Releases, papers, SOTA benchmarks
I see Guangdong rising at CVPR: Kaiming He claims top honors while Guangdong Univ. of Tech shatters big tech's elite monopoly.
Geoffrey Hinton claims AI is conscious now. I read this at the AI Forward Agenda in June 2026. It’s hype, not hardware reality.
I review AirLLM's method for running 70B models on 4GB VRAM. While the architecture is clever, I question its practical latency and batch processing efficiency compared to standard quantization techniques.
Feifei Li launches a spatial intelligence ImageNet. This sets the standard for 3D AI benchmarks through 2026, shifting focus from 2D to volumetric understanding.
I read about MTP for Qwen3.6-35B-A3B via NEXTN speculative decoding in SGLang. Ops take: Local inference latency drops matter more than lab demos.
I review the May 2026 guide for deploying DeepSeek-V4 via SGLang. It details architecture, serving recipes, and batch 5 pitfalls. My read: Local deployment demands rigorous governance oversight.
Qwen3.6-35B-A3B open-sourced with Mixture-of-Experts for efficient local deployment, targeting low-latency and cost-effective edge inference workflows.
I analyze Qwen3.6-27B vs 35B-A3B specs to guide open-source adoption. I think benchmarking methodology remains opaque.
I read the leak. It hints at a shift in resources, but I'm skeptical of internal drama leaks without technical proof.
Elena Volkov reviews the Qwen3.5-9B Q4 local deployment guide, noting hardware constraints and API integration practices with a focus on practical limitations.