Signal and Noise
Signal is the pattern that carries meaning. Noise is everything that interferes with its transmission — the static on the line, the distortion in the channel, the randomness that corrupts the pattern. Shannon formalized the distinction in 1948 with a theorem so elegant it barely seems to need proving: the capacity of any channel to transmit information is finite, and that capacity is reduced by the noise the channel introduces. The art of communication engineering is the art of maximizing the ratio of signal to noise within the constraints of the channel. Wiener extended the framework into a moral register: an amplifier carries whatever signal you feed it, noise or pattern, carelessness or care. The question at the center of Segal's Orange Pill — 'Are you worth amplifying?' — is the question of what signal-to-noise ratio you bring to the human-machine loop.
In the AI Story
An amplifier, in engineering terms, is a device that increases the magnitude of whatever passes through it. It does not evaluate; it does not distinguish between symphony and hiss. If the input is rich with signal, the output is a more powerful version of that clarity. If the input is dominated by noise, the output is a more powerful version of that confusion. This property — moral neutrality at the level of the device, combined with extraordinary consequences at the level of what the device is given to amplify — is what makes Wiener's analysis of AI as amplification so sharp. The language model is the amplifier. The human input is the signal (or the noise). The output is whatever was fed in, carried further than any previous tool has carried anything.
The most dangerous failure mode is what might be called polished noise: output that has the surface characteristics of signal but none of the substance. Segal's account in The Orange Pill of a Claude-generated passage connecting Csikszentmihalyi's flow state to Gilles Deleuze's concept of 'smooth space' illustrates the pattern precisely. The passage was eloquent. The structure was convincing. The philosophical reference was wrong in a way obvious to anyone who had read Deleuze. A high-fidelity amplifier had produced noise so smooth it was indistinguishable from signal without independent verification. The smoother the polish, the harder the detection.
The detection problem is structural. In a low-fidelity system, noise announces itself — static is audible, a badly written paragraph is visibly bad. The degradation is legible, and the human can correct for it. In a high-fidelity system, noise is invisible. The prose is smooth, the code compiles, the brief cites real cases. The human in the loop, lulled by surface quality, relaxes the vigilance that is the only defense against amplified noise. Wiener anticipated this in his 1960 Science paper, warning that automation removes from the designer's mind an effective understanding of the stages by which the machine reaches its conclusions. The opacity is the central danger: a channel whose internal noise characteristics are unknowable to the operator.
This places an irreducibly human burden on the loop: signal detection. The machine cannot evaluate its own noise; a language model trained on patterns of authoritative prose will produce authoritative-sounding prose regardless of whether the underlying claims are true. Hallucination is not a bug — it is a consequence of architecture. The model optimizes for plausibility, and plausible falsehoods are produced with the same fluency as plausible truths. The human's function in the loop is the evaluation the machine cannot perform: Is this actually correct, or does it merely look correct? The evaluation requires an independent standard — the human's own judgment, informed by experience, maintained by the willingness to reject polished output that does not survive scrutiny.
Origin
Claude Shannon's 1948 paper 'A Mathematical Theory of Communication' (published in the Bell System Technical Journal) founded information theory. Shannon proved that any communication channel has a maximum capacity, that capacity is reduced by noise, and that information itself can be quantified in bits regardless of the medium that carries it.
Wiener worked on related problems at MIT in the same period and his Cybernetics (1948) incorporates the signal-and-noise framework into a broader theory of communication and control. The two men's contributions were deeply intertwined, though Shannon's focus was more purely mathematical and Wiener's more overtly philosophical.
Key Ideas
Finite channel capacity. Every communication system has a maximum bandwidth, and that capacity is shared between signal and noise.
Amplifier is morally neutral. Amplification carries whatever it is given; the evaluation is upstream.
High fidelity conceals noise. The smoother the channel, the harder the noise is to detect from the surface alone.
Model cannot detect own noise. Hallucination is structural; confidence and correctness are independent properties of language model output.
Signal detection is the human function. Evaluation requires an independent standard that the amplifier cannot provide.
Debates & Critiques
Whether language models can eventually acquire reliable self-evaluation is an active research question. Current evidence suggests confidence calibration is improving but remains fundamentally limited by the fact that the model's internal representations do not cleanly separate 'what I was trained to say' from 'what is true.' Until that separation exists, external evaluation by humans or by other epistemic systems remains essential.
Appears in the Orange Pill Cycle
Further reading
- Claude Shannon, 'A Mathematical Theory of Communication' (Bell System Technical Journal, 1948)
- Norbert Wiener, Cybernetics (MIT Press, 1948)
- James Gleick, The Information: A History, a Theory, a Flood (Pantheon, 2011)
- Claude Shannon and Warren Weaver, The Mathematical Theory of Communication (University of Illinois, 1949)