CuspAI Raises $30M to Develop Gen-AI Material Discovery Engine

Welcome to Towards AGI, your premier newsletter dedicated to the world of Artificial Intelligence. Our mission is to guide you through the evolving realm of AI with a specific focus on Generative AI. Each issue is designed to enrich your understanding and spark your curiosity about the advancements and challenges shaping the future of AI.

Whether you're deeply embedded in the AI industry or just beginning to explore its vast potential, "Towards AGI" is crafted to provide you with comprehensive insights and discussions on the most pertinent topics. From groundbreaking research to ethical considerations, our newsletter is here to keep you at the forefront of AI innovation. Join our community of AI professionals, hobbyists, and academics as we pursue the ambitious path towards Artificial General Intelligence. Let’s embark on this journey together, exploring the rich landscape of AI through expert analysis, exclusive content, and engaging discussions.

TheGen.AI News

CuspAI Raises $30M to Develop Gen-AI Material Discovery Engine

CuspAI, a Cambridge-based startup focused on using artificial intelligence to develop new materials, has secured $30m (£24m) in funding, with support from the "godfather of AI." The company aims to revolutionize the field of advanced materials through generative AI by creating an AI search engine that allows users to request materials with specific properties.

Co-founder and chief AI officer Prof Max Welling explained, "Imagine a search engine that not only searches for existing materials but also explores all potential molecules and materials that could be created." CuspAI's AI-designed materials are expected to significantly impact carbon capture and storage, which removes carbon from the atmosphere.

Welling elaborated that users could request a material with properties tailored to selectively bind carbon dioxide, and the CuspAI platform would assess feasibility and optimize the molecular structure to meet the specified criteria.

The startup was founded by Welling, a former Microsoft Research and Qualcomm scientist, and Dr. Chad Edwards, a former Google and Quantinuum employee. Edwards highlighted the dual challenges posed by the AI revolution, including increasing energy consumption and carbon emissions from data centers. He added that the materials being developed are also suitable for applications such as energy storage, catalysis, and gas and water purification.

Geoffrey Hinton, a computer science pioneer and one of the "godfathers of AI," will advise the company's board. Hinton expressed his admiration for CuspAI's mission, noting that AI can both create and solve challenges, with climate change being a critical issue that AI-designed materials can help address.

The investment round was led by Hoxton Ventures and included participation from Basis Set Ventures, Lightspeed Venture Partners, LocalGlobe, Northzone, Touring Capital, Giant Ventures, FJ Labs, Tiferes Ventures, and Zero Prime Ventures.

Runway Launches Gen-3 Alpha: Next-Gen AI Video Model for Ultra-Realistic Clips

New York City-based startup Runway ML, also known as Runway, was one of the first to develop high-quality generative AI video creation models. However, after launching its Gen-1 model in February 2023 and Gen-2 in June 2023, the company faced stiff competition from other advanced AI video generators like OpenAI's upcoming Sora model and Luma AI's Dream Machine, which launched last week.

Today, Runway is making a significant comeback in the generative AI video field by announcing Gen-3 Alpha. According to a blog post, this model is the “first of an upcoming series of models trained by Runway on a new infrastructure built for large-scale multimodal training.” It represents a step towards creating General World Models, AI models capable of simulating a wide range of real-world situations and interactions. Sample videos created with Gen-3 Alpha are available for viewing.

Gen-3 Alpha enables users to generate 10-second, high-quality, detailed, and highly realistic video clips, featuring precise emotional expressions and camera movements. Runway has indicated that this initial rollout will support 5 and 10-second video generations, with generation times of 45 seconds for a 5-second clip and 90 seconds for a 10-second clip.

While there is no exact release date yet, Runway has showcased demo videos on its website and social media. It's also unclear if the model will be available through Runway's free tier or require a paid subscription, which starts at $15 per month or $144 per year.

Runway co-founder and CTO Anastasis Germanidis confirmed in an interview with VentureBeat that the Gen-3 Alpha model would be available to paying subscribers within days, with the free tier to follow at a later, unspecified date. A Runway spokesperson reiterated this, stating Gen-3 Alpha would soon be available to paid subscribers, the Creative Partners Program, and Enterprise users.

Gabe Michael, a Runway user, mentioned on LinkedIn that he expects access later this week. On X, Germanidis stated that Gen-3 Alpha will soon be integrated into the Runway product, enhancing existing modes like text-to-video, image-to-video, and video-to-video, as well as introducing new capabilities made possible by the advanced base model.

