ASKYM · RESEARCH · SEPTEMBER 2026
“What is this situation a means to?”
Can an AI explore all of it, well beyond what a human group can envisage? An exploratory research programme, its open questions, and what Google Research has just shown on a neighbouring subject.
The starting point
An effect can follow from what we already have and stay out of sight for years.
Our cognition tracks interactions poorly. As soon as several means combine, we envisage only a handful of connections, often the most familiar ones. Herbert Simon named this limit bounded rationality.
We call this cognitive myopia: the possibilities are there, blurred or out of frame.
The question that corrects this myopia, one situation at a time. It forces us to describe what we have before looking for what it makes possible.
This programme asks how far an AI can carry this question, first on one situation, then on many.
Two logics
Twenty-five years of research, a method practised with humans
The expert entrepreneur starts from what they have: who they are, what they know, whom they know. The goal comes later, along the way.
STEP 1
Describe the situation as a fact, through its properties, before any use. An induction: generalise its essential properties, those it cannot lack, until what they form can be named.
STEP 2
A deduction: apply a general rule to the induced properties, and infer an effect from them. Keep what the situation makes possible, directly and immediately.
STEP 3
Look for the essential function of each effect, the one it cannot lack: an induction, tested pair by pair. Does A enable B, and only B? Three to seven higher-level effects remain.
An example across the steps
CLARIFY · THE PROPERTIES, THEN THE NAME
A glove, described without its use: the shape of a hand, watertightness. Nothing about heat, nothing about washing-up. Generalised: a watertight hand-shaped container.
THE RULE
What has a shape and holds contents can serve as a mould.
EXPLORE · A DIRECT EFFECT
Hand-shaped ice cubes. A possibility that the usual use concealed.
This reasoning concerns an object, its properties, and what they make possible. That is why it can be tried outside the company, on any object whose properties can be listed.
The research hypothesis
We would give it a situation. It would describe its properties, name what they form, look for what this set makes possible, order the effects. The three steps of the method, chained without intervention.
FIRST ON ONE SITUATION
Each step would take the output of the previous one. The traces would have to stay readable from end to end.
THEN ON MANY
Ten, fifty situations, linked by their shared effects. In a workshop, a group handles one at a time.
A HUMAN, AND LIMITED TIME
The AI would generate on its own. A human would validate afterwards, or assist along the way in a variant. Their time would become the bottleneck.
This is a hypothesis. The complete chain has never run; only isolated components have been tried. What follows says how we would test it, and what would make us stop.
Step 1 · induction · two open questions
Describe the situation as a present fact, without cause, without solution, without judgement. Decide what is treated as a whole. Then an induction: from what is observed, generalise the essential properties, those the object cannot lack, until what they form can be named. The glove becomes “a watertight hand-shaped container”.
As the book puts it: “this semantic work is not incidental, it is the induction itself”. In this programme, exploration would start from this generalised description, written down in black and white.
QUESTIONS PUT TO THE AI
1. Can it induce this name without reintroducing the use? “A glove protects the hand” is already a use.
2. Who chooses the scope: a human, or the AI comparing several ways of carving it up? The properties depend on the scope, so the two questions are linked.
MEASURE fidelity of the name to the properties, absence of presupposed use, agreement of practitioners judging blind.
Step 2 · deduction · two open questions
A deduction: apply a general rule to the induced properties, and infer an effect from them. “What has a shape and holds contents can serve as a mould”: the glove makes hand-shaped ice cubes possible. We keep what the situation makes possible, directly and immediately, never what it contributes to.
For an AI, the rules would come from two sources. Connections already made by others, in texts, patents, databases. And analogy, when the object is new: start from the effects of a more general object, remove those its constraints rule out.
QUESTIONS PUT TO THE AI
1. What is actually found in the corpora, and what does the model invent?
2. An added property removes effects. Two properties joined together open some up: shape and watertightness make the mould. How do we account for both?
MEASURE number of distinct effects, share judged direct, immediate, exclusive by practitioners, already-known effects recovered.
Step 3 · induction, then test · two open questions
Set aside what contributes without enabling. For each effect, look for its essential function, the one it cannot lack. This is an induction, as in step 1, applied to an effect, and it is tested pair by pair. Does A enable B? Does A enable only B? When the answer is yes, A becomes the means to B, and B remains available as a means to a higher effect. The tree rises to three to seven higher-level effects. A disputed link is not intersubjective, and it is dropped.
