TLDR AI 2023-08-14

Supabase Studio 3.0 is here, with some huge new features, including a brand new Supabase AI integrated right into the SQL Editor, Schema Diagrams, and new Wrappers.

Recent studies suggest LLMs like Othello-GPT build world models, implying that they understand the world to an extent. The network discerned the underlying 8x8 board structure and game rules by training on game moves without explicit game rules. This challenges traditional notions, like John Searle's Chinese Room Argument, about machine understanding.

Serve Robotics, the autonomous sidewalk delivery robot startup that spun out of Uber’s acquisition of Postmates, is going public via a reverse merger with a blank-check company.

One of the best ways to control image generation is to use edges to condition the final output. The best open source model, SDXL, has a controlnet variation trained and ready to use.

One hard challenge when working with language models that write code is evaluating the quality of the generated code. This is difficult because running arbitrary code is dangerous for a system. The code could do anything, including deleting everything. MosaicML offers a secure environment for evaluating generated code.

A robotic system called "Follow Anything" can spot, trace, and follow any object in real time, even if it has never seen it before. This smart robot can recognize objects from just a text description, a picture, or even a simple click - all powered by a regular laptop!

ChainML recently launched Council — an open-source AI agent platform that makes it easy to rapidly develop and deploy customized generative AI applications using collaborative AI agents. Council integrates with a wide variety of LLMs, such as GPT-4, Llama 2, and Claude 2, to unlock generative AI for business applications, enabling use cases that were previously impossible through AI agents with sophisticated control flows.

Mirage is a platform for generating 3D scenes. It is used by game developers to speed up the process. The team has been working internally on improving a variety of 3D generation methods. This open source push is designed to leverage the community to improve its models.

Hugging Face recently removed commercial license-ability from one of its previously open packages. The package has now been forked by Preemo and will continue to be developed as an open source package.

The WizardLM group from Microsoft has released the world’s best open source math model. It’s really good at reasoning through multi-step problems. The model hallucinates a bit, which is really fun for math, and offers explanations. I think this may be one of the best models for education for those interested in learning mathematics.

This article examines the training dynamics of a tiny model and reverse engineers the solution it finds. It provides an illustration of the exciting emerging field of mechanistic interpretability.

If you zoom in on salient pieces of an image before running classification, you can get 98% on ImageNet. Fascinating and fairly easy implementation for real world computer vision systems.

Recast turns your want-to-read articles into rich audio summaries.

There are lots of models based on Llama 2 and trained on GPT-4 outputs. These are somewhat in that camp, but as non-commercial license models go, they are quite powerful.

Nvidia makes it easy to train your own vision-language assistant based on its Nemo model.