Today:

• Nakasone Joins OpenAI Board

• Microsoft Delays Recall AI

• AWS Commits $230M for AI

• Nvidia Unveils Nemotron-4 Model

• Siri Gets Major Upgrades

• Clearview AI Offers Company Stake

• Picsart, Getty Images Partnership

• AI Rodent Mimics Real Rats

• Robots Mimic Human Movements

• Nadella Builds AI Empire

OpenAI appoints Retired U.S. Army General Paul M. Nakasone to Board of Directors

OpenAI has appointed retired U.S. Army General Paul M. Nakasone to its Board of Directors. Known for his expertise in cybersecurity, Nakasone will join the Board’s Safety and Security Committee. His role will involve enhancing the security of OpenAI’s systems and contributing to the use of AI in detecting and responding to cybersecurity threats. 

Nakasone’s extensive experience, including leading U.S. Cyber Command and the National Security Agency, will help guide OpenAI in ensuring that artificial general intelligence is developed and deployed safely and beneficially.

OPENAI

Microsoft to delay release of Recall AI feature on security concerns

Microsoft is delaying the release of its new AI feature, "Recall," due to privacy concerns. Originally planned for a broad rollout, Recall will now be previewed with a smaller group via the Windows Insider Program. The feature tracks computer usage, including web browsing and voice chats, storing this data for future reference. 

Privacy issues arose after its announcement, with some fearing it could enable spying. Microsoft emphasized its commitment to security and plans to gather feedback before making Recall widely available. Billionaire Elon Musk criticized the feature, likening it to a "Black Mirror" episode.

REUTERS

AWS Announces $230 Million Commitment for Generative AI Startups

Amazon Web Services (AWS) announced a $230 million investment to support generative AI startups. This funding will provide AWS credits, mentorship, and education to early-stage companies. Part of the commitment includes the AWS Generative AI Accelerator program, which will support 80 startups with up to $1 million in credits each. 

These startups will gain access to AWS's technology, mentorship, and the NVIDIA Inception program. The initiative aims to accelerate the development of AI applications in various sectors. Applications are open until July 19, and selected startups will be announced on September 10.

AMAZON

Nvidia’s ‘Nemotron-4 340B’ model redefines synthetic data generation, rivals GPT-4

Nvidia has unveiled "Nemotron-4 340B," a groundbreaking family of open models for synthetic data generation, rivaling GPT-4. This innovation allows businesses to create powerful, domain-specific language models without costly real-world datasets. Nemotron-4 340B, trained on 9 trillion tokens and supporting over 50 natural languages and 40 programming languages, offers unmatched performance and versatility. Its commercially-friendly licensing makes it accessible for businesses of all sizes. 

The model's potential impact spans healthcare, finance, manufacturing, and more, enabling advancements in personalized medicine, fraud detection, and predictive maintenance. Nvidia’s commitment to AI innovation positions it at the forefront of the industry.

VENTUREBEAT

Apple won’t wait until next year for some Siri improvements

Apple plans to improve Siri this year with updates arriving in iOS 18. Users will soon be able to type to Siri without accessing specific settings and experience a more conversational and intuitive assistant. Additionally, Siri will gain better knowledge of Apple products. A new feature allows users to set a custom wake word for Siri, enhancing user convenience. 

Siri will also be integrated with ChatGPT, offering advanced assistance for queries it cannot handle alone. These improvements aim to enhance Siri’s usability and performance, making it a more effective digital assistant for Apple users.

THE VERGE

Clearview AI Used Your Face. Now You May Get a Stake in the Company.

Clearview AI, a facial recognition start-up, has settled a class-action lawsuit by offering a 23% stake in the company to Americans whose faces were scraped from the internet. Clearview AI, which created its app using photos from social media, faced multiple lawsuits after its practices were exposed. 

The settlement, valued at $52 million, allows class members to benefit from the company’s future success. Plaintiffs’ lawyers will also receive a portion of any proceeds. Critics argue the settlement encourages continued privacy violations, while Clearview AI is relieved to avoid potential bankruptcy. The agreement awaits court approval.

THE NEW YORK TIMES

Picsart Partners with Getty Images to Bring Commercially-Safe AI Image Generation to its Platform

Picsart has partnered with Getty Images to launch a new AI model for generating commercially safe images. This model, developed by Picsart’s AI lab, uses Getty Images' licensed content to create unique images with commercial rights, aimed at simplifying content creation for marketers. The integration allows users to generate and edit images within Picsart, supporting continuous content production. 

