X
Imagine if naked people were stupider. It turns out, naked models actually are.
@garrytanarticlex
TL;DR. LLMs aren't unreliable bullshit machines — people just keep testing the raw engine on a bench instead of building the car around it.
Takeaways
- Kyle Kingsbury's 32-page takedown catalogues real LLM failures (botched image edits, fake stock graphs), but every example is a human typing at a raw model with no harness, skills, or deterministic tools wrapped around it.
- Tan's architecture: "thin harness, fat skills" — deterministic code handles precision tasks (SQL, APIs, math), skill files teach procedures, a resolver routes requests to the right capability, and the model only does judgment.
- The "jagged frontier" of LLM capability isn't a philosophical flaw, it's a routing problem — debuggable and testable, not inherent.
- He's "testing the engine on a bench and concluding that cars are unsafe."