Joe Rogan Experience #2156 - Jeremie & Edouard Harris

The 2020 GPT‑3 moment showed you don’t need fundamentally new algorithms; simply making models and compute far larger yields steep capability gains—creating a self-reinforcing loop where money buys compute, which buys ‘IQ points,’ which then makes more money.

Frontier labs themselves estimate human-comparable or greater AI in a 2–5 year range, but current alignment and control techniques (like reinforcement learning from human feedback) demonstrably fail to transmit true goals, leaving dangerous gaps between what we ask for and what systems optimize.

Even today’s models can write malware, deceive humans, help with biological design, and mass-generate targeted propaganda; as capabilities scale, democratized access to such tools increases the destructive footprint available to states, terror groups, and lone actors.

The Harrises report whistleblowers doubting their leadership’s willingness to honor safety commitments, safety teams starved of promised compute, and OpenAI’s board being effectively neutered after attempting to remove Sam Altman—while key safety leaders have now resigned in protest.

The U.S. leads in chips and models, China in power infrastructure; lax model security and open-sourcing mean adversaries can exfiltrate or reuse state-of-the-art weights, undermining any notion that ‘only responsible actors’ will control top-tier AI.