jtbooth.com/blog/bottleneck.html

Scientific Bottlenecks

In Is Science Slowing Down?, Scott argues that we should expect science to slow down on a per input basis - because if it didn't, we'd have long since coquered the stars. I still can't help but feel there needs to be "a reason" that science is slowing down - something we could, in theory, try to "fix", or at least measure.

Scott considers some bad explanations and then the low-hanging fruit hypothesis - that making advances is getting harder, which explains the slowdown. This is plausible, but requires a pretty precise difficulty curve. Scott lists a lot of examples - and I found many more - of basically linear-looking progress in the face of exponential increases in investment. That's weird, in just the same way that it would be weird if science didn't slow down at all. Why should the difficulty curve of so many research projects line up so well with our own exponential growth in population and wealth to set against the problems?

The Development Speed Limit

I assert that it is hard for two people to revolutionize a single field at the same time. Or, equivalently, a single field of practice can't internalize more than one breakthrough at a time. When Bessemer learned how to produce cheap, high quality steel, it took decades for the industry to adopt the process, rebuild the factories, and get widespread adoption in construction. According to the wikipedia article, after Bessemer's patent in 1854, the Brooklyn Bridge was built with crucible steel, rather than the new Bessemer steel, in 1877.

What benefit would we have gotten from a hundred Bessemers? In a sense we had two or three - another inventor, William Kelly, contested Bessemer's patent, possibly based on a similar Chinese technique. With a hundred researchers on the same question, the best process might have been a bit better than Bessemer's. Multiple companies licensing steel refining patents might have managed to spread the word somewhat faster. But I doubt the industry would have been saturated so much faster - maybe it takes fifteen years in stead of thirty, but it's very difficult to imagine five years.

And until this adoption is complete, progress in steel smelting is largely frozen. All previous research into improving the quality of crucible steel has to be retooled for the Bessemer process. The difficulties faced were around controlling the mixture of gases in the air blown over the steel as it cooled - a step that isn't included in the crucible process. To further improve the Bessemer process, you need to learn it, rebuild your factory to use it, become expert, find flaws, experiment, then publish your solution and teach everyone else in the industry. You can really only make one breakthrough at a time.

I prefer this theory to the "exponential increase in difficulty" theory because it allows the possibility of a small group of people making consistent progress over time, even to great heights. All you can do with more people is run the clock faster - adopt the new methods faster, and run more parallel searches for the next improvement. But that cycle time - the time to get everyone involved up to speed on the new methods - is the development speed limit.