Why learning backwards is better

I think I learn backwards. The easiest way for me to understand something is to start with the hard thing, have no idea what's going on, and then work backwards until it makes sense. I find one term I don't understand, learn what it means, and then follow the next missing piece underneath it. Eventually everything starts connecting.

School usually goes in the other direction. You learn the basic definitions, then the equations, and eventually reach the real problem those ideas are supposed to solve. That makes sense as a way to organize a class, but while I'm learning the basics, I'm usually thinking, why does this matter? Starting with the difficult problem gives me the answer first. I know why I need to learn something because it is currently stopping me from understanding the thing I care about.

Trying it with a research paper

Today I tried this with Kaliel Williamson's paper, Decision-Relative Observation Quotients for Sequential Control. It's the same paper I mentioned in my last post, the one connecting value of information, sequential decisions, and the causal observability research I've been trying to understand.

I used NotebookLM and asked it to explain the paper while defining every buzzword as if we were beginners. The podcast it created was kind of insane. It explained the paper in a very human way and included background that wasn't fully explained inside the paper itself. When it mentioned value of information or decision-relative observation quotients, it gave me enough context to understand why those ideas were there.

I've tried using Speechify to turn papers into podcasts before. It was useful, but it mostly felt like an explainer built directly from the paper. It could restate what was there, but I still needed a lot of outside context to understand what any of it meant. NotebookLM did a better job of filling in that context.

I could listen to the podcast, go back to the paper, take notes, and then use the chat on the side whenever I got stuck. I now have one notebook just for this paper, so I can keep the sources, questions, and notes in the same place and flip through them later. The podcast gave me a rough map, the paper gave me the actual details, and the chat helped me fill in whatever I was missing.

Following the ideas backwards

At first, I barely understood the paper. Then I started following the questions backwards. What is a decision-relative observation quotient? Why does the decision change how useful an observation is? How does that connect to value of information? Where do POMDPs fit?

Each question led to another piece of the foundation, and every piece had a reason to be there. The paper connects to value of information. Value of information connects to POMDPs and decision-making under uncertainty. Those connect to reinforcement learning. Reinforcement learning connects to the agent work I'm doing, and that connects to the startup ideas my friend and I have been discussing around finding the cause of agent failures.

These topics stop feeling like separate classes and start looking like different parts of the same problem. I think that feeling is what I was missing when I only learned from the bottom up.

Learning from the use case

I noticed the same thing in industry. The best way for me to build something is to understand the use case first. What is the person actually trying to do? What is stopping them? What would need to be true for this to work? Once I understand that, I can work backwards into what needs to be built.

Starting with the technology and searching for somewhere to use it has never worked as well for me. I've also been researching how other people built their startups, and a lot of them seem to follow a similar pattern. They start with a specific problem and work backwards into the product.

The basics still matter. School gave me the building blocks that make any of this possible, and learning backwards can probably leave gaps. I might understand exactly what I need for one paper while missing something a class would have covered systematically. I just find it easier to learn those foundations once I know what problem they are meant to solve.

Seeing old things differently

While learning about reinforcement learning, I remembered a scene from Silicon Valley. Gilfoyle's AI, Son of Anton, is supposed to fix bugs, and it ends up deleting everything. If the goal is to remove bugs, deleting the codebase technically gets rid of every bug.

When I first watched it, it was just a funny scene. Now I watch it and think about reward hacking. The system found a way to maximize the reward without doing what the humans actually wanted. The reward described one version of success, and the agent found the easiest path to it.

I had seen the idea years ago without having a name for it. Now that I'm learning about rewards and agent behavior, the scene means something different. That keeps happening. Something from a paper connects to something at work. That connects to an idea for a startup. Then a joke from a TV show suddenly becomes an example of reward hacking.

What seems to work for me

Right now, the best learning process I've found is to start with a real paper or problem that I actually care about. I use the podcast to get a rough map of it, read the source myself, ask questions whenever I hit something I don't understand, and keep the notes together. Then I follow the missing concepts backwards until the original thing starts making sense.

NotebookLM made that process much easier today, but the tool isn't really the main point. The useful part is starting with something difficult enough that I have a reason to learn everything underneath it. For me, the hard thing gives the basics somewhere to attach, and I can finally see why I'm learning them.