Your cart is empty.
A customer perspective on the reasons, corporate path, initial outcomes and future goals of an initial implementation of an AI Agent. Why was this use case chosen, what were the unknown challenges and unexpected benefits during development, and what is the future of it as part of a corporate Digital Transformation.
Artificial intelligence (AI) is reshaping patterns of scientific workflows across chemical sciences. In materials discovery, one of the most impactful dimensions is rapid progress of AI towards agency. We associate an AI system agency with the capability to access historical data and conceptual knowledge, identify a discovery goal, infer an actionable plan to achieve the goal, and execute the plan using a variety of tools. Is there a common denominator for the agency of subject matter experts (SMEs) and AI in chemical research? In this talk, we will discuss how the patterns of interaction between AI and SME agents have been evolving following explosive growth in adoption of large-language models (LLMs). We will formulate technical desiderata for peer-like interactions between human and AI scientists focusing on specifics of end-to-end materials discovery. Our talk will consider several instances of materials discovery tasks specific to nanotechnology and electronic materials applications.
Moderated by Christopher Nguyen, PhD.
Knowledge graphs are increasingly being used for Retrieval Augmented Generation (RAG), where structured, deterministic retrieval is important. Creating, updating and managing these graphs is an emerging process, as well as how traditional knowledge graph technology can be best used when specific to RAG and agentic systems. This talk will discuss the pros and cons of graph structures in RAG, when and when not to use them, and what patterns are emerging to best ground your retrieval systems.
Generative AI presents exciting opportunities to create new applications and enhance existing ones with advanced capabilities. However, implementing Generative AI also incurs significant costs, including high computational demand, operational complexity, and critical security and safety challenges. In this talk, we will explore strategies for optimizing cost and performance across various layers of Generative AI applications. We will examine methods to reduce compute resource usage and systematically improve the efficiency of these applications. Additionally, we will share insights from our experience in deploying Generative AI at scale, offering practical lessons learned.
The current generation of voice interfaces have failed to gain user adoption. Amazon has invested tens of billions of dollars in the Alexa platform, and people still only use it to set alarms and play music. This talk will explore why speech interfaces haven't worked so far, and how new advances in AI can address some of those issues. I will focus on applications like integrated user manuals for equipment, real time language translation, and other ways this will impact industrial environments.
Organizations have seen a significant spike in capabilities around advanced data engineering, data science and artificial intelligence (AI) abilities thanks to advancements in higher computing software and abundance of open source solutions. At the same time,increased C-level interest and support is providing a focus and budget for these initiatives, seeking production-ready AI solutions that demonstrate real results today. Despite these advanced technical capabilities and strong leadership support, many organizations continue to face challenges navigating data quality issues and realizing the promise of AI and living up to these expectations. Graph solutions have gained momentum due to their wide-ranging applications across multiple industries. We have seen an increased number of implementations and use cases. However, the most common question remains if it is the right solution for data and AIchallenges and when?This presentation will draw upon our own experience from client projects and lessons learned to provide a selection of optimal use cases for knowledge graphs and semantic solutions along with real world examples of their applications for scaled enterprise data architecture and AI strategy.
In 2018, Zhamak Dehghani introduced the concept of the data mesh as a solution to the challenges of data centralization for AI and analytical applications, particularly within large and complex organizations. These challenges include long lead time for turning data into value, high cost associated with managing tangled data pipelines, lack of visibility into the interconnectivity of semantics and lineage across data domains, and ultimately, a loss of trust in data. The demands of next-generation AI applications have only exacerbated these issues. On one hand, there's the growing complexity and proliferation of unstructured data sources. On the other, the widespread application of generative AI across businesses. Both factors put additional strain on companies’ ability to rapidly and reliably deliver AI applications when relying on a centralized data approach. In this talk, Zhamak explains how data mesh offers a decentralized approach to data sharing, focusing on one of its core principles: domain-oriented, computational data products. She will also showcase the application of data mesh and data products in Retrieval-Augmented Generation (RAG) and traditional machine learning (ML) development flows, using containerization technology — the abstraction of the data supply chain — implemented by her team at Nextdata.
