Source
Fiderer et al. (2025), LIT-041, read in full in NOTE-032. They model agent and environment as coupled hidden-Markov channels and bound the work extracted per round. Without feedback the efficient agents are exactly the maximally predictive ones that randomise and forget their actions (Thm 4). With feedback, one binary memoryless channel makes the predictive, max-entropy-action and efficient agents three nonempty, pairwise-disjoint sets (Thm 5). From that they state, in their introduction, that “in the presence of feedback, maximally predictive agents are generally inefficient” (NOTE-032, claim C12).
What is asserted, and by whom
The thesis is the general form: once an agent’s actions affect the percepts it receives, maximal prediction of those percepts is not what thermodynamic efficiency requires, and maximally predictive agents generally extract less work than the most efficient agent. Fiderer et al. assert it.
What the record holds toward it
- THEORY-026 (Active): without feedback, through fixed kernels, the dissipation at each switch of the drive equals the memory that fails to predict the next input (Still, Sivak, Bell and Crooks), so there only nonpredictive memory is wasteful; with feedback, in Fiderer et al.’s one binary channel, no predictive agent is efficient. That channel was recomputed: the optimum is about 0.272 bits per round, and every predictive agent extracts at most zero.
- One existence example is all that is proved. NOTE-032 notes that the Lemma 10 argument extends to any memoryless channel whose action rows are pairwise distinct and whose work capacity is positive. That extension is the record’s reading and has not been checked as a proof.
What this does not say
- Not that efficiency under feedback is a trade-off between prediction and forgetting. The paper’s abstract says so, but in the only proved example the efficient agent is memoryless (NOTE-032, C13).
- Nothing about learning systems: the agent knows the full loop and nothing is learned.
- Nothing per operation: the quantities are averages.
What would confirm or reject it
Confirmed by the promote_when: a proof for a stated class of feedback channels wider than the one binary channel. Rejected for a class by a feedback channel in it whose work-efficient agents include a maximally predictive one.