html"&gt;full comparison grid here&lt;/a&gt;, but here's how the high/xhigh/max pelicans compare with the previous Sol and Astra releases:&lt;/p&gt;
&lt;p&gt;&lt;img alt="Twelve pastel illustrations compare four models drawing a pelican riding a bicycle at three effort levels. All twelve have a recognizable pelican and bicycle. The clearest differences are in the feather detail and how convincingly the bird steers and pedals. gpt-6.2-astra has the most detailed Max drawing (88.71¢), with layered feathers, a trailing red scarf, and a wing on the handlebars. Xhigh (17.70¢) adds a green cap and keeps the seaside setting. High (11.55¢) simplifies the feathers while keeping the riding pose clear. gpt-6.2-sol shows a convincing pedaling pose at both Max (13.69¢) and Xhigh (6.74¢), with one foot higher than the other and a wing reaching forward. At High (2.41¢), the bird sits lower and the handlebars are harder to distinguish. gpt-6-astra gives the bird a long wing that follows the curve of the handlebars at Max (63.21¢). Xhigh (33.85¢) shows two bent legs on the pedals. At High (18.37¢), one leg stretches almost straight down, making the pose look stiffer. gpt-6.1-sol uses nearly the same long-necked bird with a wing on the bars at Max (13.48¢) and Xhigh (5.39¢). At High (3.66¢), the bird folds its wing against its body and appears to ride hands-free." src="https://static.simonwillison.net/static/2026/gpt-6-2-grid.webp" /&gt;&lt;/p&gt;
&lt;p&gt;I also ran my most recent benchmark prompt against the two models, using "Ultra' effort levels in the Codex CLI:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;code&gt;Make a three js animation where yonkers stands up as a big mech (cool stand up motion) and stomps around manhattan and queens and Brooklyn, use openstreetmap mapping data&lt;/code&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;Here's &lt;a href="https://static.simonwillison.net/static/2026/yonkers-gpt-6.2-sol-ultra/index.html"&gt;Sol&lt;/a&gt; and here's &lt;a href="https://static.simonwillison.net/static/2026/yonkers-gpt-6.2-astra-ultra/index.html"&gt;Astra&lt;/a&gt;. Neither of them did a good enough job of morphing Yonkers into a big mech, at least in comparison to &lt;a href="https://tools.simonwillison.net/yonkers-rises"&gt;Claude Fable 5.2&lt;/a&gt; and &lt;a href="https://static.simonwillison.net/static/2026/yonkers-mech-standalone.html"&gt;Gemini 4 Carbon&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;This video showing the Astra and Sol animations playing out at the same time:&lt;/p&gt;
&lt;video controls playsinline preload="none" width="1080" height="1080" poster="https://static.simonwillison.net/static/2026/yonkers-astra-sol-20s.webp"&gt;
  &lt;source src="https://static.simonwillison.net/static/2026/yonkers-astra-sol-20s-largest.mp4" type="video/mp4"&gt;
&lt;/video&gt;

&lt;p&gt;I created that video using GPT-6.2 Astra (Ultra) in the Codex desktop app by prompting:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;code&gt;https://static.simonwillison.net/static/2026/yonkers-gpt-6.2-astra-ultra/index.html and https://static.simonwillison.net/static/2026/yonkers-gpt-6.2-sol-ultra/index.html  are two videos of the same thing by two different models&lt;/code&gt;&lt;/p&gt;
&lt;p&gt;&lt;code&gt;I need a video of the first 10 seconds where the top half of the video plays the first and the bottom half plays the second&lt;/code&gt;&lt;/p&gt;
&lt;p&gt;&lt;code&gt;I need an overlaid label of "GPT-6.2 Astra" on the top half and "GPT-6.2 Sol" on the bottom half. These labels should be tastefully positioned.&lt;/code&gt;&lt;/p&gt;
&lt;p&gt;&lt;code&gt;Make the video square and 2000x2000 pixels&lt;/code&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;Then later:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;code&gt;OK make it 20s&lt;/code&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;Here's the &lt;a href="https://gist.github.com/simonw/6b44c588a1321b60799dbbfd13db6d23"&gt;exported transcript&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The buzz around these new models is mixed at the moment. Astra 6.2 isn't scoring quite as high as Fable 5.2 on various benchmarks, and I've seen a few complaints that its communication style can be confusing. Theo Browne's &lt;a href="https://twitter.com/theo/status/2122760447103484410"&gt;review&lt;/a&gt; (in the form of a feedback letter he sent OpenAI during pre-release) has this useful example:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;When this model is working on long tasks, it can pull in a lot of subagents and manage them pretty well. It will then proceed to randomly &lt;em&gt;ask you some weird internal implementation detail bullshit that you could not under any earthly circumstance understand, much less have input on&lt;/em&gt;.&lt;/p&gt;
&lt;p&gt;I sent "/goal make a new level for Outer Wilds". 2 hours in, it's "blocked" because it wanted to ask me "does r24's reviewer write it's clips to 24_g521.mp4 or 24_p11.mp4?"&lt;/p&gt;
&lt;/blockquote&gt;


