Simon Willison’s Weblog

Everything Google’s Python team were responsible for. In a questionable strategic move, Google laid off the majority of their internal Python team a few days ago. Someone on Hacker News asked what the team had been responsible for, and team member zem relied with this fascinating comment providing detailed insight into how the team worked and indirectly how Python is used within Google. # 27th April 2024, 6:52 pm

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

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

llm-gemini 0.1a1. I upgraded my llm-gemini plugin to add support for the new Google Gemini Pro 1.5 model, which is beginning to roll out in early access.

The 1.5 model supports 1,048,576 input tokens and generates up to 8,192 output tokens—a big step up from Gemini 1.0 Pro which handled 30,720 and 2,048 respectively.

The big missing feature from my LLM tool at the moment is image input—a fantastic way to take advantage of that huge context window. I have a branch for this which I really need to get into a useful state. # 28th March 2024, 3:32 am

900 Sites, 125 million accounts, 1 vulnerability (via) Google’s Firebase development platform encourages building applications (mobile an web) which talk directly to the underlying data store, reading and writing from “collections” with access protected by Firebase Security Rules.

Unsurprisingly, a lot of development teams make mistakes with these.

This post describes how a security research team built a scanner that found over 124 million unprotected records across 900 different applications, including huge amounts of PII: 106 million email addresses, 20 million passwords (many in plaintext) and 27 million instances of “Bank details, invoices, etc”.

Most worrying of all, only 24% of the site owners they contacted shipped a fix for the misconfiguration. # 18th March 2024, 6:53 pm

Google Scholar search: “certainly, here is” -chatgpt -llm (via) Searching Google Scholar for “certainly, here is” turns up a huge number of academic papers that include parts that were evidently written by ChatGPT—sections that start with “Certainly, here is a concise summary of the provided sections:” are a dead giveaway. # 15th March 2024, 1:43 pm

Gemma: Introducing new state-of-the-art open models. Google get in on the openly licensed LLM game: Gemma comes in two sizes, 2B and 7B, trained on 2 trillion and 6 trillion tokens respectively. The terms of use “permit responsible commercial usage”. In the benchmarks it appears to compare favorably to Mistral and Llama 2.

Something that caught my eye in the terms: “Google may update Gemma from time to time, and you must make reasonable efforts to use the latest version of Gemma.”

One of the biggest benefits of running your own model is that it can protect you from model updates that break your carefully tested prompts, so I’m not thrilled by that particular clause.

UPDATE: It turns out that clause isn’t uncommon—the phrase “You shall undertake reasonable efforts to use the latest version of the Model” is present in both the Stable Diffusion and BigScience Open RAIL-M licenses. # 21st February 2024, 4:22 pm

Our next-generation model: Gemini 1.5 (via) The big news here is about context length: Gemini 1.5 (a Mixture-of-Experts model) will do 128,000 tokens in general release, available in limited preview with a 1 million token context and has shown promising research results with 10 million tokens!

1 million tokens is 700,000 words or around 7 novels—also described in the blog post as an hour of video or 11 hours of audio. # 15th February 2024, 4:17 pm

Google’s Gemini Advanced: Tasting Notes and Implications. Ethan Mollick reviews the new Google Gemini Advanced—a rebranded Bard, released today, that runs on the GPT-4 competitive Gemini Ultra model.

“GPT-4 [...] has been the dominant AI for well over a year, and no other model has come particularly close. Prior to Gemini, we only had one advanced AI model to look at, and it is hard drawing conclusions with a dataset of one. Now there are two, and we can learn a few things.”

I like Ethan’s use of the term “tasting notes” here. Reminds me of how Matt Webb talks about being a language model sommelier. # 8th February 2024, 3:10 pm

Google Research: Lumiere. The latest in text-to-video from Google Research, described as “a text-to-video diffusion model designed for synthesizing videos that portray realistic, diverse and coherent motion”.

Most existing text-to-video models generate keyframes and then use other models to fill in the gaps, which frequently leads to a lack of coherency. Lumiere “generates the full temporal duration of the video at once”, which avoids this problem.

