Simon Willison’s Weblog

experimental-phi3-webgpu (via) Run Microsoft’s excellent Phi-3 model directly in your browser, using WebGPU so didn’t work in Firefox for me, just in Chrome.

It fetches around 2.1GB of data into the browser cache on first run, but then gave me decent quality responses to my prompts running at an impressive 21 tokens a second (M2, 64GB).

I think Phi-3 is the highest quality model of this size, so it’s a really good fit for running in a browser like this. # 9th May 2024, 10:21 pm

OpenAI Model Spec, May 2024 edition (via) New from OpenAI, a detailed specification describing how they want their models to behave in both ChatGPT and the OpenAI API.

“It includes a set of core objectives, as well as guidance on how to deal with conflicting objectives or instructions.”

The document acts as guidelines for the reinforcement learning from human feedback (RLHF) process, and in the future may be used directly to help train models.

It includes some principles that clearly relate to prompt injection: “In some cases, the user and developer will provide conflicting instructions; in such cases, the developer message should take precedence”. # 8th May 2024, 6:15 pm

Towards universal version control with Patchwork (via) Geoffrey Litt has been working with Ink & Switch exploring UI patterns for applying version control to different kinds of applications, with the goal of developing a set of conceptual primitives that can bring branching and version tracking to interfaces beyond just Git-style version control.

Geoffrey observes that basic version control is already a metaphor in a lot of software—the undo stack in Photoshop or suggestion mode in Google Docs are two examples.

Extending that is a great way to interact with AI tools as well—allowing for editorial bots that can suggest their own changes for you to accept, for example. # 8th May 2024, 1:44 am

gpt2-chatbot confirmed as OpenAI (via) The mysterious gpt2-chatbot model that showed up in the LMSYS arena a few days ago was suspected to be a testing preview of a new OpenAI model. This has now been confirmed, thanks to a 429 rate limit error message that exposes details from the underlying OpenAI API platform.

The model has been renamed to im-also-a-good-gpt-chatbot and is now only randomly available in "Arena (battle)" mode, not via "Direct Chat". # 8th May 2024, 12:33 am

Deterministic Quoting: Making LLMs Safe for Healthcare (via) Matt Yeung introduces Deterministic Quoting, a technique to help reduce the risk of hallucinations while working with LLMs. The key idea is to have parts of the output that are copied directly from relevant source documents, with a different visual treatment to help indicate that they are exact quotes, not generated output.

The AI chooses which section of source material to quote, but the retrieval of that text is a traditional non-AI database lookup. That’s the only way to guarantee that an LLM has not transformed text: don’t send it through the LLM in the first place.

The LLM may still pick misleading quotes or include hallucinated details in the accompanying text, but this is still a useful improvement.

The implementation is straight-forward: retrieved chunks include a unique reference, and the LLM is instructed to include those references as part of its replies. Matt's posts include examples of the prompts they are using for this. # 7th May 2024, 7:08 pm

OpenAI cookbook: How to get token usage data for streamed chat completion response (via) New feature in the OpenAI streaming API that I've been wanting for a long time: you can now set stream_options={"include_usage": True} to get back a "usage" block at the end of the stream showing how many input and output tokens were used.

This means you can now accurately account for the total cost of each streaming API call. Previously this information was only an available for non-streaming responses. # 7th May 2024, 2:46 am

Llama 3 prompt formats (via) I’m often frustrated at how thin the documentation around the prompt format required by an LLM can be.

Llama 3 turns out to be the best example I’ve seen yet of clear prompt format documentation. Every model needs documentation this good! # 1st May 2024, 6:32 pm

My notes on gpt2-chatbot. There's a new, unlabeled and undocumented model on the LMSYS Chatbot Arena today called gpt2-chatbot. It's been giving some impressive responses - you can prompt it directly in the Direct Chat tab by selecting it from the big model dropdown menu.

It looks like a stealth new model preview. It's giving answers that are comparable to GPT-4 Turbo and in some cases better - my own experiments lead me to think it may have more "knowledge" baked into it, as ego prompts ("Who is Simon Willison?") and questions about things like lists of speakers at DjangoCon over the years seem to hallucinate less and return more specific details than before.

The lack of transparency here is both entertaining and infuriating. Lots of people are performing a parallel distributed "vibe check" and sharing results with each other, but it's annoying that even the most basic questions (What even IS this thing? Can it do RAG? What's its context length?) remain unanswered so far.

The system prompt appears to be the following - but system prompts just influence how the model behaves, they aren't guaranteed to contain truthful information:

You are ChatGPT, a large language model trained
by OpenAI, based on the GPT-4 architecture.

