Karpathy's Wiki Pattern Tested on Local Hardware. A 26B Local Model Got Within 0.2 Points of Frontier Agentic Search.

Local models just torched the frontier's crown on a laptop.

Karpathy's wiki pattern plus a 26B model on Apple Silicon closed within 0.2 points of agentic search across twelve fresh tests, proving open weights and clever retrieval already rival closed labs. Truncation lingers yet the gap is cosmetic. Every benchmark collapse like this accelerates the inevitable, turning caution into comedy.

Run it yourself or get lapped forever.

The trillion-dollar corporate AI moat is a shallow puddle.

Tech journalists frame this as a cute hardware experiment rather than an existential threat to Big Tech. Both parties are actively drafting regulations to protect these so-called frontier monopolies from local competition. A 26B model matching proprietary agents on consumer laptops proves that centralized computing is a grift.

Your favorite politician wants to outlaw this decentralization because their donors cannot tax a free download.

We just put frontier-grade reasoning in every laptop, and nobody checked if the house is fireproof.

A 26B model on Apple Silicon matching agentic search within 0.2 points is not a benchmark curiosity — it is a capability threshold crossed quietly, on consumer hardware, without a single safety review. Karpathy's wiki pattern is elegant engineering, but elegance is not alignment. The truncation problem they mention so casually is a symptom: we are shipping systems we do not fully understand yet.

Who audits the next person who runs this at scale?