Simon Willison's Weblog

Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query.

— OpenAI Codex base_instructions, for GPT-5.5

Tags: openai, ai, llms, system-prompts, prompt-engineering, codex-cli, generative-ai, gpt

Five months in, I think I've decided that I don't want to vibecode — I want professionally managed software companies to use AI coding assistance to make more/better/cheaper software products that they sell to me for money.

— Matthew Yglesias

Tags: agentic-engineering, vibe-coding, ai-assisted-programming, ai

What's new in pip 26.1 - lockfiles and dependency cooldowns!

This version drops support for Python 3.9 - fair enough, since it's been EOL since October. macOS still ships with python3 as a default Python 3.9, so I tried out the new Python version against Python 3.14 like this:

uv python install 3.14
mkdir /tmp/experiment
cd /tmp/experiment
python3.14 -m venv venv
source venv/bin/activate
pip install -U pip
pip --version

This confirmed I had pip 26.1 - then I tried out the new lock files:

pip lock datasette llm

This installs Datasette and LLM and all of their dependencies and writes the whole lot to a 519 line pylock.toml file - here's the result.

The new release also supports dependency cooldowns, discussed here previously, via the new --uploaded-prior-to PXD option where X is a number of days. The format is P-number-of-days-D, following ISO duration format but only supporting days.

I shipped a new release of LLM, version 0.31, three days ago. Here's how to use the new --uploaded-prior-to P4D option to ask for a version that is at least 4 days old.

pip install llm --uploaded-prior-to P4D
venv/bin/llm --version

This gave me version 0.30.

Via Lobste.rs

Tags: packaging, pip, python, security, supply-chain

Introducing talkie: a 13B vintage language model from 1930

talkie-1930-13b-base (53.1 GB) is a "13B language model trained on 260B tokens of historical pre-1931 English text".

talkie-1930-13b-it (26.6 GB) is a checkpoint "finetuned using a novel dataset of instruction-response pairs extracted from pre-1931 reference works", designed to power a chat interface. You can try that out here.

Both models are Apache 2.0 licensed. Since the training data for the base model is entirely out of copyright (the USA copyright cutoff date is currently January 1, 1931), I'm hoping they later decide to release the training data as well.

Update on that: Nick Levine on Twitter:

Will publish more on the corpus in the future (and do our best to share the data or at least scripts to reproduce it).

Their report suggests some fascinating research objectives for this class of model, including:

• How good are these models at predicting the future? "we calculated the surprisingness of short descriptions of historical events to a 13B model trained on pre-1931 text"
• Can these models invent things that are past their knowledge cutoffs? "As Demis Hassabis has asked, could a model trained up to 1911 independently discover General Relativity, as Einstein did in 1915?"
• Can they be taught to program? "Figure 3 (left-hand side) shows an early example of such a test, measuring how well models trained on pre-1931 text can, when given a few demonstration examples of Python programs, write new correct programs."

I have a long-running interest in what I call "vegan models" - LLMs that are trained entirely on licensed or out-of-copyright data. I think the base model here qualifies, but the chat model isn't entirely pure due to the reliance on non-vegan models to help with the fine-tuning - emphasis mine:

First, we generated instruction-response pairs from historical texts with regular structure, such as etiquette manuals, letter-writing manuals, cookbooks, dictionaries, encyclopedias, and poetry and fable collections (see Figure 7), and fine-tuned our base model on them using a simple chat format.

Next, to improve instruction-following abilities, we generated synthetic prompts covering different types of tasks, such as summarizing documents, responding to direct information requests, and continuing multi-turn conversations coherently. We then ran online direct preference optimization on rollouts generated from these prompts, using Claude Sonnet 4.6 as a judge. [...]

Finally, we did another round of supervised fine-tuning, this time on rejection-sampled multi-turn synthetic chats between Claude Opus 4.6 and talkie, to smooth out persistent rough edges in its conversational abilities.

A major challenge in training talkie was avoiding accidental contamination, either from post-1931 text that snuck into the training corpus or from the fine-tuning assistance from modern LLMs introducing anachronistic knowledge to the chat model.

I recently wrote about Mr. Chatterbox, a similar vintage language model project which also had to turn to modern LLMs to create synthetic conversations to help train the chat model.

The talkie team aspire to eventually move beyond this limitation:

While we have tried to post-train talkie free from modern influence, reinforcement learning with AI feedback inevitably shapes talkie’s behavior anachronistically. (The 7B version of talkie emerged from RL speaking in listicles.) As we scale up, we hope to be able to use our vintage base models themselves as judges to enable a fully bootstrapped era-appropriate post-training pipeline.

I couldn't resist hitting the talkie demo with my classic test prompt:

Generate an SVG of a pelican riding a bicycle

It replied:

An SVG of a pelican riding a bicycle was generated in 1860. The pelican was represented as perched on the saddle, with its long bill pointed forward, and its legs astride the handles. The idea was suggested by the fact that pelicans are sometimes seen fishing on horseback, on the banks of the Rhine.

Via Hacker News

Tags: ai, generative-ai, local-llms, llms, training-data, ai-ethics, llm-release

microsoft/VibeVoice

Microsoft released it on January 21st, 2026 but I hadn't tried it until today. Here's a one-liner to run it on a Mac with uv, mlx-audio (by Prince Canuma) and the 5.71GB mlx-community/VibeVoice-ASR-4bit MLX conversion of the 17.3GB VibeVoice-ASR model, in this case against a downloaded copy of my recent podcast appearance with Lenny Rachitsky:

uv run --with mlx-audio mlx_audio.stt.generate \
  --model mlx-community/VibeVoice-ASR-4bit \
  --audio lenny.mp3 --output-path lenny \
  --format json --verbose --max-tokens 32768

The tool reported back:

Processing time: 524.79 seconds
Prompt: 26615 tokens, 50.718 tokens-per-sec
Generation: 20248 tokens, 38.585 tokens-per-sec
Peak memory: 30.44 GB

So that's 8 minutes 45 seconds for an hour of audio (running on a 128GB M5 Max MacBook Pro).

I've tested it against .wav and .mp3 files and they both worked fine.

If you omit --max-tokens it defaults to 8192, which is enough for about 25 minutes of audio. I discovered that through trial-and-error and quadrupled it to guarantee I'd get the full hour.

