#196 - Nvidia Digits, Cosmos, PRIME, ICLR, InfAlign
Summary
This episode focuses heavily on AI hardware, detailing NVIDIA’s $3,000 personal supercomputer and its push into custom ASIC manufacturing in Taiwan to compete with Broadcom. The hosts discuss Meta’s controversial AI characters, Anthropic’s $60 billion valuation, and OpenAI’s delayed agent launch due to safety concerns. Additional topics include TSMC’s packaging capacity expansion, Microsoft pausing a data center for cooling upgrades, and Google integrating AI teams under DeepMind.
Topics discussed
Sponsors: OutShift, Box, and 988 Suicide Prevention
Intro, Host Banter, and Community Updates
NVIDIA GB200 Superchip and Data Center Infrastructure
Meta's AI Character Controversy on Instagram
NVIDIA's Custom Chip Design and R&D Expansion
Anthropic's $6B Funding Round and Valuation
AI Agent Security Risks and Prompt Injection
Anthropic Computer Use API and TSMC Chiplet Tech
Microsoft Pauses Data Center Construction for Cooling
Google Merges Brain and DeepMind Teams
NVIDIA Cosmos World Foundation Models for Robotics
Phi-4 Model and Reasoning via Reinforcement Learning
LLM Representations and In-Context Learning Mechanics
Metagenomic Sequencing and Video Generation Research
Epoch AI Report on Compute Trends and Scaling
Test-Time Alignment and Inference Scaling
Copyright Lawsuits and Anthropic's Settlement
US AI Data Export Rules and Geopolitical Policy
Listen ad-free on Castria