Daily Brief
2026-08-29 · Evening · 8 items
-
Anthropic has Claude train models itself to curb alignment failures
The method still works on models 4.7x larger than the training target without hurting general capabilities.Anthropic：Research（发表成果 · 网页）
-
Open ASR leaderboard adds first Global South language: Hindi
The dataset spans 4,888 speakers with 12 demographic attributes, marking Hindi as the leaderboard's first non-European language.Hugging Face：Blog（RSS）
-
OpenAI's In-House Chip Jalapeno Outperforms Rubin on Inference
AI-designed and taped out in nine months, it beats Rubin in tokens per megawatt — and can still run Doom.X：SemiAnalysis (@SemiAnalysis_)
-
EdgeBench Long-Horizon Test: Top Agent Scores Just 51.3 in 12 Hours
Across 134 tasks and 12 hours of run time, the best agent reaches only 51.3 out of 100, and roughly 38,000 hours of interaction still fail to yield steady gains.X：Rohan Paul (@rohanpaul_ai)
-
Andrew Ng: In the Agent Era, an Engineer's Core Skill Is Judgment
Coding has gotten cheap, but judgment has not — and that is now the engineer's real edge.X：阿易 AI Notes (@AYi_AInotes)
-
Nvidia Reportedly Eyes $13B Takeover of Hugging Face
Open-weight models sit at just 6% enterprise adoption, yet headline Nvidia's biggest rumored bid.IT之家（RSS）
-
China's Smart Computing Capacity Hits 2.45M PFLOPS by July
Eight hubs and three power-computing zones host over 85% of the capacity, with 1.45M PFLOPS now on the national scheduling platform.IT之家（RSS）
-
Nvidia's AI Edge Expands Beyond the GPU Itself
The Vera CPU inside Vera Rubin delivers roughly 3x faster flash performance, easing the orchestration crunch at gigawatt scale.TechCrunch：AI（RSS）