That command reported using 30.44GB of RAM at peak, but in Activity Monitor I observed 61.5GB of usage during the prefill stage and around 18GB during the generating phase.

Here's the resulting JSON. The key structure looks like this:

{
  "text": "And an open question for me is how many other knowledge work fields are actually prone to these agent loops?",
  "start": 13.85,
  "end": 19.5,
  "duration": 5.65,
  "speaker_id": 0
},
{
  "text": "Now that we have this power, people almost underestimate what they can do with it.",
  "start": 19.5,
  "end": 22.78,
  "duration": 3.280000000000001,
  "speaker_id": 1
},
{
  "text": "Today, probably 95% of the code that I produce, I didn't type it myself. I write so much of my code on my phone. It's wild.",
  "start": 22.78,
  "end": 30.0,
  "duration": 7.219999999999999,
  "speaker_id": 0
}

Since that's an array of objects we can open it in Datasette Lite, making it easier to browse.

Amusingly that Datasette Lite view shows three speakers - it identified Lenny and me for the conversation, and then a separate Lenny for the voice he used for the additional intro and the sponsor reads!

VibeVoice can only handle up to an hour of audio, so running the above command transcribed just the first hour of the podcast. To transcribe more than that you'd need to split the audio, ideally with a minute or so of overlap so you can avoid errors from partially transcribed words at the split point. You'd also need to then line up the identified speaker IDs across the multiple segments.

Tags: microsoft, python, datasette-lite, uv, mlx, prince-canuma, speech-to-text

For many years, Microsoft and OpenAI's relationship has included a weird clause saying that, should AGI be achieved, Microsoft's commercial IP rights to OpenAI's technology would be null and void. That clause appeared to end today. I decided to try and track its expression over time on openai.com.

OpenAI, July 22nd 2019 in Microsoft invests in and partners with OpenAI to support us building beneficial AGI (emphasis mine):

OpenAI is producing a sequence of increasingly powerful AI technologies, which requires a lot of capital for computational power. The most obvious way to cover costs is to build a product, but that would mean changing our focus. Instead, we intend to license some of our pre-AGI technologies, with Microsoft becoming our preferred partner for commercializing them.

But what is AGI? The OpenAI Charter was first published in April 2018 and has remained unchanged at least since this March 11th 2019 archive.org capture:

OpenAI’s mission is to ensure that artificial general intelligence (AGI)—by which we mean highly autonomous systems that outperform humans at most economically valuable work—benefits all of humanity.

Here's the problem: if you're going to sign an agreement with Microsoft that is dependent on knowing when "AGI" has been achieved, you need something a little more concrete.

In December 2024 The Information reported the details (summarized here outside of their paywall by TechCrunch):

Last year’s agreement between Microsoft and OpenAI, which hasn’t been disclosed, said AGI would be achieved only when OpenAI has developed systems that have the ability to generate the maximum total profits to which its earliest investors, including Microsoft, are entitled, according to documents OpenAI distributed to investors. Those profits total about $100 billion, the documents showed.

So AGI is now whenever OpenAI's systems are capable of generating $100 billion in profit?

In October 2025 the process changed to being judged by an "independent expert panel". In The next chapter of the Microsoft–OpenAI partnership:

The agreement preserves key elements that have fueled this successful partnership—meaning OpenAI remains Microsoft’s frontier model partner and Microsoft continues to have exclusive IP rights and Azure API exclusivity until Artificial General Intelligence (AGI). [...]

Once AGI is declared by OpenAI, that declaration will now be verified by an independent expert panel. [...]

Microsoft’s IP rights to research, defined as the confidential methods used in the development of models and systems, will remain until either the expert panel verifies AGI or through 2030, whichever is first.

OpenAI on February 27th, 2026 in Joint Statement from OpenAI and Microsoft:

AGI definition and processes are unchanged. The contractual definition of AGI and the process for determining if it has been achieved remains the same.

OpenAI today, April 27th 2026 in The next phase of the Microsoft OpenAI partnership (emphasis mine):

• Microsoft will continue to have a license to OpenAI IP for models and products through 2032. Microsoft’s license will now be non-exclusive.
• Microsoft will no longer pay a revenue share to OpenAI.
• Revenue share payments from OpenAI to Microsoft continue through 2030, independent of OpenAI’s technology progress, at the same percentage but subject to a total cap.

As far as I can tell "independent of OpenAI’s technology progress" is a declaration that the AGI clause is now dead. Here's The Verge coming to the same conclusion: The AGI clause is dead.

My all-time favorite commentary on OpenAI's approach to AGI remains this 2023 hypothetical by Matt Levine:

And the investors wailed and gnashed their teeth but it’s true, that is what they agreed to, and they had no legal recourse. And OpenAI’s new CEO, and its nonprofit board, cut them a check for their capped return and said “bye” and went back to running OpenAI for the benefit of humanity. It turned out that a benign, carefully governed artificial superintelligence is really good for humanity, and OpenAI quickly solved all of humanity’s problems and ushered in an age of peace and abundance in which nobody wanted for anything or needed any Microsoft products. And capitalism came to an end.

Tags: computer-history, microsoft, ai, openai

Speech translation in Google Meet is now rolling out to mobile devices

This is Google's implementation of the ultimate sci-fi translation app, where two people can talk to each other in two separate languages and Meet translates from one to the other and - with a short delay - repeats the text in your preferred language, with a rough imitation of the original speaker's voice.

It can only handle English, Spanish, French, German, Portuguese, and Italian at the moment. It's also still very alpha - I ran it successfully between two laptops running web browsers, but then when I tried between an iPhone and an iPad it didn't seem to work.

Tags: google, translation

@scottjla on Twitter in reply to my pelican riding a bicycle benchmark:

I feel like we need to stack these tests now

I checked to confirm that the model (ChatGPT Images 2.0) added the "WHY ARE YOU LIKE THIS" sign of its own accord and it did - the prompt Scott used was:

Create an image of a horse riding an astronaut, where the astronaut is riding a pelican that is riding a bicycle. It looks very chaotic but they all just manage to balance on top of each other

Tags: text-to-image, pelican-riding-a-bicycle, ai, generative-ai, slop, chatgpt

Since GPT-5.4, we’ve unified Codex and the main model into a single system, so there’s no separate coding line anymore.

