Simon Willison’s Weblog

Control your smart home devices with the Gemini mobile app on Android (via) Google are adding smart home integration to their Gemini chatbot - so far on Android only.

Have they considered the risk of prompt injection? It looks like they have, at least a bit:

Important: Home controls are for convenience only, not safety- or security-critical purposes. Don't rely on Gemini for requests that could result in injury or harm if they fail to start or stop.

The Google Home extension can’t perform some actions on security devices, like gates, cameras, locks, doors, and garage doors. For unsupported actions, the Gemini app gives you a link to the Google Home app where you can control those devices.

It can control lights and power, climate control, window coverings, TVs and speakers and "other smart devices, like washers, coffee makers, and vacuums".

I imagine we will see some security researchers having a lot of fun with this shortly.

# 2:35 pm / gemini, prompt-injection, security, google, generative-ai, ai, llms, android

Claude API: PDF support (beta) (via) Claude 3.5 Sonnet now accepts PDFs as attachments:

The new Claude 3.5 Sonnet (claude-3-5-sonnet-20241022) model now supports PDF input and understands both text and visual content within documents.

I just released llm-claude-3 0.7 with support for the new attachment type (attachments are a very new feature), so now you can do this:

llm install llm-claude-3 --upgrade
llm -m claude-3.5-sonnet 'extract text' -a mydoc.pdf

Visual PDF analysis can also be turned on for the Claude.ai application:

Also new today: Claude now offers a free (albeit rate-limited) token counting API. This addresses a complaint I've had for a while: previously it wasn't possible to accurately estimate the cost of a prompt before sending it to be executed.

# 6:55 pm / vision-llms, claude-3-5-sonnet, llm, anthropic, claude, ai, llms, pdf, generative-ai, projects

From Naptime to Big Sleep: Using Large Language Models To Catch Vulnerabilities In Real-World Code (via) Google's Project Zero security team used a system based around Gemini 1.5 Pro to find a previously unreported security vulnerability in SQLite (a stack buffer underflow), in time for it to be fixed prior to making it into a release.

A key insight here is that LLMs are well suited for checking for new variants of previously reported vulnerabilities:

A key motivating factor for Naptime and now for Big Sleep has been the continued in-the-wild discovery of exploits for variants of previously found and patched vulnerabilities. As this trend continues, it's clear that fuzzing is not succeeding at catching such variants, and that for attackers, manual variant analysis is a cost-effective approach.

We also feel that this variant-analysis task is a better fit for current LLMs than the more general open-ended vulnerability research problem. By providing a starting point – such as the details of a previously fixed vulnerability – we remove a lot of ambiguity from vulnerability research, and start from a concrete, well-founded theory: "This was a previous bug; there is probably another similar one somewhere".

LLMs are great at pattern matching. It turns out feeding in a pattern describing a prior vulnerability is a great way to identify potential new ones.

# 8:15 pm / gemini, security, sqlite, google, generative-ai, ai, llms, prompt-engineering

SmolLM2 (via) New from Loubna Ben Allal and her research team at Hugging Face:

SmolLM2 is a family of compact language models available in three size: 135M, 360M, and 1.7B parameters. They are capable of solving a wide range of tasks while being lightweight enough to run on-device. [...]

It was trained on 11 trillion tokens using a diverse dataset combination: FineWeb-Edu, DCLM, The Stack, along with new mathematics and coding datasets that we curated and will release soon.

The model weights are released under an Apache 2 license. I've been trying these out using my llm-gguf plugin for LLM and my first impressions are really positive.

Here's a recipe to run a 1.7GB Q8 quantized model from lmstudio-community:

llm install llm-gguf
llm gguf download-model https://huggingface.co/lmstudio-community/SmolLM2-1.7B-Instruct-GGUF/resolve/main/SmolLM2-1.7B-Instruct-Q8_0.gguf -a smol17
llm chat -m smol17

Or at the other end of the scale, here's how to run the 138MB Q8 quantized 135M model:

llm gguf download-model https://huggingface.co/lmstudio-community/SmolLM2-135M-Instruct-GGUF/resolve/main/SmolLM2-135M-Instruct-Q8_0.gguf' -a smol135m
llm chat -m smol135m

The blog entry to accompany SmolLM2 should be coming soon, but in the meantime here's the entry from July introducing the first version: SmolLM - blazingly fast and remarkably powerful .

# 5:27 am / llm, hugging-face, generative-ai, ai, llms, open-source, edge-llms, smollm

Please publish and share more. 💯 to all of this by Jeff Triplett:

Friends, I encourage you to publish more, indirectly meaning you should write more and then share it. [...]

You don’t have to change the world with every post. You might publish a quick thought or two that helps encourage someone else to try something new, listen to a new song, or binge-watch a new series.

Jeff shares my opinion on conclusions: giving myself permission to hit publish even when I haven't wrapped everything up neatly was a huge productivity boost for me:

Our posts are done when you say they are. You do not have to fret about sticking to landing and having a perfect conclusion. Your posts, like this post, are done after we stop writing.

And another 💯 to this footnote:

PS: Write and publish before you write your own static site generator or perfect blogging platform. We have lost billions of good writers to this side quest because they spend all their time working on the platform instead of writing.

# 3:17 pm / jeff-triplett, blogging

Claude Token Counter. Anthropic released a token counting API for Claude a few days ago.

I built this tool for running prompts, images and PDFs against that API to count the tokens in them.

The API is free (albeit rate limited), but you'll still need to provide your own API key in order to use it.

Here's the source code. I built this using two sessions with Claude - one to build the initial tool and a second to add PDF and image support. That second one is a bit of a mess - it turns out if you drop an HTML file onto a Claude conversation it converts it to Markdown for you, but I wanted it to modify the original HTML source.

