The Moonshot Hypothesis

I am not an amazing cook.

A few weeks ago, in one of the lockdown-hazes I often find myself in these days, I was cooking some garlic kale. It’s really quite simple: I stream the kale, I sauté the garlic, then I sauté the kale a bit with the garlic. Great. Except this time, I had steamed the kale and I had chopped the garlic without sautéing it, then put the kale into the hot pan.

What should I do now? The obvious thing would be to pour the kale back into its container, then saute the garlic, then add the kale back in. But when I looked at the bowl I had used to hold the kale I realized something: it was too small for me to nicely poor the kale into, because some would fall over the edges of the wide pan I was using. I would have to get a bigger plate to pour it into, and that would be another thing to wash, it just seemed like too much. I was at an entropy maxima, and if I wanted to reduce the entropy in the system I would have to put significant energy back in.

So I put the garlic directly into the kale mix without letting it sauté. After maybe 2 seconds, I regretted this decision vehemently, realizing that my lizard-brain had gone down the natural entropy-maximization path without thinking. Why did I want to have basically raw garlic with my kale? Why didn’t I just spend a little bit more time and effort getting it right? One of the things that immediately stuck out to me, is that it didn’t feel like putting the garlic into the kale was an undoable state-change, when it fact it was due to the highly increased entropy of the system once they were mistaken. My instincts didn’t have access to this information, so they tried to shortcut to having the garlic and kale in the right place, as if I could then take the kale out.

A few days ago, I made almost the same mistake when making the same dish: I poured garlic over still-raw kale in a bowl. This time, because the closely-packed shredded kale formed a semi-surface, I was able to get most of the garlic back out after spending ten extra minutes. I had increased the entropy of the system, with my mistake, but not by much.

While many of these mistakes wouldn’t have occurred if not for lockdown mode, I think they would have occurred eventually with my level of carelessness. I would argue that it is precisely developing good split-second instincts as to what kind of actions should rely on higher-level processing that allows people who are already very good at something to play around with a medium. This seems like a key property to finding interesting solutions, something which people often attribute to “intelligence”. I’m very skeptical about the idea of “intelligence” so I won’t be attributing it to anything, but I do acknowledge that humans seem to have the ability to solve problems in interesting ways.

But a key property of most problem solving in humans is that to be really good at it we have to be familiar with the building-blocks. This may seem counter to such things as Raven’s Progressive Matrices (RPM)—an IQ test based on multiple choice selection of the next item in an abstract series of objects. One might ask: “Actually, aren’t humans really good at reasoning about abstractions, for instance the ones in RPM?” to which I would respond: “Who knows? RPM is how we measure how humans compare to each other in their ability to deal with abstractions. There is no ground truth.”

But most things where we recognize human genius are things which a given individual has had reliable exposure to, because they learn how to make interesting choices that allow them to explore the environment without cutting off all their possibilities. This seems to be the delta between a toddler playing with paint for the first time and ending up with a palette of muddled browns and that same kiddo a few years later painting a fresh and inspiring self-portrait; the toddler has not yet built-up a model of how to delay the increase of entropy, but the kiddo has developed techniques to silo such entropy increases.

The “maintenance of possibilities in exploration” or “delaying of entropy” can be spun on its head—instead of being described as the minimization of statistical mechanical entropy, we should think of it from the perspective of the agent who is maximizing information theoretic entropy. Options are kept wide-open (high information entropy) until the moment when an intermediate goal is decided upon and the future is planned out.

The Entropy-Map Hypothesis: one reason humans display creativity is because we learn to construct entropy maps that describe our ability to delay collapsing the possibility-space.

With the use of such entropy-maps, humans explore possibilities without killing themselves, destroying their houses, or ruining their relationships. This has quite a strong connection to Karl Friston and co.’s ideas on the Free Energy Principle (FEP), active inference, etc. [https://royalsocietypublishing.org/doi/pdf/10.1098/rsif.2017.0792], which I largely agree with. Most of what I have read on the subject deals with maintenance and survival: fair enough, those are the key aspects of life. But once the space is mapped and a generative model of where you can go and when you can’t go back is made, a key moment happens when a plan that may fail is decided as worthy of attempt and a conscious inference happens as to what steps need to be taken. If this wasn’t the case, then we would not observe chess player’s thinking too far ahead and being nabbed by an unexpected move. But we do create and attempt specific trajectories that are less breadth-first than most examples used to describe FEP.

It is reasonable enough to assume that an organism which is aware that long-term plans can lead to greater gains (and more reduction of uncertainty, as described in FEP) but require more care to execute safely, would make specific concerted attempts to enact such plans regularly but not constantly. This is not at all a challenge of FEP, which is vague enough in its ideas to be difficult to falsify. That is because it is a lens with which to view our understanding, a meta-theory that can lead us towards falsifiable predictions. I conjecture that one key to the explanation of human problem solving, will be the circumstances under which “moonshot” attempts arise, and when people feel comfortable trying them. These moonshot attempts are not repeatable and humans cannot be exposed nearly as much to them as they are to their basic subcomponents. But through extensive knowledge of such subcomponents, humans can attempt to assign credit as to the reason why moonshots fail or succeed. Many explanations will be bogus, but those who continually succeed at moonshots are viewed as “talented” and their (internal) explanations will tend to be more robust, though often difficult to explain verbally.

The Moonshot Hypothesis: the cycle by which people learn about a domain, attempt a complex “moonshot” plan, learn from their mistakes, and try again is the key to human innovation.