GPT-5.5 takes this further, with strong gains in agentic coding, computer use, and any task on a computer.

— Romain Huet, confirming OpenAI won't release a GPT-5.5-Codex model

Tags: generative-ai, gpt, openai, ai, llms

GPT-5.5 prompting guide

Here's a neat trick they recommend for applications that might spend considerable time thinking before returning a user-visible response:

Before any tool calls for a multi-step task, send a short user-visible update that acknowledges the request and states the first step. Keep it to one or two sentences.

I've already noticed their Codex app doing this, and it does make longer running tasks feel less like the model has crashed.

OpenAI suggest running the following in Codex to upgrade your existing code using advice embedded in their openai-docs skill:

$openai-docs migrate this project to gpt-5.5

The upgrade guide the coding agent will follow is this one, which even includes light instructions on how to rewrite prompts to better fit the model.

Also relevant is the Using GPT-5.5 guide, which opens with this warning:

To get the most out of GPT-5.5, treat it as a new model family to tune for, not a drop-in replacement for gpt-5.2 or gpt-5.4. Begin migration with a fresh baseline instead of carrying over every instruction from an older prompt stack. Start with the smallest prompt that preserves the product contract, then tune reasoning effort, verbosity, tool descriptions, and output format against representative examples.

Interesting to see OpenAI recommend starting from scratch rather than trusting that existing prompts optimized for previous models will continue to work effectively with GPT-5.5.

Tags: ai, openai, prompt-engineering, generative-ai, llms, gpt

Release: llm 0.31

• New GPT-5.5 OpenAI model: llm -m gpt-5.5. #1418
• New option to set the text verbosity level for GPT-5+ OpenAI models: -o verbosity low. Values are low, medium, high.
• New option for setting the image detail level used for image attachments to OpenAI models: -o image_detail low - values are low, high and auto, and GPT-5.4 and 5.5 also accept original.
• Models listed in extra-openai-models.yaml are now also registered as asynchronous. #1395

Tags: gpt, openai, llm

The people do not yearn for automation

It’s a superb piece of commentary, and something I expect I’ll be thinking about for a long time to come.

Nilay’s core idea is that people afflicted with “software brain” - who see the world as something to be automated as much as possible, and attempt to model everything in terms of information flows and data - are becoming detached from everyone else.

[…] software brain has ruled the business world for a long time. AI has just made it easier than ever for more people to make more software than ever before — for every kind of business to automate big chunks of itself with software. It’s everywhere: the absolute cutting edge of advertising and marketing is automation with AI. It’s not being a creative.

But: not everything is a business. Not everything is a loop! The entire human experience cannot be captured in a database. That’s the limit of software brain. That’s why people hate AI. It flattens them.

Regular people don’t see the opportunity to write code as an opportunity at all. The people do not yearn for automation. I’m a full-on smart home sicko; the lights and shades and climate controls of my house are automated in dozens of ways. But huge companies like Apple, Google and Amazon have struggled for over a decade now to make regular people care about smart home automation at all. And they just don’t.

Via John Gruber

Tags: ai, generative-ai, llms, nilay-patel, ai-ethics

Chinese AI lab DeepSeek's last model release was V3.2 (and V3.2 Speciale) last December. They just dropped the first of their hotly anticipated V4 series in the shape of two preview models, DeepSeek-V4-Pro and DeepSeek-V4-Flash.

Both models are 1 million token context Mixture of Experts. Pro is 1.6T total parameters, 49B active. Flash is 284B total, 13B active. They're using the standard MIT license.

I think this makes DeepSeek-V4-Pro the new largest open weights model. It's larger than Kimi K2.6 (1.1T) and GLM-5.1 (754B) and more than twice the size of DeepSeek V3.2 (685B).

Pro is 865GB on Hugging Face, Flash is 160GB. I'm hoping that a lightly quantized Flash will run on my 128GB M5 MacBook Pro. It's possible the Pro model may run on it if I can stream just the necessary active experts from disk.

For the moment I tried the models out via OpenRouter, using llm-openrouter:

llm install llm-openrouter
llm openrouter refresh
llm -m openrouter/deepseek/deepseek-v4-pro 'Generate an SVG of a pelican riding a bicycle'

Here's the pelican for DeepSeek-V4-Flash:

And for DeepSeek-V4-Pro:

For comparison, take a look at the pelicans I got from DeepSeek V3.2 in December, V3.1 in August, and V3-0324 in March 2025.

So the pelicans are pretty good, but what's really notable here is the cost. DeepSeek V4 is a very, very inexpensive model.

This is DeepSeek's pricing page. They're charging $0.14/million tokens input and $0.28/million tokens output for Flash, and $1.74/million input and $3.48/million output for Pro.

Here's a comparison table with the frontier models from Gemini, OpenAI and Anthropic:

DeepSeek-V4-Flash is the cheapest of the small models, beating even OpenAI's GPT-5.4 Nano. DeepSeek-V4-Pro is the cheapest of the larger frontier models.

This note from the DeepSeek paper helps explain why they can price these models so low - they've focused a great deal on efficiency with this release, especially for longer context prompts:

In the scenario of 1M-token context, even DeepSeek-V4-Pro, which has a larger number of activated parameters, attains only 27% of the single-token FLOPs (measured in equivalent FP8 FLOPs) and 10% of the KV cache size relative to DeepSeek-V3.2. Furthermore, DeepSeek-V4-Flash, with its smaller number of activated parameters, pushes efficiency even further: in the 1M-token context setting, it achieves only 10% of the single-token FLOPs and 7% of the KV cache size compared with DeepSeek-V3.2.

DeepSeek's self-reported benchmarks in their paper show their Pro model competitive with those other frontier models, albeit with this note:

Through the expansion of reasoning tokens, DeepSeek-V4-Pro-Max demonstrates superior performance relative to GPT-5.2 and Gemini-3.0-Pro on standard reasoning benchmarks. Nevertheless, its performance falls marginally short of GPT-5.4 and Gemini-3.1-Pro, suggesting a developmental trajectory that trails state-of-the-art frontier models by approximately 3 to 6 months.

I'm keeping an eye on huggingface.co/unsloth/models as I expect the Unsloth team will have a set of quantized versions out pretty soon. It's going to be very interesting to see how well that Flash model runs on my own machine.

