On Rules Vs Intuition

2024.10.13

LXVIII

Building intuition and following rules are both valid and important for learning, both in humans and in AI.

Modeling the World

We try to model the world. Sometimes we do it explicitly with rules, other times we do it implicitly with intuition. 

Explicit rules* are cool because they are very easy to communicate, you can learn them quickly, and you can easily trace back to them when something goes wrong. And, a lot of the time, they’re pretty damn accurate. The drawback though is that even with a ton of difficult to navigate rules, it would be hard to capture all of the nuance of the world.

Intuition is cool because it can capture your experience and tons of nuance without having to explicitly write it out. And, after all, if something works, it doesn’t always matter why. On the other hand, when intuition goes wrong, it can be hard to see why, which can be frustrating. And you don’t really get an assurance that you can avoid the same mistake again the same way you do when you update a rule system.

While it’s tempting to turn this into a debate of rules vs intuition, it really seems that most experts I know actually blend intuition and rule following–they quickly come to really useful conclusions without navigating some convoluted 30 layer pyramid of rules. Still, when asked, they can explain why their decision lines up with a pretty straightforward rule based view of the world. 

In reality, I think rules form the skeleton that we complete with the flesh of intuition. Weirdly enough, this almost seems inverted with AI models.

*In general, I am a “rule” breaker, in the sense that I don’t particularly pay heed to “rules” imposed by others. In general, I am also a “rule” follower, in the sense that I try to follow self imposed rules I’ve largely copied from people with life outcomes I admire.

Jiu Jitsu Rules 

A very impactful rule I learned early in jiu jitsu was to protect my neck and arms at all cost. 

Humorously, we call this the “Zombie T-Rex Rule.” You hold your upper arms tight against your torso and your forearms bent up towards your neck. Speaking of the neck, that should be scrunched up with the chin tight against the chest, as you might think a zombie t rex would look like.*

Early on, the Zombie T-Rex Rule gives you a solid defensive foundation that, as a white belt, you can use to confound blue belts wearing blue belts and, ever so often, the occasional blue belt wearing a purple belt. 

However, this notion on its own is really not sufficient. It says nothing about the legs, which your opponent can also attack. Also, if you were to strictly adhere to the principle forever, you would very rarely find an appropriate counter attack, as many such movements violate the Zombie T-Rex Rule.

When you’re starting, though, these other things don’t matter so much. If you don’t start defending in some way, then your ability to attack is irrelevant.

*The less PC, less cartoonish, and more helpful version of this meme is to be like a certain theoretical physicist who had a bit of a fixation on a theory of everything; it makes it easier to picture what you should be doing with your neck. 

Update the Model

As you’re defending, you start building up an intuition for when your opponent is in a not so good position. This can help you add useful complexity:

Zombie T-Rex unless violating Zombie T-Rex is likely to advance your position. 

Something is likely to advance your position if it’s more likely that your move works than it is that it backfires and your opponent advances their position. However, this is meaningless if you don’t know if something might advance your position.

One way to get from “Zombie T-Rex” to “Zombie T-Rex Unless…” is to test the adoption of new rules. Maybe you add in one extra move: if Zombie T-Rex, but opponent does x, then I can violate Zombie T-Rex to do y. You do this until that new move is now intuitive, and then you add another new move. Rinse and repeat until you really do have “Zombie T-Rex unless violating Zombie T-Rex is likely to advance your position.”

In this way, you’ve built something of an intuitive understanding of when to violate zombie t-rex, and you can go back to one or two rules and explain your decisions. However, you’re not actively walking around with a huge decision tree in your head, you’ve compressed it into “intuition” and likely added in new paths and conditions that you never had to explicitly learn. 

AI & ML

When it comes to codifying a complex task, there is a similar interplay between using rules to get the job done and just renting out the intuition of an LLM like GPT. However, it almost feels inverted.

“Rule Based AI” is when you encode rules to help you come up with an answer. To get text out of webpage, BirdDog uses one of the BoilerPy libraries. This is a lot of coded rules for extracting text; if we had a problem with it, we could identify the rule we disagreed with, fork the repo, and rewrite it to follow our new rule. 

It’s efficient for computers and humans to run; its fast and doesn’t require manual configuration.

LLM’s like GPT fall in a subset of a subset of AI, Deep Learning, which is itself a subset of data based Machine Learning. BoilerPy is loosely data based in the sense that the rules came from a human doing a mix of reasoning with trial and error and then applied that feedback to manually change the rules. Something like an LLM is data based in the sense that the model can take data as an input to use to update itself. By updating the data, you inherently update the model.

In this sense, I think that an LLM is much more like a human’s “intuition” than rule based AI is. Which is kind of strange in that this puts a good “intuition” one api call away.

Using an LLM as good as GPT would be like loaning out an above average jiu jitsu player’s intuition. You don’t really have authority to update the intuition, but you get to see some pretty solid play. So, maybe you’re slowly trying to back solve for rules that improve or obsolesce the intuition.

I mean this both on the level of using GPT to help you do a manual task and putting GPT in a pipeline to process something. It’s insanely useful. I think it’s even more useful as you further approximate the behavior of your ideal outcome with robust rules. You don’t have an expert on call to have an expert on call, you have an expert on call to increase accuracy and reduce dependence.

GPT is great for MVP’s and seriously complex task, but I think a mature BirdDog will have a lot of small fine tuned models & rules in the pipeline. That, or the big models keep getting cheaper and more energy efficient and maybe it doesn’t matter.

Live Deeply,