Search the archive

COGVIDEO 5B LAUNCHING NEXT WEEK

CogVideo 5B model will launch next week in Hugging Face Diffusers.

COHERE COMMAND R UPDATE

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

WORKERS AT GOOGLE DEEPMIND PUSH COMPANY TO DROP MILITARY CONTRACTS

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.

APPLE'S ROBOTICS WORK AIMS TO SOLVE USER'S FIRST-WORLD PROBLEMS

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.

AUXILIARY-LOSS-FREE LOAD BALANCING FOR MOES

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.

IMPROVING LOW-LIGHT VIDEO QUALITY

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

SOUND SOURCE LOCALIZATION

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.

TRL RELEASE

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

FAST SD SUPPORT FOR SCHNELL ON CPU

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

SPIKING DIFFUSION MODELS FOR IMAGE GENERATION

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.

LAION 5B SAFETY RELEASE

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.

THE NEXT GENERATION PIXAR: HOW AI WILL MERGE FILM & GAMES

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.

CHINA'S ROBOT MAKERS CHASE TESLA TO DELIVER HUMANOID WORKERS

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.

VYBE RAISES $4.75M TO POWER AI-PHOTO FUN AND A NEW WAY TO CONNECT WITH FRIENDS

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

APPLE AND NVIDIA MAY INVEST IN OPENAI

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.

BOSTON DYNAMICS' NEW ELECTRIC ATLAS CAN DO PUSH-UPS

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.

MAGIC DEV'S 100M CONTEXT WINDOW

Yes, that's not a typo. 100 million tokens in context for reasoning and programming agents. Magic Dev also announced a partnership to build two new supercomputers on Google Cloud. This comes based on the recent $320m fundraise to accelerate its product development.

OPENAI AND ANTHROPIC WILL SHARE THEIR MODELS WITH THE US GOVERNMENT

OpenAI and Anthropic have agreed to allow the US government early access to their major new AI models before public release to enhance safety evaluations as part of a memorandum with the US AI Safety Institute.

CALIFORNIA LEGISLATURE PASSES SB 1047

California's controversial AI safety bill, SB-1047, has passed the State Assembly and is now one step away from Governor Gavin Newsom's desk. Newsom must decide by September 30 whether to sign it into law.

THE MAMBA IN THE LLAMA: DISTILLING AND ACCELERATING HYBRID MODELS

Mambda models are hard to accelerate with speculative decoding because of the lack of a KV cache for backtracking. This paper, from some of the original authors, showcases some novel acceleration algorithms and distillation methods.

EFFICIENT LLM SCHEDULING BY LEARNING TO RANK

When serving many concurrent requests to a large language model, things often get bottlenecked because we don't know how long output generation will be, so we get head of line bottlenecks. However, you can learn to rank relative lengths between requests and serve the shortest ones first, which leads to a 6.5x higher throughput for multi batch generation.

MULTI-TASK SCENE UNDERSTANDING

MTMamba++ is a new model architecture designed to enhance multi-task scene understanding. This approach uses a Mamba-based decoder featuring two core blocks, STM and CTM, to capture long-range dependencies and improve cross-task interactions.

🗓️ WHEN: Wednesday, September 4

🕺 WHO'S ATTENDING: Engineers & experts from DataStax, PayPal, NVIDIA, Vercel, Unstructured & more

HERMES FUNCTION CALLING DATASET (HUGGING FACE HUB)

Great open dataset from Nous Research for training models to call functions.

QWEN 2 VL

The Qwen team has released vision language models built on top of Qwen 2.

RAW IMAGE PROCESSING

RAW-Adapter is a new approach that adapts sRGB pre-trained models to handle camera RAW data effectively.

LLAMA USAGE DOUBLE MAY THROUGH JULY

Meta has released some statistics about Llama model usage. It found strong demand for its models being deployed in enterprise settings.

3D MEDICAL IMAGE ANNOTATION

This study has adapted the Segment Anything Model 2 (SAM 2), originally designed for video annotation, to streamline the annotation of 3D medical images.

