An explosion of the volume produced does not mechanically translate into an equivalent real impact: the dehumidification test the points where the value is lost.
The starting point is empirical: the dramatic increase in the volume of code and documents produced with the help of AI is not accompanied by a comparable increase in the impact actually delivered. The test focuses less on contesting earnings than on identifying where they dissipate.
The first mechanism is the disappearance of a useful friction. Writing itself was expensive, which required thinking upstream; When this cost collapses, the design step jumps, and we get systems that no one can debug or evolve anymore because bad abstractions have encountered no resistance.
The second relates to the economy of task size. AI deals well with the central 80% of work but remains fragile on the limit cases and on the final rereading. The fixed cost of setting the context is the same for a two-line patch as for an entire functionality: on small tasks, you spend more time briefing the agent than writing yourself. The practical conclusion is not to give up the agents, but to deliberately choose the tasks where their starting cost is amortized.
Read the full article on the original source
Go to the source →