AI Is Starting to Improve AI

In 2025, an AI found a way to make the training of Google's Gemini about one percent faster. A small number, but it means an AI improved the process that builds AI. So what happens if that loop keeps going? This idea is called recursive self-improvement, and it goes back to 1965, when Jack Good, a mathematician who had worked with Alan Turing at Bletchley Park, described an "intelligence explosion". In this video we look at how AI learns in the first place, how systems like AlphaGo Zero got stronger by playing against themselves, how AI now helps design models, chips and code, how fast the loop could really spin, and what it would take to keep it under control. Chapters 0:00 AI that builds AI 2:35 How AI learns in the first place 4:28 AI that learns from itself 7:00 AI that designs AI 9:38 How fast could the loop go? 12:07 Keeping it under control Sources include: I. J. Good (1965), DeepMind's AlphaGo Zero and AlphaZero papers, Microsoft's phi-1 paper, Anthropic's Constitutional AI paper, Shumailov et al. in Nature (2024), Zoph & Le (2016), Google DeepMind's AlphaEvolve announcement (2025), the IEA's Energy and AI report (2025), and Epoch AI. If you liked this one, subscribe for more videos on math, computers and AI. #ai #artificialintelligence #machinelearning #recursiveselfimprovement #alphago #computerscience #technology