TLDR AI 2024-07-11

The Hugging Face team has released an extremely powerful and strong math model.

Anthropic's new features in Claude allow developers to automate prompt engineering, improving AI application development by generating, testing, and refining prompts with quick feedback.

Microsoft and Apple have opted not to join OpenAI's board due to regulatory scrutiny, with Microsoft stepping down from its observer role and Apple reversing plans to become an observer.

Anole is an open autoregressive multimodal model built on Meta's Chameleon model. This work adds image generation capabilities back into the model with targeted fine-tuning.

Researchers introduce a new reinforcement learning framework to improve personalized text-to-image generation.

PerlDiff is a method that improves the precision of street view image generation by integrating 3D geometric information.

From the creator of ControlNet, IC-Light, and many other image generation systems, Paints UNDO is a system where a model creates strokes that are used to recreate an image. Interestingly, unlike previous stroke systems, this model can undo strokes and often rethinks its approach entirely part way through - very similarly to a human artist.

Scalable Rectified Flow Transformers are used, in part, for Stable Diffusion 3. This repository is a minimal implementation of this that also includes muP hyperparameter sweeps.

A framework for serving and evaluating LLM routers.

One popular area of VC investing is building the “picks and shovels” for AI (e.g. model fine-tuning, observability, and inference). VCs are betting that companies will not want to build things in-house as they add AI to their products. The TAM for AI inference is quite constrained. VCs need to believe in massive TAM expansion for their investments to pay off. While AI inference platforms are a boon for startups in the short term, they are detrimental to startups over longer periods.

SuperFlow is a new framework that enhances 3D perception in autonomous driving by using consecutive LiDAR-camera pairs for spatiotemporal pretraining.

Cyclists in San Francisco have reported close encounters and safety concerns with Waymo's and Cruise's autonomous vehicles. Nearly 200 complaints have been made to the California DMV highlighting erratic driving and near-misses involving these self-driving cars. While the companies claim their vehicles have enhanced safety features, the incidents raise questions about their readiness for broader deployment amidst increased regulatory scrutiny.

You can speed up model initialization dramatically without much loss in tokens per second by moving some lazy loading to the model on the first pass.

Remember anything with a real-time, on-device AI assistant.

OpenAI and Los Alamos National Laboratory (LANL) are collaborating to explore safe applications for AI in bioscientific research, with a focus on using models like GPT-4o to assist in laboratory tasks through multimodal capabilities such as vision and voice.

Are you interested in AI voice generation? Join Oxford Languages and WellSaid Labs for a live webinar. Sign up here: What can curated pronunciation data do for text-to-speech models?