Welcome to Replicate's weekly bulletin! Each week, we'll bring you updates on the latest open-source AI models, tools, and research. People are making cool stuff and we want to share it with you. Without further ado, here's our hacker-in-residence deepfates with an unfiltered take on the week in AI.
Hey everyone. Thanks for your lovely emails last week, and thanks especially to the reader who responded "谢谢你的来信，我已收到。" What a thoughtful sentiment, and such prompt response!
I know I promised you a signup form this week. Well, it turns out that is harder to do in an email than you would think. For now: if you're subscribed to this, forward it to someone you know. If someone forwarded this to you, go sign up for Replicate and click the "hear about stuff from Replicate" checkbox next to your email address. It's real easy, I promise.
Thanks for your patience, we're building this newsletter while flying it. See you next week!
--- deepfates
You've heard of Golden Gate Claude. Now meet Garden State Llama: a language model fine-tuned to hyperfocus on that beautiful, historical, inspirational state, the one and only, New Jersey!
Consistent Character takes one picture of a person and generates the same person in a bunch of different poses. This has been surprisingly hard to achieve up until now. Props to fofr for pulling this off.
Tooncrafter animates the motion in between static illustrations, so you can make beautiful cartoons from a few sketches. Works especially nicely with Consistent Character!
This novel method, Tree Diffusion trains a language model to code by editing random characters and seeing if it runs. A promising direction. This is how I code too.
The new line of Copilot+ computers from Microsoft will apparently take snapshots of your computer activity and store them in an unencrypted local database. TotalRecall allows you to inspect, search and summarize that data.
Copilot+ Recall will release June 18th 2024. Yes, the hackers got to it before it was even released. That's how you know you're in the cyberpunk future.
This comprehensive Applied LLMs guide shares industry lessons and best practices for building with current (and future) language models.
A long essay, or short book, in three parts: engineering tactics, operational concerns, and executive strategy. The 42 lessons are written by an all-star cast of AI engineers from companies like Amazon, Hex, Modal and Replicate.
This is a great overview of the information landscape around language models. The Generative AI Handbook gathers blogs, videos, web courses, and other resources to get you up to speed.
The course surveys sequential prediction foundations, state of the art methods, fine-tuning, multimodal models and more.
working on a portal to the multiverse
image generation on replicate is about to get REALLY fast
-- charliebholtz
What new creative tools are unlocked when you generate images in near real-time?
We're determined to find out.
What? Well, I scraped some text from the internet and stuffed it into a language model. Suddenly it claimed to be a gerbil?
It seems that Simon Willison accidentally prompt injected me.
The question here was "What is the meaning of life?"---my LLM documentation tries to be comprehensive but doesn't go as far as tackling grand philosophy! ...
In this case, my example of a gerbil talking about its love for snacks is clearly the most relevant piece of text in my documentation to that big question about life's meaning!
Systems built on LLMs consistently produce the weirdest and most hilarious bugs. I'm thoroughly tickled by this one.
--- simonw
Read the full story on his blog.
You can now disable your Replicate API tokens from the web.
No way, you actually read it? Thank you. I really like putting this thing together and I want them to let me keep sending it. Forward it to your friends, post about it on social media, sign up for Replicate if you haven't, every little bit helps.
And remember, this is an email! You can reply to it! Tell me what you think.
--- deepfates