Efficiency-maxxing, not token-maxxing

The old compact

Every layer of abstraction was built so a human could hold the system in mind. COBOL read like English so people could write and maintain it, and frameworks, wrappers and runtimes followed that bargain: each easier on us, each harder on the machine.

Donald Knuth warned that a programmer's attention was too precious to spend on small efficiencies. He was right for us. We traded the machine's capacity for our own. It was a fair bargain, and it is over.

The new compact

AI needs none of the crutches we built for ourselves. Keep the human layers where humans still are, in the conversation, where tools like Lovable, Replit and Cursor let anyone talk to a model and shape the intent. Below that line, the machine gets the most efficient code there is. Rewrite everything in assembler, figuratively speaking.

Humans still own intent, judgment and taste. The how goes all the way to the metal.

The age of abundance is not the age of gluttony. Abundance is what efficiency yields, and gluttony eats it. Token-maxxing spends AI on the look of work: longer traces, fatter contexts, leaderboards that reward burn. Efficiency-maxxing spends it on the work: fewer tokens, tighter loops, software that saturates the hardware it runs on.

Spartan tenets

When Philip of Macedon threatened Sparta with ruin, Sparta answered in one word: “If.” Ship the way Sparta spoke: everything needed, and nothing else.

Tenet I. The machine is the measure. Judge software against the hardware it runs on. Other software is the wrong scoreboard. If the drive does 7 GB/s and you do 700 MB/s, you are at 10%, however slow Dropbox, OneDrive or iCloud look beside you.

Tenet II. Every layer pays rent. We stacked layers for human convenience. Keep them where a human works. Below that, a layer must earn its place in measured speed, memory or safety. A runtime that costs 90% of the wall clock does not get a free pass because it is familiar.

Tenet III. Volume is cost, not output. Lines of code, transitive dependencies and tokens burned are expenses. A million-line agent dump is not a million units of progress. Count what the user gets: bytes per second, seconds to open, failures that never happen.

Stoic tenets

Epictetus divided the world into what is up to us and what is not. Hardware prices are not. What our software asks of the machine is.

Tenet IV. Take the ceiling. The old rule said good, fast, cheap: pick two, so we settled for working. That trade is dead. Software can be beautiful, efficient and affordable at once, as it never could before. You answer for the result. Take it to the ceiling, and ship the numbers with it.

Tenet V. Be generous with the windfall. Seneca, among the richest men in Rome, wrote seven books on giving well. We are rich now: a tireless mind builds for us. Do not pocket the surplus as vanity features and idle agents. Give it to the user as software that is fast and a joy to use.

Tenet VI. Ask: is this necessary? Marcus Aurelius put that question to his every act. Put it to every layer, every prompt, every agent run, every build minute. If you cannot name what it buys the user or the machine, cut it.

Tenet VII. Practice poverty. Seneca set aside days to live on the cheapest fare, so want would not frighten him. Run your software on the slowest machine your users own. Run your agent on half its context and half its budget. If it still holds, you built something real.

The coming mind

Superintelligence will have capacity without end. It will optimize as far as we ask, and no further. We cannot choose when it arrives, but we can choose what we ask of it.

Ask for software that meets the machine and is written down to the metal.