Automating Test Selection for AI Accelerators with PyTorch

Interested in writing fast Triton kernels for AI inference? Check out this new PyTorch blog from our team at IBM Research, written in collaboration with AMD and Red Hat. We go deep into how we optimized our Triton attention kernels to enable vLLM V1 on AMD GPUs. I'll be presenting this work, together with Aleksandr M. from AMD, at the PyTorch Conference this Thursday @ 11AM PDT (Room 2001-2003). You can also find me at the IBM Booth where I'll be happy to chat about vLLM, Triton, and AI inference acceleration. Add the talk to your schedule here: https://lnkd.in/dMgnNKHq

Raspberry Pi founder Eben Upton: “I’m an omni-geek” Eben Upton, the British technology leader and Raspberry Pi co-founder, has shared his views on coding, artificial intelligence and the risks shaping the future of technology. In the discussion, Upton described coding as a modern form of literacy, reflecting how essential computational thinking has become across education, work and everyday life. He also expressed scepticism towards some AI evangelists, urging a more measured approach to claims about the technology’s potential and risks. Rather than accepting extreme predictions, Upton’s perspective highlights the importance of examining evidence, context and real-world consequences. Addressing the idea of catastrophic AI risk, often discussed through the term “p(doom)”, Upton remarked that “the p(doom) is probably Russia” — a comment that places geopolitical developments alongside technological risks in the broader conversation about global security. His views contribute to an ongoing debate about how society should approach AI: with ambition, but also with critical thinking and a clear understanding of the wider environment in which these systems are developed and used. Artificial Intelligence School offers expert-led programs designed to help learners and professionals understand artificial intelligence, its applications and its wider implications. #aischool #artificialintelligenceschool #ArtificialIntelligence #AI #Technology #Coding #DigitalLiteracy #TechEducation #ResponsibleAI #FutureOfTechnology #Geopolitics

One of the biggest challenges in AI infrastructure is ensuring that performance gains at the frontier don’t come at the expense of openness and portability. In a new vLLM collaboration, contributors from IBM, Meta, and Hugging Face introduce hardware-agnostic layers that preserve torch.compile support and extensibility across diverse models and accelerators, while allowing specialized paths to optimize for emerging systems. This architecture is designed to help the open-source ecosystem move faster without fragmenting across hardware platforms. As deployment becomes more heterogeneous, that balance will matter even more. I’m proud of IBM’s contribution to building AI infrastructure that remains open, portable, and performance-driven. Read the blog: https://lnkd.in/er3-9s47

AI systems are becoming more capable, but one question is getting more important: How do we know they are actually behaving correctly in production? With traditional applications, we monitor logs, latency, errors, and infrastructure. With AI agents, we also need to understand decisions, tool calls, model responses, cost, failures, and whether the final output is actually useful. That is why AI observability is becoming such an interesting part of modern software engineering. Building an AI feature is one challenge. Making it reliable, measurable, and trustworthy in production is a completely different one. I’m excited to keep learning how traditional engineering practices like monitoring, tracing, and reliability are evolving for AI-driven systems. #AI #AIEngineering #AIAgents #AgenticAI #Observability #GenerativeAI #LLM #SoftwareEngineering #CloudEngineering #DevOps #SystemDesign #MLOps #SoftwareDevelopment #Technology

Don't overthink it (if you are a coding agent) ! Nice coverage of our work **CRANE** https://lnkd.in/grXSyjbT in collaboration with RPI You can read the paper here https://lnkd.in/gp-NtREt awesome work with Mingzhi Zhu, Stacy Patterson, Raju Pavuluri

Raspberry Pi founder Eben Upton: “I’m an omni-geek” Eben Upton, the British tech leader and founder of Raspberry Pi, reflects on the changing role of technology in society—and the skills people will need to navigate it. In the conversation, Upton discusses: 💻 Coding as a modern form of literacy, suggesting that understanding how technology works is becoming increasingly important. 🧠 His cautious view of AI evangelism, questioning some of the more confident predictions surrounding artificial intelligence and its future. 🌍 The risks posed by advanced technology, including his observation that “the p(doom) is probably Russia”—a reference to discussions about the probability of catastrophic outcomes from AI and geopolitical threats. 🔧 His identity as an “omni-geek,” reflecting a broad curiosity across technology, engineering and science. The discussion offers a measured perspective on AI, coding and global risk—without reducing complex issues to simple predictions. #SuperIntelligenceNews #SuperIntelligenceNewsletter #ArtificialIntelligence #AI #Technology #Coding #RaspberryPi #Innovation #TechNews #Geopolitics

Great conversation last week with Tim Huang, Kush Varshney, Chris Hay, and Sascha Brodsky in the last episode of IBM's Mixture of Experts podcast. We covered Anthropic's new Model Hardware Standard and the latest Claude updates, OpenAI's work on security, and Runway's new model. MHS was the topic that interested me the most. It lets an AI agent operate physical lab and factory equipment through a single common interface, roughly what MCP did for software tools, one layer down. The harder question was where the safety limits live. A model can propose an action, but it should not be the thing that decides whether that action is safe. Get that boundary wrong and physical AI has no business near real equipment. What do you make of the MHS preview? 👇 Episode link in the comments. #MixtureOfExperts #podcast #IBM #PhysicalAI #AIAgents #LabAutomation #AISafety

As AI infrastructure becomes more heterogeneous, performance can no longer depend on reinventing optimization techniques for every new hardware platform.    This work from IBM Research and University of California, Berkeley demonstrates how AI-driven evolutionary kernel search can transfer optimization knowledge from CUDA to Apple Silicon, achieving near-expert performance and up to 20× gains on key kernels.    The broader opportunity is accelerating AI across a diverse set of computing platforms, a key focus of IBM's AI and infrastructure strategy. Read the blog: https://lnkd.in/efBmbUF3 

We’re exploring a new abstraction for AI systems at IBM Research, introducing modularity into model design. (https://lnkd.in/eFPuUTY3)   Today’s models are monolithic. Capabilities are diffused across all the weights, which limits reuse and targeted improvement. What’s missing are composable units with defined behavior, like we see in other forms of software design.    Granite Libraries provide this through collections of task-specific adapters, including ones for retrieval, safety, and robust text generation. These are designed as reusable building blocks, where small, activated components can outperform much larger models on specialized tasks. (https://lnkd.in/ei8CPrpc)   Project Granite Switch introduces a new layer that uses the model’s attention mechanism, leveraging activated LoRAs (aLoRAs), to enable seamless switching between capabilities — without wiping the model’s short-term memory or causing latency issues. (https://lnkd.in/e3GU4X5E)   With Mellea (https://mellea.ai/), these adapters become typed functions with explicit inputs and outputs, forming the basis of a more composable, software-like system. This is a step toward AI systems you can engineer — not just train.