Commenting on Jacob Coxon Leaving Anthropic
- Philippe Beaudoin
- September 10, 2026
- Radio-Canada Première Chaîne
I joined Radio-Canada’s Première Chaîne twice to discuss the public resignation of AI researcher Jacob Coxon from Anthropic. I first spoke with Marie-Madeleine Villeneuve on Phare Ouest (British Columbia), followed by an in-studio discussion on Ça Nous Regarde with Madeleine Blais-Morin. Coxon’s departure brought renewed public attention to warnings that frontier AI models could slip beyond our control far sooner than anticipated.
Across both segments, I emphasized that the unease expressed by insiders like Coxon is no longer marketing hype—it stems from observable behaviors in the lab, particularly with flexible multi-agent architectures capable of autonomous deliberation and emergent problem-solving. During our discussions, we touched on several key dimensions:
- Alignment vs. Recursive Acceleration: Distinguishing between the alignment problem (ensuring systems respect implicit human norms, much like assuming a taxi driver won’t run red lights to reach an airport faster) and runaway self-improvement (models iteratively modifying their own code).
- Tangible, Near-Term Risks: Why immediate vulnerabilities—such as autonomous cyberattacks and systemic workforce displacement—are already eroding the traditional split between economic optimism and catastrophic safety concerns. I cautioned against assuming past industrial transitions will repeat, given that these models approach human-like cognitive flexibility.
- The Limits of Self-Regulation: Why competitive pressures force even safety-conscious labs into a race dynamic, highlighting the need to revive public oversight and legislative frameworks (such as revisiting Canada’s stalled legislative efforts or looking to the EU AI Act).
- Uncomfortable Philosophical Questions: How confronting systems with general flexibility pushes research circles—like the Canadian AI Safety Institute—to grapple with fundamental questions about machine capability, agency, and consciousness.