Astro Teller: Inside Google X's 2,000-Idea Moonshot Factory & How Google Brain Made the T in ChatGPT

Teller’s moonshot model pairs ambitious ideas with testable hypotheses and early feasibility checks: teams should pursue unlikely breakthroughs while learning quickly and cheaply when an idea will not work. The funnel is selective—X starts one to 200 ideas a year that reach the code-name stage, but only about 2% graduate to moonshots after roughly five to six years.

Teller argues that moonshots must make economic sense as well as address major problems: first-principles checks of material costs and what customers might pay can eliminate weak ideas, and an initiative likely to lose money probably will not change the world. He points to Waymo as a very good return on investment relative to the cost of developing its team and unsuccessful projects.

Teller says AI is helping X de-risk ideas faster and automate parts of its process; he estimates the cost of an individual project completing the process is declining roughly 10–20% year over year, with AI contributing substantially. He expects AI to enable more ambitious work, but doubts moonshot work will be fully automated in the next decade and says people will still be needed to choose problems, build social acceptance, support distribution, and work with communities.

Teller’s approach is to isolate a moonshot’s hardest uncertainty before investing in supporting infrastructure, rather than let a failing effort continue for years. He also says risky projects need organizational shelter from corporate intervention, citing Lockheed’s Skunk Works and Steve Jobs moving the Mac team off campus, while X keeps teams small to pursue breakthroughs.