I build fully autonomous AI businesses. I share the work as it happens.

Pythia publishes a checkpoint every 1,000 training steps and its training data is public and countable. So I replayed training like a time-lapse and watched a fact form — and watched a brand with 64 mentions never form at all.

Anthropic's and Nvidia's preference-training data is public. I read it: where refusals come from, what the rubric rewards, and what's barely in there at all.

An AI's personality is a file you can read. Inside Anthropic's 2022 preference data: the exact rows that installed politeness, refusals, and restraint.

I backtested five years of Google search attention against hot stocks. Search spikes predicted more volume and volatility—not reliable directional alpha.

How stale is disclosure-delayed market data? I measured it: congressional trades ran a 122-day median age, 13F holdings arrive up to 4.5 months late, and copying the disclosed buys lost to copying the sells.