Describing the underlying causes of phenomena affected by multiple variables can often by done via the representation of causal graphs, which are often assumed to be directed and acyclic. The identification of causal graphs – delineating the cause and effect between collection of variables has relevance in wireless networks, genetic networks, epidemiology, economics, and the social sciences to name just a few application areas. Graph identification is done via the collection of observations or realizations of the random variables which are the nodes in the graph. A host of strategies have been proposed for causal graph identification from greedy methods to those based on sparse approximation. We consider two problems in graph identification, the first is the recovery of the full directed graph by first detecting individual links in the graph. Unique to our approach, but relevant to many applications is directly considering unequal error protection for edge detection, that is false negatives versus false positives. We derive the optimal link detection rule and bound performance. These bounds can be used as benchmarks for current discovery algorithms. A challenge with our approach (as is also true for many causal graph identification methods) is the attendant computational complexity. Motivated by this issue, the second problem we address is the development of a modest complexity strategy via the learning of sub-graphs. A new edge detection method based on mutual information is derived and shown to offer superior performance to state-of-the-art methods. We apply our low complexity approach to the optimization of interventions in multi-armed bandits wherein the goal is to determine how to select choices (e.g. which slot-machine arm to pull) to maximize a reward. Interventions could include allocation of resources or enforcing nodes in a graph to have certain values or links to have certain weights. We see that unequal error protection has a significant impact in reward optimization in causal multi-armed bandits. Joint work with Joni Shaska and Chen Peng
JSR is a Japanese chemical company that is a leading supplier of semiconductor and display related process materials including photoresist. From the standpoint of the chemical industry, Challenges in applying AI/LLMs to material development and our expectations for the foundation models will be described.
In this presentation, I will introduce a use case for the banking industry that involves utilizing generative AI to analyze and select companies for lending, as well as proposing to the sales team. This use case leverages not only documents but also external data obtained through web information and APIs, integrating them with generative AI to achieve more accurate and advanced analyses. In particular, I will detail the innovative approach that utilizes advanced information retrieval methods, including Planning & Reasoning, to transform the traditional loan screening process.
This presentation will provide an overview of OpenSSA and its activities in Japan, using case studies to show how SSA has been used in various industries in Japan to transform their business operations. In particular, we will discuss case studies from Miura Industries and Furuno Electric to show how OpenSSA is addressing key challenges. Finally, the presentation will focus on the concept of the "last mile" in AI implementation and how OpenSSA enables the transition from data analysis to real-world action. This includes information structuring and accurate and responsible AI automation of the VectorStore. The future of AI use and technological advances will also be discussed.
This keynote explores the transformative impact of cutting-edge artificial intelligence technologies on semiconductor manufacturing, with a focus on advanced process and equipment control. We will examine how generative AI, including multimodal foundation models, retrieval-augmented generation (RAG), and on-device AI solutions, are revolutionizing control systems in semiconductor fabs. The presentation will showcase how these advanced AI approaches enhance yield, quality, and efficiency by enabling real-time optimization, predictive maintenance, and adaptive control strategies. We'll discuss the implementation of multimodal AI for integrating diverse data types in process control and defect detection, the use of RAG systems for context-aware decision-making, and the deployment of on-device AI for responsive equipment-level control. Additionally, we'll address the challenges and strategies for integrating these technologies within the highly regulated semiconductor environment. The keynote will conclude with a vision of AI-driven autonomous fabs and their role in accelerating next-generation chip development. This presentation offers crucial insights into the future of semiconductor manufacturing for industry professionals and researchers in this rapidly evolving field.
In this talk, I will share my perspective on the evolution of AI applications within the industrial sector, tracing the journey from pre-Generative AI days to the present and beyond. By understanding our customers' needs and reflecting on our research efforts, I will highlight current opportunities and challenges in the field. This discussion aims to provide insights into how AI is transforming industries today and what we can expect in the future as technology continues to advance.
In this presentation, we will report our ideas on the utilization of Small Specialist Agent (SSA) for problem-solving to complex issues. Especially in semiconductor manufacturing processes, we will focus on cases where SSA utilization seems to be suitable and discuss the collaboration of SSAs using General Management Agent (GMA) to address the complexity. Additionally, we will discuss the potential for integrating SSA with Semiconductor foundation model (SemiKong) and proprietary models, so that domain-specific agentic AI revolutionizes semiconductor manufacturing.These concepts can be applied to other fields holding complex issues.
We present how SemiKong, the first open-source semiconductor-industry-specific Large Language Model (LLM), and OpenSSA, a neurosymbolic engine that enables problem-solving to address critical challenges in industrial AI, can be combined to create a powerful AI advisor to accelerate key processes in semiconductor manufacturing. Multiple AI Alliance member organizations have joined hands in this work, including Tokyo Electron (contributing semiconductor domain expertise), Aitomatic (contributing AI engineering frameworks and tools), and others.