    &lt;p&gt;Tags: &lt;a href="https://simonwillison.net/tags/ai"&gt;ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/openai"&gt;openai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/generative-ai"&gt;generative-ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/llms"&gt;llms&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/llm-pricing"&gt;llm-pricing&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/pelican-riding-a-bicycle"&gt;pelican-riding-a-bicycle&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/llm-release"&gt;llm-release&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/theo-browne"&gt;theo-browne&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/gpt"&gt;gpt&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/yonkers-rising"&gt;yonkers-rising&lt;/a&gt;&lt;/p&gt;

</summary><category term="ai"/><category term="openai"/><category term="generative-ai"/><category term="llms"/><category term="llm-pricing"/><category term="pelican-riding-a-bicycle"/><category term="llm-release"/><category term="theo-browne"/><category term="gpt"/><category term="yonkers-rising"/></entry><entry><title>Gemini Decisions API</title><link href="https://simonwillison.net/2026/Nov/11/gemini-decisions-api/" rel="alternate"/><published>2026-11-11T22:03:25+00:00</published><updated>2026-11-11T22:03:25+00:00</updated><id>https://simonwillison.net/2026/Nov/11/gemini-decisions-api/</id><summary type="html">
&lt;p&gt;&lt;strong&gt;&lt;a href="https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/"&gt;Gemini Decisions API&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;
Google join &lt;a href="https://simonwillison.net/2026/Sep/21/jev/"&gt;Jev&lt;/a&gt; and &lt;a href="https://simonwillison.net/2026/Oct/6/llm-openai-decisions/"&gt;OpenAI&lt;/a&gt; in the decision model category with "Gemini Decision 1". This new model API is almost identical to Jev: it supports the same three types of questions (yes/no, choice, and rubric-scoring) and even reuses Jev's term "Noul" for a yes/no question.&lt;/p&gt;
&lt;p&gt;It's priced the same as the GPT-6 Luna decision model: 10 cents per million input tokens, and output is free. Jev remains the cheapest of the three, at 4.2 cents per million.&lt;/p&gt;
&lt;p&gt;Jev is text (and JSON) only. OpenAI can handle text and images. Gemini Decision 1 is fully multi-model: it covers text, images, audio, video, and PDFs.&lt;/p&gt;
&lt;p&gt;I tested it against video by passing these questions to it along with my &lt;a href="https://simonwillison.net/2026/Nov/10/yonkers-rises/"&gt;Yonkers stomping robot&lt;/a&gt; video file:&lt;/p&gt;
&lt;pre&gt;&lt;code&gt;dragon: Does this video have a dragon in it?
robot: Does this video have a robot in it?
france: Is this video set in France?
usa: Is this video set in the USA?
&lt;/code&gt;&lt;/pre&gt;
&lt;p&gt;Jev and OpenAI both return 1.0 or 0.0 for yes/no questions, but Gemini insists on being a little less confident than that:&lt;/p&gt;
&lt;div class="highlight highlight-source-json"&gt;&lt;pre&gt;{
  &lt;span class="pl-ent"&gt;"dragon"&lt;/span&gt;: {
    &lt;span class="pl-ent"&gt;"type"&lt;/span&gt;: &lt;span class="pl-s"&gt;&lt;span class="pl-pds"&gt;"&lt;/span&gt;yes_no&lt;span class="pl-pds"&gt;"&lt;/span&gt;&lt;/span&gt;,
    &lt;span class="pl-ent"&gt;"probability"&lt;/span&gt;: &lt;span class="pl-c1"&gt;0.04976021841520312&lt;/span&gt;
  },
  &lt;span class="pl-ent"&gt;"robot"&lt;/span&gt;: {
    &lt;span class="pl-ent"&gt;"type"&lt;/span&gt;: &lt;span class="pl-s"&gt;&lt;span class="pl-pds"&gt;"&lt;/span&gt;yes_no&lt;span class="pl-pds"&gt;"&lt;/span&gt;&lt;/span&gt;,
    &lt;span class="pl-ent"&gt;"probability"&lt;/span&gt;: &lt;span class="pl-c1"&gt;0.9234055387792485&lt;/span&gt;
  },
  &lt;span class="pl-ent"&gt;"france"&lt;/span&gt;: {
    &lt;span class="pl-ent"&gt;"type"&lt;/span&gt;: &lt;span class="pl-s"&gt;&lt;span class="pl-pds"&gt;"&lt;/span&gt;yes_no&lt;span class="pl-pds"&gt;"&lt;/span&gt;&lt;/span&gt;,
    &lt;span class="pl-ent"&gt;"probability"&lt;/span&gt;: &lt;span class="pl-c1"&gt;0.08329171353166241&lt;/span&gt;
  },
  &lt;span class="pl-ent"&gt;"usa"&lt;/span&gt;: {
    &lt;span class="pl-ent"&gt;"type"&lt;/span&gt;: &lt;span class="pl-s"&gt;&lt;span class="pl-pds"&gt;"&lt;/span&gt;yes_no&lt;span class="pl-pds"&gt;"&lt;/span&gt;&lt;/span&gt;,
    &lt;span class="pl-ent"&gt;"probability"&lt;/span&gt;: &lt;span class="pl-c1"&gt;0.9230459995157706&lt;/span&gt;
  }
}&lt;/pre&gt;&lt;/div&gt;

