Why greatness cannot be planned (Stanley and Lehman)
Saturday November 5, 2022
There are a lot of ways to misunderstand The Myth of the Objective. I take it as a useful meditation, with aspects of The Tyranny of Metrics and Against Method, encouraging exploration. Be flexible, be curious, don't follow a plan for the sake of following a plan.
I don't recall these AI researchers ever talking about local maxima or explore/exploit trade-offs explicitly, as they seem to be trying to write for a broad audience. Natural evolution, human innovation, Picbreeder, and novelty search are their examples of unplanned greatness. They point out that evolution is about exploration as much as adaptation, and critique the dominance in the AI community of the experimentalist and theoretical heuristics.
Is a goal local (we know how to get there) or is it "great" (requiring steps into the unknown)? The authors are saying that if it's the latter, we're better off exploring via other heuristics than focusing only on a particular distant imagined goal.
They talk about "interestingness" a lot, and I think a missing note is that "interesting" shouldn't necessarily exclude a sense of what takes us closer to some distant goal.
We should also be watching for opportunities to take things in a different direction: if you find a path to something great on your way to something good, you can follow that path! (Don't let the good be the enemy of the great.)
I've sometimes been apologetic about the winding path of my life and career, as if I should have had a plan from the start and followed it without distraction. Would that have been better? I'm not sure. This book makes me think that a life or career built from a kit represents a failure to identify opportunities. What are the odds there was never a way to improve on an earlier plan? Blinders make you faster, but they limit where you go.
The book applies most to the "research" setting—where we really don't know what the future is going to hold, where we don't know the mechanics or topology of the search space. Objectives such as "hovercraft like in Star Wars" or "economy like in Star Trek." There are a lot of other things we'll find before we find these, and that's not a bad thing.
"The idea that all our pursuits can be distilled into neatly-defined objectives and then almost mechanically pursued offers a kind of comfort against the harsh unpredictability of life." (page 2)
"... meeting a constraint is much different from what is usually meant by objective-driven achievement." (page 34)
"The best way to get computation is not to force great minds to waste their lives pondering a distant dream, but to let the great minds pursue their own interests in their present reality." (page 36)
This is in the example of developing computers, but the idea is one that I feel applies especially well to Effective Altruism.
"Deception applies to becoming rich just as everywhere else. For example, what sense would it make to decline an unpaid internship doing something you love simply because it doesn't make you any more rich? In fact, if you do become rich, it's probably because you did pursue your passion, not because you pursued money per se. Passion is what drives you to that point, and then one day you might realize that you are only one stepping stone away from being rich. And that is the moment, when you are one stepping stone away, when you make the move you need to make and become rich. But becoming rich didn't guide every life decision up to that point. On the contrary, a single-minded preoccupation with money is likely exactly the wrong road to abundant wealth." (page 37)
So... This works for some people. But it certainly doesn't work for everyone.
"There's much we cannot achieve by trying to achieve it." (page 38)
"... being aimless isn't always a good idea, but when it's paired with a thirst for exploration, it might indeed hint at great potential." (page 68)
On page 70, the authors mention that DeMarco, who I know as the author of Peopleware, is also responsible for "You can't control what you can't measure" in the earlier Controlling software projects. What a different perspective, between these two books! The authors here focus on his much later Software Engineering: An Idea Whose Time Has Come and Gone?, which I like a lot.
Page 71 has a bit about evaluating higher ed with the Collegiate Learning Assessment (CLA) and the Collegiate Assessment of Academic Proficiency (CAAP).
For the CAAP, the only info I can find is on individual schools' sites (example) and the link to ACT redirects to a page that doesn't mention it. Is CAAP defunct? (Are they both?)
"Splitting experts may be more of an achievement than unifying them." (page 82)
This is about expert review of proposals. May.
"Some readers will detect a hint of Paul Feyerabend in this argument—he taught that science can't be distilled to any one objective methodology." (page 87)
Feyerabend's book on this is Against Method. It's related in the following way, I think: Feyerabend is saying we can't limit the methods for getting to the next stepping stones. Stanley and Lehman are saying we can't know which stepping stones will help us reach further stepping stones. Ignoring either will limit where we can go.
"The only principle that does not inhibit progress is: anything goes." (page 119, quoting Feyerabend in Against Method)
"The successful inventor asks where we can get from here rather than how we can get there." (page 97)
I connect this to my Is it worth doing even if it fails? In the language of the book, things that are worth doing are things that can be stepping stones: "succeed" or "fail" you learn something to build on.
"When there's no destination there can't be a right path." (page 98)
"So if you're wondering how to escape the myth of the objective, just do things because they're interesting. Not everything needs to be guided by rigid objectives. If you have a strong feeling, go with it. If you don't have a clear objective, then you can't be wrong, because wherever you end up is okay. Assessment only goes so far. A great achievement is one that leads to more great achievements. If you set out to program computers but you're now making movies, you're probably doing something right." (page 99)
It's not just fun to have interests, it's useful. It expands the space of stepping stones that you can build from.
It's also about motivation, isn't it? You're more likely to do more if you're interested in doing it. And if you do more, you're more likely to do more good things (hopefully).
If you're hiring, would you rather hire somebody with broad and active interests? (Is this some part of "Googleyness"? Looks like that's supposed to include "enjoying fun" and "evidence that you've taken some courageous or interesting paths in your life," which could be related.)
"To achieve our highest goals, we must be willing to abandon them." (page 100)
"While a new interpretation could imply that the earlier interpretations were wrong, in reality the same set of data can support more than one interpretation. So just being consistent with the data isn't enough to decide between competing interpretations." (page 111)
This relates to Characteristics of good theories, which continues to surprise me by being one of the most popular things on my blog.
"Seen this way, evolution is a special kind of non-objective search: a minimal criteria search." (page 114)
Page 126 quotes from Hooker's 1995 Testing Heuristics: We Have It All Wrong: "Most experimental studies of heuristic algorithms resemble track meets more than scientific endeavors."
This is a critique applicable to modern AI work, but modern AI is also making tons of progress, often in harmony with benchmarking. New tests are introduced, and new models come out that do them really well. How separate are research and development, really?
I think Marcus makes related arguments that this focus on benchmarks may still prevent needed and very different work... Could be. It's hard to argue with the success of the benchmark-oriented work though. I guess the real message isn't "stop that" but "let's also."
"Judging by performance is a good idea for practitioners but is shaky at best for researchers—because of deception." (page 129)
"A good algorithm isn't one that performs better, but one that leads us to think of other algorithms." (page 132)