On Gut Instinct

2023.10.22

Lindy Expectancy: 34 Weeks

Go with your gut. I’ve heard it a million times; it’s one of my bigger areas for growth. In a vacuum, going with your gut is relatively easy… but, nothing is a vacuum.

Here, I talk about how I view gut instinct, and start getting into (just a little) how I’m working on leaning into mine more.

Large Language Models

You’ve probably toyed around with ChatGPT or Claude before; undoubtedly, you’ve heard of them.

They’re both Large Language Models (LLMs). A simplified explanation of what that means: you give the LLM a sentence or paragraph, and, based on its training data and fine tuning, it predicts what the response should be.

Re the training data, oh boy oh boy do these models have a lot. If you ask GPT-4 (go ahead, try it), it claims that it was probably trained on over a trillion words. I also fact checked that with the internet. Great, right? Fact checking the thing that everyone says you need to fact check by checking another thing that everyone says you need to fact check.

Remember in algebra, when we would be given 20 or 30 points and find the line of best fit?

You would draw that line to effectively predict where any other given point following that same function might land. Once you have your line of best fit, given any x value, you can (hopefully) predict the next y value.

An LLM is not too terribly different from that, but instead of inputting a number as the x value, you’re inputting text. And instead of a straight line, the line of best fit can be squiggly. And instead of one line of best fit, there are a bunch of them stacked on top of each other. And instead of just one column of lines of best fit, there’s a bunch of sequential layers of them. Oh, and each line of best fit has a lot more dimensions than two.

All of those lines emerge from training. The model learns to identify patterns that indicate a likely outcome.

And then, in many situations, the model will be fine tuned on further examples… something outside the system (often a human or group of humans) gets to observe the performance and feed it in new, more accurate examples to bring it closer to the desired output.

That’s your TL;DR for LLM’s. Not critical that you know a whole lot about them as we keep going other than the fact that they’re trained & fine tuned on a bunch of examples.

Trick Or Treat

What the hell does an LLM have to do with gut instinct? Pretty much everything.

An LLM is trained on all of these countless examples. You, if you’ve been around for long enough, are trained on examples, too. The examples we as humans get trained on are a bit more like fine tuning the model, though… we try something, we see if it succeeds or fails, we use that data to adjust our decision making model, and we try again.

Let’s say 8 year old you, is for the first time ever dropped in a US suburb on October 31st with no cultural context whatsoever. You see other kids dress up in costumes and knock on doors and get candy. That’s great, you decide to try it out. Free candy? Why not?

That first day, it works pretty well. So, you update your decision making model: now, you know that if you dress up and knock on stranger’s doors, you’ll get candy! Let’s do it again tomorrow. What could go wrong?

The second day, less neighbors open their doors and not so many kids are out anymore, but you still get some candy. By November 5th, you’re the only one trying to trick or treat anymore, and the neighbors just stop answering the door.

Okay, so maybe it was a one time thing. You fine tune your model. You update your decision making process. No more free candy. Sigh.

About a year later, when you’re 9, you see other kids start to dress up and get candy again. Wow! Here’s your chance–you follow suit again, but only for a couple days this time–remember, you’ve updated your model. Last time, the free candy only lasted for one or two days. This time, you only even try to get it on one or two days.

By the time you’re 10, eureka! You’ve discovered the pattern. Once a year, at the end of October, you can dress up and trick or treat. Your model’s now pretty damn fine tuned. Everything’s looking great… after all, you have a lifetime of free candy ahead of you.

However, now, you’re 18. Unfortunately, the adults seem a bit more reluctant to allow you to participate in this magical holiday. Your time has come. On your 10th example of Halloween, you get your most heartbreaking model update of all: only children are allowed to trick or treat. No more free candy for you.

Time marches on and on; it has no sympathy for your sugar lust.

Weights

Jumping out of a plane is more memorable than jumping down the last two steps of a staircase. Your first kiss is more memorable than your 32nd. And, in the span of one night, your first glass of wine is more memorable than your sixth, but the effects of the sixth glass on the next day are certainly more memorable than the first.

Some things stick out in your experience more than others do, and those things impact the fine tuning of your decision making model. And, yes if I haven’t made it clear yet, your decision making model is your gut.

LLM’s can have different weights for different examples, but perhaps here is a critical distinction between us and them–our learning has physiological weights added to it. Our emotional state impacts how we process things that happen to us.

In the Hour Between Dog and Wolf, John Coates talks about how our bodies are instruments that respond to the environment and danger; those emotional and physiological responses are part of how heavily we weigh data points in our own decision making models.An LLM doesn’t have that dimension of physical responses to train or fine tune on.

Of course, this can be a burden and a curse; I think PTSD can to some extent be viewed as overfitting our decision making models to the most intense emotional events we experienced.

What’s the point of this? You have emotional responses when you lead life. A good movie or book can give you an emotional response, but nothing gets the blood flowing like actually being out in the real world with lasting consequences. Our training data is incomplete (maybe we’re just LLM’s) without really living life.

Being Cerebral

An utter lack of self awareness in an 8 year old in the trick or treating example above would result in him trick or treating in perpetuity as long as enough neighbors kept giving him candy.

You absolutely need feedback from the world to make better decisions, to update your gut. You need to see how our actions affect yourselves and others, particularly when those actions are habits repeated over the long term. That means we need to be making those decisions in the real world. The more you make, the better you’ll get.

Advice and wisdom from those you know and trust is also essential to fine tuning and maybe seeing the longer term effects of our actions, as those aren’t always easy to ferret out in the moment.

However, you can’t defer to everyone all the time, especially when yourself interests are not aligned. And, sometimes, interests may be much less aligned than they seem like they are on the surface.

You need to trust your gut at some point, too. No one has the same training set and fine tuning as you do. They don’t have the same experiences and feedback; it’s impossible. You need to make decisions, and, when you’re wrong, learn from it. It’s just another data point; you made the decision on the best information you had at the time.

I’m thinking about it as jumping into an ice bath. If you sit on the precipice forever, you’ll never actually get in. You just have to go for it.

Decision making is hard. You can’t let someone else make decisions for you, though. You have to do it for yourself, and then you have to learn from it.

You need to go with your gut. I’m working on going with mine.

Live Deeply,