The API endpoint also allows you to specify a model, but as far as I can tell from running some experiments the token count was the same for Haiku, Opus and Sonnet 3.5.

# 6:52 pm / anthropic, claude, ai-assisted-programming, tools, generative-ai, ai, llms, claude-artifacts, llm-pricing

Docling. MIT licensed document extraction Python library from the Deep Search team at IBM, who released Docling v2 on October 16th.

Here's the Docling Technical Report paper from August, which provides details of two custom models: a layout analysis model for figuring out the structure of the document (sections, figures, text, tables etc) and a TableFormer model specifically for extracting structured data from tables.

Those models are available on Hugging Face.

Here's how to try out the Docling CLI interface using uvx (avoiding the need to install it first - though since it downloads models it will take a while to run the first time):

uvx docling mydoc.pdf --to json --to md

This will output a mydoc.json file with complex layout information and a mydoc.md Markdown file which includes Markdown tables where appropriate.

The Python API is a lot more comprehensive. It can even extract tables as Pandas DataFrames:

from docling.document_converter import DocumentConverter
converter = DocumentConverter()
result = converter.convert("document.pdf")
for table in result.document.tables:
    df = table.export_to_dataframe()
    print(df)

I ran that inside uv run --with docling python. It took a little while to run, but it demonstrated that the library works.

# 4:57 am / ocr, ai, pdf, python, ibm, hugging-face, uv

California Clock Change. The clocks go back in California tonight and I finally built my dream application for helping me remember if I get an hour extra of sleep or not, using a Claude Artifact. Here's the transcript.

This is one of my favorite examples yet of the kind of tiny low stakes utilities I'm building with Claude Artifacts because the friction involved in churning out a working application has dropped almost to zero.

(I added another feature: it now includes a note of what time my Dog thinks it is if the clocks have recently changed.)

# 5:11 am / claude-artifacts, ai-assisted-programming, projects, ai, timezones, llms

Nous Hermes 3. The Nous Hermes family of fine-tuned models have a solid reputation. Their most recent release came out in August, based on Meta's Llama 3.1:

Our training data aggressively encourages the model to follow the system and instruction prompts exactly and in an adaptive manner. Hermes 3 was created by fine-tuning Llama 3.1 8B, 70B and 405B, and training on a dataset of primarily synthetically generated responses. The model boasts comparable and superior performance to Llama 3.1 while unlocking deeper capabilities in reasoning and creativity.

The model weights are on Hugging Face, including GGUF versions of the 70B and 8B models. Here's how to try the 8B model (a 4.58GB download) using the llm-gguf plugin:

llm install llm-gguf
llm gguf download-model 'https://huggingface.co/NousResearch/Hermes-3-Llama-3.1-8B-GGUF/resolve/main/Hermes-3-Llama-3.1-8B.Q4_K_M.gguf' -a Hermes-3-Llama-3.1-8B
llm -m Hermes-3-Llama-3.1-8B 'hello in spanish'

Nous Research partnered with Lambda Labs to provide inference APIs. It turns out Lambda host quite a few models now, currently providing free inference to users with an API key.

I just released the first alpha of a llm-lambda-labs plugin. You can use that to try the larger 405b model (very hard to run on a consumer device) like this:

llm install llm-lambda-labs
llm keys set lambdalabs
# Paste key here
llm -m lambdalabs/hermes3-405b 'short poem about a pelican with a twist'

Here's the source code for the new plugin, which I based on llm-mistral. The plugin uses httpx-sse to consume the stream of tokens from the API.

# 6:20 pm / llm, generative-ai, llama, ai, edge-llms, llms, meta, projects, nous-research

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.

# 11:55 pm / openai, llms, ai, generative-ai, llm-pricing

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!

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

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.

# 8 pm / evals, llms, ai, generative-ai, dsl, python, nicholas-carlini

Datasette Public Office Hours, Friday Nov 8th at 2pm PT. Tomorrow afternoon (Friday 8th November) at 2pm PT we'll be hosting the first Datasette Public Office Hours - a livestream video session on Discord where Alex Garcia and myself will live code on some Datasette projects and hang out to chat about the project.

This is our first time trying this format. If it works out well I plan to turn it into a series.

# 7:10 pm / discord, open-source, datasette, alex-garcia

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.

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

uv 0.5.0. The first backwards-incompatible (in minor ways) release after 30 releases without a breaking change.

I found out about this release this morning when I filed an issue about a fiddly usability problem I had encountered with the combo of uv and conda... and learned that the exact problem had already been fixed in the brand new version!

# 11:54 pm / uv, packaging, python

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.

# 6:01 pm / windows, generative-ai, go, ai, edge-llms, llms

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.

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

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.

# 6:53 pm / ai, embeddings

How I ship projects at big tech companies (via) This piece by Sean Goedecke on shipping features at larger tech companies is fantastic.

Why do so many engineers think shipping is easy? I know it sounds extreme, but I think many engineers do not understand what shipping even is inside a large tech company. What does it mean to ship? It does not mean deploying code or even making a feature available to users. Shipping is a social construct within a company. Concretely, that means that a project is shipped when the important people at your company believe it is shipped.

Sean emphasizes communication, building confidence and gaining trust and the importance of deploying previews of the feature (for example using feature flags) as early as possible to get that crucial internal buy-in and feedback from other teams.

I think a lot of engineers hold off on deploys essentially out of fear. If you want to ship, you need to do the exact opposite: you need to deploy as much as you can as early as possible, and you need to do the scariest changes as early as you can possibly do them. Remember that you have the most end-to-end context on the project, which means you should be the least scared of scary changes.

# 11:54 pm / software-engineering, management, feature-flags