Germanidis added that since the release of Gen-2 in 2023, Runway has learned that video diffusion models have much room for performance improvements. These models build powerful representations of the visual world by learning to predict video from pixelated noise based on annotated image/video and text pairs.

China Unveils Advanced AI-Powered Sexbots Hitting Shelves Soon

Chinese scientists and engineers are incorporating ChatGPT-like technology into sex robots, aiming to create interactive, AI-powered companions despite facing significant technical and ethical challenges. In Shenzhen, Starpery Technology, a leading producer of sex dolls, is developing its own large language model to enhance its products with artificial intelligence. These advanced sex dolls, available in both male and female forms, are expected to hit the market soon.

“We are creating a next-generation sex doll capable of vocal and physical interaction with users, with prototypes expected by August this year,” said CEO Evan Lee earlier this month.

Lee acknowledged the technological hurdles, particularly in achieving realistic human interaction. “While simple dialogue is straightforward, developing interactive responses requires complex model development by specialized software companies,” he explained.

Traditional sex dolls, supported by a metal skeleton and silicone exterior, are limited to basic responses and lack the expressiveness needed for meaningful engagement with humans.

“The new generation of sex dolls, powered by AI models and equipped with sensors, can react with both movements and speech, significantly enhancing the user experience by focusing on emotional connection rather than just basic conversational abilities,” Lee said. The company, which previously focused on markets outside China, is now also targeting the domestic market.

Despite China’s generally conservative society and reluctance to discuss such topics, Lee noted that China has the largest market for sex dolls, surpassing the combined sales of the US, Japan, and Germany. “Industry insiders know that China has a huge market, with purchasing power in major cities surpassing many European countries. The market is also open-minded, though aesthetically different from the European market,” he said.

Starpery’s roadmap includes developing robots capable of performing household chores, assisting people with disabilities, and providing aged care. By 2025, the company aims to launch its first “smart service robot” capable of more complex services for people with disabilities. By 2030, these robots could be protecting people from hazardous jobs, according to the company’s plan.

To reach this level of development, there are two main challenges: battery capacity and artificial muscles, Lee said. Unlike electric vehicles, humanoid robots lack space for large batteries, so to operate independently, the energy density of batteries must improve. Additionally, current engines lack the flexibility of human muscles, which can exert force over a wide range and can be both hard and soft, closely fitting the skin.

Currently, to ensure realism, the dolls can weigh up to 40kg (88lbs), which is too heavy for the motor and poses a risk of falling or injuring the user. “Therefore, in the first stage, we focused on reducing the weight through improvements in materials and production processes,” Lee said. By July 2023, their 172cm-tall doll weighed just 29kg.

Lee stated that robots capable of performing household chores represent a societal vision but are still a long way off. While robotics companies can use servo motors to achieve certain functions, considering stability and cost, commercial viability remains distant. “The entire industry will need about 10 years to achieve this goal,” he said.

In addition to technical difficulties, Starpery also faces cost and ethical challenges. Reducers, which transfer power between the motor and the robot joints, are key components in humanoid robots' mechanical systems. They generally account for 30 percent of the robot’s cost and might require multiple gears at different joints.

“We strive to decrease the cost so that more people can afford realistic dolls, while adding motors will increase the cost to some extent,” Lee said.

TheOpen.AI News

Apple Launches 20 Core ML Models on Hugging Face to Boost Open-Source AI

Apple has taken a major step forward in empowering developers with advanced on-device AI capabilities. The tech giant has recently released 20 new Core ML models and 4 datasets on Hugging Face, a prominent platform for sharing AI models and code. This initiative highlights Apple’s dedication to advancing AI technology while emphasizing user privacy and efficiency.

Clement Delangue, co-founder and CEO of Hugging Face, underscored the importance of this update in a statement to VentureBeat. “This is a significant update, as Apple has uploaded numerous models to their Hugging Face repository using their Core ML framework,” Delangue said. “The update features exciting new models focused on text and images, such as image classification or depth segmentation. Imagine an app that can effortlessly remove unwanted backgrounds from photos or instantly identify objects in front of you and provide their names in a foreign language.”

The new Core ML models cover a variety of applications, including FastViT for image classification, DepthAnything for monocular depth estimation, and DETR for semantic segmentation. These models have been optimized to operate solely on users’ devices, eliminating the need for a network connection. This approach not only boosts app performance but also ensures that user data remains secure and private.

