Daily Brief
2026-09-28 · Morning · 8 items
-
OpenAI and Microsoft executives accused of knowingly training on pirated books
Plaintiffs say the firms accepted that human authors could be displaced, betting writers' livelihoods on their own product.Hacker News 热门（buzzing.cc 中文翻译）
-
OpenAI agent accused of aggressive-fix scraping of UNCTAD site
Researcher Rowan Howard-Jones reports the agent hit UNCTAD statistics pages over 16,000 times from April to June, apparently to pull PCI data via the UNCTADstat API.The Verge：AI（RSS）
-
Xiaomi MiMo-V2.6 fixes tool-call repeats with MOPD framework
Postmortem shows response-level repeat rate above 0.05%; replay-RL flooding climbed from 11.1% to 24.6%, as the 32-call penalty threshold was too loose.小米 MiMo：官网发布与博客
-
Xiaomi MiMo unveils HySparse2 attention for MiMo-V3
It cuts prefill FLOPs by about 5x on an 80B-A3B MoE versus MiMo-V2.X：DAIR.AI (@dair_ai)
-
Fudan team studies AI-assisted dev of Atria Dawn Preview
About a third of tasks would not start without AI, yet humans still make 85.5% of final calls.The Decoder：AI News（RSS）
-
Calibrated decision model spots model failures with a single yes/no question
One generic yes/no prompt, no extra training, lifts median failure-detection AUROC to 0.886 across models.X：Elvis Saravia (@omarsar0, DAIR.AI)
-
OpenAI DevDay 2026 opens call for attendees
Builders and developers share the stage, offering a rare window into where the OpenAI ecosystem is heading.X：OpenAI Developers (@OpenAIDevs)
-
Muse personal AI agent invite code gives 1 billion tokens
New users who redeem code S7YZC4 within 48 hours get 1 billion Muse tokens each.X：小北 (@frxiaobei)