Michael Wong
Expert of the ISO Artificial Intelligence Technical Committee, Chair of the C++ Standard Committee’s Machine Learning Group, and CTO of YetiWare
Expert of the ISO Artificial Intelligence Technical Committee, Chair of the C++ Standard Committee’s Machine Learning Group, and CTO of YetiWare. He also serves as Chair of the C++ Embedded Development Committee (SG14) and the Machine Learning Committee (SG19), as well as Chair of the C++ Evolution Working Group. He is the former Vice President of R&D at Codeplay and former CEO of OpenMP, and currently leads the Canadian Delegation to the C++ Standard Committee. Michael has extensive experience in C++ parallel computing, high-performance computing, and machine learning. He led the development of the C++ heterogeneous programming language (SYCL) standard for GPU application development, as well as OpenCL. He has profound research and insights into the performance optimization of the underlying layers of PyTorch and TensorFlow. His work specifically covers parallel programming, neural networks, computer vision, and autonomous driving. Michael previously served as a Senior Technical Expert at IBM, where he led the development of the IBM XL C++ compiler and XL C compiler.
Topic
The Platform Paradox: Why Most Open-Source AI Ecosystems Fail and How to Build One That Thrives
While the AI revolution has produced remarkable models and algorithms, the infrastructure powering this transformation tells a different story—one of fierce ecosystem battles, abandoned platforms, and a graveyard of well-intentioned projects that failed to gain traction. For every PyTorch, there are dozens of technically superior platforms that never found their community. This talk decodes the paradox of building successful open-source AI platforms: why technical excellence alone guarantees nothing, why early architectural decisions can doom or propel a project, and why the most permissive ecosystems often win over the most powerful ones. Drawing from case studies of both triumphant platforms （如CUDA、PyTorch、ONNX）and cautionary tales (the TensorFlow 2.0 migration, abandoned Intel projects, various "CUDA killers"), we'll explore the hidden dynamics that determine platform success. We'll dissect the three pillars of sustainable AI ecosystems: technical differentiation that actually matters to developers, community cultivation that goes beyond GitHub stars, and the delicate balance between openness and sustainability. Attendees will learn why PyTorch's "worse is better" philosophy defeated TensorFlow's early dominance, how OpenAI Triton carved out space in NVIDIA's walled garden, and why some platforms thrive across hardware vendors while others remain locked to single ecosystems. Whether you're building the next AI infrastructure platform, choosing a stack for your organization, or simply curious about the sociotechnical forces shaping AI's future, this talk provides a framework for understanding what separates platforms that become movements from those that become monuments to missed opportunity. You'll leave with actionable insights on developer experience design, community governance models, the strategic timing of open-sourcing decisions, and the art of building ecosystems that outlive their creators. Continuing from my ML-Summit talk from 2024 on AI infrastructures, platforms, and ecosystem, key takeaways include: the "adoption ladder" framework for community building, why documentation strategy matters more than performance benchmarks, how to navigate the build-integrate-compete decisions with existing platforms, and the sustainability models that keep open-source AI infrastructure alive without sacrificing community trust.