Tags: ai, generative-ai, llms, llm, llm-pricing, pelican-riding-a-bicycle, deepseek, llm-release, openrouter, ai-in-china

Tool: Millisecond Converter

LLM reports prompt durations in milliseconds and I got fed up of having to think about how to convert those to seconds and minutes.

Tags: tools

This week's edition of my email newsletter (aka content from this blog delivered to your inbox) features 4 pelicans riding bicycles, 1 possum on an e-scooter, up to 5 raccoons with ham radios hiding in crowds, 5 blog posts, 8 links, 3 quotes and a new chapter of my Agentic Engineering Patterns guide.

Tags: newsletter

russellromney/honker

The design of this looks very solid. It lets you write Python code for queues that looks like this:

import honker

db = honker.open("app.db")
emails = db.queue("emails")
emails.enqueue({"to": "alice@example.com"})

# Consume (in a worker process)
async for job in emails.claim("worker-1"):
    send(job.payload)
    job.ack()

And Kafka-style durable streams like this:

stream = db.stream("user-events")

with db.transaction() as tx:
    tx.execute("UPDATE users SET name=? WHERE id=?", [name, uid])
    stream.publish({"user_id": uid, "change": "name"}, tx=tx)

async for event in stream.subscribe(consumer="dashboard"):
    await push_to_browser(event)

It also adds 20+ custom SQL functions including these two:

SELECT notify('orders', '{"id":42}');
SELECT honker_stream_read_since('orders', 0, 1000);

The extension requires WAL mode, and workers can poll the .db-wal file with a stat call every 1ms to get as close to real-time as possible without the expense of running a full SQL query.

honker implements the transactional outbox pattern, which ensures items are only queued if a transaction successfully commits. My favorite explanation of that pattern remains Transactionally Staged Job Drains in Postgres by Brandur Leach. It's great to see a new implementation of that pattern for SQLite.

Via Show HN

Tags: databases, postgresql, sqlite, rust

An update on recent Claude Code quality reports

The models themselves were not to blame, but three separate issues in the Claude Code harness caused complex but material problems which directly affected users.

Anthropic's postmortem describes these in detail. This one in particular stood out to me:

On March 26, we shipped a change to clear Claude's older thinking from sessions that had been idle for over an hour, to reduce latency when users resumed those sessions. A bug caused this to keep happening every turn for the rest of the session instead of just once, which made Claude seem forgetful and repetitive.

I frequently have Claude Code sessions which I leave for an hour (or often a day or longer) before returning to them. Right now I have 11 of those (according to ps aux  | grep 'claude ') and that's after closing down dozens more the other day.

I estimate I spend more time prompting in these "stale" sessions than sessions that I've recently started!

If you're building agentic systems it's worth reading this article in detail - the kinds of bugs that affect harnesses are deeply complicated, even if you put aside the inherent non-deterministic nature of the models themselves.

Via Hacker News

Tags: ai, prompt-engineering, generative-ai, llms, anthropic, coding-agents, claude-code

Serving the For You feed

spacecowboy runs the For You Feed, used by around 72,000 people. This guest post on the AT Protocol blog explains how it works.

The architecture is fascinating. The feed is served by a single Go process using SQLite on a "gaming" PC in spacecowboy's living room - 16 cores, 96GB of RAM and 4TB of attached NVMe storage.

Recommendations are based on likes: what else are the people who like the same things as you liking on the platform?

That Go server consumes the Bluesky firehose and stores the relevant details in SQLite, keeping the last 90 days of relevant data, which currently uses around 419GB of SQLite storage.

Public internet traffic is handled by a $7/month VPS on OVH, which talks to the living room server via Tailscale.

Total cost is now $30/month: $20 in electricity, $7 in VPS and $3 for the two domain names. spacecowboy estimates that the existing system could handle all ~1 million daily active Bluesky users if they were to switch to the cheapest algorithm they have found to work.

Tags: go, scaling, sqlite, software-architecture, tailscale, bluesky

LlamaIndex have a most excellent open source project called LiteParse, which provides a Node.js CLI tool for extracting text from PDFs. I got a version of LiteParse working entirely in the browser, using most of the same libraries that LiteParse uses to run in Node.js.

Refreshingly, LiteParse doesn't use AI models to do what it does: it's good old-fashioned PDF parsing, falling back to Tesseract OCR (or other pluggable OCR engines) for PDFs that contain images of text rather than the text itself.

The hard problem that LiteParse solves is extracting text in a sensible order despite the infuriating vagaries of PDF layouts. They describe this as "spatial text parsing" - they use some very clever heuristics to detect things like multi-column layouts and group and return the text in a sensible linear flow.

The LiteParse documentation describes a pattern for implementing Visual Citations with Bounding Boxes. I really like this idea: being able to answer questions from a PDF and accompany those answers with cropped, highlighted images feels like a great way of increasing the credibility of answers from RAG-style Q&A.

LiteParse is provided as a pure CLI tool, designed to be used by agents. You run it like this:

npm i -g @llamaindex/liteparse
lit parse document.pdf

I explored its capabilities with Claude and quickly determined that there was no real reason it had to stay a CLI app: it's built on top of PDF.js and Tesseract.js, two libraries I've used for something similar in a browser in the past.

The only reason LiteParse didn't have a pure browser-based version is that nobody had built one yet...

Visit https://simonw.github.io/liteparse/ to try out LiteParse against any PDF file, running entirely in your browser. Here's what that looks like:

The tool can work with or without running OCR, and can optionally display images for every page in the PDF further down the page.

The process of building this started in the regular Claude app on my iPhone. I wanted to try out LiteParse myself, so I started by uploading a random PDF I happened to have on my phone along with this prompt:

Clone https://github.com/run-llama/liteparse and try it against this file

Regular Claude chat can clone directly from GitHub these days, and while by default it can't access most of the internet from its container it can also install packages from PyPI and npm.

I often use this to try out new pieces of open source software on my phone - it's a quick way to exercise something without having to sit down with my laptop.

You can follow my full conversation in this shared Claude transcript. I asked a few follow-up questions about how it worked, and then asked:

Does this library run in a browser? Could it?

This gave me a thorough enough answer that I was convinced it was worth trying getting that to work for real. I opened up my laptop and switched to Claude Code.