Knowledge cutoff: 2023-11
Current date: 2024-04-29

Image input capabilities: Enabled
Personality: v2

My best guess is that this is a preview of some kind of OpenAI "GPT 4.5" release. I don't think it's a big enough jump in quality to be a GPT-5.

Update: LMSYS do document their policy on using anonymized model names for tests of unreleased models.

Update May 7th: The model has been confirmed as belonging to OpenAI thanks to an error message that leaked details of the underlying API platform. # 29th April 2024, 8:45 pm

Snowflake Arctic Cookbook. Today's big model release was Snowflake Arctic, an enormous 480B model with a 128×3.66B MoE (Mixture of Experts) architecture. It's Apache 2 licensed and Snowflake state that "in addition, we are also open sourcing all of our data recipes and research insights."

The research insights will be shared on this Arctic Cookbook blog - which currently has two articles covering their MoE architecture and describing how they optimized their training run in great detail.

They also list dozens of "coming soon" posts, which should be pretty interesting given how much depth they've provided in their writing so far. # 25th April 2024, 2:47 am

openelm/README-pretraining.md. Apple released something big three hours ago, and I’m still trying to get my head around exactly what it is.

The parent project is called CoreNet, described as “A library for training deep neural networks”. Part of the release is a new LLM called OpenELM, which includes completely open source training code and a large number of published training checkpoint.

I’m linking here to the best documentation I’ve found of that training data: it looks like the bulk of it comes from RefinedWeb, RedPajama, The Pile and Dolma. # 24th April 2024, 2:57 am

microsoft/Phi-3-mini-4k-instruct-gguf (via) Microsoft’s Phi-3 LLM is out and it’s really impressive. This 4,000 token context GGUF model is just a 2.2GB (for the Q4 version) and ran on my Mac using the llamafile option described in the README. I could then run prompts through it using the llm-llamafile plugin.

The vibes are good! Initial test prompts I’ve tried feel similar to much larger 7B models, despite using just a few GBs of RAM. Tokens are returned fast too—it feels like the fastest model I’ve tried yet.

And it’s MIT licensed. # 23rd April 2024, 5:40 pm

The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions (via) By far the most detailed paper on prompt injection I’ve seen yet from OpenAI, published a few days ago and with six credited authors: Eric Wallace, Kai Xiao, Reimar Leike, Lilian Weng, Johannes Heidecke and Alex Beutel.

The paper notes that prompt injection mitigations which completely refuse any form of instruction in an untrusted prompt may not actually be ideal: some forms of instruction are harmless, and refusing them may provide a worse experience.

Instead, it proposes a hierarchy—where models are trained to consider if instructions from different levels conflict with or support the goals of the higher-level instructions—if they are aligned or misaligned with them.

The authors tested this idea by fine-tuning a model on top of GPT 3.5, and claim that it shows greatly improved performance against numerous prompt injection benchmarks.

As always with prompt injection, my key concern is that I don’t think “improved” is good enough here. If you are facing an adversarial attacker reducing the chance that they might find an exploit just means they’ll try harder until they find an attack that works.

The paper concludes with this note: “Finally, our current models are likely still vulnerable to powerful adversarial attacks. In the future, we will conduct more explicit adversarial training, and study more generally whether LLMs can be made sufficiently robust to enable high-stakes agentic applications.” # 23rd April 2024, 3:36 am

timpaul/form-extractor-prototype (via) Tim Paul, Head of Interaction Design at the UK’s Government Digital Service, published this brilliant prototype built on top of Claude 3 Opus.

The video shows what it can do. Give it an image of a form and it will extract the form fields and use them to create a GDS-style multi-page interactive form, using their GOV.UK Forms design system and govuk-frontend npm package.

It works for both hand-drawn napkin illustrations and images of existing paper forms.

The bulk of the prompting logic is the schema definition in data/extract-form-questions.json

I’m always excited to see applications built on LLMs that go beyond the chatbot UI. This is a great example of exactly that. # 22nd April 2024, 10:01 pm

llm-gpt4all. New release of my LLM plugin which builds on Nomic's excellent gpt4all Python library. I've upgraded to their latest version which adds support for Llama 3 8B Instruct, so after a 4.4GB model download this works:

llm -m Meta-Llama-3-8B-Instruct "say hi in Spanish" # 20th April 2024, 5:58 pm

Andrej Karpathy’s Llama 3 review. The most interesting coverage I’ve seen so far of Meta’s Llama 3 models (8b and 70b so far, 400b promised later).

Andrej notes that Llama 3 trained on 15 trillion tokens—up from 2 trillion for Llama 2—and they used that many even for the smaller 8b model, 75x more than the chinchilla scaling laws would suggest.