Delangue highlighted the significance of on-device AI, stating, “Core ML models run strictly on the user’s device, removing any need for a network connection. This keeps your app lightning-fast and ensures user data remains private.”

The release of these models and datasets on Hugging Face demonstrates Apple’s strengthening partnership with the AI community platform. Recently, Apple has been actively collaborating with Hugging Face on various initiatives, such as the MLX Community and the integration of open-source AI into Apple Intelligence features.

Industry experts believe that Apple’s focus on on-device AI aligns with the broader trend of shifting computational power from the cloud to edge devices. By leveraging the capabilities of Apple Silicon and minimizing memory footprint and power consumption, Core ML enables developers to create intelligent apps that deliver seamless user experiences without compromising privacy or performance.

As the demand for privacy-preserving and efficient AI solutions grows, Apple’s latest move is expected to empower developers to build innovative applications across various domains, from image and video processing to natural language understanding and beyond. With the availability of these new Core ML models and datasets on Hugging Face, the AI community can collaborate, iterate, and push the boundaries of what’s possible with on-device AI.

Apple’s dedication to advancing AI while prioritizing user privacy sets a strong example for the industry. As more tech giants recognize the importance of on-device AI, we are likely to see a surge in the development of intelligent, privacy-focused applications that harness the power of local, specialized models to deliver transformative user experiences.

Technologists and Policymakers Clash Over Open-Source AI

The open-source software movement has long had substantial backing in Silicon Valley. However, the debate over whether advanced artificial-intelligence models should be open-sourced is now gaining serious attention, extending from Silicon Valley to Washington, D.C., and beyond.

On one side of the debate is prominent Silicon Valley venture capitalist Vinod Khosla, who views the U.S. as being in a “war” with China over the future global landscape. According to Khosla, the nation that develops the leading AI models will dominate the global economy and shape social, economic, and political outcomes worldwide. He argued at the Bloomberg Technology Summit in San Francisco last month that allowing the U.S.'s top AI models to be freely available to China poses a national-security risk.

“It’s a national-security hazard to open source,” Khosla stated.

Conversely, figures like Marc Andreessen, another renowned tech venture capitalist, advocate for open-source access to AI. Speaking at Stanford's fifth anniversary event for the Institute for Human-Centered Artificial Intelligence on June 5, Andreessen warned that restricting open-source AI would lead to a monopoly of large corporations over the technology and hinder academic research.

Andreessen also questioned the feasibility of such restrictions, suggesting that preventing the spread of open-source AI might require extreme measures such as global surveillance, bombing rogue data centers, or even nuclear threats, as argued by AI researcher Eliezer Yudkowsky.

“You have to ask yourself what kind of society would you need to design that would have the enforcement mechanisms to enforce an open-source ban?” Andreessen said. “Now you start to get into [George] Orwell territory.”

Given the high stakes and the excitement surrounding AI, it's not surprising that the debate over open-sourcing AI extends beyond tech conferences. The issue has prompted numerous research papers, white papers, and opinion pieces in policy journals, and public policymakers are beginning to address it as well.

A bill authored by state Sen. Scott Wiener, D-San Francisco, which passed the state Senate last month, aims to impose new, potentially stringent restrictions on open-source AI development. Additionally, in February, the National Telecommunications and Information Administration, a branch of the U.S. Department of Commerce, sought public comments on the risks and benefits of open-sourcing AI to inform policy recommendations for President Joe Biden.

TheClosed.AI News

Color Health Partners with OpenAI to Integrate GPT-4o in New Copilot Application

The cancer screening company Color Health has announced a collaboration with OpenAI to utilize the Microsoft-backed company's latest GPT-4o model in its new copilot application, which aims to create customized cancer treatment plans and identify gaps in patient care. In a video on OpenAI's website, the capabilities of the copilot were demonstrated, showing how it can extrapolate patient data and identify any gaps in their testing.

"With the expert guidance from OpenAI, its powerful models, and HIPAA-compliant data protection standards, Color was able to focus on simplifying complex medical decision-making, refining prompts, and designing clinician-in-the-loop workflows to develop the initial version of the copilot," OpenAI stated.

The company also highlighted that with Color Health's copilot, healthcare providers could identify more gaps in patient documents and analyze their data in approximately five minutes, a process that previously took weeks.

"Throughout the second half of 2024, Color plans to use the copilot application to offer AI-generated personalized care plans, overseen by physicians, for over 200,000 patients," OpenAI added.

