Prediction is often treated as an organising principle for intelligence: remember enough of the past to reduce uncertainty about what comes next. A new theoretical result shows that this principle can conflict with thermodynamic efficiency when an agent’s actions help create the future it is trying to predict.[1]
The study concerns the fundamental energetic limits of abstract agents in perception–action loops. It is a theoretical result about extractable physical work, rather than a measured technique for lowering the electricity bill of a language model.
Passive observation and active feedback are different problems
In a passive “tape” setting, inputs arrive in a fixed stream. Better prediction reduces uncertainty, and earlier thermodynamic results connect that predictive information with efficient work extraction. The observer cannot change what appears next.
An active agent perceives, updates memory, acts and then receives a new percept partly shaped by that action. A thermostat changes the temperature it later measures. A forager changes which patch remains available. A dialog system changes what the user says next. Prediction now depends on both the environment and the agent’s own intervention.
Remembering an action has two consequences
Remembering which action it took may make the next observation easier to predict. But it also makes the action less uncertain when that memory is known. In the paper’s work bound, lower observation uncertainty helps, while lower action uncertainty can reduce extractable work. Keeping extra action information can improve the first term while worsening the second by more.
The authors define work capacity as the maximum expected rate at which an agent tailored to an environment can extract work. Their Theorem 3 identifies environments in which three objectives cannot share the same optimal agent: best prediction, maximally random actions with no retained action memory, and maximum work extraction. The work-efficient solution balances the information useful for prediction against the cost of retaining action history.[1]
A useful result can be fundamental without being immediately practical
This result does not measure a practical AI workload’s energy consumption. Turning the bound into an engineering saving would require a physical implementation and a comparison of its complete costs. The theorem instead identifies why a design principle that works for a passive observer can fail in a system that changes its own future inputs.
Its value is conceptual. If future biological or artificial systems approach fundamental limits, remembering everything relevant may cease to be a universal design rule. Information that improves prediction can still worsen the balance of conditional entropies that determines the work bound.
The agent’s objective changes what counts as useful memory
A forecasting system values information that predicts an external sequence. A work-seeking agent values information only insofar as it helps the entire perception-action cycle. Once action changes the data, predictive accuracy is no longer a complete measure of efficiency.
That distinction may eventually matter for embodied AI, adaptive control and theories of biological cognition. The paper’s most provocative implication is not that forgetting is a defect. Under the right physical objective, forgetting can be part of the optimal policy.
Sources & further reading
- Information Thermodynamics of Agents: The Work Capacity of Channels with Memory Physical Review X · 5 Oct 2026 · Accessed 15 Oct 2026