    &lt;p&gt;&lt;small&gt;&lt;/small&gt;Via &lt;a href="https://news.ycombinator.com/item?id=50456791"&gt;Hacker News&lt;/a&gt;&lt;/small&gt;&lt;/p&gt;


    &lt;p&gt;Tags: &lt;a href="https://simonwillison.net/tags/ai"&gt;ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/generative-ai"&gt;generative-ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/gemini"&gt;gemini&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/llm-release"&gt;llm-release&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/jev"&gt;jev&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/decision-models"&gt;decision-models&lt;/a&gt;&lt;/p&gt;

</summary><category term="ai"/><category term="generative-ai"/><category term="gemini"/><category term="llm-release"/><category term="jev"/><category term="decision-models"/></entry><entry><title>aws-emulator</title><link href="https://simonwillison.net/2026/Nov/11/aws-emulator/" rel="alternate"/><published>2026-11-11T21:18:30+00:00</published><updated>2026-11-11T21:18:30+00:00</updated><id>https://simonwillison.net/2026/Nov/11/aws-emulator/</id><summary type="html">
&lt;p&gt;&lt;strong&gt;&lt;a href="https://github.com/aws/aws-emulator"&gt;aws-emulator&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;
Finally, an official emulator from AWS to run tests against a local clone of their many different services without having to provision a paid API key to do so.&lt;/p&gt;
&lt;p&gt;It's written in Rust by a huge swarm of agents - 29,684 commits in the first three weeks of the project. Swaminathan (Swami) Sivasubramanian wrote about &lt;a href="https://www.allthingsdistributed.com/2026/11/aws-emulator-and-the-engineers-job-when-agents-write-the-code.html"&gt;how it came to be&lt;/a&gt; in a guest post on Werner Vogels' blog.&lt;/p&gt;
&lt;p&gt;It currently requires Docker to run on macOS or Windows. I'm hoping they fix that in the future - it would be great if you could download a single binary and use that as part of a standard test suite in other languages and on other platforms.


    &lt;p&gt;Tags: &lt;a href="https://simonwillison.net/tags/amazon"&gt;amazon&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/amazon-web-services"&gt;amazon-web-services&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/testing"&gt;testing&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/werner-vogels"&gt;werner-vogels&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/ai"&gt;ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/rust"&gt;rust&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/docker"&gt;docker&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/generative-ai"&gt;generative-ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/llms"&gt;llms&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/agentic-engineering"&gt;agentic-engineering&lt;/a&gt;&lt;/p&gt;

</summary><category term="amazon"/><category term="amazon-web-services"/><category term="testing"/><category term="werner-vogels"/><category term="ai"/><category term="rust"/><category term="docker"/><category term="generative-ai"/><category term="llms"/><category term="agentic-engineering"/></entry><entry><title>Browse tools.simonwillison.net by tag</title><link href="https://simonwillison.net/2026/Nov/11/tools-tags/" rel="alternate"/><published>2026-11-11T18:42:50+00:00</published><updated>2026-11-11T18:42:50+00:00</updated><id>https://simonwillison.net/2026/Nov/11/tools-tags/</id><summary type="html">
&lt;p&gt;&lt;strong&gt;&lt;a href="https://tools.simonwillison.net/#tags"&gt;Browse tools.simonwillison.net by tag&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;
I've been wanting a better way to browse my &lt;a href=""&gt;collection of vibe-coded tools&lt;/a&gt; (244 and counting) for a while. I had a few Fable 5.2 credits left over before my weekly reset so I set it to work using &lt;a href="https://github.com/tree-sitter/py-tree-sitter"&gt;tree-sitter&lt;/a&gt; to analyze the code for JavaScript APIs, libraries, and web API hosts used by each tool.&lt;/p&gt;
&lt;p&gt;The homepage now features tag clouds for browsing by each of those categories of tag:&lt;/p&gt;
&lt;p&gt;&lt;a href="https://tools.simonwillison.net/#tags"&gt;&lt;img alt="Screenshot of two tag cloud sections where tags are sized by count. Browse by JavaScript libraries - Libraries loaded from a CDN by two or more tools. Select a tag to list its tools. @google/generative-ai 3, @jsquash/avif 3, @jsquash/webp 4, Babel 2, Chart.js 2, Cropper.js 2, D3 2, DOMPurify 7, js-yaml 3, JSZip 3, Leaflet 4, markdown-it 4, Marked 12, pako 2, PDF.js 3, Pyodide 11, React 3, Tesseract.js 2, three.js 3. Browse by browser APIs - Web platform features used by each tool's own code. Select a tag to list its tools. &amp;lt;dialog&amp;gt; 5, AbortController 6, Blob URLs 38, BroadcastChannel 1, Built-in AI 1, Camera / microphone 9, Canvas 2D 44, Clipboard API 69, Compression Streams 3, createImageBitmap 2, crypto random values 4, DOMParser 8, Drag and drop 30, Dynamic import() 12, ES modules 92" src="https://static.simonwillison.net/static/2026/tools-tag-cloud.webp" /&gt;&lt;/a&gt;


    &lt;p&gt;Tags: &lt;a href="https://simonwillison.net/tags/javascript"&gt;javascript&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/tools"&gt;tools&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/ai"&gt;ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/generative-ai"&gt;generative-ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/llms"&gt;llms&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/claude"&gt;claude&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/vibe-coding"&gt;vibe-coding&lt;/a&gt;&lt;/p&gt;

