In the Long Run

In this episode of In The Long Run, we discuss Anthropic’s growing role in AI, from its collaboration with the Vatican on ethical questions to its compute deal with xAI. We also explore Andrej Karpathy joining Anthropic, the tension between AI safety and commercial incentives, SpaceX IPO rumors, and a new software engineering benchmark challenging how we compare AI models.

Agentic AI is moving from clever chat to real work. In this episode, Jim and João discuss GPT Image 2.0, Codex, GPT-5.5, agents in ChatGPT, and why these tools matter beyond software development. They explore how agents plan, use tools, browse, test, create, and iterate across tasks: from financial reports to Sudoku scraping. The bigger question is enterprise readiness. Large companies must now solve identity, access, data classification, governance, logging, auditing, interoperability, and security for agents working across systems, suppliers, and teams.

This week's episode opens with the Artemis II launch as a jumping-off point: progress doesn't happen automatically, it takes investment and will. Jim and João then unpack Alberto Romero's "How to Survive the Age of AI", drawing a sharp distinction between automation and systemic disruption. It's not the ATM that killed banking, digital banking did. The article's tips spark debate: don't obsess over AI skills, keep some tasks AI-free to stay sharp, and know when to stop prompting and start thinking for yourself.
On the news front: Anthropic's Claude Code source code leak (human error, apparently) raises questions about accountability as AI agents take on more sensitive roles. OpenAI's Sora social network quietly winds down, too expensive, too unfocused. A Financial Times analysis finds AI chatbots are nudging people toward the political center, a striking contrast to social media's polarizing pull. And Elon Musk's TerraFab lands with Kardashev-scale ambition: own the chip supply chain, harness the sun, build for a galactic civilization. Bold or classic Musk overreach? Time will tell.

Brain cells on a chip learning to play Doom: early signs of biological computing as an alternative to silicon chips
Embodied intelligence through a simulated fly: testing whether cognition needs a body and environment
Brain emulation and when can we upload our mind into a computer?
AI supporting personalised cancer treatment for a dog
AI running its own research loops: Karpathy’s autoresearch as a glimpse of autonomous experimentation
1M context windows: longer memory changing how AI systems handle tasks and reduce manual setup (RAG)
Jensen Huang creates "Safe" version of OpenClaw

In this episode of "In the long run", we unpack the escalating clash between the U.S. “Department of War” (Pentagon) and Anthropic, centered on whether a private AI company can restrict government use of its models, especially around mass surveillance and fully autonomous weapons.
We explore how today’s “legal” frameworks lag behind AI capabilities, making large-scale surveillance technically easy (and sometimes still lawful).
Finally, we connect the rise of AI coding agents and “vibe coding” to a shake-up in SaaS. How will this impact Software as a service companies, how many subscription apps could become trivial to recreate with personalized agents, pushing SaaS value toward support, accountability, and enterprise-grade reliability rather than basic functionality.