Summary and referral. This is my condensed retelling of an essay I published on chancejiang.com — The Knowledge Plateau for LLM Growth (March 29, 2026). Read in full →
2026 has landed LLM development at a strange crossroads. The public commons has been scraped to its bones — researchers call it the Data Wall — and the accepted remedy (a steady diet of fresh, human-generated data, at least ~30% of the training mix) only treats the symptom. Models fed on recycled exhaust go the way of the Habsburg model: inbred and nonsensical.
The essay argues F. A. Hayek saw the real constraint decades ago, in two kinds of knowledge:
1. Scientific knowledge → the easy part
Formal, published, codified knowledge is exactly what the internet-era scrape harvested. That phase is over; the marginal returns have plateaued.
2. Knowledge of time and place → the part you cannot crawl
Hayek’s second category is the fleeting, tacit, context-specific knowledge held by individuals — the plumber who knows why this pipe vibrates, the trader who feels the shift before the price moves. It is not public. It sits behind private firewalls, professional experience and intuition, and no crawler reaches it.
3. The price system is the ultimate API
Hayek’s answer transfers directly: prices are a decentralised communications network that signals the value of dispersed knowledge without anyone exposing the raw data. An LLM that needs a sovereign individual’s tacit knowledge cannot crawl their brain — it has to offer an incentive, i.e. participate in a market rather than behave like a central planner.
4. The sovereign knowledge maker
Hence the rise of decentralised AI (DeAI) protocols that treat knowledge as an asset — proof-of-knowledge-style schemes that pay the individual for what only they know. Incentive design stops being a growth hack and becomes the mechanical necessity for crossing the plateau.
Conclusion: the discovery procedure
The plateau breaks not with a bigger scrape but with a discovery procedure — markets that surface time-and-place knowledge because the knower profits from revealing it.
The full essay develops each step with the benchmarks and the protocol examples: Read in full on chancejiang.com →