</summary><category term="javascript"/><category term="tools"/><category term="ai"/><category term="generative-ai"/><category term="llms"/><category term="claude"/><category term="vibe-coding"/></entry><entry><title>Anthropic's IPO roadshow video</title><link href="https://simonwillison.net/2026/Nov/11/anthropic-ipo-roadshow/" rel="alternate"/><published>2026-11-11T15:11:59+00:00</published><updated>2026-11-11T15:11:59+00:00</updated><id>https://simonwillison.net/2026/Nov/11/anthropic-ipo-roadshow/</id><summary type="html">
&lt;p&gt;&lt;strong&gt;&lt;a href="https://ipo.anthropic.com/us/#roadshow-video"&gt;Anthropic&amp;#x27;s IPO roadshow video&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;
Anthropic kicked off their IPO roadshow yesterday, and their roadshow video has some previously unreleased information about how their business works.&lt;/p&gt;
&lt;p&gt;The section on "Approach to compute" with co-founder Tom Brown is particularly interesting, describing how Anthropic train and serve Claude across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs.&lt;/p&gt;
&lt;p&gt;I'm not sure why there's no official video transcript, but here's &lt;a href="https://gist.github.com/simonw/9806757fe8acec6221cb0a5cce4d720d"&gt;one that I extracted&lt;/a&gt; using MacWhisper. Highlights below:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;We train and serve Claude now across all four major chip families. That's only possible because early on we made an unconventional bet on a multi-chip strategy. At the time, many people thought that running AI models on different chips wasn't worth the effort. Each chip needs its own tailored software stack and inference stack, which is expensive and difficult to build. We've now been doing it for years, and we've gotten incredibly good at it. [...]&lt;/p&gt;
&lt;p&gt;As an example, this year, we started working with AMD GPUs for the first time. In the past, bringing on a new chip platform took dozens of engineers months of work &lt;strong&gt;This time, we had Mythos try it out. It got Claude running on AMD in just three days, with basically no human intervention&lt;/strong&gt;. Claude ported or wrote almost 100% of the kernels. [...]&lt;/p&gt;
&lt;p&gt;The second driver of our compute advantage is our rate of discovery of compute multipliers. We focus on total effective compute, which is the multiplicative product of physical capacity times the compute multipliers that we discover. Each one we discover makes our pool of total effective compute bigger. For example, say you need 100 units of compute to train a model of a certain intelligence. A 4x compute multiplier means you can use 25 units of compute to train the exact same model. Or you can use those hundred physical units to train an even more intelligent model that would have previously required 400 units.&lt;/p&gt;
&lt;p&gt;Compute multipliers create a competitive advantage for us that's hard to replicate. They build on each other. Each new multiplier we discover has a positive effect on the multipliers before it. They stack. &lt;strong&gt;We believe that this can accelerate even further as we approach recursive self-improvement. Claude will be able to discover compute multipliers entirely by itself&lt;/strong&gt;, allowing the number of compute multipliers to grow faster and faster.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;We got a hint at Anthropic's multi-chip strategy last September in their &lt;a href="https://www.anthropic.com/research/a-postmortem-of-three-recent-issues"&gt;postmortem of three recent issues&lt;/a&gt;, where they described bugs that affected just their TPU deployments.


    &lt;p&gt;Tags: &lt;a href="https://simonwillison.net/tags/ai"&gt;ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/generative-ai"&gt;generative-ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/llms"&gt;llms&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/anthropic"&gt;anthropic&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/gpus"&gt;gpus&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/claude-mythos-fable"&gt;claude-mythos-fable&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/recursive-self-improvement"&gt;recursive-self-improvement&lt;/a&gt;&lt;/p&gt;

