Weekly Digest: May 2, 2024
This week, the US government eyes new AI regulations, GitHub announces Copilot Workspaces, a world editor for AI Town, sewing and generating clothes from text, and a new model that can clone any voice and speak in many languages.
Welcome to our Weekly Digest, where Alex Duffy summarizes the latest updates you should know across the AI ecosystem, ComfyUI, and Salt AI. This week, the US government eyes new AI regulations, GitHub announces Copilot Workspaces, two new healthcare-focused models from Meta and Google, a new AI world builder for agents, and a new model that can clone any voice and speak in many languages.
Check out the video above, and follow along with Alex's notes and links below:
Big News
US government seemed to eye new AI regulation
Proposed in California met with angry outcry
Department of Homeland Security announced the establishment of the Artificial Intelligence Safety and Security Board
US @NIST publishes 1st draft of its "AI Risk Management Framework: Generative AI Profile
GitHub announced waitlist for technical preview of Copilot Workspace
Quick summary on Twitter
It takes a while to build tools that work, that’s why we’re building Salt, but once they start to work, they’ll just keep getting better and can deliver real value
Next we need something like this for version control! Somewhere Salt solves problems but one others are looking at too
Two new medical-focused models from Meta and Google
Meta announces Meditron a Llama LLM suite for low-resource medical settings
Google announces Med-Gemini, a family of Gemini models fine-tuned for medical tasks
MedGemini achieves SOTA on most of the tasks in these benchmarks and is preferred by clinician raters for the more challenging longform text tasks!
More and more evidence that LLM’s could really help democratize access to at least some healthcare - though of course not without risks
MIT CSAIL and @myshell_ai researchers introduce OpenVoice V2 a text-to-speech model that can clone any voice and speak in many languages
Fully open source and allows for commercial use
New in LLMs
Llama3 continues to make an impact
Researchers make finetunes and push the context limit to 32K extremely effectively and to 1M+ with various rates of success
Long context lengths are helpful for models because you can give them hundreds of examples you can do a LOT with prompting - few shot prompting is super powerfu
AI-Town seeing a resurgence in popularity thanks to Llama3 and Pinokio
Doubled down with a town world editor
Solving dependencies is always an issue and this solves it for specific user group
Simulations with agents in a game are unassuming ways to push really complex functionality, especially if they go from just chatting to playing Among Us, for example
HuggingFace released one of the best datasets for training LLMs ever
Data is very important and LLMs helping to make it is huge
The 45TB file was so popular it took down Huggingface which had intermittent outages for the next three days before becoming one of the top 10 most liked datasets on the platform
StarCoder2-Instruct - Fully Transparent and Permissive Self-Alignment for Code Generation
The team uses the model to improve the past model, another data-point highlighting the promise of synthetic data
Benchmarks
Start of a text to video benchmark
Justine Tunney continues to push the envelope
New in Generative AI
The barrier to entry to VFX for video game and movie production is lowering and tools like ComfyUI and Salt are at the core of that
BlenderAlchemy - Editing 3D Graphics with Vision-Language Models - Stanford
Sewing and Generating Clothes from Text - Shanghai Tech Univ., UPenn
A creator using ComfyUI made visuals for a Coachella set
New In ComfyUI
Transitions with control
Matteo released new nodes that can do awesome animations from people to environments
Steerable Motion is also a similar node from Banodoco that is climbing in stars and improving in quality from 1.4 to 1.5
Bonus News
LLMs might get better by planning ahead
31 years ago CERN released the source code for the World Wide Web for anyone to use
