Microsoft Unveils the 53 Family: Compact, Efficient, and Versatile Language Models
In a groundbreaking move, Microsoft has unveiled the 53 family, a collection of open small language models (SLMs) that are designed to be highly efficient and affordable. Developed through innovative training methods, these models surpass their larger counterparts in tasks such as language, coding, and math. Sonali Do, the principal product manager for generative AI at Microsoft, envisions a future where customers can choose from a range of models tailored to their specific needs.
The first member of the 53 family, the 53 mini with 3.8 billion parameters, is now accessible through various platforms like the Azure AI Model Catalog and Hugging Face. Despite its compact size, the 53 mini outperforms models twice its size. Additional models like the 53 small (7B parameters) and 53 medium (14B parameters) are on the horizon, offering a versatile range of options for diverse customer requirements.
The key advantage of these SLMs lies in their ability to be deployed on-device, enabling quick AI experiences without the need for internet connectivity. This feature caters to a wide range of applications, from smart sensors to farming equipment, while also promising privacy by keeping data on the device. While large language models excel at complex tasks, these SLMs provide a viable option for simpler tasks like query answering and summarization.
Microsoft's approach to the 53 family emphasizes curated data and specialized training, reducing computational costs while maintaining performance and reasoning abilities. Inspired by bedtime story books, the team has meticulously curated synthetic data sets like "Tiny Stories" and "Code Textbook," focusing on quality over quantity to enhance model performance.
DHS Launches an AI Safety and Security Board, Excluding Tech Giants
In a move to ensure the safe and secure use of AI across critical US infrastructure, the Department of Homeland Security (DHS) has unveiled its new Artificial Intelligence Safety and Security Board. This board aims to guide stakeholders, such as pipeline operators and internet service providers, on the responsible use of AI.
Secretary Alejandro Mayorkas emphasized the dual nature of AI, acknowledging its potential benefits while also highlighting the risks it poses. The board's role is to help the DHS anticipate and counter threats from hostile nation-states, particularly through AI-assisted cyber attacks.
The composition of the board raises questions, as it includes influential figures from Big Tech but notably excludes the likes of Mark Zuckerberg and Elon Musk. This decision by the DHS has sparked speculation about the government's stance on open-source AI models, as the ongoing debate continues over the influence of social media companies and their approach to AI development.
While some companies advocate for open-source AI, others like Meta and Anthropic have released their models as open weights, challenging the dominance of closed models. The DHS's intentional exclusion of social media companies from the board suggests a potential shift in the government's perspective on the role of these platforms in AI development and deployment.
Synthesia Introduces Emotionally Expressive AI Video Avatars
Synthesia, a company specializing in AI-powered video creation, has taken a significant step forward by introducing avatars with enhanced emotional expression, precise lip syncing, and more lifelike movements. Unlike competitors like OpenAI, which target both consumers and businesses, Synthesia's focus is solely on perfecting humanlike avatars for enterprise applications such as training and marketing.
Synthesia's new expressive avatars, generated entirely by AI, utilize large pre-trained models to mimic human speech patterns and gestures, providing a more authentic experience. Unlike traditional avatar-based video tools that stitch together pre-recorded clips, Synthesia's avatars generate responses dynamically, aiming for a less robotic and more natural appearance.
The platform's widespread adoption, with over 200,000 users across 130 languages creating more than 18 million videos using legacy avatars, highlights the growing demand for AI-powered video solutions in the enterprise sector. Synthesia's strategy of refining its avatar technology, rather than targeting a broader consumer base, has helped the company carve a niche in the crowded AI market, where sustained value matters more than initial hype.
Nvidia Acquires Run AI for $700 Million to Bolster Its AI Infrastructure Offerings
Nvidia, known for its powerful AI hardware, has made a strategic move by acquiring Run AI, a Tel Aviv-based company that simplifies the management and optimization of AI hardware infrastructure for developers and operations teams. While the deal specifics remain undisclosed, insider sources indicate a valuation of $700 million.
Nvidia plans to maintain Run AI's product offerings and enhance its roadmap, with the integration aimed at bolstering its DGX Cloud AI platform. This acquisition empowers Nvidia's enterprise customers with improved compute infrastructure and software for AI model training, particularly in the realm of generative AI scenarios across multiple data center sites.