</summary><category term="ai"/><category term="generative-ai"/><category term="llms"/><category term="anthropic"/><category term="gpus"/><category term="claude-mythos-fable"/><category term="recursive-self-improvement"/></entry><entry><title>Introducing Our Smallest, Most Efficient Model: Muse Ember</title><link href="https://simonwillison.net/2026/Nov/10/muse-ember/" rel="alternate"/><published>2026-11-10T07:28:21+00:00</published><updated>2026-11-10T07:28:21+00:00</updated><id>https://simonwillison.net/2026/Nov/10/muse-ember/</id><summary type="html">
&lt;p&gt;&lt;strong&gt;&lt;a href="https://research.meta.ai/blog/muse-ember"&gt;Introducing Our Smallest, Most Efficient Model: Muse Ember&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;
Meta AI just launched their entry in the Claude Haiku/GPT Luna category: Muse Ember, which they describe as:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;[...] our smallest and most efficient model designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;For shorter prompts it's priced the same as GPT-6 Luna and Claude Haiku 5.5 - both are $0.10/million input tokens and $0.50/million output. Today's other big model release, Claude Fable 5.2, is &lt;em&gt;80x that price&lt;/em&gt; at $8/$40.&lt;/p&gt;
&lt;p&gt;The pricing is a little less straightforward as prompts get longer, though. Luna, Ember, and Haiku all support up to a million input tokens, but they also all have a threshold after which the price goes up. Taking &lt;em&gt;that&lt;/em&gt; threshold into account Ember is priced between the two others:&lt;/p&gt;
&lt;center&gt;&lt;table&gt;
  &lt;thead&gt;
    &lt;tr&gt;
      &lt;th&gt;Model&lt;/th&gt;
      &lt;th&gt;Prompt size&lt;/th&gt;
      &lt;th&gt;Input ($/M)&lt;/th&gt;
      &lt;th&gt;Cached input&lt;/th&gt;
      &lt;th&gt;Output&lt;/th&gt;
    &lt;/tr&gt;
  &lt;/thead&gt;
  &lt;tbody&gt;
    &lt;tr&gt;
      &lt;td rowspan="2"&gt;GPT-6 Luna&lt;/td&gt;
      &lt;td&gt;≤272k&lt;/td&gt;
      &lt;td&gt;$0.10&lt;/td&gt;
      &lt;td&gt;$0.01&lt;/td&gt;
      &lt;td&gt;$0.50&lt;/td&gt;
    &lt;/tr&gt;
    &lt;tr&gt;
      &lt;td&gt;&amp;gt;272k&lt;/td&gt;
      &lt;td&gt;$0.20&lt;/td&gt;
      &lt;td&gt;$0.02&lt;/td&gt;
      &lt;td&gt;$0.75&lt;/td&gt;
    &lt;/tr&gt;
    &lt;tr&gt;
      &lt;td rowspan="2"&gt;Muse Ember 1.0&lt;/td&gt;
      &lt;td&gt;≤200k&lt;/td&gt;
      &lt;td&gt;$0.10&lt;/td&gt;
      &lt;td&gt;$0.004&lt;/td&gt;
      &lt;td&gt;$0.50&lt;/td&gt;
    &lt;/tr&gt;
    &lt;tr&gt;
      &lt;td&gt;&amp;gt;200k&lt;/td&gt;
      &lt;td&gt;$0.30&lt;/td&gt;
      &lt;td&gt;$0.012&lt;/td&gt;
      &lt;td&gt;$1.50&lt;/td&gt;
    &lt;/tr&gt;
    &lt;tr&gt;
      &lt;td rowspan="2"&gt;Claude Haiku 5.5&lt;/td&gt;
      &lt;td&gt;≤100k&lt;/td&gt;
      &lt;td&gt;$0.10&lt;/td&gt;
      &lt;td&gt;$0.01&lt;/td&gt;
      &lt;td&gt;$0.50&lt;/td&gt;
    &lt;/tr&gt; 
    &lt;tr&gt;
      &lt;td&gt;&amp;gt;100k&lt;/td&gt;
      &lt;td&gt;$0.50&lt;/td&gt;
      &lt;td&gt;$0.05&lt;/td&gt;
      &lt;td&gt;$2.50&lt;/td&gt;
    &lt;/tr&gt;
  &lt;/tbody&gt;
&lt;/table&gt;&lt;/center&gt;

&lt;p&gt;But comparing price-per-token doesn't really make sense these days, when reasoning tokens mean that different models might use dramatically different numbers of tokens to solve the same problem.&lt;/p&gt;
&lt;p&gt;For that, one measure that's useful is the Artificial Analysis. &lt;/p&gt;
&lt;p&gt;&lt;img alt="Stacked bar chart from Artificial Analysis comparing a cost in dollars across six models, each bar split into four unlabeled stacked segments, sorted from lowest to highest total. Claude Haiku 5.5 (low): $0.04* total, bottom segment $0.02, three thin unlabeled segments above. GPT-6 Luna (non-reasoning): $0.07 total, bottom segment $0.06, thin unlabeled segments above. GPT-6 Luna (xhigh): $0.10 total, segments from bottom $0.07, $0.02, $0.02, plus a thin unlabeled top segment. Muse Ember (xhigh): $0.12 total, segments from bottom $0.06, $0.02, $0.02, plus a thin unlabeled top segment. Muse Ember (max): $0.14 total, segments from bottom $0.07, $0.03, $0.03, $0.02. Claude Haiku 5.5 (xhigh): $0.25* total, segments from bottom $0.14, $0.04, $0.05, $0.02. Anthropic bars are terracotta, OpenAI bars are dark gray, Meta bars are blue, each with the company logo beneath." src="https://static.simonwillison.net/static/2026/artificial-analysis-haiku-class.webp" /&gt;&lt;/p&gt;
&lt;p&gt;And even more useful than that is their Intelligence Index vs. Cost per Task chart, which shows the following for those three models:&lt;/p&gt;
&lt;p&gt;&lt;img alt="Scatter plot from Artificial Analysis titled &amp;quot;Intelligence Index vs. Cost per Task&amp;quot;, plotting Artificial Analysis Intelligence Index (y-axis, 5 to 40) against cost per task in USD on a log scale (x-axis, $0.04 to $0.3), with a green &amp;quot;most attractive quadrant&amp;quot; in the upper left and a dotted Pareto line. Claude Haiku 5.5 (low)* scores about 19 at roughly $0.044. GPT-6 Luna (non-reasoning) scores about 9.5 at roughly $0.07. GPT-6 Luna (xhigh) scores about 23.5 at roughly $0.10. Muse Ember (xhigh) scores about 34 at roughly $0.12 and Muse Ember (max) about 36 at roughly $0.14, both on the Pareto line. Claude Haiku 5.5 (xhigh)* scores about 34.5 at roughly $0.25, well to the right of the Pareto line. Three unlabeled gray dots also sit on the Pareto line." src="https://static.simonwillison.net/static/2026/artificial-analysis-pareto.webp" /&gt;&lt;/p&gt;
&lt;p&gt;The Muse Ember models sit right on that Pareto line - the line at which you can only improve one dimension by giving up another. Meta call this out in their announcement:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;It also sits on the efficiency frontier of the full Artificial Analysis leaderboard: no model scores higher for less per task.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;(That grey dot on the Pareto line in the middle of the green area is GPT-6.1 Terra (low) - clearly I should be paying more attention to the Terra family!)&lt;/p&gt;
&lt;p&gt;Muse Ember has six reasoning levels - minimal, low, medium, high, xhigh, and max - so I &lt;a href="https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2Fc7666e8a54d84f724ac8121f9fa6cc4d"&gt;rendered six pelicans riding bicycles&lt;/a&gt;. The "high" reasoning level failed to return a full SVG, and the other SVGs were surprising in that the reasoning level didn't cleanly map to the visual quality of the response.&lt;/p&gt;
&lt;p&gt;Here's the "max" one, but the others are worth a look too:&lt;/p&gt;
&lt;p&gt;&lt;img alt="Bicycle frame is OK, basket has fish, pelican has a weird blue hat on." src="https://static.simonwillison.net/static/2026/muse-ember.webp" /&gt;&lt;/p&gt;
&lt;p&gt;The Ember pelicans are comparable in quality to those &lt;a href="https://tools.simonwillison.net/markdown-svg-renderer?url=https://gist.github.com/simonw/185cff7ff0b11379725abf7f78d5d2b2"&gt;I got from Muse Wonder&lt;/a&gt;, which is &lt;a href="https://www.llm-prices.com/#sel=muse-ember-1.0%2Cmuse-ember-1.0-200k%2Cmuse-wonder-1.0"&gt;40x the price&lt;/a&gt; of Muse Ember.


