VISION

SCIENCE

Tools for knowledge building on autopilot

Uncertain Systems is the autopilot for knowledge building. That sentence is a vision: a system that adapts around a person's mind, so difficult knowledge takes less cognitive effort to acquire, and the gap between what machines can discover and what people can understand gets smaller.

Today's software is narrower. Authors define knowledge regions and solve flows. People work through hard problems in Calibrate, Learn, Verify, and Solve sessions. The system records observable work, estimates a provisional knowledge state on a generative model, and can suggest a next step in that model. Understanding still happens in the person. The longer-term aim is knowledge acquisition with less conscious steering and less effort spent managing the path. That aim remains a research horizon, and it is not knowledge uploading.

Create your Workspace

ON THIS PAGE

Jump to a section.

Why understanding matters, the scientific influences, the measurement frame, and the research horizon.

Field depth pulling away from the people who can understand and teach it: thermodynamic computing, quantum computing, and mathematics as examples of growing nested structure.

01 · WHY

Why human understanding matters

New knowledge keeps arriving, and its nested structure keeps getting deeper. Fields such as thermodynamic computing, quantum computing, and mathematics are examples: more material, and more structure inside it. A person still has to understand, retain, and teach that structure. The capacity to do so does not automatically keep pace with the field.

The same pressure now shows up as an understanding gap. A model can produce a useful proof, a technical design, or a long argument that still exceeds someone's ability to inspect, question, and use it. We exist to close that gap: to make difficult knowledge cheaper, in cognitive demand, to acquire, while the person's understanding remains the point of the work.

Closing that gap is why this work exists. Today's software guides exploration through hard problems, stores the work, and estimates a provisional knowledge state. The larger aim is less cognitive effort spent acquiring the understanding that machines can already outrun.

EXAMPLE

A difficult proof

Take an AI-generated mathematical proof that a team cannot yet stand behind. A person still has to work through its assumptions, intermediate reasoning, and unfamiliar concepts. Human interpretability here means acquiring enough understanding to question and use the work. It is distinct from mechanistic interpretability of a model's internals, and distinct from automatic proof certification. The system's proposed role is to identify gaps in that understanding and to structure further investigation.

CONTEXTS

Hiring and company expertise

The same pattern appears in technical hiring and in company expertise. A candidate or a colleague can reach an answer, or read the documentation, and still be unable to explore an unfamiliar constraint. Verification, in that setting, is about exploration and reasoning, including work with an allowed model or harness when the problem says so. It is about the path, including the final answer. A score from that path does not, by itself, predict job performance.

02 · INFLUENCE

Uncertainty, surprise, and the Free Energy Principle.

Our point of view on knowledge building is heavily influenced by the Free Energy Principle. In that theoretical picture, adaptive systems that persist act as if they keep surprise within bounds. Surprise, or surprisal, is how unexpected an outcome is under a model. Variational free energy is a formal modelling quantity: a tractable upper bound on surprise, used in approximate inference. Prediction error is a process-level signal in some accounts of perception and action. Uncertainty is a property of beliefs about hidden states. These quantities are related in the theory. They are not interchangeable names for the same thing.

Useful exploration can reveal unknowns. An informative step can raise the number of open questions even as it improves the model of what was already in view. Learning, in this influence, is work on uncertainty about hidden structure, including the structure of a concept someone is trying to acquire. Variational free energy is not fuel the brain burns, and less effort in our product language is not free-energy minimization or reduced brain metabolism.

That is why the platform is named Uncertain Systems: we treat knowledge building as reducing uncertainty about a useful knowledge configuration, rather than as accumulating test scores.

The Free Energy Principle is an influence on this point of view. We do not claim that Uncertain Systems is Friston's full formal theory, and we do not claim completed FEP empirical results. The live Unsys model uses these Concepts for Calibrate, Learn, and Verify. It does not compute variational free energy.

POLICY

Foraging for information, not chasing scores.

Epistemic foraging is an active search for information to reduce uncertainty about an environment, rather than immediately chasing rewards. In Karl Friston's active-inference account, actions carry epistemic value: the expected information that would shrink uncertainty about hidden states.

That policy is how we think about the understanding gap. If the bottleneck is inspection, questioning, and use, then the useful next step is often the one that exposes what is still unheld, rather than immediately chasing rewards. On Uncertain Systems, Think Aloud Protocol probes, interruptions, and Proof of Work traces are the current forage: they search for signal in observable work. The Unsys model is the generative model that forage is scored against.

Epistemic foraging is an influence on this point of view. We do not claim that Uncertain Systems is Friston's full formal theory, and we do not claim completed active-inference empirical results.

A generative model you can act on.

03 · SCIENCE PATH

Knowledge tomography, then knowledge induction.

Tools for knowledge building on autopilot need a policy, a way to measure, and a longer-horizon aim. Epistemic foraging is the policy: search for information that reduces uncertainty, rather than chasing scores. Knowledge tomography reconstructs a provisional state from observable work. Knowledge induction is the later aim of changing knowledge configurations.

We keep the layers distinct on purpose. Epistemic foraging is the policy. Tomography externalizes and reconstructs current state from observable work. Knowledge induction tech uses that measurement to choose what to explore next. Without tomography, induction is blind steering; without induction, measurement never compounds into knowledge building on autopilot. A cheapest-edge score on the Unsys model is a software measurement of travel against a frozen generative model. It does not establish intelligence, hiring validity, or genuine understanding.