Run AI's co-founders, Geller and Dar, conceptualized the platform during their studies at Tel Aviv University, with the goal of creating a solution capable of distributing AI models across various hardware setups, whether on-premises, in public clouds, or at the edge. Despite facing limited direct competition, the concept of dynamic hardware allocation for AI workloads is gaining traction, with platforms like Grid.ai offering similar software for parallel AI model training.
This acquisition marks one of Nvidia's most significant moves since its $6.9 billion acquisition of Mellanox in 2019, signaling the company's commitment to strengthening its position in the AI infrastructure market.
OpenAI Faces Complaint Over Fictional Outputs, Raising GDPR Compliance Concerns
The European advocacy group Noyb has lodged a complaint against OpenAI regarding inaccuracies in the information generated by ChatGPT. They claim that these inaccuracies breach the EU's General Data Protection Regulation (GDPR), which states that personal data should be accurate and rectifiable.
Mara Degraf, a data protection lawyer at Noyb, emphasizes that false information, especially about individuals, can have serious consequences. OpenAI has acknowledged its inability to correct ChatGPT's inaccuracies or reveal its data sources due to ongoing research challenges.
Noyb cites instances where a public figure received incorrect birth dates from ChatGPT to highlight the issue. Despite requests, OpenAI refused to rectify or erase the data, citing technical limitations. The company claims it can filter certain data, but not without hindering ChatGPT's functionality. OpenAI also failed to adequately respond to access requests, breaching GDPR requirements.
Noyb urges the Austrian data protection authority to investigate OpenAI's data practices and enforce GDPR compliance, including the rectification of inaccuracies and the imposition of fines. This complaint reflects the growing concerns around the accuracy and accountability of AI-generated content, particularly in the context of personal data and privacy regulations.
Sanctuary's New Humanoid Robot: Faster Learning, Lower Costs, and Improved Capabilities
Sanctuary AI, a Canadian company, has unveiled the seventh version of its Phoenix line of humanoid robots, showcasing advancements in their upper-body design and learning capabilities. Unlike previous models that focused on incorporating legs, the latest iteration emphasizes the robot's agility and ability to handle tasks like sorting products.
Experts predict that it will take another 5 to 10 years before robots can learn tasks like humans, but Sanctuary's Phoenix robots can quickly adapt to new tasks, sometimes in less than 24 hours. CEO Jordy Rose sees this advancement as a significant step towards artificial general intelligence (AGI).
The latest 7th generation Phoenix robot, introduced a year after its predecessor, brings further improvements such as increased uptime, better motion range, lighter weight, and reduced costs. While the effectiveness of these robots varies depending on the task, Sanctuary has already deployed earlier versions and secured deals with companies like Magna for their use in manufacturing.
The company's focus on enhancing the intelligence and adaptability of its humanoid robots, rather than solely emphasizing their appearance and mechanical abilities, aligns with the industry's shift towards developing more capable and versatile robotic systems.
UK Competition Regulator Scrutinizes Microsoft and Amazon's AI Partnerships
The Competition and Markets Authority (CMA) in the UK is investigating whether partnerships and hiring practices involving Microsoft, Amazon, and three AI startups (Anthropic, Inflection, and Mila) follow merger rules and could affect competition in the UK market.
There are concerns that close collaborations between major tech players like Microsoft and Amazon might prevent new companies from challenging them, even though outright acquisitions would be closely watched by regulators. Partnerships and investments could potentially avoid such scrutiny.
The CMA is examining Microsoft's ties with OpenAI, including its recent acquisition of the team behind Inflection AI. Additionally, Microsoft has launched a new AI Hub in London, led by a former Inflection and DeepMind scientist. Amazon, on the other hand, has recently invested $4 billion in Anthropic, another AI company.
Both Microsoft and Amazon are cooperating with the CMA's inquiries, with Microsoft pledging full assistance. Amazon sees the CMA scrutiny as unusual, especially since its partnership with Anthropic does not grant it significant control over the company.
The CMA's current phase involves seeking comments until May 9th, followed by a potential phase one review engaging Microsoft and Amazon directly. This review must conclude within 40 days, after which the CMA will determine if the partnerships constitute a relevant merger.
This scrutiny reflects the growing concerns about the influence of large tech companies in the AI landscape and the potential impact on competition and innovation in the industry.
.jpg)
0 Comments