April 2023
116 posts: 17 entries, 45 links, 19 quotes, 29 beats, 6 newsletters
Wednesday, 12th April 2023
Replacing my best friends with an LLM trained on 500,000 group chat messages (via) Izzy Miller used a 7 year long group text conversation with five friends from college to fine-tune LLaMA, such that it could simulate ongoing conversations. They started by extracting the messages from the iMessage SQLite database on their Mac, then generated a new training set from those messages and ran it using code from the Stanford Alpaca repository. This is genuinely one of the clearest explanations of the process of fine-tuning a model like this I’ve seen anywhere.
Thursday, 13th April 2023
Free Dolly: Introducing the World’s First Truly Open Instruction-Tuned LLM (via) Databricks released a large language model called Dolly a few weeks ago. They just released Dolly 2.0 and it is MUCH more interesting—it’s an instruction tuned 12B parameter upgrade of EleutherAI’s Pythia model. Unlike other recent instruction tuned models Databricks didn’t use a training set derived from GPT-3—instead, they recruited 5,000 employees to help put together 15,000 human-generated request/response pairs, which they have released under a Creative Commons Attribution-ShareAlike license. The model itself is a 24GB download from Hugging Face—I’ve run it slowly on a small GPU-enabled Paperspace instance, but hopefully optimized ways to run it will emerge in short order.
GitHub Accelerator: our first cohort. I’m participating in the first cohort of GitHub’s new open source accelerator program, with Datasette (and related projects). It’s a 10 week program with 20 projects working together “with an end goal of building durable streams of funding for their work”.
Before we scramble to deeply integrate LLMs everywhere in the economy, can we pause and think whether it is wise to do so?
This is quite immature technology and we don't understand how it works.
If we're not careful we're setting ourselves up for a lot of correlated failures.
— Jan Leike, Alignment Team lead, OpenAI
The Great Flowering: Why OpenAI is the new AWS and the New Kingmakers still matter (via) James Governor discusses the potential impact of AI-assisted productivity on the wider software engineering industry, and calls me “a bellwether”!
Friday, 14th April 2023
Building LLM applications for production. Chip Huyen provides a useful, in-depth review of the challenges involved in taking an app built on top of a LLM from prototype to production, including issues such as prompt ambiguity and unpredictability, cost and latency concerns, challenges in testing and updating to new models. She also lists some promising use-cases she’s seeing for categories of application built on these tools.
Prompt injection: What’s the worst that can happen?
Activity around building sophisticated applications on top of LLMs (Large Language Models) such as GPT-3/4/ChatGPT/etc is growing like wildfire right now.
[... 2,302 words]One way to avoid unspotted prediction errors is for the technology in its current state to have early and frequent contact with reality as it is iteratively developed, tested, deployed, and all the while improved. And there are creative ideas people don’t often discuss which can improve the safety landscape in surprising ways — for example, it’s easy to create a continuum of incrementally-better AIs (such as by deploying subsequent checkpoints of a given training run), which presents a safety opportunity very unlike our historical approach of infrequent major model upgrades.
New prompt injection attack on ChatGPT web version. Markdown images can steal your chat data. An ingenious new prompt injection / data exfiltration vector from Roman Samoilenko, based on the observation that ChatGPT can render markdown images in a way that can exfiltrate data to the image hosting server by embedding it in the image URL. Roman uses a single pixel image for that, and combines it with a trick where copy events on a website are intercepted and prompt injection instructions are appended to the copied text, in order to trick the user into pasting the injection attack directly into ChatGPT.
Update: They finally started mitigating this in December 2023.
codespaces-jupyter (via) This is really neat. Click “Use this template” -> “Open in a codespace” and you get a full in-browser VS Code interface where you can open existing notebook files (or create new ones) and start playing with them straight away.
Saturday, 15th April 2023
Although fine-tuning can feel like the more natural option—training on data is how GPT learned all of its other knowledge, after all—we generally do not recommend it as a way to teach the model knowledge. Fine-tuning is better suited to teaching specialized tasks or styles, and is less reliable for factual recall. [...] In contrast, message inputs are like short-term memory. When you insert knowledge into a message, it's like taking an exam with open notes. With notes in hand, the model is more likely to arrive at correct answers.
— Ted Sanders, OpenAI
sqlite-history: tracking changes to SQLite tables using triggers (also weeknotes)
In between blogging about ChatGPT rhetoric, micro-benchmarking with ChatGPT Code Interpreter and Why prompt injection is an even bigger problem now I managed to ship the beginnings of a new project: sqlite-history.
[... 1,680 words]Sunday, 16th April 2023
Web LLM runs the vicuna-7b Large Language Model entirely in your browser, and it’s very impressive
A month ago I asked Could you train a ChatGPT-beating model for $85,000 and run it in a browser?. $85,000 was a hypothetical training cost for LLaMA 7B plus Stanford Alpaca. “Run it in a browser” was based on the fact that Web Stable Diffusion runs a 1.9GB Stable Diffusion model in a browser, so maybe it’s not such a big leap to run a small Large Language Model there as well.