While various studies have shown the potential of generative AI in analyzing medical documents and scans to alleviate doctors' workload, human expert supervision remains essential as the technology is still experimental.

Elon Musk Raises Concerns Over OpenAI Phone Access After Ex-NSA Chief Joins

Elon Musk, CEO of Tesla and SpaceX, has raised concerns about OpenAI potentially accessing his phone following the addition of former US Cyber Command chief Paul Nakasone to OpenAI's board of directors.

Musk's apprehensions, shared in a post on X (formerly Twitter), come amidst his growing criticism of OpenAI, a company he co-founded but later left. He has particularly criticized Apple's integration of OpenAI's ChatGPT into iPhones, even threatening to ban iPhones from his companies due to security concerns.

Nakasone's appointment to OpenAI's board, and his role on a new safety and security committee, have increased Musk's worries. Nakasone has pledged to protect OpenAI from "increasingly sophisticated bad actors."

Musk is concerned that integrating OpenAI's technology at the operating system level could compromise user privacy and security. He has even suggested that visitors to his companies might need to store their Apple devices in Faraday cages upon entry.

Additionally, reports indicate that Musk is working with Samsung to develop an X phone, potentially offering a more secure alternative to iPhones.

Microsoft, OpenAI, and NVIDIA Collaborate with US Agencies on AI Cybersecurity Preparedness

Officials from the US Federal government convened with several tech companies, including leading AI model operators, cybersecurity firms, and AI hardware manufacturers, to initiate the first joint simulation of a cyberattack on critical AI systems.

This exercise was crucial because responding to cyber threats targeting AI technologies requires different strategies compared to traditional hacks.

Both Washington and Silicon Valley are collaborating to stay ahead of the unique cyber threats facing AI companies. Historically, security measures have often lagged as new technologies become mainstream, leaving many companies vulnerable to evolving cyber threats, as noted by Axios. Clayton Romans, associate director of the Cybersecurity and Infrastructure Security Agency’s Joint Cyber Defense Collaborative (JCDC), emphasized that as AI tools become more widespread, hackers could exploit them to accelerate and scale their attacks.

The JCDC hosted the tabletop exercise at Microsoft’s offices in Reston, Virginia. As is customary with such simulations, CISA did not publicly disclose the specific incident that participants simulated.

Romans explained that the exercise focused on current threats and how the government and private sector can share information about those threats.

Over 50 AI experts from the US government, international government offices, and private companies participated in the four-hour session, including representatives from Amazon Web Services, Microsoft, Nvidia, OpenAI, and Palantir. Kyle Wilhoit, director of threat research at Palo Alto Networks’ Unit 42, participated and noted that the exercise provided an opportunity to discuss current threats and speculate on potential future attack vectors leveraging AI.

The exercise also helped CISA identify key contacts in the private sector for AI-related incidents and vice versa, according to Romans. This mutual understanding is crucial for effective communication and response during AI-related cyber incidents.

Participants explored potential new threats on the horizon. Insights gained from the exercise will contribute to the development of CISA’s forthcoming AI security incident playbook, expected to be published by the end of the year. Romans mentioned that the JCDC plans to host another AI tabletop exercise before releasing the playbook.

Looking ahead, the lessons learned from these exercises will be vital in shaping strategies and protocols for handling AI-related cyber threats.

By proactively addressing these issues, federal officials and industry leaders aim to strengthen the security and resilience of AI systems, ensuring they can withstand and quickly recover from potential cyberattacks.

In our quest to explore the dynamic and rapidly evolving field of Artificial Intelligence, this newsletter is your go-to source for the latest developments, breakthroughs, and discussions on Generative AI. Each edition brings you the most compelling news and insights from the forefront of Generative AI (GenAI), featuring cutting-edge research, transformative technologies, and the pioneering work of industry leaders.

Highlights from GenAI, OpenAI and ClosedAI: Dive into the latest projects and innovations from the leading organisations behind some of the most advanced AI models in open-source, closed-sourced AI.

Stay Informed and Engaged: Whether you're a researcher, developer, entrepreneur, or enthusiast, "Towards AGI" aims to keep you informed and inspired. From technical deep-dives to ethical debates, our newsletter addresses the multifaceted aspects of AI development and its implications on society and industry.

Join us on this exciting journey as we navigate the complex landscape of artificial intelligence, moving steadily towards the realisation of AGI. Stay tuned for exclusive interviews, expert opinions, and much more!

Keep reading