I wanted AI to keep working even while I play

Instead of reading every piece of news that pours in each week, I tried handing the reading and filtering to a system. The technology is not remarkable, but the change it brought was significant for me.

When I think about AI these days, I feel both curious about how far it can go and vaguely afraid of it. The possibilities are clear, but things move so fast that simply keeping up can feel like a lot. To shrink that vagueness even a little, there is one line I keep repeating to myself: “Even if I take it easy, let’s make AI do the work!”

I have been automating more of my routine tasks, handing off the tedious work and keeping the review for myself. This newsletter is actually one result of that process. To set expectations first: technically, this is nothing grand. A developer might call it a weekend project. Still, I open with this story because everything I send from here will come out of it.

I started with the job I found most tedious

Back when I worked in technology strategy, one of the most tedious tasks was tracking industry news continuously and analyzing how it might affect us. Skip it and you have nothing to base judgments on; try to do it all and it never ends.

The habit of reading news did not disappear when I changed companies. Whenever I read something, I first ask: “Does this change my judgment, or what I think something is worth?” The problem was still volume. There are hundreds of AI stories alone each week. I subscribed to several curated newsletters, but at some point I realized that someone else’s selection ultimately reflects someone else’s priorities. I needed material filtered by my own criteria.

I put a pipeline on a bare Mac mini

Worried about security, I started cautiously on an old, low-spec laptop that was sitting unused. As soon as it felt like “oh, this could work,” I ordered a Mac mini, waited a month for it, and set one up entirely for this purpose.

The setup is hybrid. Claude writes the automation code, and Gemma, running locally, handles the repetitive collection and execution. The rule for splitting the roles was simple: run tasks where mistakes are cheap locally, and reserve the stronger model for tasks where mistakes are costly. Overnight the system collects and does a first pass of filtering, and in the morning a briefing arrives.

What I learned after running it for a few months is that the hard part is not the model but operations. Pipelines die silently. So I made the system report its own status every day and leave a record of every failed job. Tokens cost more than I expected, so I keep experimenting on the thin line between “let’s see how far this goes” and a surprise bill. I plan to write up these trials every other week in the Buildlog.

What changed was not the system, but how I judge things

Looking back, the real change was not gaining a tool. It was coming to own the raw material of my own judgment.

First, the filtering criteria became my own. What surfaces and what gets filtered out — I set those criteria and I fix them. That was impossible while consuming someone else’s curation.

Second, I can now answer “why did I see it that way back then?” I still have the source records and the reasoning from the time. For someone whose job is making judgments, I think this is not a luxury but infrastructure.

Third, it became too good to keep to myself. The material lands in front of me every week anyway, so the extra effort required to share that material with my own judgment dropped close to zero. That is why this newsletter started.

In the past I mostly reviewed things from a particular vantage point and sketched the future. These days the fun is in being a user myself — building things and applying them to real work. This piece is a by-product of that fun.

What comes next

Starting with the next issue, I will send the signals this pipeline actually picked out that week — along with the four questions that separate signal from noise, and for every judgment, a note on “what would make this wrong.” I believe conviction without conditions for being wrong is closer to noise than to signal.

This newsletter is free, and it is certainly not investment advice or consulting.

Let me leave you with one question. Faced with the flood of information every day, what criteria do you use to decide what to read and what to let go?

Thoughts and other perspectives are always welcome.