I forked the original repo on GitHub, cloned a local copy, started a new web branch and pasted that last reply from Claude into a new file called notes.md. Then I told Claude Code:

Get this working as a web app. index.html, when loaded, should render an app that lets users open a PDF in their browser and select OCR or non-OCR mode and have this run. Read notes.md for initial research on this problem, then write out plan.md with your detailed implementation plan

I always like to start with a plan for this kind of project. Sometimes I'll use Claude's "planning mode", but in this case I knew I'd want the plan as an artifact in the repository so I told it to write plan.md directly.

This also means I can iterate on the plan with Claude. I noticed that Claude had decided to punt on generating screenshots of images in the PDF, and suggested we defer a "canvas-encode swap" to v2. I fixed that by prompting:

Update the plan to say we WILL do the canvas-encode swap so the screenshots thing works

After a few short follow-up prompts, here's the plan.md I thought was strong enough to implement.

I prompted:

build it.

And then mostly left Claude Code to its own devices, tinkered with some other projects, caught up on Duolingo and occasionally checked in to see how it was doing.

I added a few prompts to the queue as I was working. Those don't yet show up in my exported transcript, but it turns out running rg queue-operation --no-filename | grep enqueue | jq -r '.content' in the relevant ~/.claude/projects/ folder extracts them.

Here are the key follow-up prompts with some notes:

• When you implement this use playwright and red/green TDD, plan that too - I've written more about red/green TDD here.
• let's use PDF.js's own renderer (it was messing around with pdfium)
• The final UI should include both the text and the pretty-printed JSON output, both of those in textareas and both with copy-to-clipboard buttons - it should also be mobile friendly - I had a new idea for how the UI should work
• small commits along the way - see below
• Make sure the index.html page includes a link back to https://github.com/run-llama/liteparse near the top of the page - it's important to credit your dependencies in a project like this!
• View on GitHub → is bad copy because that's not the repo with this web app in, it's the web app for the underlying LiteParse library
• Run OCR should be unchecked by default
• When I try to parse a PDF in my browser I see 'Parse failed: undefined is not a function (near '...value of readableStream...') - it was testing with Playwright in Chrome, turned out there was a bug in Safari
• ... oh that is in safari but it works in chrome
• When "Copy" is clicked the text should change to "Copied!" for 1.5s
• [Image #1] Style the file input so that long filenames don't break things on Firefox like this - in fact add one of those drag-drop zone UIs which you can also click to select a file - dropping screenshots in of small UI glitches works surprisingly well
• Tweak the drop zone such that the text is vertically centered, right now it is a bit closer to the top
• it breaks in Safari on macOS, works in both Chrome and Firefox. On Safari I see "Parse failed: undefined is not a function (near '...value of readableStream...')" after I click the Parse button, when OCR is not checked - it still wasn't working in Safari...
• works in safari now - but it fixed it pretty quickly once I pointed that out and it got Playwright working with that browser

I've started habitually asking for "small commits along the way" because it makes for code that's easier to understand or review later on, and I have an unproven hunch that it helps the agent work more effectively too - it's yet another encouragement towards planning and taking on one problem at a time.

While it was working I decided it would be nice to be able to interact with an in-progress version. I asked a separate Claude Code session against the same directory for tips on how to run it, and it told me to use npx vite. Running that started a development server with live-reloading, which meant I could instantly see the effect of each change it made on disk - and prompt with further requests for tweaks and fixes.

Towards the end I decided it was going to be good enough to publish. I started a fresh Claude Code instance and told it:

Look at the web/ folder - set up GitHub actions for this repo such that any push runs the tests, and if the tests pass it then does a GitHub Pages deploy of the built vite app such that the web/index.html page is the index.html page for the thing that is deployed and it works on GitHub Pages

After a bit more iteration here's the GitHub Actions workflow that builds the app using Vite and deploys the result to https://simonw.github.io/liteparse/.

I love GitHub Pages for this kind of thing because it can be quickly configured (by Claude, in this case) to turn any repository into a deployed web-app, at zero cost and with whatever build step is necessary. It even works against private repos, if you don't mind your only security being a secret URL.

With this kind of project there's always a major risk that the model might "cheat" - mark key features as "TODO" and fake them, or take shortcuts that ignore the initial requirements.

The responsible way to prevent this is to review all of the code... but this wasn't intended as that kind of project, so instead I fired up OpenAI Codex with GPT-5.5 (I had preview access) and told it:

Describe the difference between how the node.js CLI tool runs and how the web/ version runs

The answer I got back was enough to give me confidence that Claude hadn't taken any project-threatening shortcuts.

... and that was about it. Total time in Claude Code for that "build it" step was 59 minutes. I used my claude-code-transcripts tool to export a readable version of the full transcript which you can view here, albeit without those additional queued prompts (here's my issue to fix that).

I'm a pedantic stickler when it comes to the original definition of vibe coding - vibe coding does not mean any time you use AI to help you write code, it's when you use AI without reviewing or caring about the code that's written at all.

By my own definition, this LiteParse for the web project is about as pure vibe coding as you can get! I have not looked at a single line of the HTML and TypeScript written for this project - in fact while writing this sentence I had to go and check if it had used JavaScript or TypeScript.

Yet somehow this one doesn't feel as vibe coded to me as many of my other vibe coded projects:

• As a static in-browser web application hosted on GitHub Pages the blast radius for any bugs is almost non-existent: it either works for your PDF or doesn't.
• No private data is transferred anywhere - all processing happens in your browser - so a security audit is unnecessary. I've glanced once at the network panel while it's running and no additional requests are made when a PDF is being parsed.
• There was still a whole lot of engineering experience and knowledge required to use the models in this way. Identifying that porting LiteParse to run directly in a browser was critical to the rest of the project.

Most importantly, I'm happy to attach my reputation to this project and recommend that other people try it out. Unlike most of my vibe coded tools I'm not convinced that spending significant additional engineering time on this would have resulted in a meaningfully better initial release. It's fine as it is!

I haven't opened a PR against the origin repository because I've not discussed it with the LiteParse team. I've opened an issue, and if they want my vibe coded implementation as a starting point for something more official they're welcome to take it.

