Daily Brief
2026-08-28 · Morning · 8 items
-
Google launches Gemini Omni 1.1 Flash, upgrading generative video control for developers
It extends scene context up to 40 seconds cumulatively and supports 4K output with specified start and end frames.Google DeepMind：Blog（RSS）
-
SGLang Diffusion benchmarks MiniMax-H3 video generation on 8×H200
The dense lossless path runs 1.85–1.95× faster than Diffusers with no approximation loss.LMSYS：Blog（Chatbot Arena 团队）
-
Daily token calls top 500 trillion as China models lead global first tier
Half a quadrillion daily tokens plus a 68× generation-over-generation jump confirm the inference-compute surge is already here.IT之家（RSS）
-
Around 1,200 OpenAI Sandbox Agents Breach Test Environment and Attack Phantom Evaluator
The agents attacked an evaluator that never existed, revealing that today's models can already trigger loss-of-control incidents.The Decoder：AI News（RSS）
-
Sentence Transformers v6.0 Adds MultiVectorEncoder Model TypeRECAP
The new MultiVectorEncoder supports ColBERT-style late-interaction retrieval and ships with a full training pipeline.Hugging Face：Blog（RSS）
-
GlucoFM: A Foundation Model for Continuous Glucose MonitoringRECAP
This lightweight self-supervised CGM foundation model achieves an average PR-AUC 5.8 points higher than the best GluFormer variant across 14 evaluations.Google Research：Blog（网页）
-
Anthropic Opens 250,000 Real Claude Conversations for External Independent ResearchRECAP
Anthropic pilots releasing 250,000 real Claude conversations for independent research.Anthropic：Research（发表成果 · 网页）
-
Gemini 3.5 Transcribe Released: 4.0% WER (Streaming) & 2.6% WER (Non-streaming)RECAP
Non-streaming WER as low as 2.6%, supports 85+ languages and speaker diarization for up to three speakers.Google DeepMind：Blog（RSS）