Signal BoostingA strong single attempt and a signal that reaches its audience are different achievements.

The Noise Floor Problem

Someone sends a carefully prepared application, publishes a polished article or makes a strong pitch, then receives no response. It is easy to interpret the silence as a judgment of the work or the person who made it.

The noise-floor account offers a different explanation: the system did not detect the attempt. Its signal was too weak to be distinguished from the other inputs.

Job openings receive thousands of applications. Millions of new pieces enter the content ecosystem each day. Sales teams receive dozens of cold calls, and dating-app users see hundreds of profiles. A single action is one observation among many, below the threshold at which the system can distinguish it from random fluctuation.

Information Theory supplies the signal-and-noise lens. The comparison stops at addition: repeats of one weak signal add together in a receiver until they cross the noise floor, while separate applications do not add up, and what accumulates across them is the number of chances taken. In Will's own applications, gym attendance and publishing, volume did more than intensity.

In each of these, one action is a small part of what it takes before anything comes back, and how many it takes differs from one domain to the next. For a particular search the number is an estimate, made before starting (see Practical Implementation). Five unanswered applications are a small part of what a job search usually takes. Stopping there means stopping before the threshold rather than establishing that five independent judgments went against you.

Reach, repetition and clearer targeting change different parts of the detection problem; none supplies a guaranteed response.

Read this diagram

Improve detectability branches into Reach across more places; Repeat through time; Reduce irrelevant noise.

Why Single Actions Feel Significant

Ordinary physical actions often have immediate, predictable effects. Pushing a rock moves it; lighting a fire produces warmth; raising an arm raises the arm. The action itself crosses the relevant threshold.

That expectation transfers poorly to a noisy stochastic system. "I took action → system should respond" treats an application like pushing a rock. But the application enters a distribution with thousands of other inputs. A single pulse is a small part of the aggregate and goes undetected.

The agency model "I act, therefore effects should follow" works under complete jurisdiction. It fails when the action competes with thousands of other signals.

The Core Algorithm: Amplification

Amplification increases a weak signal through volume, time, space or filtering until it crosses the detection threshold.

Suppose each attempt has a 0.02 chance of success. The chance of at least one success is then:

• 1 attempt: 0.02
• 50 attempts: 1 - (0.98)^50 = 0.64
• 100 attempts: 1 - (0.98)^100 = 0.87
• 200 attempts: 1 - (0.98)^200 = 0.98

The model assumes roughly independent attempts with a consistent probability of success. Actual systems are messier, but more attempts still shift the probability of obtaining a result.

Under that assumption the chance of at least one success rises more slowly than the number of attempts: fifty attempts at 0.02 give 0.64, not 1. Where attempts build on each other, through an audience, a relationship or a skill, volume can compound instead. Two hundred applications can have a much higher probability of success than five perfect ones even when each application is less polished.

Probability Space Bending describes the distribution being changed. The aim is for accumulated signal to reach the point where the system responds, without relying on one exceptional attempt.

Many cheap attemptsInteractive example

Attempts

20

Chance that one attempt is a hit

5%

Draw another batch

At 5% each, 20 attempts include at least one hit about 64% of the time. This batch produced 1.

Spatial Amplification: Broadcasting in Parallel

Spatial amplification distributes attempts across channels at the same time. One application becomes fifty good-enough applications; one platform becomes ten; one networking target becomes fifty.

The job-search comparison holds the available time fixed and gives every application the same chance of a response, 0.02:

• Intensity strategy: 5 applications × 4 hours each = 20 hours → chance of at least one response = 1 - (0.98)^5 ≈ 10%
• Volume strategy: 100 applications × 12 minutes each = 20 hours → chance of at least one response = 1 - (0.98)^100 ≈ 87%

Giving both strategies the same chance per application assumes that the extra hours add nothing to an application. If careful work made each application several times likelier to succeed, the gap would narrow, so the comparison turns on how much the hours raise that chance.

An application at the 95th percentile still competes with 500 other observations. Quality increases the probability attached to each attempt; volume increases the number of attempts. In stochastic systems, volume usually wins that comparison.

The other examples follow the same structure: publishing on 10 platforms, contacting 50 people, making 30 sales calls or messaging 100 matches instead of writing three perfect openers. Parallel distribution fits a situation with limited time and many available channels.

Temporal Amplification: Persistence Through Time

Temporal amplification repeats the signal over an extended period. It applies the 30x30 pattern to external systems as well as habits.

The publishing example develops across time. Day 1 produces one article and 12 views, a tiny fraction of the virality threshold. By Day 30, 30 articles have accumulated 500 views. At Day 100, 100 articles and 5,000 views give the algorithm a repeated pattern to detect. By Day 365, the aim is 365 articles and an audience, past the threshold for distribution effects.

