A recurring theme of this week was experiments - egregiously so. I don't mean to say bad, by any definition of the word, no. I mean experiments that challenge the very assumptions of our capabilities as humans.
Plain example: many weeks ago, I wrote about how I'd one day upgrade my journaling and digital brain dumping to include everyday audio. How could I reasonably do this, in a manner that allows me to capture serendipitous and otherwise spur-of-the-moment conversations? Bumping into a friend in line at a coffee shop or having a random-but-deep conversation with a stranger has its own constraints - pulling out your phone and saying "hold on" while you tap record is awkward.
So I did what any reasonable person would do (heavy sarcasm implied), and designed my own audio circuit. There are plenty of voice pins that exist and have existed (i.e. the failed Humane pin, Plaud, etc.), but none that actively prioritize privacy + local-first - and I can tell you precisely why.
It's not a sexy venture capital investment to say "we sell devices for $100-$120 that a user only has to pay for once." It's far more appealing to squeeze as much as you can out of your end user - have them buy a $50 device, but lock them into a $250/year subscription. Plus, you get to access all their data that they give you, and you get to sell them transcription/memory!
Because of my own need for a tiny configurable data dumper, and something I could clip to a backpack, I decided on something that solves my problem - every time I come back home, the day's conversations and audio captures would automatically flow to my vault. Plus, it's a problem I'm intricately familiar with - it follows my idiosyncratic data product thesis, and when I choose what to do, I like to think: "what can only I do?", or in other words, what passions do I hold that would carry me much further than others?
I'd like to remind you: last Monday, I touched physical dev boards and modules seriously for the first time - 1.5 weeks later, I can (almost) send this circuit design off to JLCPCB and get 5 of them printed for $100. It'll take an extra $3k to get FCC certified (the ~spooky~ regulations part to actually start selling these), and for $4k-$5k total, I have a plausible hardware startup. The embedded software, I can already do. That's what I mean by egregious - you can just do things.
The hardware was one lane. In parallel, I was building another: an app for our phones that mimics my journaling format. Downside: the obvious social blip mentioned before. Upside? I learned you can embed near instantaneous transcription and use completely local AI models (AKA: private, relatively powerful, and run on only phone hardware)!
I'm happy to release this app, by the way - it already mimics what I've done for years, and then some. Reply to this email if you want direct access to the beta and to mess with completely private local AI running on your phone - I don't see anything of yours, and I prefer to keep it that way. I'll be making the source available soon.
The speed of experiments is going drastically up, and the cost of them is going drastically down: my suggestion? Pour your AI budget into experiments that speed up your understanding. In parallel. Test multiple separate things. At once. Each lane below represents experiments - a couple are the work above, the others could be as tiny as 10-minute slices.
I want to give a special shoutout to one of my now-readers, Anupam - a walking blend of engineering and the humanities I met randomly at a cafe. I think in the past week alone, I've bumped into him at least half a dozen times - seeing him work on his own circuits, asking him his process, etcetera. Who you surround yourself with matters, and I don't think I would have been as emboldened to experiment as I am now if I didn't deliberately engage the people around me who share similar passions and desires.
Be well,
Michael Kirsanov