NVIDIA GB300 NVL72 Breaks MoE Pretraining Record, Jack Dorsey’s Buzz Merges Chat AI with Git – AI Daily Brief (July 22)
Audio in Mandarin Chinese · English transcript below
⚡ NVIDIA's GB300 NVL72 smashes MoE pre-training record; Dorsey launches Buzz (Agent + code); Suno's 55M user data breach sounds alarm.
AI agents are rapidly permeating everything from underlying infrastructure to end-user collaboration scenarios: NVIDIA is powering gigascale AI factories with a new generation of Ethernet switching systems, while Jack Dorsey’s Buzz attempts to embed AI agents into team workflows through a self-hosted model. Notably, as Suno suffers a data breach affecting 55 million users, Buzz itself has also sparked controversy over multi-user permission control. Taken together, these incidents highlight an increasingly urgent reality: every step deeper that AI capabilities take into core production environments demands a corresponding reinforcement of privacy and security.
Today's Top 3 Headlines
- Technical Breakthroughs
🚀 NVIDIA sets world record for MoE pretrain on GB300 NVL72; DeepSeek-V3: 1,648 TFLOPs/GPU
NVIDIA sets MoE pre-training record on GB300 NVL72: DeepSeek-V3 hits 1,648 TFLOPs/GPU, 671B params, 37B/token. For AI devs, this boosts MoE pre-training efficiency, cuts costs.
Source ↗ - AI Industry News
🤖 Jack Dorsey launches Buzz, integrating team chat, AI Agents, and Git code hosting
Jack Dorsey, Twitter co-founder, launches Buzz, an open-source self-hosted workspace integrating team chat, AI Agents, and Git hosting. For developers, Buzz’s AI Agents as new collaboration users pose data privacy and management permission challenges, reminding the industry to balance convenience and security.
Source ↗ - AI Industry News
🤖 NVIDIA Spectrum-6 Ethernet switch system launched, 102.4Tbps capacity powers gigabit-scale AI factory
NVIDIA released Spectrum-6 Ethernet switching system with 102.4Tbps capacity, double the previous generation, for gigabit-scale AI factories. For AI infrastructure developers, network is no longer a bottleneck, enabling full GPU cluster utilization for training Agent AI models; CoreWeave and Microsoft have already deployed it.
Source ↗
+7 more headlines
- 🛡️ Suno data breach affects 55 million users
- 🤖 OpenAI prepares for IPO, adds two independent board members
- 🤖 Poolside releases 118B parameter open-source coding model Laguna S 2.1
- 🤖 Google Gemini 3.6 Flash cuts Agent token costs by up to 65%
- 🤖 AI drones plant 50 hectares daily, aiming to restore 1 million hectares of forest.
- 🤖 Your machine learning experiments a mess? Here's the fix
- 🤖 AI companies poach top university talent: Meta, DeepMind snatch multiple professors and philosophers