[... 2,276 words]
How I Used Stable Diffusion and Dreambooth to Create A Painted Portrait of My Dog (via) I like posts like this that go into detail in terms of how much work it takes to deliberately get the kind of result you really want using generative AI tools. Jake Dahn trained a Dreambooth model from 40 photos of Queso—his photogenic Golden Retriever—using Replicate, then gathered the prompts from ten images he liked on Lexica and generated over 1,000 different candidate images, picked his favourite, used Draw Things img2img resizing to expand the image beyond the initial crop, then Automatic1111 inpainting to tweak the ears, then Real-ESRGAN 4x+ to upscale for the final print.
Monday, 17th April 2023
MiniGPT-4 (via) An incredible project with a poorly chosen name. A team from King Abdullah University of Science and Technology in Saudi Arabia combined Vicuna-13B (a model fine-tuned on top of Facebook’s LLaMA) with the BLIP-2 vision-language model to create a model that can conduct ChatGPT-style conversations around an uploaded image. The demo is very impressive, and the weights are available to download—45MB for MiniGPT-4, but you’ll need the much larger Vicuna and LLaMA weights as well.
Latest Twitter search results for “as an AI language model” (via) Searching for “as an AI language model” on Twitter reveals hundreds of bot accounts which are clearly being driven by GPT models and have been asked to generate content which occasionally trips the ethical guidelines trained into the OpenAI models.
If Twitter still had an affordable search API someone could do some incredible disinformation research on top of this, looking at which accounts are implicated, what kinds of things they are tweeting about, who they follow and retweet and so-on.
RedPajama, a project to create leading open-source models, starts by reproducing LLaMA training dataset of over 1.2 trillion tokens. With the amount of projects that have used LLaMA as a foundation model since its release two months ago—despite its non-commercial license—it’s clear that there is a strong desire for a fully openly licensed alternative.
RedPajama is a collaboration between Together, Ontocord.ai, ETH DS3Lab, Stanford CRFM, Hazy Research, and MILA Québec AI Institute aiming to build exactly that.
Step one is gathering the training data: the LLaMA paper described a 1.2 trillion token training set gathered from sources that included Wikipedia, Common Crawl, GitHub, arXiv, Stack Exchange and more.
RedPajama-Data-1T is an attempt at recreating that training set. It’s now available to download, as 2,084 separate multi-GB jsonl files—2.67TB total.
Even without a trained model, this is a hugely influential contribution to the world of open source LLMs. Any team looking to build their own LLaMA from scratch can now jump straight to the next stage, training the model.
What’s in the RedPajama-Data-1T LLM training set
RedPajama is “a project to create leading open-source models, starts by reproducing LLaMA training dataset of over 1.2 trillion tokens”. It’s a collaboration between Together, Ontocord.ai, ETH DS3Lab, Stanford CRFM, Hazy Research, and MILA Québec AI Institute.
[... 1,077 words]Wednesday, 19th April 2023
LLaVA: Large Language and Vision Assistant (via) Yet another multi-modal model combining a vision model (pre-trained CLIP ViT-L/14) and a LLaMA derivative model (Vicuna). The results I get from their demo are even more impressive than MiniGPT-4. Also includes a new training dataset, LLaVA-Instruct-150K, derived from GPT-4 and subject to the same warnings about the OpenAI terms of service.
Inside the secret list of websites that make AI chatbots sound smart. Washington Post story digging into the C4 dataset—Colossal Clean Crawled Corpus, a filtered version of Common Crawl that’s often used for training large language models. They include a neat interactive tool for searching a domain to see if it’s included—TIL that simonwillison.net is the 106,649th ranked site in C4 by number of tokens, 189,767 total—0.0001% of the total token volume in C4.
Stability AI Launches the First of its StableLM Suite of Language Models (via) 3B and 7B base models, with 15B and 30B are on the way. CC BY-SA-4.0. “StableLM is trained on a new experimental dataset built on The Pile, but three times larger with 1.5 trillion tokens of content. We will release details on the dataset in due course.”
Thursday, 20th April 2023
Data analysis with SQLite and Python for PyCon 2023
I’m at PyCon 2023 in Salt Lake City this week.
[... 347 words]Friday, 21st April 2023
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.
The AI Writing thing is just pivot to video all over again, a bunch of dead-eyed corporate types willing to listen to any snake oil salesman who offers them higher potential profits. It'll crash in a year but scuttle hundreds of livelihoods before it does.
Other tech-friendly journalists I know have been going through something similar: Suddenly, we’ve got something like a jetpack to strap to our work. Sure, the jetpack is kinda buggy. Yes, sometimes it crashes and burns. And the rules for its use aren’t clear, so you’ve got to be super careful with it. But sometimes it soars, shrinking tasks that would have taken hours down to mere minutes, sometimes minutes to seconds.
