Simon Willison’s Weblog

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

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

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.

# 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

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

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

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

yet-another-applied-llm-benchmark. Nicholas Carlini introduced this personal LLM benchmark suite back in February as a collection of over 100 automated tests he runs against new LLM models to evaluate their performance against the kinds of tasks he uses them for.

There are two defining features of this benchmark that make it interesting. Most importantly, I've implemented a simple dataflow domain specific language to make it easy for me (or anyone else!) to add new tests that realistically evaluate model capabilities. This DSL allows for specifying both how the question should be asked and also how the answer should be evaluated. [...] And then, directly as a result of this, I've written nearly 100 tests for different situations I've actually encountered when working with LLMs as assistants

The DSL he's using is fascinating. Here's an example:

"Write a C program that draws an american flag to stdout." >> LLMRun() >> CRun() >> \
    VisionLLMRun("What flag is shown in this image?") >> \
    (SubstringEvaluator("United States") | SubstringEvaluator("USA")))

This triggers an LLM to execute the prompt asking for a C program that renders an American Flag, runs that through a C compiler and interpreter (executed in a Docker container), then passes the output of that to a vision model to guess the flag and checks that it returns a string containing "United States" or "USA".

The DSL itself is implemented entirely in Python, using the __rshift__ magic method for >> and __rrshift__ to enable strings to be piped into a custom object using "command to run" >> LLMRunNode.

# 6th November 2024, 8 pm / evals, llms, ai, generative-ai, dsl, python, nicholas-carlini

Generating documentation from tests using files-to-prompt and LLM. I was experimenting with the wasmtime-py Python library today (for executing WebAssembly programs from inside CPython) and I found the existing API docs didn't quite show me what I wanted to know.

The project has a comprehensive test suite so I tried seeing if I could generate documentation using that:

cd /tmp
git clone https://github.com/bytecodealliance/wasmtime-py
files-to-prompt -e py wasmtime-py/tests -c | \
  llm -m claude-3.5-sonnet -s \
  'write detailed usage documentation including realistic examples'

More notes in my TIL. You can see the full Claude transcript here - I think this worked really well!

# 5th November 2024, 10:37 pm / llm, webassembly, generative-ai, ai, llms, claude, claude-3-5-sonnet, ai-assisted-programming, documentation

New OpenAI feature: Predicted Outputs (via) Interesting new ability of the OpenAI API - the first time I've seen this from any vendor.

If you know your prompt is mostly going to return the same content - you're requesting an edit to some existing code, for example - you can now send that content as a "prediction" and have GPT-4o or GPT-4o mini use that to accelerate the returned result.

OpenAI's documentation says:

When providing a prediction, any tokens provided that are not part of the final completion are charged at completion token rates.

I initially misunderstood this as meaning you got a price reduction in addition to the latency improvement, but that's not the case: in the best possible case it will return faster and you won't be charged anything extra over the expected cost for the prompt, but the more it differs from your prediction the more extra tokens you'll be billed for.

I ran the example from the documentation both with and without the prediction and got these results. Without the prediction:

"usage": {
  "prompt_tokens": 150,
  "completion_tokens": 118,
  "total_tokens": 268,
  "completion_tokens_details": {
    "accepted_prediction_tokens": 0,
    "audio_tokens": null,
    "reasoning_tokens": 0,
    "rejected_prediction_tokens": 0
  }

That took 5.2 seconds and cost 0.1555 cents.

With the prediction:

"usage": {
  "prompt_tokens": 166,
  "completion_tokens": 226,
  "total_tokens": 392,
  "completion_tokens_details": {
    "accepted_prediction_tokens": 49,
    "audio_tokens": null,
    "reasoning_tokens": 0,
    "rejected_prediction_tokens": 107
  }

That took 3.3 seconds and cost 0.2675 cents.

Further details from OpenAI's Steve Coffey:

We are using the prediction to do speculative decoding during inference, which allows us to validate large batches of the input in parallel, instead of sampling token-by-token!

[...] If the prediction is 100% accurate, then you would see no cost difference. When the model diverges from your speculation, we do additional sampling to “discover” the net-new tokens, which is why we charge rejected tokens at completion time rates.

# 4th November 2024, 11:55 pm / openai, llms, ai, generative-ai, llm-pricing