The article that appears to succeed alone is almost never the first. It comes late in a long run of pieces from someone whose accumulated signal has begun receiving amplification.

Four effects explain the pattern. Repetition adds exposure. Algorithms, relationships and skill reward consistency. More samples reduce variance and make a pattern clearer. Thresholds delay the response until enough signal has accumulated.

The same account applies to a habit: it needs enough repetitions before its pattern takes hold, and tracking it over a month counts them.

Filter Amplification: Improving Signal-to-Noise Ratio

A signal can also become more detectable by matching what the receiver looks for.

In the hiring example, a recruiter scans 500 résumés with about 6 seconds for each. A 95th-percentile résumé with the wrong keywords is filtered out, while a 50th-percentile résumé with the expected keywords gets through.

The work is to identify the receiver's keywords, formats and channels, match the relevant patterns, remove material that obscures the main signal and concentrate information in the dimensions being evaluated.

Filtering improves the sender's own perception too. Five customer interviews may conceal a pattern that becomes apparent across 100. Pattern recognition develops through the larger sample.

Reading 1,000 examples calibrates taste in a domain. Fifty failed-startup post-mortems reveal recurring failures. Thirty days of recorded behavior reveals patterns memory misses. More samples improve the filter, which raises the effective signal-to-noise ratio.

The Asymmetry Table

The relevant constraint determines the kind of amplification:

Many channels and little time favor parallel attempts. Few channels and more time favor persistence. A poor signal-to-noise ratio favors enough samples to calibrate a filter. The methods can also be combined.

Intensity wins only when three conditions hold together: outlier-quality output can be produced reliably, the system rewards outliers disproportionately, and capacity for volume is severely limited. These conditions rarely hold, so applying an intensity strategy everywhere usually misallocates effort.

The Outlier Simulation Trap

"I'll be so good they can't ignore me" assumes one 99th-percentile attempt can replace one hundred 50th-percentile attempts.

There are four problems with that strategy. Nobody can consistently produce 99th-percentile work. Visible outliers generally iterated before their breakthrough. The time spent perfecting one attempt could produce dozens of adequate ones. In numbers, one attempt has about a 1% chance of landing in the top 1%, while a hundred independent attempts give about a 63% chance that at least one does (1 - (0.99)^100 ≈ 0.63), and more attempts raise it further.

Steve Jobs's sequence—Apple I, Apple II, Apple III, Lisa, Macintosh, NeXT and the return to Apple—contains repeated attempts behind the story of one perfect product.

The story remains attractive because one attempt seems efficient, because failure can be explained as not having really tried, and because visible success stories omit their earlier volume. In stochastic systems, the resulting strategy usually underperforms volume.

The Cybernetic Reframe

"Did this action contribute to signal strength?" measures something different from "did this action succeed?"

An unanswered application can become "1/100 samples collected, 1% toward threshold." An article with 12 views can become "Day 1/365, signal accumulating." A sales call that does not convert can become "1/30 conversations, 3% toward pattern detection."

The cybernetic account tracks the accumulating input rather than treating each attempt as an isolated verdict. Single actions almost never succeed alone, but each contributes to the aggregate. Tracking cumulative signal makes that contribution visible.

Rejection then reads as a signal still below threshold, silence as an insufficient sample, and effort as accumulated rather than wasted. The immediate objective is to build signal strength.

Agency as Felt Causal Potential

"My actions don't matter" can be an accurate description of a single attempt. One application has near-zero influence on a job-search outcome; one article does not create an audience; one call does not establish a deal pattern.

It does not follow that accumulated action has no effect. The micro-level observation and the macro-level conclusion concern different quantities.

Agency increases when this distinction becomes usable. Individual actions are expected to have little effect, while enough accumulated actions are expected to change the distribution above the threshold.

The felt sense of causal potential comes from knowing how that accumulation works. Someone who expects most individual attempts to fail, while expecting the aggregate to succeed, keeps going. Someone who reads each failure as evidence of powerlessness stops before collecting enough signal to test the larger effect.

Why People Give Up

A person can be a small fraction of the way to the threshold while feeling as though they have tried as hard as they can. If invisibility is interpreted as rejection, it becomes a conclusion about personal inadequacy.

The diagnostic questions identify the estimated threshold, count how much signal was sent, compute signal ÷ threshold, and distinguish volume from intensity.

The failure patterns differ. Someone may stop a small fraction of the way there, spend most of the effort on one perfect attempt, remember 15 attempts as though they were 100, or use feelings in place of tracking. Without tracking, the felt amount of signal sent runs ahead of the real count.