Twenty effects make 380 ordered pairs. Our internal benchmark in September, on fifteen models, flagged a drop on these tasks. It did not evaluate the complete chain.
QUESTIONS PUT TO THE AI
1. Is sharing a function enough to establish “A enables B”?
2. Does a closed-answer judge, yes or no with its probability, do better than a model that writes free text?
MEASURE trees already labelled by practitioners, compared branch by branch; overall consistency, absence of cycles.
A neighbouring system, published on 26 August 2026
A system from Google Research, PPE for short. It is asked a question about the planet in everyday language: “Which areas of Nigeria are at risk of food shortage?” It chooses the data, fuses them, trains a model and writes the report. Nobody intervenes between the question and the report.
Three stages orchestrated by an LLM: data selection, curation, model and prediction. The details are in the appendix.
WHY MENTION IT HERE
1. A major laboratory reports that an entire chain runs without a human, on real data. This is the kind of autonomy our hypothesis assumes.
2. Its chain starts from a target and looks for the data that predict it. Ours would start from the means and look for what they make possible. Same architecture, opposite direction.
Planetary Prediction Engine · Google Research · 26 August 2026
Three questions put to the system, three reports, compared with the baselines of each domain.
NIGERIA · FOOD SECURITY
66.1%
of variance explained when moving down from state level to local level, against 31.5% for the baseline.
DRC · EBOLA 2026
15 of 18
newly affected health zones recovered across five weekly forecasts, ten points better than the best published model.
UNITED STATES · 21 HEALTH INDICATORS
76.8%
of variance explained on average, against 60.0% for pipelines hand-tuned by experts.
Our reading of its scope
The engine looks for the means to an already known end. It cannot discover an end. Anything that does not predict the target is set aside, even if it is valuable for something else.
It does not say what to do. The report stops where the decision begins. It is up to humanitarian actors to turn a risk map into action.
We know after the fact where the shortage occurred, so its predictions are measured automatically. Where nothing settles the matter after the fact, this loop does not exist.
Google's chain, and ours in the opposite direction
Same layers as the PPE's Nigerian and Congolese cases · market prices · price anomalies · rainfall · vegetation · night-time lights · roads · buildings · health centres · mobility
PPE OUTPUT · illustrated form
Signals retained: maize price anomaly, rainfall deficit, low night-time activity.
The decision belongs to those who read it.
ENGINE OUTPUT · illustrated form, not computed
Properties: price gap, link, storage capacity. Effect: a pooled purchasing reserve between the two areas.
Properties: care capacity, link, flow. Effect: a forward screening point, with no new infrastructure.
The two could be chained. The PPE says where it will hurt. The engine would say what the place already has to respond with. The goal “food security” would keep the pooled purchasing reserve and set aside the screening point, without having dictated either. Local actors would validate, one phone call at a time.
What the engine would return, and to whom
Each line would hold four elements: the means we own, the property we use, the effect we reach, the stakeholder who can confirm it.
Three filters would make it usable as is. Direct, held by a means already there. Immediate, actionable without acquiring anything. Exclusive, specific to this situation.
For humans. Faced with a problem beyond them, seeing an action within their reach that they had not seen.
For an AI. An agent that knew what its means make possible would choose better what to attempt. The engine would explore and propose. Execution would fall to another system.
ENTREPRENEUR
An action plan drawn from their means, with no prior business plan.
TERRITORY, INCUBATOR
An inventory of means read as a reservoir of possibilities to confirm.
LABORATORY
Latent protein functions to test, ranked by information value.
INDUSTRIAL PROPERTY
Unclaimed uses of a patent, to file or to license.
World model, action model, LLM: where the engine sits
LLM
“What is usually said about this?”
INPUT a text
OUTPUT the likeliest continuation
GOAL in the prompt
Proposes by association. The sorting remains to be done.
WORLD MODEL
“What happens if I do X?”
INPUT a state, an action
OUTPUT the next state
GOAL elsewhere, in the planner
Simulates the consequences of a choice already made.
DreamerV3 · V-JEPA 2 · Genie 3
ACTION MODEL
“How to execute goal X?”
INPUT a goal
OUTPUT clicks, tool calls
GOAL given at the start
Executes a goal it is given.
Computer use · Large Action Models
PPE
“What value does the target take, here?”
INPUT a target, layers
OUTPUT a prediction
GOAL at the start: the target
Predicts where and how much, stops there.
EFFECTUAL ENGINE
“What do these means make possible?”
