April 10, 2026

TL;DR. Frontier LMs are "mismanaged geniuses" — the next capability leap won't come from bigger models, but from training them to decompose tasks themselves instead of leaning on brittle human-engineered agent scaffolds.

Takeaways

• Today's agents are hand-crafted decomposition strategies that are narrow, brittle, and make models look dumber than they are.
• The bottleneck isn't model intelligence — it's the "space of decompositions" we give LMs access to.
• Bitter-lesson move: stop scaling the genius, start training the manager (i.e. learned task decomposition).
• Provocation: maybe LMs can't play certain games or do long-horizon work not because they can't — but because nobody built them a decent scaffold.

TL;DR. Bet on companies whose AI-fueled "Ontology Velocity" lets free cash flow per share compound exponentially — and buy them when the market is most confused.

Takeaways

• Stock prices track free cash flow per share long-term; the only real margin of safety is buying companies you believe can grow FCF/share far beyond what's priced in.
• Modern moats = the Costco Algorithm + network effects: massive engaged user bases create proprietary data moats that train models competitors can't replicate.
• "Ontology Velocity" — even a 0.1% edge in value-per-token compounds into an uncrossable gap as better outputs attract more data, which trains better models.
• Author's track record: bought AMD at $4.2, Tesla at $13, Palantir at $7 — all during "peak confusion." You only need to get it right once.