Contents
On method
Programming has been through this twice before, and the arguments made against each transition are on the record — made by intelligent people, for defensible reasons, and mostly correct at the time. Knowing how those arguments went is the fastest way to see the shape of this one.
Each cost is shown on a worked example small enough to check by hand and reproduce from the labs — an existence proof that the cost is real and has the shape claimed. Where a number is owed and has not been taken, the chapter says so rather than estimating one.
The labs compile and run. Nothing in the book is pseudocode, though Appendix A.12 is explicit about the fragment of computation the language covers today — which is smaller than a reader might assume from the argument around it.
It specifies a structured diagnostic record: a stable code, the fragment at fault, the obligation not met, whether it is false or merely unproven, the finite set of legal repairs, and how much of the program may be touched. Anyone building a coding agent can adopt that record next week, against Python or Rust or anything else, without writing a line of AIR.
Three eras, two transitions, and the same reason each time. What each era's programmer actually held — and who writes it now.
Why models emit Python, and what it costs to have a machine write in a language shaped around human working memory. Each cost is shown on a worked example.
The languages machines already speak to machines — SSA, LLVM IR, MLIR — what they genuinely do, and what they were never asked to do.
What a machine author does not need, and what it does: intent and freedom, constraints and contracts, effects, capabilities and obligations.
AIR, in full. It parses, it verifies, it compiles, and it runs — and every laboratory exercise is reproducible with Python and Clang.
Eight observations that would prove the book wrong — including two measurements it has not taken — then the AIR reference manual, the labs, sources and a glossary.
Who this is for
You should be comfortable reading a program. You do not need to know how a compiler works — Part III explains what you need, and no more than that. If you have never seen assembly, you will by page 30. If you write compilers for a living, Part III will be familiar and Part IV will not.
What you need to run the labs: a machine with Python 3 and Clang on it.
27 chapters · 4 appendices · 15 figures · 178 pages