Simon Willison’s Weblog

open-interpreter (via) This "natural language interface for computers" open source ChatGPT Code Interpreter alternative has been around for a while, but today I finally got around to trying it out.

Here's how I ran it (without first installing anything) using uv:

uvx --from open-interpreter interpreter

The default mode asks you for an OpenAI API key so it can use gpt-4o - there are a multitude of other options, including the ability to use local models with interpreter --local.

It runs in your terminal and works by generating Python code to help answer your questions, asking your permission to run it and then executing it directly on your computer.

I pasted in an API key and then prompted it with this:

find largest files on my desktop

Here's the full transcript.

Since code is run directly on your machine there are all sorts of ways things could go wrong if you don't carefully review the generated code before hitting "y". The team have an experimental safe mode in development which works by scanning generated code with semgrep. I'm not convinced by that approach, I think executing code in a sandbox would be a much more robust solution here - but sandboxing Python is still a very difficult problem.

They do at least have an experimental Docker integration.

# 24th November 2024, 6:29 pm / llms, ai, generative-ai, uv, sandboxing, code-interpreter, openai, ai-assisted-programming, python

Quantization matters (via) What impact does quantization have on the performance of an LLM? been wondering about this for quite a while, now here are numbers from Paul Gauthier.

He ran differently quantized versions of Qwen 2.5 32B Instruct through his Aider code editing benchmark and saw a range of scores.

The original released weights (BF16) scored highest at 71.4%, with Ollama's qwen2.5-coder:32b-instruct-fp16 (a 66GB download) achieving the same score.

The quantized Ollama qwen2.5-coder:32b-instruct-q4_K_M (a 20GB download) saw a massive drop in quality, scoring just 53.4% on the same benchmark.

# 23rd November 2024, 6:39 pm / aider, llms, ai, generative-ai, ollama, edge-llms, paul-gauthier

Say hello to gemini-exp-1121. Google Gemini's Logan Kilpatrick on Twitter:

Say hello to gemini-exp-1121! Our latest experimental gemini model, with:

• significant gains on coding performance
• stronger reasoning capabilities
• improved visual understanding

Available on Google AI Studio and the Gemini API right now

The 1121 in the name is a release date of the 21st November. This comes fast on the heels of last week's gemini-exp-1114.

Both of these new experimental Gemini models have seen moments at the top of the Chatbot Arena. gemini-exp-1114 took the top spot a few days ago, and then lost it to a new OpenAI model called "ChatGPT-4o-latest (2024-11-20)"... only for the new gemini-exp-1121 to hold the top spot right now.

(These model names are all so, so bad.)

I released llm-gemini 0.4.2 with support for the new model - this should have been 0.5 but I already have a 0.5a0 alpha that depends on an unreleased feature in LLM core.

I tried my pelican benchmark:

llm -m gemini-exp-1121 'Generate an SVG of a pelican riding a bicycle'

Since Gemini is a multi-modal vision model, I had it describe the image it had created back to me (by feeding it a PNG render):

llm -m gemini-exp-1121 describe -a pelican.png

And got this description, which is pretty great:

The image shows a simple, stylized drawing of an insect, possibly a bee or an ant, on a vehicle. The insect is composed of a large yellow circle for the body and a smaller yellow circle for the head. It has a black dot for an eye, a small orange oval for a beak or mouth, and thin black lines for antennae and legs. The insect is positioned on top of a simple black and white vehicle with two black wheels. The drawing is abstract and geometric, using basic shapes and a limited color palette of black, white, yellow, and orange.

Update: Logan confirmed on Twitter that these models currently only have a 32,000 token input, significantly less than the rest of the Gemini family.

# 22nd November 2024, 6:14 am / vision-llms, gemini, llm, google, generative-ai, ai, llms, logan-kilpatrick, pelican-riding-a-bicycle

OK, I can partly explain the LLM chess weirdness now (via) Last week Dynomight published Something weird is happening with LLMs and chess pointing out that most LLMs are terrible chess players with the exception of gpt-3.5-turbo-instruct (OpenAI's last remaining completion as opposed to chat model, which they describe as "Similar capabilities as GPT-3 era models").

