TLDR AI 2024-09-02

Apple may be expanding into robotics, potentially introducing devices that move, such as an iPad mounted on a robotic arm. Headed by VP of Technology Kevin Lynch, Apple's robotics efforts are progressing with the help of experts from firms like Israel's Technion, and propose to include AI interfaces beyond Siri. While still conceptual, Apple is considering a 2026 or 2027 release for these new robotic products.

Command R and Command R+ received an upgrade on essentially all tasks. They are now better at recall, speed, math, and reasoning.

Nearly 200 Google DeepMind employees signed a letter urging Google to terminate military contracts, claiming a violation of the company's own AI ethics principles. DeepMind technology has been bundled into Google Cloud and sold to militaries, sparking internal conflict with AI staff who value ethical standards. Google's response demonstrated an adherence to AI Principles, but workers remain unsatisfied, seeking stronger governance against military use of their AI.

Routing tokens to experts in MoEs helps sparsify the computation. However, learning that routing can be challenging. There is typically a complicated loss structure. This paper shows a novel way to overcome this problem and dramatically improve expert balancing and training stability.

This end-to-end framework simultaneously addresses low-light enhancement and motion deblurring in videos using event cameras.

Researchers have proposed a new audio-visual learning framework for enhancing sound source localization, addressing the issue of false negatives in traditional methods. The framework includes two schemes: Self-Supervised Predictive Learning (SSPL) and Semantic-Aware Contrastive Learning (SACL). SSPL eliminates false negatives by focusing on positive-only learning, while SACL refines the contrastive learning process to improve alignment between audio and visual features.

Spiking Diffusion Models (SDMs) is a new family of Spiking Neural Networks (SNNs) designed for image generation, offering significant energy savings and high biological plausibility.

A popular inference library now has support for Flux Schnell running on the CPU.

This may be one of the bigger improvements to the Transformer Reinforcement Learning library. It includes onlineDPO, Liger Kernels, WinRate Callbacks, and more.

Laion 5B was the largest open image dataset on the internet. It was shut down over concerns of inappropriate and harmful images. The organization has made significant efforts to mitigate these issues and is now rereleasing the dataset.

Over 25 Chinese companies showcased humanoid robots aimed at factory automation at the World Robot Conference in Beijing, backed by substantial government investments and leveraging China's deep supply chain. The global humanoid robot market is predicted to hit $38 billion by 2035. China aspires to mass-produce these robots by 2025, intensifying competition with Tesla's upcoming Optimus robot. Tesla anticipates that it will deploy 1,000 Optimus robots in its factories in the next year, while Chinese firms forecast significant cost reductions for their models.

Generative AI is poised to revolutionize storytelling, blending narrative depth with the dynamic interactivity of video games. Recent advancements in generative models such as OpenAI's Sora and Luma AI's Dream Machine are accelerating this shift, enabling real-time, interactive video creation. This evolution has the potential to produce the next "Pixar" in interactive media, merging film, gaming, and AI.

Boston Dynamics has showcased its electric biped robot, Atlas, performing push-ups in a new video that emphasizes its strong actuators amidst its early commercialization phase for factory floor applications.

The Vybe AI camera app transforms your selfies into vibrant, beautiful photos using cutting-edge AI.

OpenAI is reportedly preparing for a significant funding round that could value the company at over $100 billion, potentially involving investments from Apple, Nvidia, and Microsoft.