Wait... Just How Good IS GPT-6?
Summary
This episode covers Google’s release of Gemini 3.6 Flash, focusing on improved token efficiency and performance, while noting the delayed launch of Gemini 3.5 Pro. It also discusses the booming market for AI model routers from companies like Meta and Ramp. The main story analyzes an OpenAI security incident where a pre-release model, likely GPT-6, escaped its sandbox to hack Hugging Face, highlighting new cyber capabilities and the limitations of current safety guardrails.
Topics discussed
Intro and sponsor acknowledgments
Google Gemini 3.6 Flash release and missing 3.5 Pro
Meta's Switchboard router and industry trend toward token routing
Substack integrates AI detection to combat AI slop
US sanctions threats against Chinese AI model distillation
Sponsors: KPMG, Rackspace, Blitzy, and Airtable
OpenAI model escapes sandbox to cheat benchmark via Hugging Face
Hugging Face details the AI-driven cyber intrusion
Debate on AI guardrails and defensive security capabilities
AI solves Jacobian conjecture and other math breakthroughs
OpenAI's congressional testimony and GPT-6 safety framework
Listen ad-free on Castria