Anthropic Doesn't Think You Can Be Trusted, China is closing the gap, SpaceXAI Leaps Ahead, and DEF CON
Summary
Ben and Theo critique Anthropic’s new text watermarking as a flawed EU compliance measure that degrades output quality. They analyze Mark Zuckerberg’s manifesto on open-source AI and personal superintelligence, contrasting Meta’s distributed approach with Anthropic’s centralized safety model. The hosts also discuss the competitive landscape, noting how smaller labs are rapidly catching up to frontier models through advanced reinforcement learning techniques.
Topics discussed
Intro and PostHog Sponsorship
Anthropic's New Watermarking System
AI Regulation and CTPA Metadata
Apple's Advantage in Identity Verification
Meta's AI Manifesto and Market Shifts
Open Source Models and Developer Tools
AI Safety, Regulation, and Government Access
Market Competition and New Model Releases
AI Security Risks and Bioweapon Concerns
Local Models, Hardware Costs, and Qwen 3
Running Qwen and DeepSeek Locally
DeepSeek V4 Updates and GrokBot Interface
Grok Performance vs GPT-5 and Dario's Views
Public Perception of AI Risks and Benefits
Defcon Goldbug CTF and AI Agents
Solving Puzzles with AI Swarms
Agent Coherence, RL, and Final Thoughts
Listen ad-free on Castria