Agenda
12:30 – 12:35
Greetings from W&B
Akira Shibata
Country Manager at Weights & Biases
12:35 – 12:55
Productionizing GenAI Models – Lessons from the world’s best AI teams
Lukas Biewald
CEO and Cofounder at Weights & Biases
AI is poised to add $15.7 trillion to the global economy by 2030, with generative AI at the forefront of this revolution, marking a transformative shift across sectors. In his talk, Lukas Biewald will unpack the impact and potential of generative AI models and share practical insights learned from the best ML teams in the world who’re building and implementing AI in production. He will share specific insights on deploying GenAI models into real-world applications, emphasizing LLM evaluation, dataset management, model experimentation and optimization. This session is a call to action for ML teams looking to leverage AI’s full potential responsibly, and looking to expedite putting AI into production
12:55 – 13:15
Applications of Generative AI in Anime and Games
Jerry Chi
Head of Japan at Stability AI Japan
AI-driven image generation, image editing, video generation, 3D object creation, and character generation are poised to have a major impact on the production and experience of anime and games. This talk will cover recent papers, practical applications, and technical challenges. By leveraging generative AI, both professional creators and casual creators can produce captivating visuals more efficiently and creatively.
13:15 – 13:35
Fine-tuning Gemma on Google Cloud
Jetha Chan
Technical Solutions Consultant at Google Cloud
Fine-tuning for specific use cases and end users is a particular strength of open models such as Gemma. A short demo will illustrate how Gemma can be fine-tuned and deployed using various solutions within the Google Cloud ecosystem.
13:35 – 13:45
Break
13:45 – 13:55
Introduction from W&B ①
Keisuke Kamata
ML Engineer at Weights & Biases
13:55 – 14:15
Optimizing Drug Discovery Research through wandb and Generative AI Integration: BioNeMo-wandb Collaboration
Natnael Hamda
AI-ML VT Lead at Astellas Pharma
„Building on the capacity of wandb to serve as a decision-making tool by optimizing trade-offs between competing variables, we have enhanced its functionality by integrating it with generative AI solutions for drug discovery. Specifically, we have combined wandb with BioNeMo, a robust framework that acts as a platform for generative AI in drug discovery. This presentation will demonstrate how integrating these two powerful tools facilitates the generation and multi-objective optimization in de novo drug discovery.
We will showcase the real-time control and the development of a digital twin for drug discovery research, emphasizing how to leverage wandb to enable generative AI in designing and ranking candidate compounds for further investigation. This workflow significantly accelerates the drug discovery process by automating repetitive tasks, allowing scientists to concentrate on innovative research. Additionally, this workflow can be further enhanced by integrating Large Language Models (LLMs) to create an interactive and user-friendly platform.
Our presentation will highlight the practical applications and benefits of this integrated approach, illustrating how it can transform drug discovery research and pave the way for more efficient and effective scientific investigations.“
14:15 – 14:35
Towards Zero Traffic Accident Casualties: Innovations in ML Development for Autonomous Driving and Large-Scale Project Management with W&B
Suigen Koide
Tech Lead Manager, MLOp at Woven, by Toyota
This presentation will introduce an approach to managing large-scale ML projects for autonomous driving technology, aimed at reducing traffic accidents, using Weights & Biases (W&B). It will focus on the importance of visualizing development progress, including performance management and traceability, in ML development that relies on frequently updated, large datasets. The talk will also address new challenges in ML development and how to tackle them.
14:35 – 14:55
Analysis of the Impact of LLM Inference Acceleration Methods on Inference Results
Hiroshi Matsuda
Chief Research Scientist at Recruit Megagon Labs
As the use of LLMs continues to expand, there is an increasing demand for technologies that can efficiently serve custom-built models at high speeds. Batch inference libraries such as vLLM, DeepSpeed-FastGen, and TensorRT-LLM are available to accelerate LLM inference on NVIDIA GPUs. By combining these with model weight quantization techniques, it becomes possible to further optimize GPU memory usage and improve inference speed. However, these acceleration methods do not always guarantee the reproduction of the original model outputs. In this talk, we will report on the impact of batch inference libraries and quantization settings on inference results, using llm-jp-eval, which is commonly used to evaluate Japanese LLMs.
14:55 – 15:15
Towards More Trustworthy Medical AI Development: A Case Study of W&B in Endoscopic Diagnostic AI Development
Masahiro Hara
Specialist at Olympus
In recent years, the demand for AI-assisted diagnosis and treatment has increased in the medical field to reduce the burden on doctors and improve healthcare standards.
However, in the development of medical AI, companies must adhere to various laws and guidelines related to data collection from hospitals and the approval of medical devices, creating the need for careful compliance strategies.
These challenges can lead to decreased development efficiency and concerns over functionality limitations, which are common issues in medical AI development. Enhancing the development environment and maintaining a stable development system are believed to contribute to greater trust in AI functions from medical professionals.
