Feb 16, 2026

From 2005 to around 2014, the best AI-assisted human chess players were better than the best AI and human players alone. More occupations are experiencing their own centaur age. What economic changes should we expect?

Here’s a minimal framework for thinking about this. There are two tasks, A (automatable tasks) and B (bottleneck tasks). A can be performed by AI, B cannot. Output is Leontief, Y = min(A, B). Some implications follow:

  1. Two kinds of labor prosper.
    • Those who use AI to produce unprecedented amounts of A, e.g. delegating to swarms of coding agents.
    • Those who produce B.
  2. What is B? Everything which results in the firm being maximally focused on what the market rewards. There is much here. Important subcomponents are:
    • the taste and product sense to know what to produce,
    • the charisma, moral authority, trustworthiness, credibility, and communication skills to reduce organizational frictions and interpersonal misalignment.
    • Both are facets of leadership. Both are challenging for conventional educational institutions to certify or credential.
  3. Firms will be disadvantaged if they have large process and communication overhead, and fragmented, undifferentiated culture. This disproportionately affects large firms. Previously they may have been able to succeed because small firms could not produce enough of A for a viable product. Now that advantage has disappeared.
  4. As the cost of A drops to zero, the market becomes flooded with high A, low B products. We have a name for this: slop. Its defining characteristic is fluency of form put to no good use.
  5. Bottlenecks are endogenous. AI progress takes tasks in B and puts then in A.
    • There will be large economic incentives to remove bottlenecks, one by one. Previously insignificant costs will become intolerable. If you spend 50% of your time programming (A) and 50% of your time on alignment and communications (B), halving the latter only increases output by one-third. If AI drives the time spent programming to zero, halving the time on B doubles output.
    • This applies recursively, as AI progressively recategorizes tasks from B to A. Costs which pre-AI would have been trivial nuisances become critical bottlenecks. Large rewards exist for those who can automate them. They fall one by one, until the end of the centaur age.