Game-changing Clinical AI, Humanoid Robot Responds to Natural Speech, Is Nvidia Scraping ‘Human Lifetime’ of Videos … and more

Welcome to AI Weekly Breakthroughs, a roundup of the news, technologies, and companies changing the way we work and live.

'Game Changing' AI Can Detect Hidden Heart Inflammation

A new AI model developed by Oxford University spinout Caristo Diagnostics is being hailed as "game changing" for its ability to detect hidden inflammation in the heart that could lead to heart attacks within the next decade. Unlike traditional CT scans, this AI can identify biological processes that are invisible to the human eye, allowing for earlier intervention. The technology, currently being piloted in several UK hospitals, has shown that patients with coronary inflammation have a 20 to 30 times higher risk of a fatal cardiac event. As a result, nearly half of these patients were advised to make lifestyle changes or prescribed medication to prevent heart attacks. The AI model has already been approved in Europe and Australia.

Paige and Microsoft’s Advanced AI Models for Cancer Diagnosis

Paige and Microsoft have launched Virchow2 and Virchow2G, advanced AI models designed to revolutionize cancer diagnosis and treatment. Trained on over 3 million pathology slides from 225,000 patients worldwide, these models cover more than 40 tissue types and utilize diverse staining methods, making them highly versatile for cancer pathology. Virchow2G, the largest pathology model ever created, boasts 1.8 billion parameters and sets new standards in AI performance. These models aim to improve the accuracy, efficiency, and precision of cancer diagnosis, while also offering tools for life sciences and pharmaceutical companies to enhance therapeutic targeting and clinical trials. Virchow2 is available on Hugging Face for research purposes.

Figure’s New Humanoid Robot Can Respond to Natural Speech

Figure has introduced its latest humanoid robot, the Figure 02, which builds on its predecessor with a comprehensive redesign that includes enhanced AI-driven communication capabilities. A partnership with OpenAI enables the robot to understand and respond to natural speech, making it more effective in collaborative work environments. The Figure 02 features improved hardware such as six RGB cameras, an advanced visual language model, and upgraded hands with 16 degrees of freedom. The robot has already begun pilot projects with companies like BMW, with broader applications anticipated in both commercial and domestic settings in the future.

Palantir and Microsoft Partner on AI for National Security

Palantir and Microsoft have formed a strategic partnership to provide AI services for U.S. defense and intelligence agencies, integrating Microsoft's Azure OpenAI Service into Palantir's platforms within classified cloud environments. This collaboration marks a significant milestone in the application of AI for national security, potentially transforming critical defense operations. While details of the services remain unclear, the partnership has been well-received by investors, reflecting optimism about AI's role in national security. However, it also raises important ethical questions regarding the use of AI in surveillance and defense.

AI Safety Expert Joins OpenAI’s Board

OpenAI has added Zico Kolter, a Carnegie Mellon professor specializing in AI safety, to its board of directors. His appointment follows the departure of key OpenAI figures focused on AI safety, including co-founder Ilya Sutskever. Kolter will serve on the Safety and Security Committee, which advises on safety issues across OpenAI’s projects, though its effectiveness has been questioned due to its insider composition. Kolter brings significant expertise, with a background in AI safety research, and previous roles at C3.ai, Bosch, and AI startup Gray Swan.

Another OpenAI Co-founder Exits, Joins Anthropic

John Schulman, a co-founder of OpenAI, has left the company to join rival AI startup Anthropic, focusing on AI alignment research. Schulman's departure follows other key exits at OpenAI, including president Greg Brockman, who is taking an extended leave, and product manager Peter Deng. Schulman played a crucial role in developing ChatGPT and led OpenAI's alignment science efforts. Despite ongoing controversies at OpenAI, Schulman emphasized that his decision was personal and not due to a lack of support for alignment research. With his exit, only 3 of OpenAI’s 11 original founders remain.

New Nvidia Tools Use Apple Vision Pro to Control Humanoid Robots

Nvidia has introduced a suite of tools to advance the development of humanoid robots, with a focus on using the Apple Vision Pro for control and simulation. These tools, including Nvidia's NIM microservices and OSMO orchestration service, enable developers to train and control robots through AI and simulation workflows. A key feature allows the Apple Vision Pro to translate user movements, such as hand gestures, into corresponding robot actions, enabling precise teleoperation. This approach aims to accelerate and simplify the complex process of developing humanoid robots by leveraging advanced simulation and AI-generated data.

Whistleblower Accuses Nvidia of Scraping ‘Human Lifetime’ of Videos

Nvidia is facing accusations of scraping millions of videos online to train its AI systems, including those for Omniverse, self-driving cars, and Digital Humans avatars. An anonymous former employee claimed Nvidia was instructed to gather vast amounts of video data from various sources, including YouTube, Netflix, and other platforms. The company allegedly aims to build a substantial video data library to enhance its AI models, a move raising legal and ethical concerns. Despite Nvidia’s claims of having broad approval for data use, the practice echoes similar controversies faced by other AI companies. As AI training laws are still evolving, Nvidia’s aggressive data collection approach highlights ongoing debates about the legality of using online content for AI development.