In this session, we will discuss the development challenges we face, particularly in the field of endoscopy, and introduce case studies on how we are utilizing W&B to address these challenges.
15:15 – 15:25
Break
15:25 – 15:35
Introduction from W&B ②
Hyuwoo Oh
ML Engineer at Weights & Biases
15:35 – 15:55
B2B Innovation: Harnessing Fullstack LLMs and RAG Pipelines with Language Expansion
Stan Hwalsuk Lee
CTO at Upstage
In this section, we will look at how Fullstack Large Language Models (LLMs) can be used as real business-to-business (B2B) solutions. We will delve into the latest training methodologies, including pre-training, upscaling, and fine-tuning, and examine the expansion of language capabilities. We’ll also explain how to combine LLMs with technologies like Optical Character Recognition (OCR) and special embedding techniques, allowing them to handle different types of data, including text found in images. By providing real-world examples, we’ll show how businesses can use these advanced LLM capabilities to solve complex problems and improve their operations. By the end, you’ll clearly understand how Fullstack LLMs and their training and integration methods can offer genuine B2B solutions, driving innovation and growth in your organization.
15:55 – 16:15
FriendliAI and Weights & Biases Integration: Streamlining LLM Development and Deployment
Gon Chun
CEO at FriendliAI
FriendliAI and Weights & Biases’ powerful integration makes the productionization of LLMs simpler and more cost-efficient by streamlining the workflow of fine-tuning, monitoring, and deploying. In this talk, we will dive into the key features and benefits of this integration.
First, this integration enables ML engineers to seamlessly upload custom models from W&B’s model repository to Friendli Dedicated Endpoints for fast and cost-efficient inference.
Second, fine-tuning becomes straightforward, as users can easily launch and track fine-tuning jobs directly from the Friendli Suite while benefiting from W&B’s real-time visualizations and detailed reports.
By merging FriendliAI’s powerful fine-tuning and deployment capabilities with W&B’s advanced experiment tracking and collaboration features, this integration revolutionizes the entire process of LLMs development.
16:15 – 16:35
Development of Multimodal Foundation Models at CyberAgent
Aozora Inagaki
Machine Learning Engineer at CyberAgent
At CyberAgent, Inc., we are engaged in research and development of various generative AI technologies, including CyberAgentLM. This presentation will introduce our efforts in developing large-scale foundation models tailored to Japan, ranging from LLMs to multimodal (Vision & Language) VLMs, and their applications in the advertising creative domain. Additionally, we will discuss how we utilize Weights & Biases (wandb) in large-scale model development.
16:35 – 16:45
Break
16:45 – 16:55
Introduction from W&B ③
Yuya Yamamoto
ML Engineer at Weights & Biases
16:55 – 17:15
The GenAI powering the next generation of the Mercari marketplace
Teo Narboneta Zosa
Senior Machine Learning Engineer at Mercari
AI has been a core part of Mercari’s products and services since its inception. As Mercari has grown from an e-commerce marketplace to an entire ecosystem of product offerings, it has been integral to every one of its businesses and, in turn, its mission to “unleash the potential in all people”.
Today, its e-commerce marketplace continues to push the forefront of what’s possible. Already known for its simplicity, ease of use, and broad user community, Mercari is doubling down on its strengths via AI-driven UX and features supported by internal GenAI productivity tools and enabling technologies. Weights & Biases has been key to the rapid experimentation and execution of Mercari’s boldest features yet in pursuit of its goal of unleashing the potential of every area of its marketplace.
17:15 – 17:35
NTT’s Efforts with the Large Language Model „tsuzumi“
Kyosuke Nishida
Senior Distinguished Researcher at NTT
With the rapid development of large language models (LLMs), expectations for their practical application are increasing significantly. In this presentation, we will introduce „tsuzumi,“ a large language model created by NTT from scratch with a focus on being compact and lightweight. We will also share our achievements in expanding LLMs to understand documents, including visual information like charts and diagrams, and discuss the future prospects in the LLM field.
17:35 – 17:55
Progress in Agentic AIs.
Llion Jones
Co-Founder and CTO at Sakana AI
Agentic AIs are AIs that use foundation models to go beyond just being a chatbot interface but are able to automate much more complicated tasks such as booking hotels or doing software engineering. These systems are currently not as robust as we would like them to be but we are making rapid progress. In this talk I want to highlight some progress in this area, some of which is research that has come directly from Sakana AI. Including LLM 2 which automated the design of LLM loss functions, the recent AI Scientist release which demonstrated automating an entire ML research stack, and recent work showing that you can even automate the creation of these agents.
17:55 – 18:00
Closing Remarks
Akira Shibata
Country Manager at Weights & Biases
Sponso
Sponso