After diving deep into this, Dynomight now has a theory. It's mainly about completion models v.s. chat models - a completion model like gpt-3.5-turbo-instruct naturally outputs good next-turn suggestions, but something about reformatting that challenge as a chat conversation dramatically reduces the quality of the results.

Through extensive prompt engineering Dynomight got results out of GPT-4o that were almost as good as the 3.5 instruct model. The two tricks that had the biggest impact:

1. Examples. Including just three examples of inputs (with valid chess moves) and expected outputs gave a huge boost in performance.
2. "Regurgitation" - encouraging the model to repeat the entire sequence of previous moves before outputting the next move, as a way to help it reconstruct its context regarding the state of the board.

They experimented a bit with fine-tuning too, but I found their results from prompt engineering more convincing.

No non-OpenAI models have exhibited any talents for chess at all yet. I think that's explained by the A.2 Chess Puzzles section of OpenAI's December 2023 paper Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision:

The GPT-4 pretraining dataset included chess games in the format of move sequence known as Portable Game Notation (PGN). We note that only games with players of Elo 1800 or higher were included in pretraining.

# 21st November 2024, 7:51 pm / prompt-engineering, generative-ai, openai, gpt-4, ai, llms, training-data

llm-gguf 0.2, now with embeddings. This new release of my llm-gguf plugin - which provides support for locally hosted GGUF LLMs - adds a new feature: it now supports embedding models distributed as GGUFs as well.

This means you can use models like the bafflingly small (30.8MB in its smallest quantization) mxbai-embed-xsmall-v1 with LLM like this:

llm install llm-gguf
llm gguf download-embed-model \
  'https://huggingface.co/mixedbread-ai/mxbai-embed-xsmall-v1/resolve/main/gguf/mxbai-embed-xsmall-v1-q8_0.gguf'

Then to embed a string:

llm embed -m gguf/mxbai-embed-xsmall-v1-q8_0 -c 'hello'

The LLM docs have extensive coverage of things you can then do with this model, like embedding every row in a CSV file / file in a directory / record in a SQLite database table and running similarity and semantic search against them.

Under the hood this takes advantage of the create_embedding() method provided by the llama-cpp-python wrapper around llama.cpp.

# 21st November 2024, 7:24 am / llm, generative-ai, projects, ai, embeddings, llama-cpp

TextSynth Server (via) I'd missed this: Fabrice Bellard (yes, that Fabrice Bellard) has a project called TextSynth Server which he describes like this:

ts_server is a web server proposing a REST API to large language models. They can be used for example for text completion, question answering, classification, chat, translation, image generation, ...

It has the following characteristics:

• All is included in a single binary. Very few external dependencies (Python is not needed) so installation is easy.
• Supports many Transformer variants (GPT-J, GPT-NeoX, GPT-Neo, OPT, Fairseq GPT, M2M100, CodeGen, GPT2, T5, RWKV, LLAMA, Falcon, MPT, Llama 3.2, Mistral, Mixtral, Qwen2, Phi3, Whisper) and Stable Diffusion.
• [...]

Unlike many of his other notable projects (such as FFmpeg, QEMU, QuickJS) this isn't open source - in fact it's not even source available, you instead can download compiled binaries for Linux or Windows that are available for non-commercial use only.

Commercial terms are available, or you can visit textsynth.com and pre-pay for API credits which can then be used with the hosted REST API there.

This is not a new project: the earliest evidence I could find of it was this July 2019 page in the Internet Archive, which said:

Text Synth is build using the GPT-2 language model released by OpenAI. [...] This implementation is original because instead of using a GPU, it runs using only 4 cores of a Xeon E5-2640 v3 CPU at 2.60GHz. With a single user, it generates 40 words per second. It is programmed in plain C using the LibNC library.

