François CholletVerified account

Sometimes it’s enough to be approximately right (better than to be exactly wrong!)

Excellent viewpoint, Francois! As you mentioned, This is especially true for AI as blackbox algorithms has to go a long way before being truely "intelligent".

I see this every day in the space of modeling mechanical failures. You can capture 80% of the failures with quick work, but catching the remaining 20% is a grind. Cost benefit analysis has to be in mind when deciding where/when to make that effort.

I remember how well speech recognition worked in 80s. Took another 30y.

The first 80% of anything is the easy bit

80/20 rule - 80% of output is achieved w/ 20% of effort. The remaining 20% of output takes the remaining 80% of effort. The origin of "Devil is in the details". W/ history of resources & finances lacking AI industry pushed hard to do the 20% to show results, hence black box proof

The new 20/1980 rule. 20/80 rule applies only for physical problems ?

Indeed it may even be more than 100x longer, based on what you built to achieve that 20%. One can lengthen the overall timeline by building a simplistic structure around the easy wins.

The even more subtle trap is to assume that solving 20% of the task isn't valuable. Most of the time, it is better to just stop there and move on to solving 20% of another task. Widely applicable to AI successes.