    &lt;p&gt;Tags: &lt;a href="https://simonwillison.net/tags/ai"&gt;ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/generative-ai"&gt;generative-ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/llms"&gt;llms&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/meta"&gt;meta&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/llm-pricing"&gt;llm-pricing&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/llm-release"&gt;llm-release&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/artificial-analysis"&gt;artificial-analysis&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/muse"&gt;muse&lt;/a&gt;&lt;/p&gt;

</summary><category term="ai"/><category term="generative-ai"/><category term="llms"/><category term="meta"/><category term="llm-pricing"/><category term="llm-release"/><category term="artificial-analysis"/><category term="muse"/></entry><entry><title>datasette #3046: Don't serve link: rel=next HTTP headers to Firefox for HTML pages</title><link href="https://simonwillison.net/2026/Nov/7/link-rel-next/" rel="alternate"/><published>2026-11-07T23:57:46+00:00</published><updated>2026-11-07T23:57:46+00:00</updated><id>https://simonwillison.net/2026/Nov/7/link-rel-next/</id><summary type="html">
&lt;p&gt;&lt;strong&gt;&lt;a href="https://github.com/simonw/datasette/issues/3046"&gt;datasette #3046: Don&amp;#x27;t serve link: rel=next HTTP headers to Firefox for HTML pages&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;
I was looking at the Firefox network panel today when I noticed that requests to a Datasette table page were triggering an &lt;a href="https://developer.mozilla.org/en-US/docs/Glossary/Prefetch"&gt;HTTP prefetch&lt;/a&gt; to the next page of rows in that table.&lt;/p&gt;
&lt;p&gt;This was happening because Datasette pages serve an HTTP &lt;code&gt;link&lt;/code&gt; header like this:&lt;/p&gt;
&lt;p&gt;&lt;code&gt;link: &amp;lt;https://latest.datasette.io/fixtures/compound_three_primary_keys?_next=a%2Cd%2Cv&amp;gt;; rel="next"&lt;/code&gt;&lt;/p&gt;
&lt;p&gt;This is bad. A tiny fraction of the visitors to one of those pages will click the "next" link, which means for everyone else their browser is fetching unnecessary data and the Datasette server is having to render an entire page that then gets thrown away.&lt;/p&gt;
&lt;p&gt;I dug in and learned that Firefox has &lt;a href="https://bugzilla.mozilla.org/show_bug.cgi?id=175418"&gt;an issue&lt;/a&gt; that has been open for 24 years about this behavior...&lt;/p&gt;
&lt;p&gt;And one of the early comments on that issue links to &lt;a href="https://simonwillison.net/2002/Oct/17/mozillaPrefetching/"&gt;my blog post from 2002&lt;/a&gt; complaining about this feature. I hadn't just forgotten that I had written about it, I had forgotten the feature entirely!


