Ironwood: The first Google TPU for the age of inference
Ironwood is Google’s seventh-generation TPU, engineered specifically for the age of inference, enabling advanced AI models to generate insights proactively rather than merely responding to queries.
The Agent2Agent Protocol (A2A)
Agent2Agent (A2A) Protocol enables AI agents to communicate and collaborate across diverse platforms, enhancing productivity and reducing costs through interoperability among over 50 technology partners including Atlassian and Salesforce.
Colossus for Rapid Storage
Rapid Storage leverages the Colossus stateful protocol to achieve sub-millisecond read/write latency and support for unlimited appends, enhancing performance for data-intensive workloads like AI and ML.
[D] Yann LeCun Auto-Regressive LLMs are Doomed
Yann LeCun argues that auto-regressive LLMs are fundamentally flawed, suggesting that their reliance on sequential data processing limits their potential for true understanding and reasoning.
Visual Reasoning Is Coming Soon
OpenAI's GPT-4o model revolutionizes image manipulation by integrating full conversational context into image generation, allowing for more accurate and consistent results, such as dressing a specific cat in a detective hat.
The next generation of Bazel builds
Bonanza represents a radical shift in build systems, leveraging cloud capabilities to handle all aspects of builds, including dependency management and graph construction, thus eliminating local execution complexities inherent in Bazel.
The 2025 AI Index Report
The 2025 AI Index Report reveals significant advancements in AI research, highlighting a 50% increase in published papers compared to the previous year, indicating a robust growth in the field.
Controlling Language and Diffusion Models by Transporting Activations
Activation Transport (AcT) is a novel framework that enables fine-grained control over generative models' outputs with minimal computational overhead, addressing challenges in reliability and safety without altering model parameters.
TVMC: Time-Varying Mesh Compression
TVMC introduces a novel approach for time-varying mesh compression using volume-tracked reference meshes, enhancing efficiency in 3D data representation.
[P] B200 vs H100 Benchmarks: Early Tests Show Up to 57% Faster Training Throughput & Self-Hosting Cost Analysis
B200 GPUs demonstrate up to 57% higher training throughput than H100s in computer vision tasks, indicating a significant performance leap for model training workloads.
Trustworthy AI Without Trusted Data
EPFL researchers have developed ByzFL, a tool that enhances federated learning by benchmarking models against adversarial threats, ensuring AI systems can be built without trusting individual data sources.
EGPU: Extending eBPF Programmability and Observability to GPUs
eGPU introduces a groundbreaking framework that dynamically offloads eBPF bytecode onto GPUs via PTX injection, enhancing observability and programmability for AI and HPC applications without significant overhead.
Self-Steering Language Models
The paper introduces DisCIPL, a method for self-steering language models that allows a Planner model to create task-specific inference programs executed by Follower models, enhancing reasoning capabilities. Link to article
Google Cloud Rapid Storage
AI Hypercomputer integrates over a decade of Google’s AI expertise, delivering 24x higher intelligence per dollar than GPT-4o, thanks to innovations like the Ironwood TPU, which offers 5x more peak compute capacity and 6x the HBM of its predecessor.
[P] Reducing Transformer Training Time Without Sacrificing Accuracy — A Dynamic Architecture Update Approach
A novel method for optimizing transformer models enables dynamic architecture updates during training, significantly reducing training time while preserving accuracy.