I’ve been a collector and connoisseur of fine concepts for some time now and it doesn’t happen often that I come across one that is entirely new to me, but this happened yesterday.
Adam Mastroianni on Experimental History writes about strong and weak link problems, and rather than try to summarize his excellent post, I’m going to copy the table he made to define the two kinds of problems and add examples for each.
| When you have a Strong Link Problem, you want to: | When you have a Weak Link Problem, you want to: |
|---|---|
| Increase outliers/variance/weirdness because you’ll benefit from having more very good things | Decrease outliers/variants/weirdness because you’ll be harmed by having more very bad things |
| Don’t gatekeep because you might accidentally keep the best out | Gatekeep because it keeps the worst out |
| Ignore the worst | Improve the worst |
| Improve the best | Ignore the best |
| Accept risk, because the downside doesn’t matter | Avoid risk, because the downside is all that matters |
| Examples: 1. Music (you listen to what you like and ignore the rest) 2. Olympic success (a country’s medal count hinges on its top athletes) 3. Venture capital (one billion-dollar winner can offset many failures) 4. Breakthrough research (one paradigm-shifting discovery drives progress) 5. Literary reputation (a single classic novel can define an author’s legacy) 6. Art exhibitions (one masterpiece elevates the whole show) 7. Sports teams (star players often dictate the game’s outcome) | Examples: 1. Food safety (a single contaminated ingredient spoiling a dish) 2. Engineering (a broken transmission in a car engine ruins a car even if the rest of the engine is in great working order) 3. Nuclear security (one lax facility can endanger everyone) 4. Construction (the weakest beam determines overall safety) 5. Cybersecurity (a lone vulnerability can compromise the entire system) 7. Supply chains (one failing supplier can disrupt the whole production line) |
For many of the most interesting problems, reasonable people will disagree if they should be strong or weak link. Education, the economy and the welfare state come to mind.
One response to “Strong and Weak Link Problems”
[…] that getting AI to be more powerful (both in the sense of more useful and more dangerous) is a weak link problem. Improving the abilities they already excel at, such as math, reasoning or writing, matters less […]
LikeLike