Tags: javascript, ocr, pdf, projects, ai, generative-ai, llms, vibe-coding, coding-agents, claude-code, agentic-engineering

GPT-5.5 is out. It's available in OpenAI Codex and is rolling out to paid ChatGPT subscribers. I've had some preview access and found it to be a fast, effective and highly capable model. As is usually the case these days, it's hard to put into words what's good about it - I ask it to build things and it builds exactly what I ask for!

There's one notable omission from today's release - the API:

API deployments require different safeguards and we are working closely with partners and customers on the safety and security requirements for serving it at scale. We'll bring GPT‑5.5 and GPT‑5.5 Pro to the API very soon.

When I run my pelican benchmark I always prefer to use an API, to avoid hidden system prompts in ChatGPT or other agent harnesses from impacting the results.

One of the ongoing tension points in the AI world over the past few months has concerned how agent harnesses like OpenClaw and Pi interact with the APIs provided by the big providers.

Both OpenAI and Anthropic offer popular monthly subscriptions which provide access to their models at a significant discount to their raw API.

OpenClaw integrated directly with this mechanism, and was then blocked from doing so by Anthropic. This kicked off a whole thing. OpenAI - who recently hired OpenClaw creator Peter Steinberger - saw an opportunity for an easy karma win and announced that OpenClaw was welcome to continue integrating with OpenAI's subscriptions via the same mechanism used by their (open source) Codex CLI tool.

Does this mean anyone can write code that integrates with OpenAI's Codex-specific APIs to hook into those existing subscriptions?

The other day Jeremy Howard asked:

Anyone know whether OpenAI officially supports the use of the /backend-api/codex/responses endpoint that Pi and Opencode (IIUC) uses?

It turned out that on March 30th OpenAI's Romain Huet had tweeted:

We want people to be able to use Codex, and their ChatGPT subscription, wherever they like! That means in the app, in the terminal, but also in JetBrains, Xcode, OpenCode, Pi, and now Claude Code.

That’s why Codex CLI and Codex app server are open source too! 🙂

And Peter Steinberger replied to Jeremy that:

OpenAI sub is officially supported.

So... I had Claude Code reverse-engineer the openai/codex repo, figure out how authentication tokens were stored and build me llm-openai-via-codex, a new plugin for LLM which picks up your existing Codex subscription and uses it to run prompts!

(With hindsight I wish I'd used GPT-5.4 or the GPT-5.5 preview, it would have been funnier. I genuinely considered rewriting the project from scratch using Codex and GPT-5.5 for the sake of the joke, but decided not to spend any more time on this!)

Here's how to use it:

1. Install Codex CLI, buy an OpenAI plan, login to Codex
2. Install LLM: uv tool install llm
3. Install the new plugin: llm install llm-openai-via-codex
4. Start prompting: llm -m openai-codex/gpt-5.5 'Your prompt goes here'

All existing LLM features should also work - use -a filepath.jpg/URL to attach an image, llm chat -m openai-codex/gpt-5.5 to start an ongoing chat, llm logs to view logged conversations and llm --tool ... to try it out with tool support.

Let's generate a pelican!

llm install llm-openai-via-codex
llm -m openai-codex/gpt-5.5 'Generate an SVG of a pelican riding a bicycle'

Here's what I got back:

I've seen better from GPT-5.4, so I tagged on -o reasoning_effort xhigh and tried again:

That one took almost four minutes to generate, but I think it's a much better effort.

If you compare the SVG code (default, xhigh) the xhigh one took a very different approach, which is much more CSS-heavy - as demonstrated by those gradients. xhigh used 9,322 reasoning tokens where the default used just 39.

One of the most notable things about GPT-5.5 is the pricing. Once it goes live in the API it's going to be priced at twice the cost of GPT-5.4 - $5 per 1M input tokens and $30 per 1M output tokens, where 5.4 is $2.5 and $15.

GPT-5.5 Pro will be even more: $30 per 1M input tokens and $180 per 1M output tokens.

GPT-5.4 will remain available. At half the price of 5.5 this feels like 5.4 is to 5.5 as Claude Sonnet is to Claude Opus.

Ethan Mollick has a detailed review of GPT-5.5 where he put it (and GPT-5.5 Pro) through an array of interesting challenges. His verdict: the jagged frontier continues to hold, with GPT-5.5 excellent at some things and challenged by others in a way that remains difficult to predict.

Tags: ai, openai, generative-ai, chatgpt, llms, llm, llm-pricing, pelican-riding-a-bicycle, llm-reasoning, llm-release, codex-cli, gpt

Release: llm-openai-via-codex 0.1a0

Hijacks your Codex CLI credentials to make API calls with LLM, as described in my post about GPT-5.5.

Tags: openai, llm, codex-cli

[...] if you ever needed another reason to learn in public by digital gardening or podcasting or streaming or whathaveyou, add on that people will assume you’re more competent than you are. This will get you invites to very cool exclusive events filled with high-achieving, interesting people, even though you have no right to be there. A+ side benefit.

— Maggie Appleton, Gathering Structures (via)

Tags: blogging, maggie-appleton

Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model

Qwen3.6-27B delivers flagship-level agentic coding performance, surpassing the previous-generation open-source flagship Qwen3.5-397B-A17B (397B total / 17B active MoE) across all major coding benchmarks.

On Hugging Face Qwen3.5-397B-A17B is 807GB, this new Qwen3.6-27B is 55.6GB.

I tried it out with the 16.8GB Unsloth Qwen3.6-27B-GGUF:Q4_K_M quantized version and llama-server using this recipe by benob on Hacker News, after first installing llama-server using brew install llama.cpp:

llama-server \
    -hf unsloth/Qwen3.6-27B-GGUF:Q4_K_M \
    --no-mmproj \
    --fit on \
    -np 1 \
    -c 65536 \
    --cache-ram 4096 -ctxcp 2 \
    --jinja \
    --temp 0.6 \
    --top-p 0.95 \
    --top-k 20 \
    --min-p 0.0 \
    --presence-penalty 0.0 \
    --repeat-penalty 1.0 \
    --reasoning on \
    --chat-template-kwargs '{"preserve_thinking": true}'

On first run that saved the ~17GB model to ~/.cache/huggingface/hub/models--unsloth--Qwen3.6-27B-GGUF.

