AI Daily — August 28, 2026
2026-08-28
anthropic.com
Previewing the Model Hardware Standard
Anthropic has published a research preview of a 'Model Hardware Standard,' suggesting an initiative to define or certify hardware requirements for running AI models at specific capability levels. This kind of standard, if adopted broadly, could have significant implications for deployment infrastructure decisions and regulatory frameworks around AI compute.
deepmind.google
Google DeepMind pilots world's first double-blind AI evaluations
DeepMind is piloting a double-blind evaluation methodology for AI systems, where neither the evaluators nor the models' developers know which model is being assessed during benchmarking. This approach directly targets evaluation contamination and developer-side bias, two well-documented problems undermining the reliability of current AI benchmarks.
deepmind.google
Gemini Omni 1.1 Flash lets you build with more control
Google DeepMind released Gemini Omni 1.1 Flash, an updated version of its efficient multimodal model with enhanced developer control features. The 'Omni' designation indicates full-spectrum multimodal capabilities, and the Flash tier targets low-latency, cost-sensitive production use cases.
rss.arxiv.org
TreeGraft: Adaptive Multi-Drafter Grafting for Tree-Based Speculative Decoding
TreeGraft introduces a speculative decoding framework that uses multiple draft models of different sizes to collaboratively build a shared draft tree, with the stronger drafter rescoring candidates from the weaker one. This addresses the core latency-quality tradeoff in tree-based speculative decoding and demonstrates improved accepted token length without proportional latency increases — a practically significant result for LLM inference optimization.