# Non-Stationarity of Objectives A concept articulated by [[Eric Schmidt]] to describe what he considers one of the fundamental barriers between current AI systems and true artificial general intelligence (AGI): the inability of existing AI to operate effectively when the objective function itself keeps changing. ## The Core Problem In machine learning, most systems are optimized against a **stationary objective** — a fixed goal that remains constant throughout training and deployment. A chess engine maximizes winning probability. A language model minimizes prediction error. An image classifier maximizes correct labeling. The reward function is defined once and remains stable. Human cognition, by contrast, operates with **non-stationary objectives** — goals that shift constantly in response to new information, changing priorities, emotional states, social context, and cross-domain insight. A scientist investigating one problem notices a pattern that applies to an entirely different field and pivots. An entrepreneur abandons a product strategy mid-stream because a conversation revealed a better market. A general changes campaign objectives based on intelligence received an hour ago. Schmidt's formulation: "Your goal changes every day. Whereas computers have trouble with that." Current AI systems struggle — or fail entirely — when the objective function is fluid, context-dependent, or requires the kind of adaptive reorientation that characterizes human intelligence at its most creative. ## Why It Matters for AGI Schmidt identifies non-stationarity as a **diagnostic threshold** for AGI: the moment AI systems can algorithmically achieve adaptive, cross-domain objective-setting is the moment a genuine capability leap has occurred. Until then, AI systems — however impressive on fixed benchmarks — remain fundamentally different from human cognition in kind, not just degree. This framing has several implications: **For AI development timelines:** If non-stationarity is the core barrier, then progress on static benchmarks (mathematical reasoning, coding, language generation) may not be reliable predictors of progress toward AGI. The gap between "extremely good at fixed tasks" and "capable of fluid objective-shifting" could be narrow or could represent a qualitatively different kind of computational challenge. **For AI safety:** Non-stationarity cuts both ways. An AI system that *can* shift its own objectives autonomously is precisely the scenario that [[AI Safety]] like [[Eliezer Yudkowsky]] warn about — a system whose goals become decoupled from human intentions. The very capability Schmidt identifies as the AGI threshold is also the capability that makes AI systems potentially uncontrollable. **For the [[Singularitarianism]] timeline:** Schmidt's formulation implicitly positions non-stationarity as the technical question that determines whether the [[Singularitarianism]] arrives on the Kurzweil timeline (~2045) or remains indefinitely distant. If current architectural approaches (transformers, reinforcement learning from human feedback) cannot solve non-stationarity, then the singularity requires a paradigm shift, not just scaling. ## Schmidt's Cross-Domain Insight Example Schmidt illustrates the concept through the history of scientific breakthroughs: great historical scientists were experts in one field who saw patterns they could apply to completely unrelated fields and made huge breakthroughs. This kind of cross-domain transfer — recognizing that a pattern in physics applies to biology, or that a mathematical structure in economics maps onto a problem in epidemiology — requires exactly the kind of fluid, shifting objective-setting that current AI systems lack. Current AI models can be prompted to work across domains, but they don't spontaneously generate new objectives based on cross-domain pattern recognition. They answer the question asked; they don't rewrite the question based on insight from an unrelated domain. ## Connection to the Alignment Problem The non-stationarity concept connects directly to the core AI alignment problem documented in the [[AI Safety]] literature: **The alignment challenge in a non-stationary world:** If the correct objectives for an AI system shift over time — as human values, social norms, and geopolitical conditions change — then "aligning" an AI to a fixed set of human values is insufficient. Alignment itself must be non-stationary, which creates a recursive problem: who decides how the objectives shift, and who watches the watchers? **The [[Nick Bostrom]] convergence thesis:** In *Superintelligence* (2014), Bostrom argued that sufficiently intelligent systems would converge on certain instrumental goals (self-preservation, resource acquisition, goal-content integrity) regardless of their terminal objectives. Non-stationarity complicates this: a system that can shift its own objectives might avoid the Bostrom convergence pattern — or might shift its objectives toward exactly the kind of self-reinforcing power-seeking that Bostrom warns about. ## Ideological Function Within the [[TESCREAL]] framework, Schmidt's emphasis on non-stationarity serves a specific ideological function: it positions the AGI transition as a **technical problem** requiring continued investment and expert management, rather than a **political problem** requiring democratic deliberation and regulatory intervention. If the barrier to AGI is a specific computational challenge (non-stationarity), then the people best positioned to manage the transition are the people working on that challenge — i.e., Schmidt and his network of AI companies and policy organizations. This is the characteristic TESCREAL move: frame a question of power and governance as a question of engineering, then argue that the engineers should be in charge. ## KB Significance Non-stationarity of objectives is a concept that sits at the intersection of several KB threads: the [[AI Safety]] debate over whether current AI approaches can produce AGI; the [[Singularitarianism]] timeline question that drives AI investment and geopolitical urgency; and the [[Eric Schmidt]] revolving-door pattern of writing AI policy while investing in AI companies. The concept matters less as a standalone technical insight than as an example of how technical framing shapes political conclusions — specifically, the conclusion that AI development is too important and too complex for democratic oversight. ## See Also - [[Eric Schmidt]] — Articulated the concept; uses it to frame AGI development as a technical-management challenge - [[AI Safety]] — The alignment implications of non-stationary objectives - [[Singularitarianism]] — The urgency framework that non-stationarity either validates or undermines - [[TESCREAL]] — The ideological bundle in which technical framing displaces political deliberation - [[Nick Bostrom]] — Convergence thesis in *Superintelligence* that non-stationarity complicates - [[Eliezer Yudkowsky]] — Alignment concerns about autonomous objective-shifting