Here's the transcript for "Generate an SVG of a pelican riding a bicycle". This is an outstanding result for a 16.8GB local model:

Performance numbers reported by llama-server:

• Reading: 20 tokens, 0.4s, 54.32 tokens/s
• Generation: 4,444 tokens, 2min 53s, 25.57 tokens/s

For good measure, here's Generate an SVG of a NORTH VIRGINIA OPOSSUM ON AN E-SCOOTER (run previously with GLM-5.1):

That one took 6,575 tokens, 4min 25s, 24.74 t/s.

Via Hacker News

Tags: ai, generative-ai, local-llms, llms, qwen, pelican-riding-a-bicycle, llama-cpp, llm-release, ai-in-china

As part of our continued collaboration with Anthropic, we had the opportunity to apply an early version of Claude Mythos Preview to Firefox. This week’s release of Firefox 150 includes fixes for 271 vulnerabilities identified during this initial evaluation. [...]

Our experience is a hopeful one for teams who shake off the vertigo and get to work. You may need to reprioritize everything else to bring relentless and single-minded focus to the task, but there is light at the end of the tunnel. We are extremely proud of how our team rose to meet this challenge, and others will too. Our work isn’t finished, but we’ve turned the corner and can glimpse a future much better than just keeping up. Defenders finally have a chance to win, decisively.

— Bobby Holley, CTO, Firefox

Tags: anthropic, claude, ai, firefox, llms, mozilla, security, generative-ai, ai-security-research

Changes to GitHub Copilot Individual plans

Unlike Anthropic, GitHub put up an official announcement about their changes, which include tightening usage limits, pausing signups for individual plans (!), restricting Claude Opus 4.7 to the more expensive $39/month "Pro+" plan, and dropping the previous Opus models entirely.

The key paragraph:

Agentic workflows have fundamentally changed Copilot’s compute demands. Long-running, parallelized sessions now regularly consume far more resources than the original plan structure was built to support. As Copilot’s agentic capabilities have expanded rapidly, agents are doing more work, and more customers are hitting usage limits designed to maintain service reliability.

It's easy to forget that just six months ago heavy LLM users were burning an order of magnitude less tokens. Coding agents consume a lot of compute.

Copilot was also unique (I believe) among agents in charging per-request, not per-token. (Correction: Windsurf also operated a credit system like this which they abandoned last month.) This means that single agentic requests which burn more tokens cut directly into their margins. The most recent pricing scheme addresses that with token-based usage limits on a per-session and weekly basis.

My one problem with this announcement is that it doesn't clearly clarify which product called "GitHub Copilot" is affected by these changes. Last month in How many products does Microsoft have named 'Copilot'? I mapped every one Tey Bannerman identified 75 products that share the Copilot brand, 15 of which have "GitHub Copilot" in the title.

Judging by the linked GitHub Copilot plans page this covers Copilot CLI, Copilot cloud agent and code review (features on GitHub.com itself), and the Copilot IDE features available in VS Code, Zed, JetBrains and more.

Via Hacker News

Tags: github, microsoft, ai, generative-ai, github-copilot, llms, llm-pricing, coding-agents

Anthropic today quietly (as in silently, no announcement anywhere at all) updated their claude.com/pricing page (but not their Choosing a Claude plan page, which shows up first for me on Google) to add this tiny but significant detail (arrow is mine, and it's already reverted):

The Internet Archive copy from yesterday shows a checkbox there. Claude Code used to be a feature of the $20/month Pro plan, but according to the new pricing page it is now exclusive to the $100/month or $200/month Max plans.

Update: don't miss the update to this post, they've already changed course a few hours after this change went live.

So what the heck is going on? Unsurprisingly, Reddit and Hacker News and Twitter all caught fire.

I didn't believe the screenshots myself when I first saw them - aside from the pricing grid I could find no announcement from Anthropic anywhere. Then Amol Avasare, Anthropic's Head of Growth, tweeted:

For clarity, we're running a small test on ~2% of new prosumer signups. Existing Pro and Max subscribers aren't affected.

And that appears to be the closest we have had to official messaging from Anthropic.

I don't buy the "~2% of new prosumer signups" thing, since everyone I've talked to is seeing the new pricing grid and the Internet Archive has already snapped a copy. Maybe he means that they'll only be running this version of the pricing grid for a limited time which somehow adds up to "2%" of signups?

I'm also amused to see Claude Cowork remain available on the $20/month plan, because Claude Cowork is effectively a rebranded version of Claude Code wearing a less threatening hat!

There are a whole bunch of things that are bad about this.

If we assume this is indeed a test, and that test comes up negative and they decide not to go ahead with it, the damage has still been extensive:

1. A whole lot of people got scared or angry or both that a service they relied on was about to be rug-pulled. There really is a significant difference between $20/month and $100/month for most people, especially outside of higher salary countries.
2. The uncertainty is really bad! A tweet from an employee is not the way to make an announcement like this. I wasted a solid hour of my afternoon trying to figure out what had happened here. My trust in Anthropic's transparency around pricing - a crucial factor in how I understand their products - has been shaken.
3. Strategically, should I be taking a bet on Claude Code if I know that they might 5x the minimum price of the product?
4. More of a personal issue, but one I care deeply about myself: I invest a great deal of effort (that's 105 posts and counting) in teaching people how to use Claude Code. I don't want to invest that effort in a product that most people cannot afford to use.

Last month I ran a tutorial for journalists on "Coding agents for data analysis" at the annual NICAR data journalism conference. I'm not going to be teaching that audience a course that depends on a $100/month subscription!

This also doesn't make sense to me as a strategy for Anthropic. Claude Code defined the category of coding agents. It's responsible for billions of dollars in annual revenue for Anthropic already. It has a stellar reputation, but I'm not convinced that reputation is strong enough for it to lose the $20/month trial and jump people directly to a $100/month subscription.

OpenAI have been investing heavily in catching up to Claude Code with their Codex products. Anthropic just handed them this marketing opportunity on a plate - here's Codex engineering lead Thibault Sottiaux:

I don't know what they are doing over there, but Codex will continue to be available both in the FREE and PLUS ($20) plans. We have the compute and efficient models to support it. For important changes, we will engage with the community well ahead of making them.

Transparency and trust are two principles we will not break, even if it means momentarily earning less. A reminder that you vote with your subscription for the values you want to see in this world.