    &lt;p&gt;Tags: &lt;a href="https://simonwillison.net/tags/firefox"&gt;firefox&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/html"&gt;html&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/mozilla"&gt;mozilla&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/datasette"&gt;datasette&lt;/a&gt;&lt;/p&gt;

</summary><category term="firefox"/><category term="html"/><category term="mozilla"/><category term="datasette"/></entry><entry><title>I discovered a sound that leaks your exact location (and it's everywhere) - YouTube</title><link href="https://simonwillison.net/2026/Nov/3/a-sound-that-leaks-your-exact-location/" rel="alternate"/><published>2026-11-03T05:32:25+00:00</published><updated>2026-11-03T05:32:25+00:00</updated><id>https://simonwillison.net/2026/Nov/3/a-sound-that-leaks-your-exact-location/</id><summary type="html">
&lt;p&gt;&lt;strong&gt;&lt;a href="https://www.youtube.com/watch?v=UytYrhYLOKI"&gt;I discovered a sound that leaks your exact location (and it&amp;#x27;s everywhere) - YouTube&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;
Genius piece of OSINT by Colsto, who demonstrates how you can take almost imperceptible background noise of planes from a video soundtrack, figure out their Doppler curves, then cross reference them against ADS-B data to find the exact location of the video based on their speed, distance, and gaps between different flights.&lt;/p&gt;
&lt;p&gt;&lt;lite-youtube videoid="UytYrhYLOKI" js-api="js-api"
  title="I discovered a sound that leaks your exact location (and it's everywhere)"
  playlabel="Play: I discovered a sound that leaks your exact location (and it's everywhere)"
&gt; &lt;/lite-youtube&gt;

    &lt;p&gt;&lt;small&gt;&lt;/small&gt;Via &lt;a href="https://lobste.rs/s/s1btqo/this_background_sound_leaks_your_exact"&gt;Lobste.rs&lt;/a&gt;&lt;/small&gt;&lt;/p&gt;


    &lt;p&gt;Tags: &lt;a href="https://simonwillison.net/tags/privacy"&gt;privacy&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/youtube"&gt;youtube&lt;/a&gt;&lt;/p&gt;

</summary><category term="privacy"/><category term="youtube"/></entry><entry><title>How We Took Over ChatGPT Accounts</title><link href="https://simonwillison.net/2026/Oct/31/how-we-took-over-chatgpt-accounts/" rel="alternate"/><published>2026-10-31T23:15:22+00:00</published><updated>2026-10-31T23:15:22+00:00</updated><id>https://simonwillison.net/2026/Oct/31/how-we-took-over-chatgpt-accounts/</id><summary type="html">
&lt;p&gt;&lt;strong&gt;&lt;a href="https://www.parameter.ai/labs/we-found-a-gap-at-openai"&gt;How We Took Over ChatGPT Accounts&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;
Parameter Labs describe an exploit they found against ChatGPT's implementation of the artifacts pattern (OpenAI used to call this "Canvas" but it looks like the latest revision is called &lt;a href="https://help.openai.com/en/articles/20001246-working-with-writing-blocks-and-code-blocks-in-chatgpt"&gt;HTML previews&lt;/a&gt;), where interactive custom HTML apps are displayed in a sandboxed iframe.&lt;/p&gt;
&lt;p&gt;It's important that HTML in these stays tightly sandboxed to avoid untrusted code interacting with OpenAI APIs and accessing private data.&lt;/p&gt;
&lt;p&gt;Parameter found quite a complex attack chain. OpenAI use the MathLive JavaScript library as part of their graph support, and that library includes &lt;a href="https://mathlive.io/mathfield/virtual-keyboard/"&gt;its own virtual keyboard&lt;/a&gt;, which can be triggered in the parent page via &lt;code&gt;postMessage()&lt;/code&gt; from the sandboxed iframe.&lt;/p&gt;
&lt;p&gt;That message receiving code uses &lt;code&gt;innerHTML&lt;/code&gt;, so anything that sends a message can add custom HTML to the parent - and there was no mechanism to prevent other sandboxed frames, including HTML preview frames, from calling that method.&lt;/p&gt;
&lt;p&gt;This meant untrusted code in an iframe could insert HTML, and while &lt;code&gt;innerHTML&lt;/code&gt; doesn't directly execute script elements it could be used to inject a new iframe &lt;em&gt;without&lt;/em&gt; sandbox restrictions.&lt;/p&gt;
&lt;p&gt;OpenAI’s CSP rules blocked inline scripts entirely, and only allowed script loading from a handful of domains, but one of those was YouTube - and the YouTube oEmbed endpoint could be used to reflect JavaScript back in a way that executed it.&lt;/p&gt;
&lt;p&gt;Once the attack had JavaScript execution on ChatGPT, it was trivial to fetch private user data from APIs and exfiltrate it via an image reference.&lt;/p&gt;
&lt;p&gt;The end result: you could create a ChatGPT share URL which, when visited by another signed-in user, could exfiltrate their private data or take actions on their behalf.&lt;/p&gt;
&lt;p&gt;OpenAI closed the hole and paid out a $5,000 bug bounty, which I think was a bit cheap for a vulnerability this serious.