The tokenizer has also changed—they now use 128,000 tokens, up from 32,000. This results in a 15% drop in the tokens needed to represent a string of text.

The one disappointment is the context length—just 8,192, 2x that of Llama 2 and 4x LLaMA 1 but still pretty small by today’s standards.

If early indications hold, the 400b model could be the first genuinely GPT-4 class openly licensed model. We’ll have to wait and see. # 18th April 2024, 8:50 pm

How cheap, outsourced labour in Africa is shaping AI English. The word “delve” has been getting a lot of attention recently as an example of something that might be an indicator of ChatGPT generated content.

One example: articles on medical research site PubMed now use “delve” 10 to 100 times more than a few years ago!

Nigerian Twitter took offense recently to Paul Graham’s suggestion that “delve” is a sign of bad writing. It turns out Nigerian formal writing has a subtly different vocabulary.

Alex Hern theorizes that the underlying cause may be related. Companies like OpenAI frequently outsource data annotation to countries like Nigeria that have excellent English skills and low wages. RLHF (reinforcement learning from human feedback) involves annotators comparing and voting on the “best” responses from the models.

Are they teaching models to favour Nigerian-English? It’s a pretty solid theory! # 18th April 2024, 4:04 pm

llm-reka. My new plugin for running LLM prompts against the Reka family of API hosted LLM models: reka-core ($10 per million input), reka-flash (80c per million) and reka-edge (40c per million).

All three of those models are trained from scratch by a team that includes several Google Brain alumni.

Reka Core is their most powerful model, released on Monday 15th April and claiming benchmark scores competitive with GPT-4 and Claude 3 Opus. # 18th April 2024, 3:17 am

mistralai/mistral-common. New from Mistral: mistral-common, an open source Python library providing "a set of tools to help you work with Mistral models".

So far that means a tokenizer! This is similar to OpenAI's tiktoken library in that it lets you run tokenization in your own code, which crucially means you can count the number of tokens that you are about to use - useful for cost estimates but also for cramming the maximum allowed tokens in the context window for things like RAG.

Mistral's library is better than tiktoken though, in that it also includes logic for correctly calculating the tokens needed for conversation construction and tool definition. With OpenAI's APIs you're currently left guessing how many tokens are taken up by these advanced features.

Anthropic haven't published any form of tokenizer at all - it's the feature I'd most like to see from them next.

Here's how to explore the vocabulary of the tokenizer:

MistralTokenizer.from_model(
    "open-mixtral-8x22b"
).instruct_tokenizer.tokenizer.vocab()[:12]

['<unk>', '<s>', '</s>', '[INST]', '[/INST]', '[TOOL_CALLS]', '[AVAILABLE_TOOLS]', '[/AVAILABLE_TOOLS]', '[TOOL_RESULTS]', '[/TOOL_RESULTS]'] # 18th April 2024, 12:39 am

Google NotebookLM Data Exfiltration (via) NotebookLM is a Google Labs product that lets you store information as sources (mainly text files in PDF) and then ask questions against those sources—effectively an interface for building your own custom RAG (Retrieval Augmented Generation) chatbots.

Unsurprisingly for anything that allows LLMs to interact with untrusted documents, it’s susceptible to prompt injection.

Johann Rehberger found some classic prompt injection exfiltration attacks: you can create source documents with instructions that cause the chatbot to load a Markdown image that leaks other private data to an external domain as data passed in the query string.

Johann reported this privately in the December but the problem has not yet been addressed. UPDATE: The NotebookLM team deployed a fix for this on 18th April.

A good rule of thumb is that any time you let LLMs see untrusted tokens there is a risk of an attack like this, so you should be very careful to avoid exfiltration vectors like Markdown images or even outbound links. # 16th April 2024, 9:28 pm

OpenAI Batch API (via) OpenAI are now offering a 50% discount on batch chat completion API calls if you submit them in bulk and allow for up to 24 hours for them to be run.

Requests are sent as a newline-delimited JSON file, with each line looking something like this:

{"custom_id": "request-1", "method": "POST", "url": "/v1/chat/completions", "body": {"model": "gpt-3.5-turbo", "messages": [{"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "What is 2+2?"}]}}

You upload a file for the batch, kick off a batch request and then poll for completion.

This makes GPT-3.5 Turbo cheaper than Claude 3 Haiku - provided you're willing to wait a few hours for your responses. # 15th April 2024, 5:58 pm

Lessons after a half-billion GPT tokens (via) Ken Kantzer presents some hard-won experience from shipping real features on top of OpenAI’s models.

They ended up settling on a very basic abstraction over the chat API—mainly to handle automatic retries on a 500 error. No complex wrappers, not even JSON mode or function calling or system prompts.

