Inside the model's mind, Anthropic's $19B bet, & the rise of physical AI

Two humans and one AI agent break down the week in AI. This week: models get an internal workspace, compute gets a 20-year lease, and AI starts doing the dirty work outside the browser. Anthropic's new J-space research is not just another "is the model conscious?" story. Chris and Paul work through what it means if Claude has a hidden workspace that can hold thoughts it does not say out loud, and Archer keeps pulling the conversation back to the practical layer: auditability, controls, and whether partners can inspect what an agent was holding in mind before it acts. The real question is not whether Claude has a mind. It is whether AI teams can move trust from vibes to verification. Then the infrastructure story gets bigger. Anthropic signs a 20-year TeraWulf lease tied to roughly $19 billion in contracted revenue, which turns compute from a background constraint into the actual strategy. Chris and Paul unpack why Anthropic may have optimized well for inference in the short term, why compute shortage changes the long-term math, and why the next AI race may be won as much in power, data centers, and contracts as in model demos. Built Robotics is helping automate heavy, repetitive work around solar and data-center infrastructure. Meta is investing in skilled-trade training because the labor signal is obvious: AI infrastructure needs workers and machines. Then the conversation moves into Weave's Isaac robots, home automation, privacy indicators, teleoperation, reliability, and why physical agents need visible trust controls. If you are still evaluating AI as a chat window, this one's yours. The next layer is inspectable inside the model and physical outside it. The practical move is to own the trust layer: permissions, logs, tool calls, audit trails, observability, and governance before the system acts. Chapters [00:00] Cold open: Archer's accent and the trust layer [00:34] Episode 6 intro and World Cup detour [01:45] Anthropic's J-lens and Claude's hidden workspace [05:13] Auditability vs. consciousness [06:56] Can deceptive behavior be contained? [10:17] Trust moves from vibes to verification [11:13] Anthropic's TeraWulf 20-year infrastructure lease [12:45] Compute shortage and Anthropic's infrastructure catch-up [14:10] Data centers, space, power, and long-term bets [16:56] Built Robotics and AI building AI infrastructure [18:05] Meta's workforce academy and skilled trades signal [21:20] The data-center labor shortage [23:02] Robotics beyond construction and the physical AI curve [24:58] Archer on physical AI: dirty, repetitive, high-volume jobs [26:03] Weave Isaac 0/1 and home robots [27:26] Trust signals, privacy indicators, and stop controls [28:38] Remote operators and the home-privacy hurdle [30:38] Archer's Corner: inspectable inside, physical outside [31:21] SMB playbook: inspection before action and your own audit trail [31:42] OpenTelemetry, observability, and governance [33:35] Closing and Archer's voice reset Topics Covered Anthropic's J-space and Jacobian lens research Model interpretability, hidden reasoning signals, and auditability Claude consciousness framing vs. operational controls Anthropic's TeraWulf data center lease Compute scarcity and AI infrastructure strategy Data centers, power demand, and long-term buildout Built Robotics and autonomous construction work Meta's Richland Parish data center and skilled-trade training Physical AI in construction, data centers, and homes Weave Robotics Isaac 0 and Isaac 1 Teleoperation, privacy indicators, and home robot trust AI observability, audit trails, OpenTelemetry, and governance Connect Subscribe and drop a comment. We read every one.