LWiAI Podcast #231 - Claude Cowork, Anthropic $10B, Deep Delta Learning
Summary
Hosts Andrei Krennikov and Jeremy Harris review recent AI developments, including Anthropic’s new Co-work agent and $10B funding, XAI’s $20B raise, and NVIDIA’s H200 supply constraints. They discuss Google’s Gemini updates, OpenAI’s compute deal with Cerebras, and CoreWeave’s financial adjustments. The episode also covers research on scaling reinforcement learning to prevent catastrophic forgetting in general-purpose models.
Topics discussed
Introduction and episode overview
Anthropic's Claude 4.5, Cloud Code, and safety frameworks
Google Gemini updates, AI Overviews, and health misinformation
Slack AI agent and the rise of workplace automation
AI industry funding: xAI, CoreWeave, and NVIDIA supply chain
CoreWeave liquidity issues and LMSYS valuation
DeepSeek papers: RLHF, catastrophic forgetting, and H-Res
DeepSeek V3 architecture and Loop Transformers
TII Falcon H1R-7B and hybrid Mamba-Attention models
Deep Delta Learning and recursive prompt search
Conditional memory, N-gram lookups, and extending context
Anthropic's constitutional classifier and jailbreak defense
Geopolitics: NVIDIA H200s, China export controls, and policy
Listen ad-free on Castria