INPUT means, their properties
OUTPUT effects, then combinations
GOAL afterwards, to filter
Would look for what can be done before knowing what to aim at.
Stakeholders would validate. Each effect they keep would become a means again.
The programme · three questions that nest together
QUESTION 1
The three steps, each with its measure. The effectuation literature does not compute means → effects; biochemistry does, in the forward direction (network expansion, underground metabolism), without redescription, without exclusivity, without hierarchy. It is these three steps that we would measure, against that baseline.
MEASURE known-answer corpora.
QUESTION 2
The goal would come after the effects, to filter them. Part of the budget would stay blind to the goal, so that the unexpected keeps coming.
MEASURE the cost of the goal: what is gained in speed, what is lost in discoveries.
QUESTION 3
Agreement between judges would establish a shared judgement; feasibility would be verified elsewhere. At what throughput does the human start to limit the chain?
MEASURE links validated per hour, depending on the prioritisation strategy.
A first test, to be built
DIRECT GENERATION
The question is put to an LLM as is, without the method.
AUTONOMOUS CHAIN
The same model runs clarify, explore, hierarchise, without intervention.
ASSISTED CHAIN
The same chain, with a human at the checkpoints.
WHAT WE WOULD LOOK AT
The effects that practitioners retain when judging blind, and how many.
WHAT WE WOULD COUNT
The errors: reintroduced uses, false links, invented properties.
WHAT WE WOULD TIME
Human time, with assistance and validation counted separately.
And now
DOES THE QUESTION HOLD?
Well-founded or naive: either answer helps us.
WHAT ALREADY EXISTS?
Related work we may not have read, on induction, analogy or hierarchisation.
WHERE TO START?
A first test, tiny and real: one situation, its means, and a phone call to check.
Quentin Tousart · ASKYM · askym.ai
Going further
A
Three stages orchestrated by an LLM, and its Feature Gate.
B
Everything is a situation: the planet, the territory, the company, the molecule.
C
Structural biology.
D
Industrial property.
Planetary Prediction Engine · three stages orchestrated by an LLM
STAGE 1 · DATA SELECTION
Territory, grid resolution, period. The system forms hypotheses about useful signals, checks them against the literature, then fetches the data from Data Commons, Earth Engine and, failing that, from public portals.
STAGE 2 · MULTIMODAL CURATION
The data found are fused with geospatial embeddings of population and imagery. A Feature Gate removes anything that would leak the answer into the data: sub-components of the target, downstream effects, future data.
STAGE 3 · MODEL AND PREDICTION
Regularised linear models, boosted trees, perceptrons. An overfitting guard reruns the computation when generalisation fails. The report is written at the end.
Keep the Feature Gate in mind. A filter that prevents the answer from already being in the data. The effectual engine needs a filter of the same kind, turned the other way: at step 1, no property may contain the use being sought.
Everything is a situation
For the method, everything is a situation: the state of the planet, the territory, the company, the protein, and they nest within one another. At each level, the same question and the same three steps: clarify, explore, hierarchise.
ZOOM IN
Enter a situation contained within this one, and draw out what it makes possible. From the planet to the molecule, our hypothesis is that the reasoning does not change, and that only the means change.
ZOOM OUT
Carry a selected effect back up as a means of the situation that contains this one. This is the recursion of the method.
The protein is the case where the generalisation can be tested most directly. Its properties are measured rather than induced, and nature settles what it makes possible. The uncertainty is then concentrated in the inference.
Generalisation · structural biology
Drug discovery asks “which molecule for this target?” The engine would turn the question around: a protein described by its properties, pockets, charges, flexibility, without its known function, and what those properties make possible.
KNOWN-ANSWER CORPUS · MOONLIGHTING PROTEINS
One and the same protein, two documented functions. The crystallins of the eye lens are enzymes recycled as optical proteins. We mask the second function and measure whether the inference recovers it.
In systems biology, the inversion exists at the scale of the whole network: underground metabolism computes the nutrients a network would make possible without having been selected for them, then laboratory evolution settles the matter. We want to test the same inversion on a component described without its function, with explicit chains that can be criticised link by link.
Generalisation · industrial property
A patent described by its properties, without its claim. The engine would propose the uses those properties make possible. A clue to the surplus already exists: a patent later cited by patents from another technological class has served a purpose other than the one intended.
The measure already exists in the literature, after the fact. The engine would aim to produce it beforehand. Mandatory safeguard: de-identify the patents, so that the model infers rather than remembers.