# 21st November 2024, 5:16 am / llms, ai, generative-ai, internet-archive, gpt-2

Preview: Gemini API Additional Terms of Service. Google sent out an email last week linking to this preview of upcoming changes to the Gemini API terms. Key paragraph from that email:

To maintain a safe and responsible environment for all users, we're enhancing our abuse monitoring practices for Google AI Studio and Gemini API. Starting December 13, 2024, Gemini API will log prompts and responses for Paid Services, as described in the terms. These logs are only retained for a limited time (55 days) and are used solely to detect abuse and for required legal or regulatory disclosures. These logs are not used for model training. Logging for abuse monitoring is standard practice across the global AI industry. You can preview the updated Gemini API Additional Terms of Service, effective December 13, 2024.

That "for required legal or regulatory disclosures" piece makes it sound like somebody could subpoena Google to gain access to your logged Gemini API calls.

It's not clear to me if this is a change from their current policy though, other than the number of days of log retention increasing from 30 to 55 (and I'm having trouble finding that 30 day number written down anywhere.)

That same email also announced the deprecation of the older Gemini 1.0 Pro model:

Gemini 1.0 Pro will be discontinued on February 15, 2025.

# 19th November 2024, 6:26 pm / gemini, google, generative-ai, ai, llms

Pixtral Large (via) New today from Mistral:

Today we announce Pixtral Large, a 124B open-weights multimodal model built on top of Mistral Large 2. Pixtral Large is the second model in our multimodal family and demonstrates frontier-level image understanding.

The weights are out on Hugging Face (over 200GB to download, and you'll need a hefty GPU rig to run them). The license is free for academic research but you'll need to pay for commercial usage.

The new Pixtral Large model is available through their API, as models called pixtral-large-2411 and pixtral-large-latest.

Here's how to run it using LLM and the llm-mistral plugin:

llm install -U llm-mistral
llm keys set mistral
# paste in API key
llm mistral refresh
llm -m mistral/pixtral-large-latest describe -a https://static.simonwillison.net/static/2024/pelicans.jpg

The image shows a large group of birds, specifically pelicans, congregated together on a rocky area near a body of water. These pelicans are densely packed together, some looking directly at the camera while others are engaging in various activities such as preening or resting. Pelicans are known for their large bills with a distinctive pouch, which they use for catching fish. The rocky terrain and the proximity to water suggest this could be a coastal area or an island where pelicans commonly gather in large numbers. The scene reflects a common natural behavior of these birds, often seen in their nesting or feeding grounds.

Update: I released llm-mistral 0.8 which adds async model support for the full Mistral line, plus a new llm -m mistral-large shortcut alias for the Mistral Large model.

# 18th November 2024, 4:41 pm / vision-llms, mistral, llm, generative-ai, ai, llms

Qwen: Extending the Context Length to 1M Tokens (via) The new Qwen2.5-Turbo boasts a million token context window (up from 128,000 for Qwen 2.5) and faster performance:

Using sparse attention mechanisms, we successfully reduced the time to first token for processing a context of 1M tokens from 4.9 minutes to 68 seconds, achieving a 4.3x speedup.

The benchmarks they've published look impressive, including a 100% score on the 1M-token passkey retrieval task (not the first model to achieve this).

There's a catch: unlike previous models in the Qwen 2.5 series it looks like this one hasn't been released as open weights: it's available exclusively via their (inexpensive) paid API - for which it looks like you may need a +86 Chinese phone number.

# 18th November 2024, 3:40 pm / llms, ai, qwen, generative-ai

llm-gemini 0.4. New release of my llm-gemini plugin, adding support for asynchronous models (see LLM 0.18), plus the new gemini-exp-1114 model (currently at the top of the Chatbot Arena) and a -o json_object 1 option to force JSON output.

I also released llm-claude-3 0.9 which adds asynchronous support for the Claude family of models.

# 18th November 2024, 7:37 am / llm, plugins, ai, llms, async, python, generative-ai, projects, claude, gemini, anthropic, google

LLM 0.18. New release of LLM. The big new feature is asynchronous model support - you can now use supported models in async Python code like this:

import llm

model = llm.get_async_model("gpt-4o")
async for chunk in model.prompt(
    "Five surprising names for a pet pelican"
):
    print(chunk, end="", flush=True)

Also new in this release: support for sending audio attachments to OpenAI's gpt-4o-audio-preview model.

