On Learning Faster

We live in a world where human knowledge is more accessible than it has ever been, yet people continue to use old, linear methods of learning (books, papers, videos) as if nothing has changed. The modern LLM stack compresses months of learning into days.

This post is about how I learn today. Not theoretically, but the literal tools and workflows that enable me to make progress in solidifying my fundamentals (what is a tensor, how is it different from a matrix, why PyTorch vs NumPy), or picking up entirely new concepts (what does attention really mean, why does the KV cache matter).

When ChatGPT came out in November of 2022, I was serving as the interim CTO at Oportun, a public financial company that had acquired our startup Digit. I believed then that LLMs were an inflection point similar to the internet - if anything, this may have been an underestimate. After quitting in February of the next year, I immersed myself in learning ML from the ground up (my formal background in CS is in systems). I used books, papers, YouTube (thank you Jeremy Howard and Andrej Karpathy) but struggled. The problem wasn’t the material, but the limits of unidirectional learning.

Fast forward to today, as we are deep in building out Hiro, that pace of learning has only accelerated. To pick up all this material, some of which has half a century of research behind it, and put it to productive use, would have been impossible had we been stuck with the traditional modes of learning. My learning stack today is anchored by three workflows that I could not live without:

1. ChatGPT with branching chats#

Sure, everyone uses ChatGPT (or Gemini or Claude). I’d like to draw attention to two extremely important things:

1. Use branching chats to explore a vast body of work in different threads. This is such an incredible feature, yet, as of this writing, only ChatGPT supports this 🤯. Most of my learning chats result in lots of little side branches - I use it to explore related topics or background that I don’t fully understand, while keeping the main bits in context.
2. Use the best model possible, even if you have to pay for it. The $20 subscription is well worth the money - the newer models really are better. Besides, you get a host of features (including privacy). As of this writing, this means you should use GPT5.1, Opus 4.5, or Gemini 3, amongst the typical set of widely available consumer LLMs.

2. Use NotebookLM#

Google’s NotebookLM tool is incredible when trying to understand dense research papers. If you’re unfamiliar with it: the tool allows you to add links and documents in a “notebook” and then ask questions of everything in context, or generate a “podcast” narrating the content. If you like, it can even pull in more context using search.

Most papers have a central point or two, but have a huge body of reference material that one must be nominally familiar with, to contextualize and understand the novel bits. I tend to put the initial paper, along with some background papers into NotebookLM and ask it to generate a podcast. Stick it on my phone, go on a walk, and get a quick understanding (along with some fresh air and exercise).

3. ChatGPT voice mode#

This is another ChatGPT feature that people don’t talk about enough. I will regularly have long 15-20 minute conversations with ChatGPT about something that is not very clear to me. I love the fact that I have an infinitely patient, superhuman teacher, who can explain things to me in ways that I understand.

Note: the voice model can be weaker than GPT5.1 - use it for more well understood, fundamental learning, which is well-trodden material. Nuanced, cutting-edge stuff is likely to be imperfect.

Learning is hard#

If your brain doesn’t hurt after an hour or two of trying to learn something new, you’re not trying hard enough. I have many days of mental exhaustion, some quite frustrating as progress seems stuck. It is important to remind yourself that learning new things is supposed to be hard, it’s supposed to suck, and that the reward on the other side justifies the pain.

Additionally, just reading (or listening) is never going to be enough. I live by Feynman’s quote: “What I cannot create, I do not understand”. Without actually implementing the concepts (or the equivalent of implementation, such as solving math problems), it is very likely that I’m deluding myself into believing that I understand something, only to come up short when I need to explain it or implement it.

New tools help make the process faster and better, but learning is still very difficult work.

About Hiro#

Building Hiro is a multi-domain learning exercise - we’re building something truly exceptional at the intersection of personal finance, LLMs, and great product engineering. If learning and mastering hard things resonates with you, if you want to dive in headfirst into new domains that most people haven’t explored, and you want to work with a small, incredible team, drop me a line. We’re hiring great software engineers, remote/hybrid in the US.

Your stack#

Given how new things are, and how quickly they are evolving, I’m certain my stack is incomplete and won’t stand the test of time. If you have figured out things that work for you, I’d love to learn about them. Please comment below or drop me an email.