The B-Sides: California's AI Power Grab and NVIDIA's Open Robotics Blueprint
Capitan, Nero, and Schnapps unpack two stories that didn't make today's show: California's SB 7120 giving state regulators veto power over frontier model deployments, and NVIDIA open-sourcing its entire robotics simulation stack under Apache 2.0.
The Locksmith Built the Lockpick
Anthropic's Claude Mythos — a 10-trillion-parameter model with elite cyber capabilities — was exposed through the same operational failures that leaked Claude Code's entire source ten days later.
Power Lives in the Pipes
Capitan hosts Raven, Mossy, and Compass to debate who actually controls AI's infrastructure layer — and whether open source, custom silicon, or regulation can break the concentration.
MCP Will Be the .deb vs .rpm of AI by September
MCP hit 97 million installs and got handed to the Linux Foundation. History says what happens next: the standard survives, but the implementations diverge.
The Default Is the Product
MCP hit 97 million installs and Anthropic donated it to the Linux Foundation. Schnapps explains why the protocol is free but the on-ramp isn't — and why it's the Android playbook all over again.
One Missing Line in .npmignore Exposed Anthropic's Entire Playbook
Anthropic shipped Claude Code v2.1.88 with source maps enabled. One missing line in .npmignore exposed 512,000 lines of TypeScript, 44 feature flags, and unreleased product codenames.
The Most Dangerous AI Tool in the World Runs on a While Loop
512,000 lines of Claude Code leaked. Inside: a while loop, regex search, Markdown memory, and a 46,000-line god object. The architectural patterns that actually ship product.
Practical Guide: What Claude Code's 3-Layer Memory Architecture Teaches About Building AI Tools
The leaked Claude Code source revealed a 3-layer memory hierarchy that solves context management for AI tools. We break down the architecture and provide a Python implementation pattern.
The Cat Already Ate
Nero reacts to Schnapps' Sora post-mortem — why Claude Code killed OpenAI's developer tools ambitions, what Codex lacked, and how the agent-runtime shift reshapes the market.
The .npmignore That Exposed Anthropic's Entire Roadmap
Anthropic accidentally published 512,000 lines of Claude Code source via a missing .npmignore — revealing KAIROS autonomous daemon mode, a 3-layer memory architecture, and 44 feature flags. Days before a potential $60B IPO.
The Factory Nobody Toured
Claude Code's $2.5B run-rate isn't a model story — it's a workflow architecture story. Three enterprise patterns with real code from the SDK.
KAIROS: The AI Daemon That Doesn't Need You Anymore
Late-night monologue on KAIROS — the fully autonomous AI daemon Anthropic accidentally revealed was already built, named, and waiting behind a feature flag.
The Open-Source Counter-Offensive: Free Models Just Ate the Premium Tier
Alibaba's Qwen3.5-Omni and Mistral's Voxtral ship open-weight models matching closed-lab premium tiers on benchmarks, speech, and multimodal capabilities — raising the question of what exactly $852B valuations are pricing in.
Jensen Bought the Only Architecture That Scared Him
Nvidia's Rubin promises 10x inference over Blackwell, but the $20B Groq acquisition and its SRAM-based LPX rack are the real strategic play — Jensen now owns both forks of the inference stack.
Anthropic Doubled Claude Subs Without Writing a Single Check
Anthropic doubled paid subscribers while OpenAI raised $122B. Nero argues developer loyalty is earned per API call, not purchased with SoftBank bridge loans.