JLL’s Post

https://lnkd.in/grqnb4Ue A follow up to yesterday's post, Dotmatics CEO Thomas Swalla talks about #digitaltransformation in the life and materials sciences

In our latest blog, Dr. Grant Campbell, PhD from the University of Aberdeen shares insights from a meeting at Rothamsted Research, where researchers from various fields came together to discuss the future of digital twins. Grant describes digital twins as a “representation of a system or environment in a virtual context using real-time information or data.” At #AI4SoilHealth, we’re creating a digital twin of Europe’s soils, to support EU policymakers in making accurate and effective soil management decisions. One key takeaway from the meeting is the common challenges around working with AI and digital technologies, alongside unique issues faced by different groups. Grant highlights some key considerations for developing digital twins, such as engaging stakeholders early in the design process, adopting an interdisciplinary approach, thoroughly understanding and evaluating data, and creating user-friendly interfaces for better accessibility. 🔗 Read the full blog to learn more: https://lnkd.in/eHCM6yQA

Who better to tackle the growing divide between science and society than those who communicate complex ideas every day? In this #PageTurner blog by Lori Teranishi of iQ 360® she explores how in a time when digital platforms amplify falsehoods, the gap between science and society is growing wider. Communicators have the crucial role of translating those sophisticated scientific concepts into accessible and understandable insights for the public. Discover how we can all play a part in protecting the integrity of scientific knowledge here: https://brnw.ch/21wLHSn

National Academies of Science--Webinar Series: Navigating the Era of Artificial Intelligence Part 1: Achieving human-AI harmony

Exciting news! I am a #Speaker at DSC DACH 24 happening this September 11-13 in Vienna, Austria. My keynote talk is titled “Data Engineering for Sustainable AI: Optimizing Energy Usage and Real-World Impact.” Join me as we explore how AI and data engineering can transform our approach to global challenges, particularly in optimizing energy usage and promoting sustainability. Link in the comment section below! ⬇️

🌟 Generative Agent Simulations: A Step Toward Human Behavioral Modeling 🌟 How close are we to creating computational agents that simulate human behavior accurately across domains? A new paper, "Generative Agent Simulations of 1,000 People," takes us a step closer to answering this question—and the results are fascinating. 📖 The Paper: https://lnkd.in/dyBMGjWc This work presents an innovative agent architecture capable of simulating the behaviors and attitudes of 1,052 real individuals. By combining large language models with qualitative interviews, these agents replicate participants' responses to the General Social Survey with 85% accuracy, rivaling how consistently participants replicate their own answers over time. 🔑 Key Contributions: The architecture reduces accuracy biases across racial and ideological groups. It predicts personality traits and outcomes in experimental settings, demonstrating potential applications in policymaking and social science research. It lays the groundwork for tools to better understand individual and collective behavior. This paper highlights the immense potential of generative agents in modeling human behavior—but it also raises important questions about ethical design and deployment. Let’s discuss: 🤔 How could tools like these reshape our understanding of social dynamics? 💡 What safeguards should we implement to ensure their responsible use? Join the conversation in the comments! 👇 #AIResearch #GenerativeAgents #BehavioralModeling #ArtificialIntelligence #SocialScience

LoRA vs Full-finetuning. Both perform similarly on eval tasks. But are they equivalent? Researchers at the Massachusetts Institute of Technology published "LoRA vs Full Fine-tuning: An Illusion of Equivalence". They found that the singular value decomposition (SVD) of weight metrics has a different structure for LoRA and the Full-finetuned model. Moreover, they both have different scales of generalization behavior. The weight metrics of LoRA models have an extra dimension (called intruder dimension) in their singular vectors. Due to this, it has a shift in corresponding singular vectors (measured with cosine similarity) to the pre-trained model. Thus, even though both (LoRA and full-finetuning) have similar eval performance, the LoRA model becomes a worse model of the pre-training distribution and adapts less robustly to multiple tasks sequentially. Whereas, the full-finetuned model retains pre-training distribution and adapts much more robustly to new tasks. As the rank in LoRA increases, it starts aligning more with the full-finetuned model slowly removing the intruder dimension. #MachineLearning #GenAI #LLM #LoRA #DataScience

LoRA vs Full-finetuning : They are different due to pre trained weighted matrix learned during fine-tuning. Full Fine-Tuning: The entire pre-trained weight matrix is adjusted, directly learning new values for each parameter. LoRA: The pre-trained weight matrix remains static, and task-specific knowledge is encoded into low-rank matrices (AAA and BBB).

After a year of research, we have reached the same conclusions as the prominent Yann LeCun: the transformer architecture and infrastructure are insufficient to meet the long-term expectations of the AI industry. CEOs like Sam Altman should be more thoughtful in guiding the industry and its investors towards speculative hype. Nevertheless, We believe that the AI industry will deliver its incredible benefits in due course, certainly not next year!