Releases, papers, SOTA benchmarks
I read the claim that a new RNN treats hidden states as models, beating Transformers. Ops take: If it doesn't fit our GPU budget, it's just another lab demo.
I read the Kimi paper revealing an inference architecture that handles 80% of traffic. It’s a pragmatic approach to scaling AI workloads efficiently.
Kaiming He now leads his first MIT team for an AI gen project, featuring double Olympiad gold medalist Mingyang Deng. I note this work sits outside the standard 2025-2026 publication cycle.
I see Tsinghua & Ant's pure MLP beating Transformers for time series forecasts, challenging the status quo.
I read about a domestic gen-AI model matching AlphaFold3 for antibody design. My read: Domestic capability is closing the gap.
BaiChuan's new model tops Chinese benchmarks and launches 'Bai Xiaoying,' its first AI assistant. I note this release sits outside the primary 2025–2026 development window.