Advanced Language Models for Complex Math

Alibaba Cloud's Qwen team has launched Qwen2-Math, a series of advanced language models designed to solve complex mathematical problems, outperforming previous industry leaders like GPT-4 and Claude 3.5. Built on a diverse mathematics-specific corpus and evaluated on multiple benchmarks, the Qwen2-Math models, especially the flagship Qwen2-Math-72B-Instruct, showcase exceptional proficiency in arithmetic and mathematical challenges. Rigorous decontamination methods ensure the model's accuracy, and future plans include expanding its capabilities to multilingual support, making advanced math problem-solving accessible globally.

Google Meet Will Take Notes for You

Google Meet is set to roll out a new AI-powered feature called Take Notes for Me, which will automatically take notes during meetings, helping users focus more on discussions rather than multitasking. Powered by the Gemini AI platform, admins can already test and configure the feature and a general launch of the feature is expected soon. Google Meet will be available to users with specific licenses, such as Gemini Enterprise and AI Meetings and Messaging add-ons, as part of Google Workspace's broader efforts to enhance collaboration tools.

AI Can Predict 3D Structures of Receptors to Enhance Drug Development

A new study demonstrates that AI can significantly enhance drug development by predicting the three-dimensional structures of important receptors. Researchers from Uppsala University applied AI to model the TAAR1 receptor, a target for drugs treating mental health disorders like schizophrenia and depression. By using supercomputers, they identified potential drug molecules from vast chemical libraries that fit the AI-generated model, leading to successful activation of the receptor in tests. The accuracy of AI-predicted structures surpassed traditional methods, showing promise for future drug discovery efforts.

Tesla's Voxel-Based Vision System: Revolutionizing Autonomous Robots

Tesla's patent for "Artificial intelligence modeling techniques for vision-based occupancy determination" introduces a new approach to autonomous robot navigation. The system uses a single neural network to process raw camera data and predict 3D environmental occupancy without additional sensors like LiDAR or radar. By dividing the space into 3D voxels, the technology allows real-time understanding of surroundings through features such as binary occupancy, shape details, semantic classifications, and motion information. This development is expected to advance Tesla's vision of versatile, efficient humanoid robots operating safely in complex environments.

Cohere Co-founder Says AI Is Not a Bubble

Nick Frosst, co-founder of Cohere, argues that while the AI industry is attracting significant investment and high valuations, it should not be considered a bubble. He acknowledges the industry's excitement but emphasizes that companies like Cohere are delivering real, tangible value through their AI models. But Frosst is skeptical about the prospect of achieving AGI soon, contrasting with some AI leaders who are more optimistic about the possibility. He advocates for a realistic understanding of AI's capabilities, stressing that while AI is powerful and useful, it is not a panacea or a threat to humanity. Frosst believes companies can use the technology effectively by focusing on practical applications and understanding AI's limitations.

Table Tennis Robot Praised as an ‘Interesting Practice Partner’

A team at Google has engineered a table tennis-playing robot to challenge human competitors. Detailed in a paper on ArXiv, the project uses twenty motion-capture cameras, dual 125 FPS cameras, a 6 DOF robot on linear rails, and a specialized paddle. Central to its operation is a high-level controller that directs multiple low-level controllers trained on a large annotated dataset through convolutional neural networks. The robot underwent testing with players from a local club. It easily defeated beginners but struggled against intermediate players and was outmatched by advanced players.?

New Machine Learning Model Boosts Disease Prediction Accuracy

A recent study showcases a groundbreaking ensemble feature selection model for predicting multiple diseases using Electronic Health Records. Combining statistical, deep, and optimal features through the Stabilized Energy Valley Optimization with Enhanced Bounds (SEV-EB) algorithm, the model significantly enhances prediction accuracy and stability, capturing both short- and long-term health patterns. It achieves 95% accuracy and a 94% F1-score, outperforming traditional methods and marking a significant advancement in healthcare.

Mechanical Orchard, led by ex-Pivotal CEO, scores $50M round led by Alphabet’s GV

SoundHound acquires Amelia AI for $80M after it raised $189M+

Anysphere, a GitHub Copilot rival, has raised $60M Series A at? $400M valuation from a16z, Thrive

AI4 - Las Vegas - August 12 - 14?

The AI Conference 2024 - San Francisco - September 10 - 11

Dreamforce - San Francisco - September 17-19

World Summit AI - Amsterdam - October 9 - 10?

Gitex Global - Dubai - October 14 - 18?

Big Data Conference Europe - Vilnius - November 19 - 22

AWS re:Invent 2024 - Las Vegas - December 2-6