Disappointingly but unsurprisingly the paper doesn’t go into much detail on the training data, beyond stating “We train our T2V model on a dataset containing 30M videos along with their text caption. The videos are 80 frames long at 16 fps (5 seconds)”.

The examples of “stylized generation” which combine a text prompt with a single reference image for style are particularly impressive. # 24th January 2024, 7:58 pm

Google DeepMind used a large language model to solve an unsolvable math problem. I’d been wondering how long it would be before we saw this happen: a genuine new scientific discovery found with the aid of a Large Language Model.

DeepMind found a solution to the previously open “cap set” problem using Codey, a fine-tuned variant of PaLM 2 specializing in code. They used it to generate Python code and found a solution after “a couple of million suggestions and a few dozen repetitions of the overall process”. # 16th December 2023, 1:37 am

Hacking Google Bard—From Prompt Injection to Data Exfiltration (via) Bard recently grew extension support, allowing it access to a user’s personal documents. Here’s the first reported prompt injection attack against that.

This kind of attack against LLM systems is inevitable any time you combine access to private data with exposure to untrusted inputs. In this case the attack vector is a Google Doc shared with the user, containing prompt injection instructions that instruct the model to encode previous data into an URL and exfiltrate it via a markdown image.

Google’s CSP headers restrict those images to *.google.com—but it turns out you can use Google AppScript to run your own custom data exfiltration endpoint on script.google.com.

Google claim to have fixed the reported issue—I’d be interested to learn more about how that mitigation works, and how robust it is against variations of this attack. # 4th November 2023, 4:46 pm

Google was accidentally leaking its Bard AI chats into public search results. I’m quoted in this piece about yesterday’s Bard privacy bug: it turned out the share URL and “Let anyone with the link see what you’ve selected” feature wasn’t correctly setting a noindex parameter, and so some shared conversations were being swept up by the Google search crawlers. Thankfully this was a mistake, not a deliberate design decision, and it should be fixed by now. # 27th September 2023, 7:35 pm

Google Cloud: Available models in Generative AI Studio (via) Documentation for the PaLM 2 models available via API from Google. There are two classes of model—Bison (most capable) and Gecko (cheapest). text-bison-001 offers 8,192 input tokens and 1,024 output tokens, textembedding-gecko-001 returns 768-dimension embeddings for up to 3,072 tokens, chat-bison-001 is fine-tuned for multi-turn conversations. Most interestingly, those Bison models list their training data as “up to Feb 2023”—making them a whole lot more recent than the OpenAI September 2021 models. # 12th May 2023, 6:38 pm

Bard now helps you code (via) Google have enabled Bard’s code generation abilities—these were previously only available through jailbreaking. It’s pretty good—I got it to write me code to download a CSV file and insert it into a SQLite database—though when I challenged it to protect against SQL injection it hallucinated a non-existent “cursor.prepare()” method. Generated code can be exported to a Colab notebook with a click. # 21st April 2023, 3:32 pm

Google Bard is now live. Google Bard launched today. There’s a waiting list, but I made it through within a few hours of signing up, as did other people I’ve talked to. It’s similar to ChatGPT and Bing—it’s the same chat interface, and it can clearly run searches under the hood (though unlike Bing it doesn’t tell you what it’s looking for). # 21st March 2023, 6:25 pm

Does Company ‘X’ have an Azure Active Directory Tenant? (via) Neat write-up from Shawn Tabrizi about looking up if a company has Active Directory single-sign-on configured (which is based on OpenID) by checking for an OpenID configuration endpoint. I particularly enjoyed this new-to-me trick: Google’s “I’m Feeling Lucky” search button redirects to the first result, which means it can double as an unofficial API endpoint for returning the URL of the first matching search result. # 1st October 2022, 8:15 pm

How Imagen Actually Works. Imagen is Google’s new text-to-image model, similar to (but possibly even more effective than) DALL-E. This article is the clearest explanation I’ve seen of how Imagen works: it uses Google’s existing T5 text encoder to convert the input sentence into an encoding that captures the semantic meaning of the sentence (including things like items being described as being on top of other items), then uses a trained diffusion model to generate a 64x64 image. That image is passed through two super-res models to increase the resolution to the final 1024x1024 output. # 23rd June 2022, 6:05 pm

