#250 - Mythos Mess, GPT 5.6-Sol, GLM 5.2
Summary
Hosts Andrei Kurenkov and Jeremy Harris discuss the US government’s emerging AI licensing regime, highlighted by Anthropic’s restricted release of MIFOS and OpenAI’s gated launch of GPT-5.6. They analyze benchmarking controversies, including allegations of model cheating, and the political dynamics behind these regulatory moves. The episode also covers OpenAI’s new Jalapeno chip, Amazon’s potential chip sales, and Anthropic’s ambient AI agent, Claude Tag.
Topics discussed
Intro: AI treaties, national security, and enforceability
Ad: Context layer for AI agents
Anthropic, US Govt relations, and Tom Brown's role
OpenAI GPT-5.6 launch and evaluation benchmarks
AI safety, licensing regimes, and cyber risk thresholds
Google Gemini, Meta, and political dynamics in AI policy
Hardware: Custom AI chips, Amazon Trainium, and memory
Memory market: HBM supply, SK Hynix, and Micron
Ads: Progressive and Allstate
Open Source: GLM-5, Kimi 2.7, and sparse attention
AI impact on jobs and the 'Skill Act' initiative
DeepMind safety framework and loss of control taxonomy
AI lobbying, PACs, and conservative protests
Research: Model convergence and streaming architectures
Amazon cancels AI film and podcast outro
Listen ad-free on Castria