ANTHROPIC SAYS IT IS CLOSER TO BACKING CALIFORNIA'S AI BILL AFTER LAWMAKERS MADE SOME FIXES

Anthropic has shifted to a cautious endorsement of California's AI regulation bill SB 1047 after key amendments. Changes include reduced penalties, clearer language, and increased reporting requirements. The bill aims to ensure safety in AI practices, gaining support from Anthropic yet facing opposition from OpenAI and Meta.

WISPR FLOW (PRODUCT LAUNCH)

Flow is a Mac dictation app that lets you speak naturally and writes in your style in every application, 3x faster than a keyboard.

OPENAI SAYS CHATGPT USAGE HAS DOUBLED SINCE LAST YEAR

OpenAI announced that ChatGPT has surpassed 200 million weekly active users, doubling its user base from a year ago, and noted that 92% of Fortune 500 companies are using its products.

TIKTOK OWNER BYTEDANCE LAUNCHES NEW VIDEO SEARCH TOOL

ByteDance has introduced Douyin Search, an app for searching short video content on China's version of TikTok, directly competing with Baidu's search dominance.

See you there!

MIDJOURNEY SAYS IT'S ‘GETTING INTO HARDWARE'

Midjourney is expanding into hardware with a new team based in San Francisco. The move, hinted at by the hiring of ex-Neuralink and Apple Vision Pro engineer Ahmad Abbas, suggests that the hardware could be related to its ongoing development of AI models for video and 3D generation.

GOOGLE ROLLS OUT GEMS AND IMAGEN 3 TO GEMINI ADVANCED

Google is rolling out its new features, Gems and Imagen 3, to Gemini Advanced subscribers. Gems allows users to create custom versions of Gemini tailored to specific tasks, offering premade options like learning coaches and coding partners, while Imagen 3 is Google's latest image generation model, now available to generate detailed and artistic images.

OPENAI IN TALKS FOR FUNDING ROUND VALUING IT ABOVE $100B

OpenAI is in discussions to raise several billion dollars in a new funding round led by Thrive Capital that would value the company above $100 billion, with Microsoft also expected to participate.

GENERATIVE VERIFIERS: REWARD MODELING AS NEXT-TOKEN PREDICTION

Reward models are generally trained as discriminative classifiers. This work from DeepMind uses a language model's logits for yes/no as the reward signal. It found that by allowing a model to use CoT and ensembling, researchers improved performance by 16%.

MULTILINGUAL ARBITRAGE

Cohere's Aya model was able to dramatically improve its win rate vs baseline models by taking advantage of the difference between oracle model performance and routing synthetic data generation.

INTRODUCING TABLE-AUGMENTED GENERATION

Table-Augmented Generation is a new paradigm that combines language models with databases to answer complex natural language questions.

session next Wednesday (September 4) for the best experience →

FAST AND HIGH-QUALITY 3D SHAPE GENERATION WITH DIFFUSION MODELS

OctFusion is a method for generating 3D shapes with diffusion models that achieves both efficiency and high quality. It can produce 3D shapes at arbitrary resolutions in just 2.5 seconds on a single Nvidia 4090 GPU.

DISTILLING TRANSFORMERS INTO EFFICIENT LINEAR RNNS

EVENT-GUIDED VIDEO DEBLURRING

Researchers have developed a new approach to video deblurring by integrating an event camera, which captures motion with microsecond temporal resolution, to enhance the clarity of motion-blurred videos.

REAL TIME ON DEVICE AI

Cartesian has announced a number of model and system improvements. It has also released an open hybrid State space model.

STEPHEN WOLFRAM THINKS WE NEED PHILOSOPHERS WORKING ON BIG QUESTIONS AROUND AI

Stephen Wolfram emphasizes the importance of incorporating philosophical rigor into AI research, addressing core ethical questions. As AI's influence grows, these questions, traditionally part of philosophy, are increasingly relevant. Wolfram advocates for deeper, classical philosophical thinking in AI development to better navigate its implications for humanity.

THE TOP AI DEALS IN EUROPE THIS YEAR

Despite general headwinds for startups, AI ventures continue to secure substantial funding. U.S. AI startups have achieved nearly 30 deals over $100M in 2024, with Europe not far behind. Major investments include WAYVE ($1B), Mistral AI (~$1B), Helsing ($484M), Poolside ($400M), DeepL ($320M), H ($220M), and Flo Health ($200M).