•
April 25, 2024
•
4 mins
Hello Niuralogists!
Step into this week's edition as we navigate the dynamic realm of artificial intelligence to present you with the most recent breakthroughs. Our primary focus is to dissect the implications of these updates on various aspects of our lives, ranging from workplaces and businesses to policies and individual experiences. In this issue, we'll unveil compelling advancements, featuring highlights such as Microsoft's introduction of Phi-3: a formidable language model, and an exploration of visual memory through the mapping of brain pathways.
For deeper insights, continue reading…
Microsoft has just unveiled Phi-3, a formidable new lineup of compact language models that surpass larger competitors, setting new standards in the realm of smaller AI models. The Phi-3 series comprises three variants: Phi-3-mini boasting 3.8 billion parameters, Phi-3-small with 7 billion parameters, and Phi-3-medium with 14 billion parameters. Despite its diminutive size, Phi-3-mini matches the performance of Mixtral and GPT 3.5, boasting an impressive 128k context window. Moreover, its modest parameter count allows for on-device deployment without compromising quality or cost efficiency. The 7-billion-parameter Phi-3-small and 14-billion-parameter Phi-3-medium are currently undergoing training and are slated for release in the coming weeks. Microsoft's breakthrough underscores the potential of small language models, demonstrating that with appropriate training data and methodologies, they can deliver significant impact. Particularly noteworthy is Phi-3-mini's capabilities, heralding a major advancement and paving the way for high-performing models to operate efficiently on mobile devices.
Researchers at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) have used a pioneering combination of magnetoencephalography (MEG) and functional magnetic resonance imaging (fMRI) to map how the brain processes memorable visual images. Their study, published in PLOS Biology, reveals that highly memorable images elicit stronger and more sustained brain responses across key regions involved in visual processing and memory encoding. This breakthrough challenges previous assumptions about memory formation and offers potential insights for diagnosing and treating memory-related disorders like Alzheimer's disease. Led by CSAIL Senior Research Scientist Aude Oliva, the team's innovative approach promises to reshape our understanding of cognitive processes and opens new avenues for neuroscience research.
Adobe has unveiled the latest iteration of its Firefly image generation model, introducing enhancements in image quality, stylization features, processing speed, and fine details, along withthe integration of new AI functionalities. Dubbed Firefly Image 3, the upgrade promises advancements in photorealism, text rendering, prompt comprehension, and illustration capabilities. Notable additions include Structure and Style Reference tools for precise control, and updates to Photoshop tools such as Generative Fill, Generate Image, Generate Similar, Generate Background, and Enhance Detail. Adobe highlights the model's training on licensed content, with Firefly images automatically receiving AI metadata tags. This upgrade positions Firefly as a strong contender in the image generation landscape, while Adobe's integration of powerful AI into its flagship platforms signifies a strategic shift towards redefining creativity as an enhanced tool rather than a mere replacement.
OpenAI strides ahead in the AI race, unfazed by Meta's Llama 3's surge, as it introduces new enterprise AI features. With a focus on bolstering security and administrative control, OpenAI expands its enterprise-grade offerings for API customers, enhancing its Assistants API with improved file retrieval capabilities and real-time streaming for conversational responses. Additionally, OpenAI introduces Projects, enabling organizations to manage roles and API keys at the project level for better oversight and resource allocation. Cost management features like discounted rates and asynchronous workloads through the Batch API further aid organizations in scaling their AI operations economically. As OpenAI continues to prioritize enterprise-grade security and cost-effective solutions, it aims to provide a more accessible experience for businesses, countering the competition from open models like Llama 3 and Mistral.
Meta has unveiled Llama 3, the latest iteration of its open-source LLM, boasting 8B and 70B parameter versions that surpass competitors in various evaluation benchmarks. Tests conducted by Meta demonstrate the superiority of Llama 3 8B and 70B over similarly-sized models like Google's Gemma, Mistral 7B, and Anthropic's Claude 3 Sonnet. These models were trained on a dataset seven times larger than its predecessor, Llama 2, comprising 15 trillion tokens and four times more code. The release of the largest 400B+ parameter version is anticipated in the "coming months," poised to rival GPT-4. Llama 3 is seamlessly integrated across Meta's platforms, including the Meta AI assistant available on Facebook, Instagram, WhatsApp, Messenger, and the new meta.ai website. This development underscores Meta's strategic vision, extending top-tier AI capabilities to over three billion users worldwide, thus shaping the future of AI accessibility and integration.
📬 Receive our amazing posts straight to your inbox. Get the latest news, company insights, and Niural updates.
MIT has developed a groundbreaking security solution for power-intensive AI models, ensuring protection against common attacks while maintaining efficient computing on smartphones. Addressing the challenge of slow and energy-inefficient health-monitoring apps, which rely on large machine-learning models shuttled between smartphones and central memory servers, engineers have introduced a machine-learning accelerator resistant to prevalent attacks. This innovation enables the safeguarding of sensitive user data, such as health records and financial information, while facilitating the efficient operation of massive AI models on devices. Through meticulous optimizations, the chip offers robust security without compromising computational accuracy, promising significant benefits for demanding AI applications like augmented reality and autonomous driving.
MIT researchers have devised a method to model the behavior of both human and AI agents, considering their computational limitations, to build more effective AI systems. This technique, aimed at predicting the actions of agents with suboptimal decision-making, enables the anticipation of future behaviors based on their inference budgets, which represent their problem-solving constraints. Demonstrating its utility, the researchers applied this approach to infer navigation goals from past routes and predict chess players' moves, showcasing its efficacy in decision-making modeling. By understanding and inferring human behavior, AI systems can better collaborate with their human counterparts, offering improved assistance and adaptation to human weaknesses.
🤖 Open Agent Studio is a powerful no-code agent editor
💬 DuckDuckGO AI Chat is a private AI chat, with no AI training on your conversations
🌿 Healax is a mental health solution for students
🗺️ Journey AI converts customer research into journey maps
🎨 Pietra Product Design Studio designs a best-selling product using AI
