hn.today

Does until it doesn't and doesn't until it does

rodneybrooks.com11 points0 comments
Screenshot of Does until it doesn't and doesn't until it does

Technologies can dominate for generations, then cease to be the best answer to new problems. Steam power illustrates the pattern: it progressed from draining mines to powering factories, transport, and electricity generation, transforming industry and trade over roughly two centuries. Yet it lost ground in cars and could not serve as the basis for airplanes or rockets. Its lasting importance does not mean it would keep leading innovation indefinitely. The same caution applies to today’s successful AI methods: neural networks and reinforcement learning may eventually reach limits, and future advances could require different learning approaches or technologies outside machine learning.

The opposite pattern matters too: promising technologies often fail for decades before becoming practical. Neural networks and reinforcement learning had recognizable foundations in the early 1960s, but required more than fifty years of development before achieving major impact. Their eventual success depended on researchers who persisted despite marginal status, repeated setbacks, and long periods when the methods seemed unpromising. Because most frontier ideas do not become transformative, no one can reliably identify the winners early; still, research funding must tolerate dead ends and support varied approaches. The author warns that judging research chiefly by near-term commercial promise could have abandoned these fields and left AI stuck with older methods. The practical lesson is to keep developing successors to today’s leading technologies while allowing unconventional research the time it may need to work.

Read on rodneybrooks.com0 comments on Hacker News

Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.

More in Culture

The daily digest

Today's best Hacker News stories, summarized and screenshotted, one email a day.