How do I learn new things?
I am curious about a lot of things across domains. This has expressed itself in the scholarship I received in high school (Science, with a research internship) degree I chose (Law), the career paths I pivoted to (tech), and my current focus (AI), and my general affinity towards a diverse range of books I read (see: my bookshelf).
Through exposure to different domains (humanities vs science), trial and error, and unspooling my own thread of curiosity, here’s roughly how I learn new things:
- The syllabus-driven ‘Conscientious Student’ approach
- The hypothesis-driven ‘Operator’ approach
Syllabus-driven, Conscientious Student approach
This approach is about learning something in a methodical way. I might start with a syllabus or construct my own. I start with the foundations, identify settled doctrine, state of the art, and unsettled areas.
When I taught myself to code, I completed a Coursera course on Data Structures and Algorithms to get a foundational understanding of the building blocks of software.
If I had to study a new area of law, I’d start with a textbook to get a lay of the land, zone in on settled law, identify edge cases and find the controversial questions.
This is straightforward, logically coherent and emotionally satisfying. I also think this is a very unrealistic way of learning anything to make an impact fast, unless you’re doing a PhD. I reserve this approach for passion projects to pursue things for the sake of joy.
Hypothesis-driven, Operator approach
This is hypothesis-driven, problem-first learning. What I mean is, the starting point isn’t to know but to be blocked.
I start by doing, say building an AI app using whatever means I have available. When I feel like I’m hitting a ceiling of understanding, I look for the reasonably available source of truth.
For example: I was working on a fallacy detection app in early 2026. I was unhappy with the kind of results I was getting from pre-Fable class models back then. I started reading research papers to understand how machines are taught to detect fallacies.
As a result, I ended speaking with one of the authors and discussed some ideas. Then, the models got considerably better and the problem was not significant enough.
Second example: I noticed that people, mostly non-engineers like consultants & lawyers and some technical students, were relying on heuristics like best-in-class model or cheapest model to use for all of their tasks. People weren’t thinking in effort-levels enough. So I went to figure out whether there was a difference between model outputs at different efforts.
I used 1 prompt to build the same game at 36+ different model-effort settings, controlling the conditions as much as reasonably possible with a 30+ step workflow. Then I analysed the outputs and traces to figure out how the models worked and what led to differences in outputs. These insights were pretty cool and led me to share my learnings with the Community.
The trick is to find reasonably available source of truth. In an ideal world, all your information should be derived from a source of truth. But in reality, you have to trade-off time or cost to find information. In such cases, I default to triaging sources of truth by access. This is a disciplined yet practical approach. The bias is to find sources of truth or primary sources as much as possible; but not to the extent that it becomes an excuse to delay or let the problem remain unsolved. The overarching principle is that knowledge is a means to solve a problem or unblock, it’s not the end in itself.
Beware the dilettante
For people who enjoy learning new things, there’s a real risk of lapsing into intellectual masturbation or dilettantism. I write more about it here: Why it's the Polymath's moment