Blog
Tracking the development of an AI Economist
The Macro Training Data Only One AI Lab Will Own
Chib and Tan showed a transformer trained mostly on DSGE simulations can match a VAR. RoboMacro's 30-country shock simulations take that idea global, are licensed for AI training, and will be sold to one AI lab only.
Same Bet, Different Engine: Chib & Tan's DSGE Transformer Next to Our Multi-Country Sims
Chib and Tan train a small transformer on DSGE simulations of seven US series. We build 30-country shock simulations with 45 series each, including FX, yields, equities and commodities. Here's how the two compare: data, splits, how we built a leak-proof eval, and why wide, cross-country data is what language models need.
From Consensus to Alpha: Harness Engineering for Macro Strategy
Every word of our Pre-MPC briefing is AI-generated. The analysis follows a multi-step logic chain, all comments are tied into data and all data ties back to a numbered chart. The reason the note reads well isn't because the LLM is smart — it is because the harness around it is.
Refining the Logic of an AI Economist
The AI wasn't just mistaken; it was, at times, completely backwards. After a deep-dive review, we found that in about 20-30% of the notes, the AI was inverting the core logic of economics.
First Attempt at an AI Daily Note
This newsletter (US Daily Macro) mirrors my human approach to writing daily notes 25 years ago. It's entirely "Vibe Coded"—I haven't written a word of code myself.