    &lt;p&gt;&lt;small&gt;&lt;/small&gt;Via &lt;a href="https://news.ycombinator.com/item?id=50314699"&gt;Hacker News&lt;/a&gt;&lt;/small&gt;&lt;/p&gt;


    &lt;p&gt;Tags: &lt;a href="https://simonwillison.net/tags/iframes"&gt;iframes&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/javascript"&gt;javascript&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/sandboxing"&gt;sandboxing&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/security"&gt;security&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/xss"&gt;xss&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/openai"&gt;openai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/chatgpt"&gt;chatgpt&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/content-security-policy"&gt;content-security-policy&lt;/a&gt;&lt;/p&gt;

</summary><category term="iframes"/><category term="javascript"/><category term="sandboxing"/><category term="security"/><category term="xss"/><category term="openai"/><category term="chatgpt"/><category term="content-security-policy"/></entry><entry><title>dbreunig/nesformer</title><link href="https://simonwillison.net/2026/Oct/30/nesformer/" rel="alternate"/><published>2026-10-30T21:59:22+00:00</published><updated>2026-10-30T21:59:22+00:00</updated><id>https://simonwillison.net/2026/Oct/30/nesformer/</id><summary type="html">
&lt;p&gt;&lt;strong&gt;&lt;a href="https://github.com/dbreunig/nesformer"&gt;dbreunig/nesformer&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;
Inspired by &lt;a href="https://divagr.com/blog/qwenputer"&gt;qwenputer&lt;/a&gt; by Divyansh Agrawal, Drew Breunig baked a working emulator for a Nintendo Entertainment System 6502 CPU into the weights of Qwen3.5-4B. You can now play NES ROMs such as Super Mario Bros (very slowly) running inside Qwen!&lt;/p&gt;
&lt;p&gt;In &lt;a href="https://news.ycombinator.com/item?id=50306442#50307402"&gt;a comment on Hacker News&lt;/a&gt; Drew reveals the Opus 5.5 prompts he used to build this system:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;code&gt;Read this article and any relevant source code: https://divagr.com/blog/qwenputer. Then explore if you could emulate the chip and ram in a Nintendo Entertainment System (as people do today with FPGAs in MISTRs) using this technique. If so, what size model would you have to build and if you could use smaller faster ones, could you get it to be at the standard clockspeed.&lt;/code&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;Followed by:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;code&gt;I think we should build it! Just for fun. It's okay if it's not fast. We want to be able to have the weights, and a cli to run them that we can point at a ROM.&lt;/code&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;Apparently Opus found and then used &lt;a href="https://github.com/christopherpow/nes-test-roms/blob/master/other/nestest.txt"&gt;this collection of conformance tests&lt;/a&gt; for the NES CPU released 15 years ago by Christopher Pow.

    &lt;p&gt;&lt;small&gt;&lt;/small&gt;Via &lt;a href="https://news.ycombinator.com/item?id=50306442"&gt;Hacker News&lt;/a&gt;&lt;/small&gt;&lt;/p&gt;


    &lt;p&gt;Tags: &lt;a href="https://simonwillison.net/tags/games"&gt;games&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/ai"&gt;ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/generative-ai"&gt;generative-ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/llms"&gt;llms&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/drew-breunig"&gt;drew-breunig&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/qwen"&gt;qwen&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/conformance-suites"&gt;conformance-suites&lt;/a&gt;&lt;/p&gt;

</summary><category term="games"/><category term="ai"/><category term="generative-ai"/><category term="llms"/><category term="drew-breunig"/><category term="qwen"/><category term="conformance-suites"/></entry><entry><title>API credits benefit for eligible ChatGPT subscriptions</title><link href="https://simonwillison.net/2026/Oct/29/api-credits/" rel="alternate"/><published>2026-10-29T22:48:29+00:00</published><updated>2026-10-29T22:48:29+00:00</updated><id>https://simonwillison.net/2026/Oct/29/api-credits/</id><summary type="html">
&lt;p&gt;&lt;strong&gt;&lt;a href="https://help.openai.com/en/articles/20001626-api-credits-benefit-for-eligible-chatgpt-subscriptions"&gt;API credits benefit for eligible ChatGPT subscriptions&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;
Earlier this month Anthropic &lt;a href="https://simonwillison.net/2026/Oct/7/claude-haiku-5-5/#api-credits"&gt;announced a generous new benefit&lt;/a&gt; for their Max and Team subscriptions: $100, $200, and $500 per month subscribers would get the equivalent quantity of (expiring) credits to use with the Anthropic API.&lt;/p&gt;
&lt;p&gt;OpenAI have now launched pretty much exactly the same deal: if you have a $100, $200, or $500 subscription plan with them you now get that same quantity of credits to apply to one of your API organizations.&lt;/p&gt;
&lt;p&gt;I maintain subscriptions with both companies, so this is great news for me! I may even ship some public-facing demos now that the credits are already covered by my existing subscriptions.


    &lt;p&gt;Tags: &lt;a href="https://simonwillison.net/tags/ai"&gt;ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/openai"&gt;openai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/generative-ai"&gt;generative-ai&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/llms"&gt;llms&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/anthropic"&gt;anthropic&lt;/a&gt;, &lt;a href="https://simonwillison.net/tags/llm-pricing"&gt;llm-pricing&lt;/a&gt;&lt;/p&gt;

</summary><category term="ai"/><category term="openai"/><category term="generative-ai"/><category term="llms"/><category term="anthropic"/><category term="llm-pricing"/></entry></feed>