A Fairly Distracted Investor
An update on where AI sits in my investing process
I spent a lot of this year being an excellent systems administrator and a fairly distracted investor. I’ve messed around with custom agents, Ollama, Hermes and, of course, Claude and NotebookLM, and in the process spent more time building systems and less time thinking about businesses. Slightly wiser now ( I hope), I thought I’d briefly touch on a few things I’ve learned over the past few months.
Conviction takes time, and there is no “right” process here. At least for me, anyway. Watch a long video, run through a sector note, upload a few documents to NotebookLM, chat with Claude Cowork and, of course, talk to someone who understands it better than you. Some or all of that may need to happen for me to get comfortable enough to act or not act. And so I need to budget my time and resources accordingly, with most of my time spent doing this rather than building things.
The useful conversations are often pretty long. Which means I don’t want to waste premium usage on things that don’t need it. If I’m monitoring my portfolio through transaction statements, holding statements and other external sources, I could have Claude do it every time. But if the job is repetitive, I’d rather spend a weekend writing some code that does it for me forever. The same goes for a lot of boring work. Depending on your desktop, local models like Qwen or Mistral can handle plenty of classification, extraction and other grunt work. They don’t need to be brilliant. They just need to be good enough. Save the good models for when you actually want to “think”.
There is always a never-ending list of things to read or watch. Books, blog posts, podcasts, YouTube videos, and I haven’t even gotten to sector notes and company related reading. To try and solve this problem, at least a little, I built an agent that tracks a bunch of finance and business writers I like. It reads and summarises their articles and presents everything in a single document, almost like a newspaper, with sections. I run this weekly and only dip into the articles that actually interest me. It’s less about reading more and more about filtering some of the noise. And this is another place where I use local models. Something like Qwen 2.5 7B can handle this perfectly well. You really don’t need a frontier model for summarising a bunch of blog posts.
Create a digital repository. Record your thoughts on businesses, on the markets, or even just create sector notes. I use Obsidian and Markdown files, but the process is more important than the tools here. With AI, it later becomes an investing brain that you can query and ask questions of.
And there’s so much you can do here. For instance, if you’re writing about a paints business, you can tag everything related to Titanium Dioxide and have a separate page for it. Over time, you start building these little linkages across businesses, industries and ideas, which can later be queried. It becomes a way of building your own database of things you’ve thought about, rather than starting from scratch every time you look at something. And even without AI, it helps minimise hindsight bias.
“Power of the sun, in the palm of my hands.” I just watched Spider-Man, so I had to squeeze this in. But I suppose what I mean is that just because we can build something doesn’t mean we should. There are plenty of free and paid resources that do an incredible job using AI, and if a task is common, we don’t always need to build it ourselves. For instance, concall summaries is something multiple external sources already do pretty well, and I don’t think I have a particularly unique way of doing it that requires me to build my own thing. Doesn’t hurt, of course, but if 90% of the job can be handled externally, without costing a lot and saving me a lot of time, I’m perfectly happy with that. Screener, Perivis, Stock Scan and Wrapped all do a pretty good job with a lot of this stuff. (Perivis is from a friend and it’s pretty great. Yes, I am biased, but it actually is.)
None of this is a case for stopping. I’m still building. Right now it’s a tagging system that marks names by macro sensitivity, so that when something moves in the world I can open one file and see what it touches. Also working on a cleaner mutual fund dashboard that tracks common names, new additions and quiet exits. Half of what I build turns out to be useless, and I build it anyway, because that is how you find the two or three things that stick. I’m just a little more wary now of losing sight of what all the building is for. Conviction takes time. AI should be handing me more of that time, not quietly eating it.
Thanks for reading.
Cover image taken from Unsplash.