Dropping to log-level
I’ve been building more, posting less lately. Why? AI is moving incredibly fast, and the metagame is moving even faster. The only way I’ve been able to make sense of it is by building to think.
In the last 4 months, my workflow has changed repeatedly, from code completion, to coding agents, to multiple agents in multiple worktrees, to agent factories. I’m about 3x faster, and I suspect I can ratchet this up to 10x with factories. At the same time, I feel keenly aware that I am just a hair’s breadth ahead in the Red Queen Race, with selection pressure nipping at my heels.
Now, here, you see, it takes all the running you can do, to keep in the same place. If you want to get somewhere else, you must run at least twice as fast as that!
(The Red Queen in Lewis Carroll’s Through the Looking-Glass)
This is what it feels like to live through disruption. The situation gets into your OODA loop and it’s hard to stay oriented. You just keep moving as fast as you can.
This poses a challenge when writing. See, when I write, it is usually to capture a speed run through an idea maze. For example, I explored the idea maze around decentralized software for about 5 years, and came away with a map of paths that work (user-owned keys and hash addressed data), and paths that seem like they should work, but are dead-ends (flat p2p network topologies). It took a long time to map these paths. They are highly compressed models of the problem space.
I struggle to develop this kind of highly-compressed model for AI. A model is a map of the parts of a system that aren’t changing. But during a disruption, everything changes. Value chains dissolve and are reformed, and the competitive landscape goes from clear to chaotic. If you try to build a complex model in a chaotic environment, it will end up being overfitted, and will get obsoleted faster than you can build it anyway.
God is very cruel. He only gives us data about the past.
-Paraphrasing Clayton Christensen
When data is really bad, you should use the simplest model at hand. When data is very good, you can use complicated models.
-SFI Podcast Apr 6, 2020
Really, when things are moving this quickly, we don’t want a model, we want a log. It’s the simplest narrative structure that could possibly work. Just lines of timestamped comments, one after the other: “this happened, then this happened, then this happened”. No higher-level analysis. In software, we log everything that happens during program execution, because logs are invaluable when things spin out of control. We can comb through the log to find patterns, and begin to form higher-level hypotheses about what is happening.
I think this is the right level of sense-making for the moment. It’s time to drop to log level. You’ll see me shifting away from essays, toward work in progress, technical posts, fragmentary ideas, and raw logging.
In/out:
Out: essays.
In: logging what I learn, as I learn it.
Out: theory.
In: practice.
Out: good writing.
In: good ideas.
Out: “having an audience”.
In: hoarding things I know how to do.