The intervention replaces the outlier strategy with amplification and tracks the accumulated percentage. It expects no response before detection, with most attempts feeling wasted along the way.

Practical Implementation

Before beginning, the process estimates the required volume by researching the domain and asking people who succeeded. A tracking system counts attempts as well as successes, and the planned rate establishes a timeline. A threshold of 100 at 5 attempts per week requires at least 20 weeks.

During execution, the recorded measure is the percentage of the threshold reached. No response is expected until most of the estimate has been reached. Each attempt adds its share to the aggregate, and evaluation waits until the threshold is crossed because earlier results measure noise.

When stopping becomes attractive, the relevant comparison is 10% versus 200% of the estimate. Below 100%, the prescribed response is to continue. At 200%, the domain itself may be wrong. Above 200%, the estimated threshold may be wrong or the quality may require better filtering. Individual outcomes below threshold are treated as uninformative.

The example record is:

When Volume Strategy Doesn't Apply

The framework assumes a stochastic system with a high noise floor. Some settings violate that assumption.

Art markets, research breakthroughs and viral content can reward a singular outlier disproportionately. Small networks remember damage, so repeated bridge-burning accumulates negative signal. Strong quality gates may reject an attempt before volume matters.

Even there, volume often still wins, but the result has to be tested. If 100 attempts show no signal after the estimated threshold, either the estimate is wrong or quality needs improvement before more attempts.

Signal Boosting as Intelligence Design

Signal boosting also applies to building intelligent systems, AI agents among them.

The Paradigm Shift: Prompting vs Signal Engineering

The deterministic view treats a prompt like a program: input, function, output. If the agent fails, the prompt must be repaired.

An LLM instead samples from a distribution of possible outputs. The same prompt can produce different results. Hallucination, drift and misinterpretation are intrinsic properties of that substrate rather than individual defects that a perfect prompt removes.

"How do I make the chance of X high enough across N attempts that correct output emerges reliably?" asks about the surrounding system. It replaces the single-call question "how do I make the agent do X?"

The Core Algorithm: Generate + Filter

Suppose each call is correct with a chance of 0.7, independently of the others. The success rates are then:

• 1 call: 70% success
• 3 calls + majority vote: 78% success
• 5 calls + majority vote: 84% success

Prompting improves the 0.7 component. Signal engineering designs the system that produces 84% from those components.

Each row pairs a way of generating candidates with a way of filtering them until useful results emerge. AlphaCode's generation of millions of candidate programs followed by tests applies the same algorithm as the 200-application example.

Signal Function Taxonomy

Different calls occupy different positions in the signal flow. Their failure consequences determine how much reliability they need.

Source functions create material or an initial plan:

Routing functions determine where the material goes and coordinate execution:

Transformation functions change, compress, extract or combine the material:

Filtering functions check it and handle failures:

A generator can remain noisy because later stages filter its output. A router needs high accuracy because sending the signal down the wrong path can corrupt everything afterward. Reliability investment follows that difference in downstream consequences.

Asymmetric Verification: When Signal Boosting Works

Filtering is useful when checking an output is cheaper than producing it.

Other domains make the check as uncertain or difficult as generation:

If checking an output is much cheaper than generating one, repeated generation and filtering has the needed asymmetry. If checking costs about as much as generating, the extra check doubles computation for only marginal gain.

The Meta-Principle

An agent system designs information flow through unreliable channels. The design identifies the intended signal, locates where hallucination or drift enters, chooses amplification through volume, clarity or filtering, measures the output distribution through evaluations, and checks whether it crosses the required threshold.

The same account applies to behavior: willpower as a resource and probability distributions over behavior are variables in the system. Biological neurons and transformer weights differ, but the same design questions apply to both.

For a gym habit, prevention architecture reshapes the chance of going on an ordinary day rather than forcing each visit. For an agent, the surrounding design reshapes the chance of a correct output rather than depending on one perfect prompt. The target is the distribution over instances.

Related Concepts

• Intelligence Design applies generation and filtering to agent architecture.
• Information Theory concerns signal-to-noise ratio and uncertainty.
• Signal Theory distinguishes authentic signal from noise.
• Probability Space Bending describes accumulated changes to a distribution.
• Agency connects micro-actions with felt power over aggregate effects.
• Tracking measures attempts and cumulative signal.
• 30x30 Pattern applies repeated execution across 30 days.
• Cybernetics measures the response to accumulating signal.
• Expected Value compares probability changes from volume and persistence.
• Ladder of Agency concerns capability demonstrated through accumulated evidence.
• Skill Acquisition applies repeated practice to learning.
• Reliability describes the point at which an output can be used without checking it.
• Optimal Foraging Theory allocates search resources.
• Free Will relates individual actions to aggregate causal outcomes.