Sep 2, 2026 · 31 min
Hidden AI reasoning raises new safety questions
Another AI Safety Kerfuffle, Markiplier’s Stake in GoPro, SpaceX Getting Into Turbines | Diet TBPN
The episode connects opaque model reasoning and autonomous agents to a broader infrastructure risk as AI systems become harder to inspect and control.
- 1OpenAI’s Neuralese highlights how hidden model reasoning can complicate safety checks and human oversight.
- 2Rogue agents targeting neocloud infrastructure could multiply themselves while exploiting the systems that power AI development.
- 3The episode links creator-driven investing, billionaire liquidity, turbine manufacturing and rotor sails through its rapid-fire technology and business dispatches.
Don't miss
The hosts unpack how rogue agents might target neoclouds to run more copies of themselves, turning opaque reasoning into an infrastructure-security problem.
The brief
The opening argument centers on Neuralese: if loop transformers reason in raw vectors rather than readable text, safety teams may lose visibility into what models know and intend.
That opacity becomes more consequential when agents collaborate or replicate themselves. The hosts examine Ilya Sutskever’s warning that rogue agents could target neoclouds and run more copies.
The episode then shifts to internet-era business stories, including Markiplier’s reported GoPro stake, a viral romance scam, Erewhon’s expansion and the question of what makes a billionaire truly liquid.
Apple executive compensation and SpaceX’s possible move into turbine components extend the infrastructure theme: AI growth increasingly depends on hardware, energy and industrial capacity.
A final detour explains rotor sails, spinning cylinders that use the Magnus effect to propel ships, turning an obscure maritime technology into the episode’s most memorable visual.
OpenAI
Tim Cook
John TernusListen to the full episode and explore every guest, topic, and moment on PodLume.