You cannot match the machine's variety

The law that ends the guilt

In 1956, the cybernetician W. Ross Ashby proved something that sounds abstract and lands personally: only variety can absorb variety. A controller — a thermostat, a pilot, a mind — can only govern a system if it can match the range of states that system can produce. He called it the Law of Requisite Variety, and it is as close to a physical law as management science owns.

Now count states. A modern AI assistant can produce more distinct, plausible, decision-relevant-looking paragraphs in an afternoon than you can evaluate in a month. Your feeds, channels, and inboxes multiply on top. The variety of what arrives at you exceeds the variety of what your attention can absorb — not slightly, but by an order of magnitude that grows every year.

Ashby's law says what follows: you, unaided, cannot control this. Not because you are undisciplined. Because arithmetic.

Amplify or attenuate — there is no third option

Cybernetics offers exactly two responses to a variety mismatch, and Stafford Beer built his Viable System Model on them. You can amplify your own variety — tools that let one decision govern many cases. And you can attenuate the incoming variety — filters that compress what arrives down to what matters.

Reading faster is neither. It is the doomed third option everyone tries first: matching variety point-for-point with a biological system that does not scale. Beer spent a career showing that viable systems — firms, brains, bodies — survive precisely because they refuse this option. Between every level of a viable system sit attenuators, so that each level receives only the variety it can govern.

Your cognition is a viable system, or it isn't

Beer designed his model for corporations, but the recursion goes all the way down. A single mind managing AI-scale information flows is a control system facing exactly the mismatch his attenuators exist to solve.

The design translates directly. Raw streams carry unbounded variety. Between them and your working attention belong compressors: relevance gates that discard differences that make no difference (Signal is a difference that makes a difference), format contracts that collapse essays into claims, validation layers that hold candidates away from trusted knowledge, review cycles that decay what expired. Each layer absorbs variety so the next one doesn't have to. By the time anything reaches the level where you deliberate, its variety is small enough to govern.

Why AI makes the law non-negotiable

Before generative tools, a determined person could almost fake it. Information arrived at human speed, produced by other humans with their own variety limits. Heroic reading could feel like control, sometimes even be it.

That era is over. The machine's variety is not human-limited, and — the sharper point — you commissioned it. The noise now scales with your own use of leverage: the more work you delegate, the more output returns. Any filtering strategy that depends on your reading rate now has a built-in expiry date. The only strategies that survive are the ones where the mechanics, not the reader, absorb the volume.

Attenuation is not loss

The objection arrives on schedule: won't compression lose something? Yes — that is its function. Attenuation deliberately destroys variety, and the design question is only which variety dies at which layer.

A well-engineered cascade kills variety that makes no difference to your decisions and preserves what does — which is why the relevance definition (Signal is a difference that makes a difference) had to come first. Attenuation without a definition of signal is censorship of random information. Attenuation aimed by your decision portfolio is the difference between a mind that governs its inputs and one that drowns in them politely.

The delegation you cannot refuse

Here is the uncomfortable completion of the thought. Refusing to build attenuators does not mean you avoided delegating your filtering. It means you delegated it to whoever produces the loudest inputs — the notification designers, the feed algorithms, the model's verbosity defaults. Some variety-reducer always decides what reaches your attention. The only choice you have is whether it is one you engineered.

Protocol: the mismatch audit

Once a quarter, redo the arithmetic from the exercise: output volume versus attention volume. The ratio will have grown — model capability compounds, and so does your own use of it. Each time it grows, the cascade needs another layer or a tighter contract, not another resolution to read more carefully. Treat a rising mismatch the way an engineer treats rising load: as a capacity-design problem with known solutions, none of which involve asking the bridge to try harder.

Sources

• Ashby, W. R. (1956). An Introduction to Cybernetics — the Law of Requisite Variety.
• Beer, S. (1972). Brain of the Firm — variety engineering, attenuators and amplifiers, the Viable System Model.
• Simon, H. A. (1971). "Designing Organizations for an Information-Rich World" — attention as the scarce resource.