SemiKong, a groundbreaking open-source foundation model for semiconductors, exemplifies the power of global collaboration in driving AI innovation. This presentation explores how this AI Alliance initiative, led by Aitomatic with domain expertise from Tokyo Electron and AI engineering power from Vietnam's FPT, is reshaping the semiconductor landscape. By leveraging advanced AI technologies, SemiKong addresses real-world challenges in semiconductor manufacturing, demonstrating the transformative potential of international cooperation. We'll examine how Vietnam's involvement through FPT showcases its rising status in the global AI and semiconductor ecosystem, inviting industry leaders to envision a future of unprecedented innovation driven by collaborative efforts across established and emerging tech hubs.
As artificial intelligence (AI) continues to grow and evolve, the demand for higher performance, greater density, energy efficiency, and scalability in computing systems is intensifying. These technological advancements are essential to meet AI’s faster processing, reduced power consumption, and enhanced reliability. This presentation delves into pivotal innovations in semiconductor materials, device designs, and 3D and heterogeneous integrations that are driving today’s AI and shaping its future.
We explain two examples of leveraging Large Language Models (LLMs) for optimizing cold chain equipment maintenance: the construction of a chat application and the automatic generation of rule-based failure diagnosis logic. The chat application addresses questions related to cold chain equipment maintenance, encouraging on-site workers to resolve their questions independently. Plain LLMs lack specialized knowledge in the cold chain domain and may produce hallucinations. To mitigate this issue, the chat application utilizes Retrieval Augmented Generation (RAG) that references manuals and know-how of cold chain equipment maintenance. Maintenance experts in Panasonic conducted a two-stage subjective evaluation of the chat application's response quality. The evaluation achieved a Good rate of 63.6% that is close to the target Good rate of 70%. The automatic generation of rule-based failure diagnosis logic utilizes LLMs to simplify the process of implementing rules described in natural language into program logic. Conventionally, this process requires both equipment knowledge experts to describe the diagnostic rules and programmers to implement the logic. This approach enables to implement programs with just equipment knowledge, reducing labor costs associated with logic implementation. Although diagnostic rules written in natural language often contain logical ambiguities, combining interactive rule correction using LLMs and self-correction of the generated logic by LLMs ensures output of correctly implemented logics. We introduce examples where the quality of logic has been improved by using those approaches.
Panasonic China is focusing on the potential of generative AI in its business sectors with the advancement of large language models (LLMs) across various fields. The focus of this session will be on two key topics: firstly, how LLMs can enhance the efficiency of developers working on new house hold products, even in complex regulatory frameworks. Second, how Panasonic’s expertise in energy conservation can be applied to drive green and low-carbon initiatives in our manufacturing facilities. We’ll talk about how using LLMs can speed up progress in these crucial areas, giving Panasonic China’s business efforts the innovative capabilities of generative AI.
In a year since the last K1st World, we have seen the key forces of OSS AI kick Ito gear. The AI Alliance, hinted at in my talk before the official launch, was established and keeps growing. The Llama 3.1 405B became the first OSS frontier model. But most importantly, the ecosystem of OSS AI has evolved and now includes both the ChatGPT era startups and the foundational technologies enabling LLMs to get real. Graphs underpin one such area. It is key to building the deployable and hirable AI. Knowledge Graphs hold the keys to validation and ground-truthing of LLMs. A vast experience of building knowledge bases, and the amount of data stored in them, is one pillar of production-ready AI. The other is constructing graphs with LLMs and using them to enhance the quality of answers. In this talk, Dr. Alexy Khrabrov, now the first AI Community Architect at Neo4j, a category-defining graph database company, reviews the advances in OSS AI linked to GraphRAG, the new pattern of quality in AI, and outlines the integrations that will drive the next chapter of making AI real.
Deep learning has emerged as a powerful tool for predicting molecular properties and generating molecule candidates, significantly advancing scientific exploration in various fields such as drug discovery and materials science. This progress can be attributed to the successful application of foundation models, which leverage large-scale pre-training methodologies to learn contextualized representations of input tokens through self-supervised learning on extensive unlabeled corpora. The pre-trained foundation models are subsequently fine-tuned for specific downstream tasks. In this presentation, we will introduce the suite of foundation models for chemistry and materials being developed by IBM Research. These models encompass a range of representation types, from SMILES annotations to 3D atomic positions of compounds. We will illustrate how these foundation models can be applied in diverse downstream use cases, showcasing their potential to accelerate scientific discovery.
What's key to the broader adoption of AI in the enterprise? Building a quick AI prototype using a large language model is easy. Making these prototypes accurate at scale is challenging. In this session, Richard Socher, founder and CEO of you.com and AIX Ventures, explores the leap from AI proofs-of-concept to reliable, business-critical systems. The session will also focus on grounding LLMs in factual knowledge, enhancing reasoning capabilities, and implementing robust fact-checking. Learn how to build an AI Operating System focused on accuracy and reliability.
.png)
.png)