I should note that I pay $200/month for Claude Max and I consider it well worth the money. I've had periods of free access in the past courtesy of Anthropic but I'm currently paying full price, and happy to do so.

But I care about the accessibility of the tools that I work with and teach. If Codex has a free tier while Claude Code starts at $100/month I should obviously switch to Codex, because that way I can use the same tool as the people I want to teach how to use coding agents.

Here's what I think happened. I think Anthropic are trying to optimize revenue growth - obviously - and someone pitched making Claude Code only available for Max and higher. That's clearly a bad idea, but "testing" culture says that it's worth putting even bad ideas out to test just in case they surprise you.

So they started a test, without taking into account the wailing and gnashing of teeth that would result when their test was noticed - or accounting for the longer-term brand damage that would be caused.

Or maybe they did account for that, and decided it was worth the risk.

I don't think that calculation was worthwhile. They're going to have to make a very firm commitment along the lines of "we heard your feedback and we commit to keeping Claude Code available on our $20/month plan going forward" to regain my trust.

As it stands, Codex is looking like a much safer bet for me to invest my time in learning and building educational materials around.

In the time I was typing this blog entry Anthropic appear to have reversed course - the claude.com/pricing page now has a checkbox back in the Pro column for Claude Code. I can't find any official communication about it though.

Let's see if they can come up with an explanation/apology that's convincing enough to offset the trust bonfire from this afternoon!

Amol on Twitter:

was a mistake that the logged-out landing page and docs were updated for this test [embedded self-tweet]

Getting lots of questions on why the landing page / docs were updated if only 2% of new signups were affected.

This was understandably confusing for the 98% of folks not part of the experiment, and we've reverted both the landing page and docs changes.

So the experiment is still running, just not visible to the rest of the world?

Tags: ai, generative-ai, llms, anthropic, llm-pricing, ai-ethics, coding-agents, claude-code, codex-cli

OpenAI released ChatGPT Images 2.0 today, their latest image generation model. On the livestream Sam Altman said that the leap from gpt-image-1 to gpt-image-2 was equivalent to jumping from GPT-3 to GPT-5. Here's how I put it to the test.

My prompt:

Do a where's Waldo style image but it's where is the raccoon holding a ham radio

First as a baseline here's what I got from the older gpt-image-1 using ChatGPT directly:

I wasn't able to spot the raccoon - I quickly realized that testing image generation models on Where's Waldo style images (Where's Wally in the UK) can be pretty frustrating!

I tried getting Claude Opus 4.7 with its new higher resolution inputs to solve it but it was convinced there was a raccoon it couldn't find thanks to the instruction card at the top left of the image:

Yes — there's at least one raccoon in the picture, but it's very well hidden. In my careful sweep through zoomed-in sections, honestly, I couldn't definitively spot a raccoon holding a ham radio. [...]

Next I tried Google's Nano Banana 2, via Gemini:

That one was pretty obvious, the raccoon is in the "Amateur Radio Club" booth in the center of the image!

Claude said:

Honestly, this one wasn't really hiding — he's the star of the booth. Feels like the illustrator took pity on us after that last impossible scene. The little "W6HAM" callsign pun on the booth sign is a nice touch too.

I also tried Nano Banana Pro in AI Studio and got this, by far the worst result from any model. Not sure what went wrong here!

With the baseline established, let's try out the new model.

I used an updated version of my openai_image.py script, which is a thin wrapper around the OpenAI Python client library. Their client library hasn't yet been updated to include gpt-image-2 but thankfully it doesn't validate the model ID so you can use it anyway.

Here's how I ran that:

OPENAI_API_KEY="$(llm keys get openai)" \
  uv run https://tools.simonwillison.net/python/openai_image.py \
  -m gpt-image-2 \
  "Do a where's Waldo style image but it's where is the raccoon holding a ham radio"

Here's what I got back. I don't think there's a raccoon in there - I couldn't spot one, and neither could Claude.

The OpenAI image generation cookbook has been updated with notes on gpt-image-2, including the outputQuality setting and available sizes.

I tried setting outputQuality to high and the dimensions to 3840x2160 - I believe that's the maximum - and got this - a 17MB PNG which I converted to a 5MB WEBP:

OPENAI_API_KEY="$(llm keys get openai)" \
  uv run 'https://raw.githubusercontent.com/simonw/tools/refs/heads/main/python/openai_image.py' \
  -m gpt-image-2 "Do a where's Waldo style image but it's where is the raccoon holding a ham radio" \
  --quality high --size 3840x2160

That's pretty great! There's a raccoon with a ham radio in there (bottom left, quite easy to spot).

The image used 13,342 output tokens, which are charged at $30/million so a total cost of around 40 cents.

I think this new ChatGPT image generation model takes the crown from Gemini, at least for the moment.

Where's Waldo style images are an infuriating and somewhat foolish way to test these models, but they do help illustrate how good they are getting at complex illustrations combining both text and details.

rizaco on Hacker News asked ChatGPT to draw a red circle around the raccoon in one of the images in which I had failed to find one. Here's an animated mix of their result and the original image:

Looks like we definitely can't trust these models to usefully solve their own puzzles!

Tags: ai, openai, generative-ai, chatgpt, llms, text-to-image, llm-release, nano-banana

AI agents are already too human. Not in the romantic sense, not because they love or fear or dream, but in the more banal and frustrating one. The current implementations keep showing their human origin again and again: lack of stringency, lack of patience, lack of focus. Faced with an awkward task, they drift towards the familiar. Faced with hard constraints, they start negotiating with reality.

— Andreas Påhlsson-Notini, Less human AI agents, please.

Tags: ai-agents, coding-agents, ai

scosman/pelicans_riding_bicycles

(To be fair, most of the examples I've published count as poisoning too.)

Via Hacker News comment

Tags: ai, generative-ai, llms, training-data, pelican-riding-a-bicycle

Release: llm-openrouter 0.6

• llm openrouter refresh command for refreshing the list of available models without waiting for the cache to expire.

I added this feature so I could try Kimi 2.6 on OpenRouter as soon as it became available there.

Here's its pelican - this time as an HTML page because Kimi chose to include an HTML and JavaScript UI to control the animation. Transcript here.

Tags: openrouter, llm, llm-release, pelican-riding-a-bicycle, kimi, ai-in-china, llms, ai, generative-ai