This partnership promises high-quality, commercially viable AI-generated imagery directly within the Picsart platform. The release, expected later this year, will also include an API for broader access.

PICSART

A virtual rodent predicts the structure of neural activity across behaviors - Nature

The study presents a 'virtual rodent' model where an artificial neural network controls a realistic rat model in a physics simulator, trained through deep reinforcement learning to imitate real rats' behavior. By comparing the neural activity of the virtual rodent to real rats, researchers found that the network's activity better predicted neural activity in specific brain regions associated with motor control, suggesting the regions implement inverse dynamics.

The virtual rodent's latent variability predicted neural variability across behaviors, aligning with optimal feedback control principles. Overall, the results demonstrate how simulating virtual animals can aid in understanding neural activity across behaviors and its relation to motor control principles.

NATURE

Humanoid Shadowing and Imitation from Humans

"HumanPlus" introduces a comprehensive system for humanoid robots to learn from human motion data. The system uses reinforcement learning to train robots in simulation, allowing them to mimic human movements in real time using only an RGB camera. This shadowing enables human operators to control humanoids and collect motion data for various tasks. 

The data is then used to train the robots to perform tasks autonomously, such as walking, unloading objects, and folding clothes, with high success rates. The system demonstrates significant advancements in humanoid robot training and autonomous skill execution.

GITHUB

Microsoft’s Nadella Is Building an AI Empire. OpenAI Was Just the First Step.

Microsoft CEO Satya Nadella is aggressively expanding the company’s AI capabilities beyond its partnership with OpenAI. While the OpenAI deal, which created ChatGPT, positioned Microsoft as a frontrunner in AI, Nadella is investing widely in AI talent, tools, and technology. Recently, he invested $1.5 billion in an AI startup based in Abu Dhabi. 

This strategy aims to diversify Microsoft's AI portfolio and secure its leadership in the rapidly evolving AI industry. Nadella’s efforts are transforming Microsoft into a powerhouse of AI innovation through global partnerships and strategic investments.

THE WALL STREET JOURNAL

🧠RESEARCH

Depth Anything V2

Depth Anything V2 introduces a more efficient and accurate monocular depth estimation model. Key improvements include using synthetic images, enhancing the teacher model, and leveraging pseudo-labeled real images. The model is over 10 times faster than its predecessors and offers various scales, improving generalization and precision in depth predictions.

An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels

This paper reveals that vanilla Transformers can perform well by treating each pixel as a token, challenging the necessity of local inductive bias in computer vision models. This approach, although less efficient, shows strong results in object classification, self-supervised learning, and image generation, suggesting a potential shift in future neural architecture designs.

Transformers meet Neural Algorithmic Reasoners

The paper introduces TransNAR, a hybrid model combining Transformers and graph neural network-based neural algorithmic reasoners (NARs). This approach enhances Transformers' ability in algorithmic reasoning tasks by allowing them to cross-attend to node embeddings from NARs. Evaluations on the CLRS-Text benchmark show significant improvements over Transformer-only models in both in-distribution and out-of-distribution scenarios.

MotionClone: Training-Free Motion Cloning for Controllable Video Generation

MotionClone is a training-free framework for motion-based controllable text-to-video generation. Unlike previous methods requiring model training or fine-tuning, MotionClone clones motions from a reference video using temporal attention and location-aware semantic guidance. This approach ensures high motion fidelity, textual alignment, and temporal consistency, outperforming traditional models in generating realistic video motions.

Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing

The paper introduces Magpie, a method for synthesizing large-scale, high-quality alignment data for large language models (LLMs) without needing human labor. By leveraging the auto-regressive nature of aligned LLMs like Llama-3-Instruct, Magpie generates user queries and responses, producing 4 million instructions. Fine-tuning LLMs with Magpie data results in performance comparable to officially aligned models, demonstrating the effectiveness and scalability of this self-synthesis approach.

🛠️TOP TOOLS

Jace -  AI assistant to do long-running tasks in the digital world

OTTO SEO - A first-of-its-kind AI SEO tool that will revolutionize the way you do SEO

GenType -Make an alphabet out of anything

Mapify - Quickly transform overwhelming content into simple, clear mind maps

Suno - Make a song about anything

📲SOCIAL MEDIA