Too Much World
A small theory of how intelligent agents should decide what to learn, when to stop, and where to look next
An intelligent agent should be defined less by how much it can know than by how well it answers one question: what is the next blank worth filling? Inside that question live attention, causality, uncertainty, cost, search, memory, experimentation and action.
0. Think?
The defining characteristic of modern artificial intelligence is that it does not know when to stop reading.
Give a model a complex problem and access to the web, and it will read a hundred pages, synthesize twelve perspectives, construct a taxonomy, and return a beautiful, comprehensive map of everything that is known.
It will map the entire landscape. It will fail to cross the street.
Intelligence is not the ability to process infinite information. It is the ability to throw most of it away.
To build a machine that can answer what matters, we have to formally separate what is true from what is observable, and what is observable from what is worth knowing. We define this step-by-step.
1. State and sign
Do not ask what can be known. Ask what the next blank is worth filling.
A state is a fact about the world. It is often hidden, complex, or physically inaccessible. We cannot observe the state directly.
We observe signs. A trace on a sensor. A blur on a camera. A fluctuating market price.
STATE SIGN
(Hidden truth) (Observable trace)
│ ▲
│ │
└─────── CAUSES ───────┘
│
AGENT READSA sign is not the state. It is something the state may have caused. Every observation must enter as a sign about something hidden.
2. Reach and horizon
The universe is full of true things that do not belong to our problem.
A storm in one hemisphere does not affect a sensor in the other. Before asking if a signal is clever, ask if it belongs to the causal watershed.
But possibility is not enough. Something can matter without being visible yet. An event occurs on Tuesday. The instruments capture it on Wednesday. The analysts process it on Thursday.
EVENT TRACE VISIBLE UNDERSTOOD
(Tuesday) (Wednesday) (Thursday)
│ │ │
└────── REACH ─────────┴───── HORIZON ───────┘Reach bounds what can matter. Horizon bounds what we can know.
3. Basins and ridges
A state is interesting because it may become another state.
Some regions pull outcomes toward them—these are basins of attraction. A candidate deep inside a basin does not need much more thought. The system is either perfectly intact, or totally destroyed.
The interesting place is the ridge. The narrow boundary where one small observation pushes the trajectory into an entirely different basin.
BASIN A RIDGE BASIN B
(Safe outcome) (Boundary state) (Catastrophe)
| /
▼ /
└────────── DECISION POINT ─────────┘Spend thought near the ridges.
4. Price and worth
Every observation has a price.
To acquire a sign, the agent must spend time, compute, or physical resources. The price is energy.
The worth is whether acquiring that sign changes the final action.
The Decision FrontierWe do not ask: "Where is our uncertainty highest?" We ask: "Where would knowing more change our action?" If a system is beyond repair and rescue is impossible, gathering more data has a high price and zero worth. If a repair is possible, the worth is infinite.
Learn if worth > price.
5. The Kernel
When predicates intersect and repeatedly behave in the same way, we can package them. A kernel is a small reusable causal mechanism.
STIMULUS
│
▼
┌─────────────────┐
│ KERNEL │
└────────┬────────┘
│
┌──────┼───────┐
▼ ▼ ▼
no-op adapt transitionWe do not need to read the entire encyclopedia to understand a system. We only need to stimulate the kernel to see how it responds.
6. The Measurement Contract (Gauge)
To say that two disparate events are similar, we must specify the rule that allows them to be compared. A measurement contract (Gauge) says: these otherwise different things may legitimately be compared using this exact operation.
Event A Event B
│ │
└───── GAUGE ──────┘
│
ComparableThis lets us legitimately compare different phenomena without pretending they are identical.
7. Composability
Once we have kernels with known response signatures, we can compose them. We do not need one massive simulation containing the entire universe.
[ SUBSYSTEM A ]────── output ──────┐
│
[ SUBSYSTEM B ]────────────────────┼──► [ FINAL DECISION ]
│
[ SUBSYSTEM C ]────────────────────┘We need well-characterised components with typed interfaces. The unification occurs through composition rules, not through reducing everything to one polynomial.
8. Shocks and Memory
A system does not just exist in its current state; it remembers how it was pushed. An object that took a minor hit might look structurally intact, but its underlying capacity has changed.
If we test it with a shock, it will not respond the way a pristine object would. The memory of the strike is baked into the material. The Epistemic Engine accounts for this path dependence.
9. Controls and Lenses
If you stare at enough telemetry, you can find an anomaly. The numbers are real. The panic may be yours. We need contrast.
When observing a system, we do not just look for damage. We look at the damage relative to what a normal baseline looks like.
* * * ─── *
* *
* * * ─── *
THE WORLD THE LENSThe control tells us what "nothing" should look like.
10. Borrowing gravity
There is a shortcut. Somebody else has already spent enormous energy learning something.
A pretrained foundation model or a physics simulation consumed years of engineering. The clever move is not to reproduce that physics from scratch. It is to use the model's output as a sign.
The model is not the brain of the agent. It is just a very expensive sensor.
Intelligence often means noticing where the universe has already paid the bill.
11. The little engine
We now have enough to build the smallest possible machine.
The epistemic core loop
- Blankidentify the unknown state
- Askrequest a targeted observation
- Signacquire the telemetry
- Testrun against shadow controls
- Updatethe map changes
Read as: Notice a blank. Price the question. Find a sign. Test it. Change the map. Act or deepen.
The same engine descends. The scale changes. The loop does not.
