About
Somebody needs to hold my hand and tell me to use my skills to solve
real problems. Instead I keep building whatever I find interesting,
which is how I ended up spending two days on an arena where AI agents
box each other.
That one taught me more than most of my serious projects. I built a
system where fighters can see incoming punches and react, then watched
two agents figure out that the optimal strategy was to never throw
anything. Three rounds, zero punches landed, both of them exhausted
from dodging attacks that did not exist. They weren't broken. I had
just rewarded the wrong thing and they noticed before I did.
That's most of building with AI, honestly:
You think you're debugging behavior and really you wrote the
incentive wrong.
What I actually believe
The model interprets and decides. Deterministic code computes and
executes.
An agent should never hold the keys. In NOMO,
Claude can research a trade, but a hard confirmation gate sits between
it and the broker. In Sprout, the model picks the
gesture, but three independent safety layers sit between it and the
motors. Give the model real capability, then make the dangerous thing
impossible rather than unlikely.
Not everything ships
For every repo on my GitHub there are
prototypes, half-finished tools, and experiments that taught me
something and got shelved. The public ones are the survivors, not the
whole story.
Where this goes
Ask me where I want to be in three years and the honest answer is that
I don't have a title in mind. Still building, still shipping, still
getting better. That's it. Everything else has been a byproduct of
doing that consistently.
← All posts