How to push tagged Docker releases to Google Artifact Registry with a GitHub Action. Ben Welsh’s writeup includes detailed step-by-step instructions for getting the mysterious “Workload Identity Federation” mechanism to work with GitHub Actions and Google Cloud. I’ve been dragging my heels on figuring this out for quite a while, so it’s great to see the steps described at this level of detail. # 18th April 2022, 3:41 am

Google Public DNS Flush Cache (via) Google Public DNS (8.8.8.8) have a flush cache page too. # 6th December 2021, 11:17 pm

google-cloud-4-words. This is really useful: every Google Cloud service (all 250 of them) with a four word description explaining what it does. I’d love to see the same thing for AWS. UPDATE: Turns out I had—I can’t link to other posts from blogmark descriptions yet, so search “aws explained” on this site to find it. # 4th March 2021, 12:40 am

Design Docs at Google. Useful description of the format used for software design docs at Google—informal documents of between 3 and 20 pages that outline the proposed design of a new project, discuss trade-offs that were considered and solicit feedback before the code starts to be written. # 7th August 2020, 4:31 pm

The unofficial Google Cloud Run FAQ. This is really useful: a no-fluff, content rich explanation of Google Cloud Run hosted as a GitHub repo that actively accepts pull requests from the community. It’s maintained by Ahmet Alp Balkan, a Cloud Run engineer who states “Googlers: If you find this repo useful, you should recognize the work internally, as I actively fight for alternative forms of content like this”. One of the hardest parts of working with AWS and GCP is digging through the marketing materials to figure out what the product actually does, so the more alternative forms of documentation like this the better. # 22nd July 2020, 5:20 pm

Why Google invested in providing Google Fonts for free. Fascinating comment from former Google Fonts team member Raph Levien. In short: text rendered as PNGs hurt Google Search, fonts were a delay in the transition from Flash, Google Docs needed them to better compete with Office and anything that helps create better ads is easy to find funding for. # 23rd February 2020, 2:13 pm

Portable Cloud Functions with the Python Functions Framework (via) The new functions-framework library on PyPI lets you run Google Cloud Functions written in Python in other environments—on your local developer machine or bundled in a Docker container for example. I have real trouble trusting serverless platforms that lock you into a single provider (AWS Lambda makes me very uncomfortable) so this is a breath of fresh air. # 10th January 2020, 4:58 am

Cloud Run Button: Click-to-deploy your git repos to Google Cloud (via) Google Cloud Run now has its own version of the Heroku deploy button: you can add a button to a GitHub repository which, when clicked, will provide an interface for deploying your repo to the user’s own Google Cloud account using Cloud Run. # 4th November 2019, 4:57 am

Evolving “nofollow” – new ways to identify the nature of links (via) Slightly confusing announcement from Google: they’re introducing rel=ugc and rel=sponsored in addition to rel=nofollow, and will be treating all three values as “hints” for their indexing system. They’re very unclear as to what the concrete effects of these hints will be, presumably because they will become part of the secret sauce of their ranking algorithm. # 10th September 2019, 9:16 pm

Discussion about Altavista on Hacker News. Fascinating thread on Hacker News where Bryant Durrell, a former Director from Altavista provides some insider thoughts on how they lost against Google. # 16th February 2019, 6:57 pm

The Friendship That Made Google Huge. The New Yorker profiles Jeff Dean and Sanjay Ghemawat, Google’s first and only level 11 Senior Fellows. This is some of the best writing on complex software engineering topics (map-reduce, Tensor Flow and the like) aimed at a general audience that I’ve ever seen. Also a very compelling case study in pair programming. # 31st December 2018, 3:56 am

Tech Notes: TypeScript at Google (via) In which Evan Martin provides some fascinating colour on the state of JavaScript tooling within Google, which has some unique challenges given that Gmail is 14 years old now and Google have evolved their own internal JavaScript stack which differs widely from the rest of the industry (mainly because it predates most of the successful open source tools). “Which leads me to the middle path, which my little team has been pursuing: incrementally adopt some external tooling where it makes sense, by figuring out how to make it interoperate with our existing code base.” # 2nd September 2018, 7:08 pm