LWiAI Podcast #235 - Sonnet 4.6, Deep-thinking tokens, Anthropic vs Pentagon
Summary
Hosts Andrei Kerenkov and Jeremy Harris discuss recent AI news, including Anthropic’s Sonnet 4.6 and Google’s Gemini 3.1 Pro, which achieved high scores on the ARC-AGI 2 benchmark. They cover Meta’s $100 billion chip deal with AMD, funding for startups Metax and World Labs, and delays in OpenAI’s Stargate data centers. The episode also analyzes China’s chip production goals and a research paper on adaptive optimizers.
Topics discussed
Intro, startup update, and listener feedback
Anthropic Claude 4.6 release and benchmarks
Google Gemini 3.1 Pro and xAI Grok 4.2
AI agents: Cloud Code, OpenClaw, and multi-agent systems
Meta-AMD chip deal and NVIDIA's MATX processor
World Labs funding, data centers, and SMIC chips
Research: Cautious optimizers and sparse updates
Research: Deep thinking tokens and layer divergence
Research: Attractor states in LLM conversations
Research: Geometry of counting and task horizons
Research: Scaling laws, efficiency, and persona models
Policy: Anthropic-DOD conflict and distillation attacks
OpenAI disrupts scam operations and outro
Listen ad-free on Castria