We Should Be Cautious When Harnessing Evolution

A while back I wrote that Evolutionary Algorithms Could be More Significant Than Machine Learning. The reason for this is that they keep improving towards a specific goal, without additional input, by leveraging the powers of evolution, i.e., reproduction, variation, heritability, and differential success.

Reinforcement Learning is considered a subset of Machine Learning by many, so this isn’t technically a one-vs-other scenario.

This means that evolutionary algorithms, properly configured, can produce extraordinary results in a very short time, and the problem is that we won’t always what they’re going to create—or what impact that creation will have on society—before it happens.

Two recent examples:

  1. Deep Blue took over 10 years, and required many experts and millions of dollars to create before it could beat a human at Chess. Alpha Go beat the best human at Go, but it took dozens of engineers giving their experience and programming, required 140 Google CPUs, and took hundreds of hours to finish. Alpha Go Zero beat Alpha Go in 3 days using only 4 CPUs and didn’t require any initial training whatsoever.

  2. Facebook uses many of the concepts of an evolutionary algorithm to constantly adjust content and UI/UX to make people spend more time in the platform. We thought we were building a useful social media platform, but what we actually built was a giant sinkhole of human attention that seems to be causing significant mental health issues.

Bret Weinstein talked about this point of social media and addiction situation on Sam Harris’ podcast.

The obvious and much discussed case is Artificial General Intelligence (AGI), which will be able to use these same techniques to improve itself very quickly, but there are many more present, near-term, and realistic examples that we need to watch for as well.

The takeaway here is quite simple: we need to be very careful about creating anything that both self-improves and has significant interaction with humanity. Failing to do so can product all manner of harm to us, from an annoyance (really good advertising) to catastrophe (AGI without morality).

Use. Evolutionary Algorithms. Wisely.

Related posts: