Oxford University Press’ Post

It's ever so simple - a decent future arises if people are moral, rational and wisdom pervades. These are all characteristics of the human mind. The problem with data is that it's infinite in scope. We can't possibly store it all - and we can't possibly identify the dataset that's worthy of storing until it's too late and there's no further point in storing data. In simple terms - the point of data is to help the mind achieve enlightenment (wisdom is a simpler word to understand) - and arises upon deep Morality acquisition. The core problem with data ('all science is physics or stamp-collecting') is that it contravenes the principle of Science which is to accept/reject one of two plausible hypotheses on an outcome. If a thought experiment identifies no alternative hypothesis - then there's no point in performing an experiment or storing the data on it. The experiment itself is moot if there's only 1 imaginable path which could give rise to it. Everything falls into place as soon as somebody asks the question Why? does obesity strongly predispose to cancer. And then Why? are we set to have ~1.5 billion T2Diabetics on the planet in our life-time. An epic fail on the part of the extremely well funded global Biomedical complex.

Algorithmic Sovereignty is the most viewed item of the University of Plymouth digital library: this is my doctoral dissertation, published as Creative Commons of course, six years ago. It is still worth revisiting for insights on the generative AI discourse on #algorethics, which is lead by Paolo Benanti, today.

Key points of "From Google Gemini to OpenAI Q* (Q-Star): A Survey of Reshaping the Generative Artificial Intelligence (AI) Research Landscape": **1. Transformative technologies:** The paper focuses on three advancements in generative AI: Mixture of Experts (MoE), multimodal learning, and potential strides towards Artificial General Intelligence (AGI). It explores how these innovations, exemplified by models like Google's Gemini and the anticipated OpenAI Q*, are reshaping research priorities and applications across various fields. **2. Impact on research taxonomy:** The research analyzes the impact of these technologies on how generative AI research is categorized and understood. It highlights how MoE, multimodal learning, and AGI advancements necessitate revisions to the existing research taxonomy to fully capture the evolving landscape. **3. Real-world implications:** Beyond theoretical analysis, the paper delves into the practical implications of these advancements across various domains. It examines how MoE can optimize resource allocation in AI systems, how multimodal learning can enhance robot perception and interaction, and how AGI, if achieved, could revolutionize countless fields. **4. Computational challenges and scalability:** The research acknowledges the significant computational challenges associated with these cutting-edge technologies. It discusses strategies for addressing these challenges to ensure scalability and real-world practicality of the new models and algorithms. **5. Ethical considerations:** Recognizing the potential ethical concerns surrounding powerful generative AI, the paper emphasizes the importance of incorporating ethical and human-centric approaches into AI development. It advocates for aligning AI advancements with societal norms and prioritizing well-being. **6. Future research directions:** The paper concludes by outlining future research directions to harness the potential of MoE, multimodal learning, and AGI advancements responsibly. It emphasizes the need for balanced and conscientious use of these technologies to ensure positive societal impact. **Additionally, the paper:** * Discusses the growing field of AI-generated preprints and the challenges it poses for peer-review and academic communication. * Examines the impact of these advancements on the generative AI research community and identifies emerging areas of investigation.

The Mixture of Experts (MoE) architecture holds immense potential for optimizing resource allocation in robots and revolutionizing their efficiency in several ways: **1. Modular Expertise and Dynamic Task Allocation:** * MoE divides a complex task into subtasks and assigns each subtask to a specialized "expert" model within the overall system. This allows robots to dynamically allocate resources based on the immediate task demands. If a subtask requires more intensive processing, resources can be temporarily shifted towards the corresponding expert, while less demanding tasks can be handled by smaller experts, saving computational power. **2. Efficient Learning and Adaptation:** * Each expert model within the MoE architecture can be individually trained on specific tasks or data sets, leading to improved learning efficiency and generalization. This avoids the need for training a single, monolithic network for all tasks, which can be resource-intensive and less effective. Additionally, as robots encounter new environments or challenges, specific MoE experts can be updated or swapped for new ones, enabling dynamic adaptation and continuous improvement. **3. Enhanced Scalability and Resource Utilization:** * MoE enables building larger and more complex AI systems while maintaining computational efficiency. By dividing tasks among specialized experts, the overall system can handle a broader range of situations and adapt to diverse environments without requiring a corresponding increase in hardware resources. This opens up possibilities for deploying highly capable robots on platforms with limited computing power, such as drones or mobile robots. **4. Reduced Power Consumption and Improved Battery Life:** * Efficient resource allocation through MoE can significantly reduce the overall power consumption of robots. By directing processing power only towards the relevant experts for each task, robots can minimize unnecessary calculations and extend their battery life, which is crucial for autonomous operations in the real world. **5. Increased Parallelism and Real-time Performance:** * MoE architectures enable parallel processing of subtasks by the individual expert models. This can significantly improve the real-time performance of robots, allowing them to react and adapt to dynamic situations much faster. Overall, MoE presents a revolutionary approach for optimizing resource allocation in robots, leading to: * **Enhanced efficiency:** Robots can perform complex tasks with minimal power consumption and computational resources. * **Improved adaptability:** Robots can dynamically adapt to different tasks and environments. The future of robot intelligence is closely intertwined with advancements in MoE architectures. By utilizing this modular and dynamic approach, we can unlock a new generation of robots that are efficient, adaptable, and capable of operating in increasingly complex real-world scenarios.

I am pleased to share an insightful blog post on the complementary contributions of academia and industry to AI research. This post delves into the recent advances made by industry and the impact of academic research in the field over the last 25 years, establishing intriguing patterns. It highlights that while industry teams tend to produce state-of-the-art models with high citation rates, academic teams contribute significantly to the novelty and unconventionality of AI research. The findings underscore the irreplaceable contributions of both sectors to the progress of AI. Read the full post at https://bit.ly/47ZsMHf [cs.CY].

I've noticed that the citation behavior of some (many) researchers is gradually becoming questionable. Apart from the usual suspects (such as excessive self-citation, citation cartels, impact factor and h-index manipulation, salami slicing and so forth), one road that is less traveled and not gathering sufficient scholastic discussion is the issue of "out of context citation." In the picture below, some Asian authors cited my work "out-of-context." Although the citation has increased my cumulating metric, the context of that citation does not fit what we wrote in that paper last year. One possible reason is the pervasive use of AI among scholars. This may form a new line of research for me.

"Eventually, [machines] could be our natural evolution. But to play safe let’s start mimicking first human intelligence and giving room to less biological and unsupervised processes within controlled frameworks. By doing so we would crucially ease the convergence between the regulators and the industry needs." Read more about it in our new paper: Advances in AI: When Applied Science is not Science Applied https://lnkd.in/dtMHQWjF

A Good initiative to acquire more knowledge and practical skills as a Librarian

In this era of AI, reviewers must execute their roles with precision and care.

White Paper - "Generative AI and Disinformation: Recent Advances, Challenges, and Opportunities" "According to the authors, the goal of the publication is to ‘deepen the understanding of disinformation-generation capabilities of state-of-the-art Artificial Intelligence, as well as the use of AI in the development of new disinformation detection technologies’. It includes pointing to associated ethical and legal challenges. Ultimately, challenges and opportunities brought about by generative AI in the context of disinformation production, spread, detection, and debunking are assessed and portrayed." It is a great and comprehensive read.