Think twice / 06
The Quiet Machine
Four needles, all connected. Can a machine find its balance when its own reactions get in the way?
How to play
You operate four connected needles and try to hold all of them in the green band for twelve consecutive steps. Some starting responses make disturbances worse. Your challenge is to find a workable response, then test whether it survives another disturbance. A quiet streak is your first milestone; the full trial gives you a comparison.
- Press Run machine with the starting challenge. Watch the quiet-streak indicator. The machine pauses on a newly completed twelve-step hold, so you can inspect what worked.
- Tap a dial to inspect its response, nudge it or lock its automatic settings. Switch views to compare the same start without automatic retuning. Manual nudges and counter-pushes affect both machines.
- Try A small disturbance, then A constant push. In the latter, try a negative counter-push on A and compare it with a one-off nudge. Use Run to comparison for the complete 180-step trial; explain whether recovery required new settings, ongoing compensation or neither.
What to notice
A thermostat should heat a cold room. Imagine it accidentally heats a hot one instead. Pushing it harder will make the problem worse. To recover, you may have to change how it responds, not merely push the temperature back once.
Here four connected needles should stay in a green band. Run the machine and watch where they go. Select one unit and give it a one-off nudge, then compare that short intervention with changing a response setting. A persistent counter-push is different again: it keeps acting after you stop touching the controls.
Ross Ashby's Homeostat, developed in the 1940s, explored machines that could retune their own responses when important variables went out of range. Ashby called that search ultrastability. Our four dials and discrete choices make the distinction visible; they are not a reconstruction of his electrical machine.
A setting may happen to bring the needles back for a while, and another change may disturb a recovery already under way. This matters in engineering and in recurring organisational problems: do you keep correcting the outcome, or change the rule that produces it? A twelve-step hold is a useful comparison, not proof that the system will stay steady forever. Did the machine find a better response, or only a temporary quiet moment?
Read more: W. Ross Ashby, Design for a Brain, chapter 8 (PDF).
The connected instruments
Step 0 / 180One response is wrong. Watch the needle and the settings it tries.
No clock is running. Press Run machine, or inspect a dial first.
Unit A
A nudge moves this needle once, by 0.55. Both machines receive it.
Response settings & persistent counter-push
The needle still moves. The lock only stops automatic setting changes in the self-tuning machine.
Inspect the current response
“Own” scales this needle’s current value. “Incoming” scales its neighbour’s value. Negative reverses a contribution; positive keeps its direction. All four values update together.
A counter-push stays on every step. It changes the steady input, not just the starting position.
The path so far
Response changes and nudges - latest 64 events
Did the machine find a workable response?
Detailed trial comparisons
| Trial | First 12-step hold: no auto | First hold: self-tuning | Automatic changes | Your interventions |
|---|
Rules & model
Your goal and the display
Each needle has an invented, unitless value between −1 and +1. The green band is −0.20 to +0.20, including its edges. Hold all four inside it for twelve consecutive steps. Running pauses when you first complete a new twelve-step streak, so you can disturb the machine again. A trial has at most 180 steps.
How the needles move
The needles form a ring: D affects A, A affects B, B affects C, and C affects D. Each response setting supplies an “own” weight and an “incoming” weight. Values change together from the previous step, not one after another. A value that reaches an end stop stays bounded there; being still at an end stop does not count as being in the green band.
What automatic retuning does
If a needle stays outside the band for the selected waiting time, the self-tuning machine tries one different response for an unlocked unit. It chooses from eight stated settings using the trial’s repeatable random sequence. It does not predict which setting is best. At most one unit changes per step, and a changed setting starts affecting movement on the next step. Locks block that unit’s automatic changes, not its movement or your manual controls.
The comparison machine uses the same starting state and receives the same nudges, manual response settings and counter-pushes. It never retunes itself. Turning retuning off makes both follow their current settings; it does not rewind earlier changes. Switching the view only changes what you inspect.
Different kinds of intervention
A nudge adds or subtracts 0.55 once. A counter-push adds a constant input on every later step. The constant-push challenge adds +0.65 to A at every step. Trying another response changes that unit’s setting in both machines; it does not reset their needle positions. Replay restores the starting positions, settings and trial sequence and clears your manual interventions.
Show the exact calculation and eight settings
next[i] = clamp(0.8 × old[i] + 0.2 × (own × old[i] + incoming × old[previous] + challengePush[i] + counterPush[i]), −1, +1)
Available (own, incoming) settings: (−0.8, +0.25), (−1.2, +0.6), (−0.6, −0.4), (+0.2, +0.5), (+1.4, +0.8), (+0.8, −0.9), (−0.2, +1.1), (+1.2, −0.4). Retuning draws uniformly among the seven settings different from the current one. Outside-band counters reset when a unit enters the band, receives a manual change, or is automatically retuned.
A nudge, counter-push change or manual response change restarts the quiet-streak measurement for both machines. Previously recorded first holds stay in the trial record.
What the result means
A quiet streak is a measured observation, not a proof of long-term stability or intelligence. The safe band, weights, trial limit, counter-push and challenges are our design. Some constraints prevent recovery. This is a small feedback experiment inspired by Ashby, not a numerical reconstruction of his electrical apparatus or a model of a human brain. Nothing is uploaded or saved between visits.
Where this began
W. Ross Ashby’s Design for a Brain, chapter 8 (PDF) describes the Homeostat’s interacting units and setting changes triggered when variables leave acceptable limits. His family archive documents the original machine. We borrow the distinction between correction within a response and changing that response, using our own discrete model and controls.