The Slower, More Expensive, and Safer Way to Develop Artificial Intelligence

Bill Hibbard   1 October 2026

The race to advanced AI among nations and coorporations is motivating them to take the fastest, cheapest path, which is large language models trained via reinforcement learning. This fast, cheap path is dangerous. There is a slower, more expensive path that is also safer. I discuss this in two opinion articles that I submitted to The New York Times and The Guardian.

My opinion articles say that AI should choose actions that lead to a more highly valued world, or make the world a better place. The articles don't go into detail about how these valued/better futures are defined. There is much more about that in my AGI-12 paper, Avoiding Unintended AI Behaviors, and in my 2014 on-line book, Ethical Artificial Intelligence. There is an easy overview of these sources in my comparison of them with Stuart Russell's book, Human Compatible: Artificial Intelligence and the Problem of Control.