TLDR AI 2023-09-18

Mixture-of-Experts models are sparse models, which means they save on computation while not sacrificing accuracy. They're tricky to train. IBM has released several along with their architecture and some code.

Industry heavyweights like Musk, Gates, and Zuckerberg debated AI regulation in a private session, discussing concerns about open-source misuse and job threats. A consensus emerged on the need for government oversight.

Adobe's Firefly generative AI models, previously in beta, are now available in Creative Cloud, Adobe Express, and Adobe Experience Cloud. Users will be charged via "generative credits" based on model interaction. Paid Adobe plans come with various credit amounts. Firefly will still work (but run slower) once credits run out.

This recently updated work dives into the many angles of AI copyright. AI copyright can broadly be classified into Input and Output legal challenges. The paper doesn’t present any opinions - it is merely a summary of many of the common issues.

Researchers have developed a method called EncodecMAE that borrows ideas from text-based models to make a universal audio model. This new approach works well for different kinds of audio tasks like understanding speech, identifying music, or recognizing sounds in the environment.

SyntheWorld is a groundbreaking collection of fake but super-detailed images that help researchers study the Earth from above without the usual big costs and headaches. With 40,000 images, it's perfect for learning how land changes over time.

Researchers have created Repilot, a system that makes AI copilots better at helping programmers fix bugs in their code.

Recent improvements in the field of 3D generation have been blistering. This work is the most stable and general. You can input an image and the model will synthesize many views and a resulting 3D model.

Transformers are smart algorithms usually used for text or images. This project tweaked them to work with EEG data to better understand sleep stages.

Consultants who use GPT-4 vastly outperform consultants who don’t. A recent study provided further evidence that knowledge workers who utilize AI will be the winners in the future.

A set of spiral-based control net images recently went semi-viral. This space was made shortly thereafter to reproduce that style.

This article breaks down the risks and challenges surrounding AI security and looks at the opportunities for companies in this new world order.

Generative dynamics uses both modern and classic AI to synthesize motion in a still image. It is completely interactive as well.

Contrary to popular opinion, GPT-4’s performance is not systematically declining.

Trickle produces AI-generated insightful summaries for each uploaded screenshot.