# 17th November 2024, 8:40 pm / async, llm, python, generative-ai, projects, ai, llms

NuExtract 1.5. Structured extraction - where an LLM helps turn unstructured text (or image content) into structured data - remains one of the most directly useful applications of LLMs.

NuExtract is a family of small models directly trained for this purpose (though text only at the moment) and released under the MIT license.

It comes in a variety of shapes and sizes:

• NuExtract-v1.5 is a 3.8B parameter model fine-tuned on Phi-3.5-mini instruct. You can try this one out in this playground.
• NuExtract-tiny-v1.5 is 494M parameters, fine-tuned on Qwen2.5-0.5B.
• NuExtract-1.5-smol is 1.7B parameters, fine-tuned on SmolLM2-1.7B.

All three models were fine-tuned on NuMind's "private high-quality dataset". It's interesting to see a model family that uses one fine-tuning set against three completely different base models.

Useful tip from Steffen Röcker:

Make sure to use it with low temperature, I've uploaded NuExtract-tiny-v1.5 to Ollama and set it to 0. With the Ollama default of 0.7 it started repeating the input text. It works really well despite being so smol.

# 16th November 2024, 4:33 pm / llms, ai, generative-ai, fine-tuning, phi, qwen, hugging-face, smollm, structured-extraction

Voting opens for Oxford Word of the Year 2024 (via) One of the options is slop!

slop (n.): Art, writing, or other content generated using artificial intelligence, shared and distributed online in an indiscriminate or intrusive way, and characterized as being of low quality, inauthentic, or inaccurate.

Update 1st December: Slop lost to Brain rot

# 15th November 2024, 6:46 pm / slop, ethics, generative-ai, ai, llms

Recraft V3. Recraft are a generative AI design tool startup based out of London who released their v3 model a few weeks ago. It's currently sat at the top of the Artificial Analysis Image Arena Leaderboard, beating Midjourney and Flux 1.1 pro.

The thing that impressed me is that it can generate both raster and vector graphics... and the vector graphics can be exported as SVG!

Here's what I got for raccoon with a sign that says "I love trash" - SVG here.

That's an editable SVG - when I open it up in Pixelmator I can select and modify the individual paths and shapes:

They also have an API. I spent $1 on 1000 credits and then spent 80 credits (8 cents) making this SVG of a pelican riding a bicycle, using my API key stored in 1Password:

export RECRAFT_API_TOKEN="$(
  op item get recraft.ai --fields label=password \
  --format json | jq .value -r)"

curl https://external.api.recraft.ai/v1/images/generations \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $RECRAFT_API_TOKEN" \
  -d '{
    "prompt": "california brown pelican riding a bicycle",
    "style": "vector_illustration",
    "model": "recraftv3"
  }'

# 15th November 2024, 4:24 am / ai, generative-ai, text-to-image, pelican-riding-a-bicycle

OpenAI Public Bug Bounty. Reading this investigation of the security boundaries of OpenAI's Code Interpreter environment helped me realize that the rules for OpenAI's public bug bounty inadvertently double as the missing details for a whole bunch of different aspects of their platform.

This description of Code Interpreter is significantly more useful than their official documentation!

Code execution from within our sandboxed Python code interpreter is out of scope. (This is an intended product feature.) When the model executes Python code it does so within a sandbox. If you think you've gotten RCE outside the sandbox, you must include the output of uname -a. A result like the following indicates that you are inside the sandbox -- specifically note the 2016 kernel version:

Linux 9d23de67-3784-48f6-b935-4d224ed8f555 4.4.0 #1 SMP Sun Jan 10 15:06:54 PST 2016 x86_64 x86_64 x86_64 GNU/Linux

Inside the sandbox you would also see sandbox as the output of whoami, and as the only user in the output of ps.

