What this utterly misses is the huge worth of “non-code” AI work. Some of the highest-value interactions with an AI assistant contain debugging stack traces, analyzing fuzzy enterprise necessities, or understanding edge instances. An engineer would possibly spend an hour conversing with an LLM to pinpoint a root trigger, in the end leading to a clear, three-line repair. That is sensible, environment friendly engineering. Yet, beneath a tokenmaxxing regime, that engineer seems to be much less productive on the supervisor’s dashboard than the one who simply requested the agent to spit out 500 strains of spaghetti code.
Some of the perfect makes use of of AI are in considerate studying, discovery, and design. Like the brainstorming, whiteboarding, and conversations which have all the time supplied the substratum for creating software program, the worth of that type of AI use is nearly unattainable to estimate or distill right into a metric.
The economics of waste
When the measurement we use to know the well being of software program engineering turns into so divorced from the reality, the monetary mechanics flip violently towards the enterprise. From an financial viewpoint, tokens are an enter value, not an output worth. Worse, output tokens can value as much as 5 instances as a lot as enter tokens throughout top-tier frontier fashions. When you incentivize code era, you might be urgent on an extremely leveraged a part of the IT funds.
