Daily Brief
2026-09-14 · Morning · 8 items
-
Musk: Grok 4.8 finishes training this week with 2.5T parameters
2.5 trillion parameters and a new C++ software stack, with RL starting this week.IT之家（RSS）
-
Musk: Grok 4.8, a 2.5T model, finishes training this week
Grok 4.8 uses a new C++ stack with 2.5T parameters and wraps training this week, likely clearing the way for Grok 4.7 next week.X：Haider (@haider1)
-
Google Cloud CEO: AI servers pay back in under 2 years; in-house chip halves payback vs GPUs
Custom AI accelerator deployment is over 2x the second-largest hyperscaler, with up to 80% better cost-performance than GPUs.IT之家（RSS）
-
MIT's HardFlow makes generative AI meet hard constraints
A plug-and-play sampling method enforces safety and physics constraints at deployment, with no retraining.MIT News（RSS）
-
JD releases JoyAI-Echo 1.5 and open-sources interactive audio-visual world model EchoWM
EchoWM enables interactive audio-visual generation with open weights and code, ready for downstream app development.公众号：京东JoyAI
-
Developer names Muse Spark 1.3 and agent as default stack
Near-instant responses and strong browser-based control are why this developer made it the everyday default.X：Alexandr Wang（Scale AI 创始人/Meta 首席 AI 官） (@alexandr_wang)
-
Meta's Muse smart glasses spotted at the US Open
A textbook case of product-market fit as wearables enter major sports venues.X：Alexandr Wang（Scale AI 创始人/Meta 首席 AI 官） (@alexandr_wang)
-
Muse sparks nationwide buzz within 24 hours
Its single-day spread rate has been widely documented as a cultural phenomenon.X：Alexandr Wang（Scale AI 创始人/Meta 首席 AI 官） (@alexandr_wang)