[api]

For quite a while I've suspected that you might be able to support Feyerabend's argument using modern learning theory. In particular I suspect that this is relevant:

https://en.wikipedia.org/wiki/No_free_lunch_in_search_and_op...

Science is essentially a learning process. Since every learning method has a performance envelope and fares better against some fitness landscapes than others, restricting science to a singular method effectively prohibits learning over fitness landscapes whose structures lie outside that method's envelope.

In layman's terms: a single scientific method will be unable to learn certain things, or at least unable to learn them in a reasonable amount of time.

This is also why I am eternally skeptical of all business and management "methodologies." If there were a closed-form methodology that always yielded successful businesses there would be no entrepreneurs. Large companies and investment funds would simply execute this closed-form method deterministically and reliably pump out successes while retaining 100% ownership. Entrepreneurs exist because creating successful companies is an "AI complete problem" that requires the full multi-approach multi-paradigm multi-methodology capabilities of a human intellect... and even then it's hard.

(E.g. I put down Lean Startup when I realized I was just reading a description of gradient descent in the business domain. Gradient descent only works over very regular fitness landscapes with clear peaks and well-connected paths to those peaks. In a rocky fitness landscape you will get stuck at a local maximum almost instantly and never go any further.)

[vslira]

>(E.g. I put down Lean Startup when I realized I was just reading a description of gradient descent in the business domain. Gradient descent only works over very regular fitness landscapes with clear peaks and well-connected paths to those peaks. In a rocky fitness landscape you will get stuck at a local maximum almost instantly and never go any further.)

Funny, that's exactly how I interpreted the book, but I didn't see that as a bad thing. Of course, a naive gradient descent won't solve everything, but will help on a lot of things satisfactorily. Maybe LS won't help you build the next Airbnb, but not all business must reach Unicorn Status. Pretty clever on Ries' part

[EdHominem]

> Science is essentially a learning process.

Science is a testing process which we engage in to learn. Minor, but critical difference.

> a single scientific method

"Hypothesize and test." That's it, there is only one.

Now, which strategy do you use to pick the best hypothesis, and to test faster, etc? That's the part you don't want to get dogmatic about. (eg gradient descent.)