You Should Seek Entropy In Your Life

I was recently listening to an interview on the Dwarkesh podcast with Andrej Karpathy and there was a little idea mentioned in that interview that I think deserves a bit of meditating on. This is that you should seek entropy in your life. These are words to live by.

According to Karpathy, today’s LLMs (ChatGPT, Claude, Gemini) exhibit something called model collapse. Model collapse is the degradation of AI models as they are increasingly trained on similar data, rather than diverse data. Getting trained on homogenous texts (similar ideas, logic, structure, etc.) results in models giving uniform, lower-quality outputs across a variety of prompts and tasks. It happens particularly when the data fed to models is AI-generated content, which tends to be significantly less diverse than human-generated text from the internet. As more and more of the data being fed to a model is generated by it, you get stuck in a kind of garbage feedback loop, leading to outputs that degrade in quality and collapse around a certain set of points.

Karpathy made an interesting point about model collapse, which is that humans tend to also undergo collapse. As kids, we start out with unlimited curiosity across several subjects, thinking about all sorts of things and constantly moving from one idea to another as we get stimulated by the rich environment around us. As we get older, though, we become significantly more uniform and are exposed to significantly less diversity: we reflect on the same old set of ideas, the same kinds of thoughts run in our minds over and over again, we speak about the same things, and rarely do things outside our normal distribution. This results in degradation too: it becomes harder to learn newer things, we become less adaptable, we become less competent, and eventually, senile.

To prevent model collapse, engineers must ensure that the model is constantly exposed to new, diverse data. Exposure to new data that disrupts a current distribution — data that adds diversity, or entropy, to the character of the training set — you can generate progress rather than degradation. By contrast, if the model is the source of new training data, collapse is hard to avoid.

The same is true of humans. If we want to prevent degradation in our thoughts, as adults we cannot narrow our exposure and become the dominant generator of the thoughts that run in our heads. Instead, we have to seek entropy. We need to be constantly exposed to novel “data,” ie, we need to be constantly reading/writing/speaking/engaging with new material, ensuring that many of the thoughts we are considering in our heads are not being generated by us, but rather by external, novel sources. This gets harder as we get older, but it is key to avoiding collapse and degradation.

There’s something really deep about this idea. It doesn’t matter how smart you are or how good the “internal weights” of your neural network/brain are. If you aren’t exposing your brain to new kinds of data — data that shakes up your world model and distribution of ideas — you will inevitably face degradation and collapse on a negative feedback loop. You can’t just create totally new ideas ex nihilo that disturb your own way of seeing the world. You need fresh data that adds something new to the soup of ideas your brain is already simmering on.

Meditating on this should motivate one to seek out these sources of entropy: talk to new people, read new material, and try new things. Importantly, make sure this stuff is new. If you’ve read lots of novels, expose yourself to microeconomics. If you’re a mathematician, pick up a book on normative ethics. If you read a lot of history, maybe try and learn about some of the new breakthroughs in geophysics. If you want to stay in the same general vicinity, at least try and read from different communities or authors. And make sure you’re talking to new people, as much as you can.

I also do not hesitate to mention that this adds to the problematic nature of echo chambers. Not only do echo chambers create polarization, but they inevitably lead to degradation, as they are prime spots for getting stuck in collapse. Nothing screams homogenous data like a proper echo chamber.

One last point: adding entropy isn’t just key to avoiding degradation. It’s ultimately key to continuing to generate new ideas and being perennially creative, even within one’s field of speciality.

So get out there and seek some entropy to your training data. Please don’t collapse.