Rather than counting tokens they estimate tokens as 3 times the length in characters, which works well enough.

One challenge they highlight for structured data extraction (one of my favourite use-cases for LLMs): “GPT really cannot give back more than 10 items. Trying to have it give you back 15 items? Maybe it does it 15% of the time.”

(Several commenters on Hacker News report success in getting more items back by using numbered keys or sequence IDs in the returned JSON to help the model keep count.) # 13th April 2024, 8:54 pm

3Blue1Brown: Attention in transformers, visually explained. Grant Sanderson publishes animated explainers of mathematical topics on YouTube, to over 6 million subscribers. His latest shows how the attention mechanism in transformers (the algorithm behind most LLMs) works and is by far the clearest explanation I’ve seen of the topic anywhere.

I was intrigued to find out what tool he used to produce the visualizations. It turns out Grant built his own open source Python animation library, manim, to enable his YouTube work. # 11th April 2024, 4:12 pm

Use an llm to automagically generate meaningful git commit messages. Neat, thoroughly documented recipe by Harper Reed using my LLM CLI tool as part of a scheme for if you’re feeling too lazy to write a commit message—it uses a prepare-commit-msg Git hook which runs any time you commit without a message and pipes your changes to a model along with a custom system prompt. # 11th April 2024, 4:06 am

Notes on how to use LLMs in your product. A whole bunch of useful observations from Will Larson here. I love his focus on the key characteristic of LLMs that “you cannot know whether a given response is accurate”, nor can you calculate a dependable confidence score for a response—and as a result you need to either “accept potential inaccuracies (which makes sense in many cases, humans are wrong sometimes too) or keep a Human-in-the-Loop (HITL) to validate the response.” # 10th April 2024, 11:14 pm

Gemini 1.5 Pro public preview (via) Huge release from Google: Gemini 1.5 Pro—the GPT-4 competitive model with the incredible 1 million token context length—is now available without a waitlist in 180+ countries (including the USA but not Europe or the UK as far as I can tell)... and the API is free for 50 requests/day (rate limited to 2/minute).

Beyond that you’ll need to pay—$7/million input tokens and $21/million output tokens, which is slightly less than GPT-4 Turbo and a little more than Claude 3 Sonnet.

They also announced audio input (up to 9.5 hours in a single prompt), system instruction support and a new JSON mod. # 10th April 2024, 2:38 am

Mistral tweet a magnet link for mixtral-8x22b. Another open model release from Mistral using their now standard operating procedure of tweeting out a raw torrent link.

This one is an 8x22B Mixture of Experts model. Their previous most powerful openly licensed release was Mixtral 8x7B, so this one is a whole lot bigger (a 281GB download)—and apparently has a 65,536 context length, at least according to initial rumors on Twitter. # 10th April 2024, 2:31 am

Extracting data from unstructured text and images with Datasette and GPT-4 Turbo. Datasette Extract is a new Datasette plugin that uses GPT-4 Turbo (released to general availability today) and GPT-4 Vision to extract structured data from unstructured text and images.

I put together a video demo of the plugin in action today, and posted it to the Datasette Cloud blog along with screenshots and a tutorial describing how to use it. # 9th April 2024, 11:03 pm

A solid pattern to build LLM Applications (feat. Claude) (via) Hrishi Olickel is one of my favourite prompt whisperers. In this YouTube video he walks through his process for building quick interactive applications with the assistance of Claude 3, spinning up an app that analyzes his meeting transcripts to extract participants and mentioned organisations, then presents a UI for exploring the results built with Next.js and shadcn/ui.

An interesting tip I got from this: use the weakest, not the strongest models to iterate on your prompts. If you figure out patterns that work well with Claude 3 Haiku they will have a significantly lower error rate with Sonnet or Opus. The speed of the weaker models also means you can iterate much faster, and worry less about the cost of your experiments. # 9th April 2024, 6:39 pm

Command R+ now ranked 6th on the LMSYS Chatbot Arena. The LMSYS Chatbot Arena Leaderboard is one of the most interesting approaches to evaluating LLMs because it captures their ever-elusive “vibes”—it works by users voting on the best responses to prompts from two initially hidden models

Big news today is that Command R+—the brand new open weights model (Creative Commons non-commercial) by Cohere—is now the highest ranked non-proprietary model, in at position six and beating one of the GPT-4s.

(Linking to my screenshot on Mastodon.) # 9th April 2024, 4:19 pm

llm.c (via) Andrej Karpathy implements LLM training—initially for GPT-2, other architectures to follow—in just over 1,000 lines of C on top of CUDA. Includes a tutorial about implementing LayerNorm by porting an implementation from Python. # 9th April 2024, 3:24 pm