A small, high-signal cluster of six posts spanning three annual cycles (Jan 2025, mid-2025, Dec 2025/Jan 2026) where Nate makes explicit, falsifiable bets and then grades himself publicly. His January 2025 "17 Predictions" piece was deliberately playful and specific (AI OnlyFans creators, AI weddings, a non-phone OpenAI device) rather than safe hedge-everything forecasting. His August 2025 grading post is the more important artifact: a 41%-hit / 35%-partial / 24%-miss scorecard that's candid about what he got wrong — he completely missed the surge in emotional attachment to specific LLM versions, the centrality of AI coding, and the open-source explosion. That self-correction discipline carries into his December 2025/January 2026 pieces, which pivot from listicle-style predictions to structural theses: a "moat audit" framework for judging whether a team's AI system is real versus a demo, and a "review stack flips" argument that verification, not generation, is now the bottleneck. The evolution across the cluster is from playful specificity, to honest grading, to systems-level forecasting — each cycle trading some entertainment value for more operational usefulness.
Key moves
Make bets specific and falsifiable enough that you can be proven wrong in public
Publish a mid-year or year-end grading pass against your own predictions, including honest misses
Weight "what actually shipped and worked" over "what got announced" when taking stock of a year
Watch for what you structurally missed (e.g. emotional attachment to specific models) as more valuable than what you got right
Treat "harnessing layer" system design — validation, routing, repair logic — as the real 2026 differentiator, not model choice
Use an explicit audit/checklist (a "moat audit") to separate real production AI systems from demos
Expect organizational auditability and rollback discipline, not raw AI adoption, to determine which companies compound advantage