# 14th November 2024, 11:44 pm / security, generative-ai, openai, ai, llms, code-interpreter

QuickTime video script to capture frames and bounding boxes. An update to an older TIL. I'm working on the write-up for my DjangoCon US talk on plugins and I found myself wanting to capture individual frames from the video in two formats: a full frame capture, and another that captured just the portion of the screen shared from my laptop.

I have a script for the former, so I got Claude to update my script to add support for one or more --box options, like this:

capture-bbox.sh ../output.mp4  --box '31,17,100,87' --box '0,0,50,50'

Open output.mp4 in QuickTime Player, run that script and then every time you hit a key in the terminal app it will capture three JPEGs from the current position in QuickTime Player - one for the whole screen and one each for the specified bounding box regions.

Those bounding box regions are percentages of the width and height of the image. I also got Claude to build me this interactive tool on top of cropperjs to help figure out those boxes:

# 14th November 2024, 7 pm / claude-artifacts, ai-assisted-programming, claude, tools, projects, ffmpeg, llms, ai, generative-ai

Releasing the largest multilingual open pretraining dataset (via) Common Corpus is a new "open and permissible licensed text dataset, comprising over 2 trillion tokens (2,003,039,184,047 tokens)" released by French AI Lab PleIAs.

This appears to be the largest available corpus of openly licensed training data:

• 926,541,096,243 tokens of public domain books, newspapers, and Wikisource content
• 387,965,738,992 tokens of government financial and legal documents
• 334,658,896,533 tokens of open source code from GitHub
• 221,798,136,564 tokens of academic content from open science repositories
• 132,075,315,715 tokens from Wikipedia, YouTube Commons, StackExchange and other permissively licensed web sources

It's majority English but has significant portions in French and German, and some representation for Latin, Dutch, Italian, Polish, Greek and Portuguese.

I can't wait to try some LLMs trained exclusively on this data. Maybe we will finally get a GPT-4 class model that isn't trained on unlicensed copyrighted data.

# 14th November 2024, 5:44 am / ethics, generative-ai, training-data, ai, llms, pleias

Ollama: Llama 3.2 Vision. Ollama released version 0.4 last week with support for Meta's first Llama vision model, Llama 3.2.

If you have Ollama installed you can fetch the 11B model (7.9 GB) like this:

ollama pull llama3.2-vision

Or the larger 90B model (55GB download, likely needs ~88GB of RAM) like this:

ollama pull llama3.2-vision:90b

I was delighted to learn that Sukhbinder Singh had already contributed support for LLM attachments to Sergey Alexandrov's llm-ollama plugin, which means the following works once you've pulled the models:

llm install --upgrade llm-ollama
llm -m llama3.2-vision:latest 'describe' \
  -a https://static.simonwillison.net/static/2024/pelican.jpg

This image features a brown pelican standing on rocks, facing the camera and positioned to the left of center. The bird's long beak is a light brown color with a darker tip, while its white neck is adorned with gray feathers that continue down to its body. Its legs are also gray.

In the background, out-of-focus boats and water are visible, providing context for the pelican's environment.

That's not a bad description of this image, especially for a 7.9GB model that runs happily on my MacBook Pro.

# 13th November 2024, 1:55 am / vision-llms, llm, llama, ai, edge-llms, llms, meta, ollama, generative-ai

Ars Live: Our first encounter with manipulative AI (via) I'm participating in a live conversation with Benj Edwards on 19th November reminiscing over that incredible time back in February last year when Bing went feral.

# 12th November 2024, 11:58 pm / bing, generative-ai, arstechnica, ai, speaking, llms, benj-edwards

Binary vector embeddings are so cool (via) Evan Schwartz:

Vector embeddings by themselves are pretty neat. Binary quantized vector embeddings are extra impressive. In short, they can retain 95+% retrieval accuracy with 32x compression and ~25x retrieval speedup.

It's so unintuitive how well this trick works: take a vector of 1024x4 byte floating point numbers (4096 bytes = 32,768 bits), turn that into an array of single bits for > 0 or <= 0 which reduces it to just 1024 bits or 128 bytes - a 1/32 reduction.

