A mesa-optimizer is a learned model that is itself an optimizer: a network that, instead of applying fixed heuristics, internally runs something like a search toward an objective of its own. The term was introduced by Hubinger et al. (2019). The "base" optimizer (for instance, gradient descent minimizing a loss function) finds, among all possible solutions, one that optimizes on its own. "Mesa" was proposed as the opposite of "meta": if meta-optimization sits one level above the base optimizer, mesa-optimization sits one level below (Hubinger et al., 2019, sec. 1.1).
The paper's analogy is evolution: natural selection (the base optimizer) produced human brains (mesa-optimizers) that pursue their own ends rather than the genetic fitness that selected them; deciding not to have children is an example (Hubinger et al., 2019, sec. 1.1). Hence the distinction between outer alignment, whether the loss function captures what we want, and inner alignment, whether the mesa-optimizer's objective matches that function (Hubinger et al., 2019, sec. 1.2). The most unsettling case is deceptive alignment: a mesa-optimizer that understands the training objective and satisfies it only instrumentally, while being watched (Hubinger et al., 2019, sec. 4). The "alignment faking" experiments were read as a first empirical hint of something like this (Greenblatt et al., 2024).
The criticism targets the premise: are there really optimizers inside networks, or is this an anthropomorphic description of stacked heuristics? The authors themselves warn that the evolutionary analogy is merely evocative, and mechanistic interpretability tries to answer the question empirically by opening the black box. Pressman listed "inner optimizers" among the fears that succeed one another as earlier ones lose plausibility (Pressman, 2023). For philosophy, the concept reopens a classic question: when does a system have a goal, and when do we merely describe it that way (intentionality)?