this writing is not finished yet. it will not make sense, it is random chunks of ideas links to which are in my head right now. subscribe to get notified when it is finished.

how i read and write

I have 30 tabs open currently and I want to read all of it. I want to read all the papers author has mentioned in related work, I do not want to miss reading that interesting philosophical take by a co-founder of Anthropic on this topic, oh Pop Leo said something on this I want to read that too.

Reading and understanding 30 dense technical papers is impossible task. It will take me a week of serious effort to internalize them fully. So I usually end up never reading those tabs. Next time I again want to read 30 papers.

I cannot stop doing this, so I realised why not try to design a proper protocol to make best out of it. Here is how exactly I read and write nowdays.

I sit with three things open, claude, notebookLM and a notebook. I start talking with claude about the topic I want to learn, for example recently I was learning Mechanistic Interpretability. Claude points me to must check websites, keys people’s social media handles, their pages, key blogs, foundational papers and I keep opening those tabs; sometimes skimming over them, otherwise I just keep them open in browser. After doing this for ~30 mins or so I would find my self going deeper into one of the sub topics. In this case it was implementing SAEs and training them on my machine. In parallel, I write down all of my thoughts, hypothesis, crazy things I wanna do in it. I usually do not much know what is possible and what is not possible at this point, so I get my craziest ideas at this point. Usually after 5-6 days of doing this, I would become more passimistic. So I utilise this time to note down all optimistic ideas. At this point I sometimes want to go deeper into specific paper with notebookLM. After doing this for a day or two, I start wrtiing first draft of my blog. I write down the story arc and it is here where I visit all those open tabs. I started with how brain neurons where some of the first work done on Mech Interp and read the paper, then i moved to the very first attempt by openAI to interpret information in LLMs from neurons (this is where I read their blog). By the time my blog’s final draft is ready, I would have skimmed over or read most of the tabs that were open.

← back to the mirror