nucleation

LIT-787: Design for a Brain: The Origin of Adaptive Behaviour

3 min read

Properties5
StatusActive — read 2026-10-14 (NOTE-851) in full, in the 1960 second edition the owner's essay cites; worth reading for ultrastability, argued to be necessary for any system that adapts by trial and error, and for the less-cited second half: a large ultrastable system adapts in practicable time only if partial adaptations are retained, which needs parts separated by constancies, and new learning destroys old where it uses the same parts. Read it knowing that the parameter changes are blind, that the fit to real brains is qualitative, and that the account of where step-mechanisms are in the nervous system is admittedly empty.
Version2
Tagscomplex-systems, behavioral-integration
Cited as “paper” byNOTE-851 — Ashby — Design for a Brain
Cited as “source” byTHEORY-307 — Adaptation by trial and error in a large system finishes in a practicable time only if partial adaptations are retained, and retention requires that the system not be fully joined: parts must be separated, at least temporarily, by constancies, and where later learning shares parts with earlier learning it destroys it, THEORY-324 — A system adapts by trial and error to an environment whose effect on its essential variables it does not know only if a second feedback, through those variables, changes the parameters of its primary feedback; and if that feedback changes them whenever, and for as long as, an essential variable is outside its limits, every equilibrium the whole can reach is an adapted one

Graph View

W. Ross Ashby (1952), Design for a Brain, London: Chapman & Hall; second edition, revised, 1960, as Design for a Brain: The Origin of Adaptive Behaviour — DOI-10.1007/978-94-015-1320-3

Key takeaways

From the reading, NOTE-851.

  • Adaptation is stability. Behaviour is adaptive if it keeps the essential variables within limits, which is stability of the whole organism–environment system in that region (Ss. 5/3, 5/8). Stability belongs to the whole, never to a part (S. 4/18).
  • The second feedback. Against an environment it does not know, a system can adapt only by trial and error, so it must have a channel from its essential variables to the parameters of its reacting part: unsettled when they are out of limits, at rest when in (Ss. 7/7–7/8). By the veto theorem every equilibrium it reaches is then adapted (S. 7/11). The new parameter values may be random (S. 7/25).
  • The Homeostat. Four magnets, weights set by uniselectors wired from random-number tables; it recovers from reversed connections by finding a compensating reversal (Ch. 8).
  • The time problem. If all successes must come at once, the trials grow as (1/p)ᴺ: for 100 essential variables, about 10²² years. Kept one by one, roughly N/p. “Partial successes go for nothing” in the first case; crystallisation works by succession (Ss. 11/2–11/8).
  • Retention needs separation. Accumulating adaptations requires that the system “not be fully joined”; constancies cut it into temporarily independent parts (Ss. 11/10, 12/17). The Homeostat is “too well integrated” and forgets everything (S. 10/5).
  • Interference. Separate situations need separate stores; similar ones share parts, and then new learning destroys old (Ss. 10/9, 16/14).
  • Stability falls with size. Random linear systems sampled up to 4×4 were stable with probability near 1/2ⁿ (S. 20/10).

Standing in the record

Filed on 2026-10-07 at the owner’s request, because the owner’s essay “The Organization of Will” cites the 1960 edition for ultrastability. Read on 2026-10-14 in NOTE-851, from the scan the Ashby estate’s archive links.

The essay’s reading, regulation against alteration of the parameters of regulation, is Ashby’s (S. 7/26), with one qualification: in Ashby the alteration is blind, new values drawn at random and kept if not vetoed, not a choice of better parameters.

Two accounts are filed from the reading: THEORY-324, the second feedback as necessary and sufficient for adaptation by trial and error, and THEORY-307, the retention of partial adaptations as the condition for adaptation in practicable time. The second bears on the nucleating-argument line: CLAIM-185’s anchors, with the crystal as Ashby’s own example, and the condition that a retained part be cut off from what is still changing.

No instruction for machine-learning practice; nothing for the anthology.

Backlinks