What happens when AI builds itself?

当人工智能自我构建时,会发生什么?

Babbage from The Economist

2026-06-17

49 分钟
PDF

单集简介 ...

What happens if AI models get so good that they can build their own successors? So model A builds a better model B, which then builds an even better model C, without any humans to slow things down? The answer could be a high-speed rush towards an AI superintelligence that is beyond human understanding, and possibly beyond human control. This process, known as recursive self-improvement, may be much closer than people realise – perhaps just a couple of years away. So is the world ready for an artificial intelligence explosion? Guests and hosts: Max Tegmark, MIT physics professor, co-founder and leader of the Future of Life InstituteAlex Hern, The Economist's AI writerAlok Jha, The Economist's science and technology editor  Topics covered: Recursive Self Improvement in AIUS export ban on Anthropic's Fable 5 model Transcripts of our podcasts are available via economist.com/podcasts. Listen to what matters most, from global politics and business to science and technology—subscribe to Economist Podcasts+. For more information about how to access Economist Podcasts+, please visit our FAQs page or watch our video explaining how to link your account.
更多

单集文稿 ...

  • This episode of Babbage is supported by IDA Island.

  • With the highest share of STEM graduates per capita in the EU,

  • IDA Island can help source the skills you need to internationalise and thrive.

  • Visit IDAisland.com to learn more.

  • AI models can already do lots of useful things.

  • You probably used them already to summarise meetings or books.

  • Mathematicians are finding them useful in solving age-old problems.

  • Scientists are using them to understand complex proteins and design new drugs.

  • Each iteration of the frontier models brings what can seem frankly like astonishing, superhuman new capabilities.

  • The cleverness of these models is approaching a point that computer scientists have long wondered about.

  • Could an AI model ever design and build an improved version of itself?

  • Just think what that would mean.

  • Imagine AI models working night and day to improve themselves without people slowing them down.

  • So version one designs version two, version two designs version three and so on.

  • Now this closed loop known as recursive self-improvement would mean that everything

  • you've heard so far about progress in AI could accelerate even more.

  • What might have taken decades could become possible in just years or a few months.

  • A self-improving AI might be the last machine humans would ever need to build.

  • Optimists predict that self-improvement would lead to something called an intelligence explosion.

  • Pessimists worry that it would be the first step towards something a lot darker, a technological singularity.