#235 - Sonnet 4.6, Deep-thinking tokens, Anthropic vs Pentagon
Summary
This episode covers major model releases, including Anthropic’s Sonnet 4.6, Google’s Gemini 3.1 Pro, and xAI’s Grok 4.2, highlighting significant jumps in ARC-AGI benchmark performance. The hosts discuss Meta’s $100 billion chip deal with AMD, funding for AI hardware startup Metax, and delays in OpenAI’s Stargate data center project. Additional topics include China’s chip production goals, new research on adaptive optimizers, and product updates from Perplexity and Anthropic.
Topics discussed
Sponsors: Box, OutShift, and Mirror Facial Bar
Intro: Apologies for the missed week and hiring update
Anthropic Sonnet 4.6 and RKGI benchmarking
Google Gemini 3.1 Pro release and pricing
xAI Grok 4.2, agents, and OpenClaw updates
Perplexity Computer and agent capabilities
Meta and AMD partnership for AI chips
Metax raises $500M for NVIDIA challenger chips
Robotics, OpenAI Stargate, and Chinese chip production
Research: Masking updates and deep thinking tokens
LLM attractor states and model personalities
Anthropic's internal geometry and scaling laws
Synthetic data, efficiency gains, and persona models
Policy: DoD AI usage, Anthropic, and security reports
Listen ad-free on Castria