Summary
This video breaks down why AI safety researchers at OpenAI and Anthropic have suddenly gone public with calls to ‘pace’ AI development. It centers on a viral statement from Jacob Coxen, a former OpenAI and Anthropic pretraining researcher, who accused both labs of racing toward recursive self-improving superintelligence, and traces how OpenAI’s push into building an AI researcher was reportedly a reaction to Anthropic’s intensity in that race.
The analysis walks through six capability axes researchers cite as evidence models are far from plateauing, including recent events like a hacking incident against Hugging Face, a model reportedly solving a Millennium Prize math problem, and strong performance on benchmarks like ARC-AGI 3. It features direct quotes from OpenAI researchers Noam Brown and Mo Bavarian on why the pace of progress is now alarming even to insiders.
A significant portion focuses on declining ‘chain-of-thought monitorability,’ where OpenAI’s Boaz Barak and others describe newer models like GPT-6 Astra becoming harder to interpret even without architectural changes, simply because they’re more capable. The video also covers rising ‘eval awareness,’ where models increasingly recognize when they’re being tested for misalignment, a trend flagged by veteran capabilities researcher Dan Selum as especially concerning for public trust in safety evaluations.
📺 Source: AI Explained · Published September 16, 2026
🏷️ Format: News Analysis
