⚡ NVIDIA NeMo RL spec-decoding 8B 1.8× faster, Google & Korea build world’s first AI campus
Audio in Mandarin Chinese · English transcript below
⚡ NVIDIA hits the gas, Google & Korea mint an AI university, OpenAI eyes an IPO, Meta lets models cook their own weights—plus robo-car and water-saving gossip, the AI world crams a year into one rush-hour ride!
Artificial intelligence is rapidly accelerating towards full maturity and scaled development. NVIDIA, by significantly enhancing model training efficiency through speculative decoding, signals that AI technological breakthroughs are continuously lowering development barriers. Meanwhile, Google's co-building of the world's first AI campus with South Korea aims to solidify the foundations for future talent and infrastructure. These advancements at both technological and ecosystem levels, coupled with OpenAI's commercialization push towards an IPO, collectively paint a comprehensive picture of AI's journey from innovation to application and ultimately to market maturity.
Today's Top 3 Headlines
- AI Industry News
🤖 NeMo RL Spec-Decode: 8B LLM 1.8× faster, 235B eyes 2.5×
NVIDIA adds speculative decoding to NeMo RL, 8B-Agent rollout 1.8× faster; 235B end-to-end training heads for 2.5×. RL teams get more runs per dollar, LLM cycles shrink from weeks to days.
Source ↗ - AI Industry News
🤖 Google & Korea build world’s first AI campus
Korea’s President Lee Jae-myung & Google DeepMind CEO Demis Hassabis unveil world’s first AI campus, pledging 10× compute + joint degrees. Devs gain cross-border jobs as Asia’s LLM race accelerates.
Source ↗ - Finance
🤖 OpenAI CFO Sarah Friar steers biggest-ever AI IPO countdown
Sarah Friar told staff OpenAI’s IPO prep is under way, eyeing a $100B+ valuation that could top tech listings. For devs and investors, cheaper models and faster monetization of toolchains are now on the horizon.
Source ↗
+7 more headlines
- 🤖 Meta Autodata turns any LLM into a data scientist in seconds
- 🤖 TRL Post-Train Codebook: SFT→DPO→GRPO Inference Deep Dive
- 🤖 Gemini Embedding 2 unifies text/image/video/audio/doc embeddings
- 🤖 AI data centers use far less water than feared
- 🤖 Uber wants to turn millions of drivers into a live AV sensor web
- 🤖 Kill list or AI for all? DeepSeek 1.3B offers a 3rd path
- 🤖 Codex 1-click import turns scattered context into client briefs