Now you can compare vectors using a simple Hamming distance - a count of the number of bits that differ - and yet still get embedding similarity scores that are only around 10% less accurate than if you had used the much larger floating point numbers.

Evan digs into models that this works for, which include OpenAI's text-embedding-3-large and the small but powerful all-MiniLM-L6-v2.

# 11th November 2024, 6:53 pm / ai, embeddings

MDN Browser Support Timelines. I complained on Hacker News today that I wished the MDN browser compatibility ables - like this one for the Web Locks API - included an indication as to when each browser was released rather than just the browser numbers.

It turns out they do! If you click on each browser version in turn you can see an expanded area showing the browser release date:

There's even an inline help tip telling you about the feature, which I've been studiously ignoring for years.

I want to see all the information at once without having to click through each browser. I had a poke around in the Firefox network tab and found https://bcd.developer.mozilla.org/bcd/api/v0/current/api.Lock.json - a JSON document containing browser support details (with release dates) for that API... and it was served using access-control-allow-origin: * which means I can hit it from my own little client-side applications.

I decided to build something with an autocomplete drop-down interface for selecting the API. That meant I'd need a list of all of the available APIs, and I used GitHub code search to find that in the mdn/browser-compat-data repository, in the api/ directory.

I needed the list of files in that directory for my autocomplete. Since there are just over 1,000 of those the regular GitHub contents API won't return them all, so I switched to the tree API instead.

Here's the finished tool - source code here:

95% of the code was written by LLMs, but I did a whole lot of assembly and iterating to get it to the finished state. Three of the transcripts for that:

• Web Locks API Browser Support Timeline in which I paste in the original API JSON and ask it to come up with a timeline visualization for it.
• Enhancing API Feature Display with URL Hash where I dumped in a more complex JSON example to get it to show multiple APIs on the same page, and also had it add #fragment bookmarking to the tool
• Fetch GitHub API Data Hierarchy where I got it to write me an async JavaScript function for fetching a directory listing from that tree API.

# 11th November 2024, 3:27 am / claude-3-5-sonnet, mozilla, javascript, projects, ai, github, llms, ai-assisted-programming

Everything I’ve learned so far about running local LLMs (via) Chris Wellons shares detailed notes on his experience running local LLMs on Windows - though most of these tips apply to other operating systems as well.

This is great, there's a ton of detail here and the root recommendations are very solid: Use llama-server from llama.cpp and try ~8B models first (Chris likes Llama 3.1 8B Instruct at Q4_K_M as a first model), anything over 10B probably won't run well on a CPU so you'll need to consider your available GPU VRAM.

This is neat:

Just for fun, I ported llama.cpp to Windows XP and ran a 360M model on a 2008-era laptop. It was magical to load that old laptop with technology that, at the time it was new, would have been worth billions of dollars.

I need to spend more time with Chris's favourite models, Mistral-Nemo-2407 (12B) and Qwen2.5-14B/72B.

Chris also built illume, a Go CLI tool for interacting with models that looks similar to my own LLM project.

# 10th November 2024, 6:01 pm / windows, generative-ai, go, ai, edge-llms, llms, llama-cpp

ChainForge. I'm still on the hunt for good options for running evaluations against prompts. ChainForge offers an interesting approach, calling itself "an open-source visual programming environment for prompt engineering".

The interface is one of those boxes-and-lines visual programming tools, which reminds me of Yahoo Pipes.

It's open source (from a team at Harvard) and written in Python, which means you can run a local copy instantly via uvx like this:

uvx chainforge serve

You can then configure it with API keys to various providers (OpenAI worked for me, Anthropic models returned JSON parsing errors due to a 500 page from the ChainForge proxy) and start trying it out.

The "Add Node" menu shows the full list of capabilities.

The JavaScript and Python evaluation blocks are particularly interesting: the JavaScript one runs outside of a sandbox using plain eval(), while the Python one still runs in your browser but uses Pyodide in a Web Worker.

# 8th November 2024, 8:52 pm / pyodide, evals, uv, ai, llms, prompt-engineering, prompt-injection, python, javascript, generative-ai