Why Subplots
I dumped my undigested brain content into a prediction machine and pulled the lever.
I did it a few times, with ChatGPT, Gemini and finally Claude Code (Opus). The last one hallucinated output noticeably better than others, so I continued asking and discussing and refining. Claude kept circling back to what I didn’t want to admit. Being acknowledged by smart people. I tried to omit this answer, since I’m very introverted and hate depending on others. But Claude kept asking how I feel about small ideas which it asked me to build — what are my emotions at each stage, from idea inception to a working demo. About previous projects and ideas.
I realized my blindspot — I want to be recognized for what I build and write. Not by the wide public, but by people who I share values with: experimenters, hackers, builders. I really like the LessWrong blog and forum for its integrity and opinionated communication style. I want to build something with that texture but for tech experiments, emergent systems, decentralized mechanics and economics.
The finishing problem
Here I am, following the recommendations of an overcomplicated text predictive machine. But honestly, I think it assessed me right. I’ve never been diagnosed with ADHD or similar conditions, but I have a damn hard time finishing projects. Never really understood why.
One thing if it’s a job-related task, with bosses, colleagues, salary and responsibilities. That I never failed and I can even call myself a reliable and good engineer. But things I do for myself, those I fail constantly. For years I’ve been collecting a huge list of ideas, ranging from small weekend projects to crazy (impossible) ones. Most ideas weren’t touched. Those that I built ended up in the ideas graveyard.
Many times I’ve made the mistake of following the path: “It’s a promising idea. It can earn me money/investments.”. Or I rushed to the idea on pure enthusiasm. An idea that had no prospects, and I had very little expertise in. Spent months building to just find out it’s useless and no one cares.
One arc at a time
I’ve decided to learn a lesson and follow a different path. Thus this blog “Subplots” — where each idea or writing will be a subplot. Arc with idea, plan, product and validation. Based on the progress of each one I’ll be deciding if I want to push it further and make A BIG GREAT THING. Or just cut it and proceed. I’m taking the freedom to choose what to do and at the same time to understand if it’s something that has value. Being public about what I make, and also sharing my thoughts and experiments, will help me to move forward.
Two subplots, already dead
For the last couple of weeks I’ve done 2 demos of projects that seemed cool and utilized possibilities not available just a few years back (LLMs obviously).
The first one was the “Quiz game” — years back I had a fun time playing online quizzes with my girlfriend and friends. Choose a topic, quickly answer a bunch of questions within a time limit. My theory was that LLMs became good enough to produce decent quizzes on various topics.
After a few days of tinkering I arrived at the point where it became clear:
- It’s freaking hard to make LLMs produce high quality questions and more importantly answers. Even with a complex pipeline and multiple verification steps it produced statements like “knight’s armor had drainage to pee”, which I’ve found no proof of. I quickly got tired of fighting constant hallucinations.
- I found out that I don’t actually like quizzes. At least not that much to sign up for a prolonged period of manually checking hallucinated slop on a variety of topics.
Next please.
The second idea was the “Detective game”. I found out that there are public online archives of criminal cases. Hundreds of thousands of criminal cases with all details in them — evidence, transcripts, documents, etc. I’ve decided that I could process them with an LLM pipeline and produce interesting stories. After a couple of days I made a prototype. Played it, tweaked mechanics, tried different stories. It felt like a shitty boring game. It didn’t tick — too much text, too boring a process of connecting evidence. I’m sure there is an audience for that — it’s just not me, and I don’t want to make a game that I wouldn’t play.
Next please.