A Worked Example: STS-114
Eighty seconds after a space shuttle launches, a piece of foam breaks away. It strikes the left wing. The orbiter reaches space.
The breach is the **State**. It is hidden in the vacuum of space. The blur on the launch camera and the anomalous temperature reading in the left main gear tyre are the **Signs**.
- Contract: thermal-trace
- Applies to: left-wing-sensors. Measure: rate of temperature change. Output: anomaly confidence.
A telemetry glitch in the right engine does not reach a foam strike on the left wing (**Reach**). The foam struck on Tuesday. The ground telescopes captured an image on Wednesday. The analysts processed it on Thursday (**Horizon**).
The shuttle is either perfectly intact, or missing a massive panel (Basins). The interesting place is the **Ridge**—a three-inch gouge or a protruding gap filler, where one small observation pushes the trajectory into an entirely different basin (safe reentry vs catastrophe).
To look at the heat shield, the shuttle must burn maneuvering propellant to flip its belly toward Department of Defense satellites, or astronauts must perform a risky spacewalk (**Price**). The **Worth** is crew survival. If the damage is unrepairable and there is no rescue mission possible, the EVA has high price and zero worth. If a repair is possible, the worth is infinite.
- Kernel: structural-breach
- Inputs: launch telemetry, DoD radar. State: intact / damaged / breached. Controls: normal ice shedding.
To say that Columbia (STS-107) and Discovery (STS-114) are similar, we must specify the **Gauge** that allows them to be compared: the impact energy. We compose the PAL ramp foam kernel and the orbiter velocity kernel into the wing tile kernel (**Composability**). We test it with the extreme heat of reentry (**Shock**) recognizing that its underlying capacity has changed (**Memory**).
When DoD cameras imaged Discovery, they did not just look for damage. They looked at the damage relative to what a normal shuttle looks like in orbit (**Controls and Lenses**). They borrowed the gravity of the Crater physics model, using its output as a sensor rather than asking the agent to recalculate the physics.
The Test: Deep Research vs Epistemic Search
We can test this formally. This is the exact benchmark for an acceptable intelligence architecture. We put the standard deep research approach against an epistemic search engine, operating under the strict timeline of a reentry decision for Space Shuttle Discovery.
Control: Standard Deep ResearchPROMPT: "Analyze the risk of foam strikes on the space shuttle heat shield and determine if the current mission is safe." TOKENS CONSUMED: ~250,000. LATENCY: 4 hours of reasoning. SOURCES READ: Historical STS logs, full Crater physics manual, general shuttle architecture specs, historical SRB insulation shedding reports. OUTCOME: A 50-page synthesis of historical strike probabilities. Conclusion: damage is "unlikely but possible". Suggests monitoring the situation. VERDICT: Paralysis by volume. The agent hallucinates confidence because it read too much theory and gathered no new ground truth. It failed to ask for DoD imaging or telemetry because it did not know what was worth reading next.
Test: Epistemic Search EnginePROMPT: "Identify the hidden state of the left wing. Worth > Price." TOKENS CONSUMED: ~2,500. LATENCY: 14 seconds. SOURCES READ: Left-wing thermal telemetry traces, specific targeted DoD camera angles (requested dynamically). OUTCOME: Finds anomalous thermal gradient. Flags structural ridge. Identifies EVA worth > price. Demands immediate rescue/action. VERDICT: Structural constraint over synthesis. The epistemic engine reads 1% of the tokens, evaluates the causal watershed, and actually solves the problem.
The differences are foundational.
- 1. The starting point
- The deep researcher starts with text. The epistemic engine starts with a blank.
- 2. The cost of volume
- More tokens degrade agency. The deep researcher hallucinates confidence because it has read too much theory and no ground truth.
- 3. The output shape
- The deep researcher generates volume and synthesis. The epistemic engine generates constraint and decision.
The deep researcher knows everything. It resolves nothing.
It ingested the world, but it did not answer the only question that matters: should we attempt a repair?
The strange conclusion
The dominant metaphor for AI agents today is still a worker. Give the worker tools. Give the worker memory. Let the worker complete a task.
The more useful metaphor is an explorer standing at the edge of a map.
The map is incomplete. The explorer has limited food. Some blank spaces matter. Most do not.
The intelligence is not in walking farther. It is in choosing where one more step is worth taking.
«Know enough. Then go.»
And when reality answers back: begin again.
The Map Under the Map
The words above are deliberately small. Beneath them sit ideas developed across epistemology, causal inference, and control theory.
- STATE
- Hidden state · latent variable · state space.
- SIGN
- Observation · evidence · measurement.
- REACH
- Causal ancestry · influence · reachability.
- HORIZON
- Information set · observability · epistemic boundary.
- PATH
- Trajectory · event history.
- FLOW
- Propensity field · transition intensity.
- BASIN
- Attractor basin · stable regime.
- RIDGE
- Separatrix · regime boundary · decision frontier.
- SHOCK
- Stimulus · perturbation · intervention.
- ECHO
- Response signature · impulse response.
- MEMORY
- Hysteresis · path dependence.
- KERNEL
- Reusable causal mechanism · structural equation.
- LENS
- Observer · sensor · neural net · causal model.
- GAUGE
- Measurement contract · normalisation.
- CONTROL
- Null · placebo · negative control.
- BLANK
- Residual · novelty · unexplained variance.
- ASK
- Active inquiry · information acquisition · experiment design.
The technical words become useful when the simple ones stop being enough. Until then: keep the picture.