Epistemic foraging A generative model you can act on.

01 · Policy

Epistemic foraging

Epistemic foraging (Friston / active inference) is an active search for information that reduces uncertainty about an environment, rather than immediately chasing rewards. Tomography measures the state being foraged; induction is later transformation.

02 · Measurement

Knowledge tomography

Knowledge tomography is the family of methodologies that prompt human as well as agentic entities to try to reproduce their state of knowledge: multi-angle projections of what appears held, missing, and transferable, not finals alone. The reconstruction is provisional. It does not see all of a mind's knowledge.

03 · Long-horizon aim

Knowledge induction tech

Knowledge induction is the longer-horizon aim: technology that guides transformation through knowledge configuration space, raising proximity to useful states. Tomography measures; induction transforms. This is an Uncertain Systems aim, still ahead of a completed capability.

04 · APPROACHES

Approaches to cognitive enhancement

If the aim is difficult knowledge at lower cognitive demand, cognitive enhancement research is part of the landscape. It is a research literature. Association with that literature does not validate Uncertain Systems.

In Hacking the Brain: Dimensions of Cognitive Enhancement (Dresler, Sandberg, Bublitz, Ohla, Trenado, Mroczko-Wąsowicz, Kühn, and Repantis, ACS Chemical Neuroscience, 2019), Martin Dresler and colleagues group strategies by main mode of action into biochemical, physical, and behavioral. They also stress other dimensions: the cognitive domain targeted, personal factors, time scale, side effects, availability, and social acceptance. Benefits can be domain-specific, vary by person, and involve trade-offs. A plausible mechanism is not an established useful effect. No route is uniformly effective or appropriate.

Our headings follow that paper for the first two labels. "Neural interfaces and neuromodulation" is our narrower, product-relevant adaptation of part of their physical category, with overlaps. Physical approaches in the paper are broader than brain-computer interfaces or stimulation alone. These three headings are ours. They are overlapping research approaches. They are not a ladder, an intelligence scale, a cost ranking, or an inevitable product sequence.

Behavioral cognitive enhancement

Behavioral strategies change what people do: tasks, habits, training, and the structure of interaction. Dresler and colleagues discuss examples including mnemonic strategies and meditation. They also note that computerized training has faced serious questions about broad transfer: gains often stay close to the trained task.

Uncertain Systems' solve flows and structured problem exploration sit in this research area as software. That is a starting point, not validation by association, and not evidence of general cognitive enhancement.

Think-aloud can make aspects of reasoning observable. Observing cognition is distinct from enhancing it. Effects, and measurement reactivity, depend on the task and the protocol. We do not treat think-aloud as a proven general-purpose attention booster.

Biochemical cognitive enhancement

Biochemical strategies influence cognition through biochemical mechanisms. The research category includes pharmacological agents and other biological approaches, discussed in the literature at a high level. Psychological priming is a narrower construct than chemical or biological intervention, and we do not use priming as an umbrella for this category. Biological mechanisms do not all belong in one clean box.

This is a possible research direction. It is not an Uncertain Systems service, recommendation, or announced development program. Any future exploration would need appropriate expertise, consent, ethical oversight, and applicable regulation.

Neural interfaces and neuromodulation

Cinel, Valeriani, and Poli (2019) survey neurotechnologies that observe brain activity and neurotechnologies that influence it. Recording can support interaction or adaptation. Stimulation and other neuromodulation can change neural activity. Systems may combine both. Recording is distinct from stimulation. Neither implies that understanding can be uploaded.

Approaches may be invasive or non-invasive. Non-invasive is a placement description. It is not a safety conclusion. Neuromodulation can also involve biochemical mechanisms, so these headings overlap.

Cinel and colleagues is a 2019 review with forecasts, not a current product-readiness assessment. Uncertain Systems does not read neural activity in today's software.

05 · FRAME

Human-AI symbiosis

The unifying frame is human-AI symbiosis: people and computational systems adapting in response to each other during knowledge building.

J. C. R. Licklider (1960), in Man-Computer Symbiosis, described a hoped-for coupling in which people set goals, formulate hypotheses, determine criteria, and perform evaluations, while computing machines do the preparative work that makes insight and decision possible. He was writing about formulative thinking: bringing the machine into the posing of the question, as well as into the solution of already-posed problems. That paper is a historical influence on how we talk about partnership. It is not evidence that Uncertain Systems has achieved measured co-adaptation.

A design and research framing for a closed loop is: observe work, estimate a provisional knowledge state, adapt the interaction, and check subsequent understanding.

Parts of that loop exist in software as projections of taps, speech, and canvas work onto a frozen generative model, with Learn showing preferred next edges and Verify recording distance from a cheapest policy. That is an engineering mechanism on author-seeded structure. It is not a readout of neural activity, and it is not evidence of measured human-AI co-adaptation. The loop as a whole remains a research framing wherever those stronger claims are unconfirmed.

Cognition remains in the person. The ambition is less conscious steering and lower cognitive demand while that cognition is happening. Understanding is still the person's work. The ambition is also larger than keeping every step under fully manual control.

06 · NOW AND NEXT

Software now, and the research horizon

Current software

Authors define a knowledge region as a shared target of problems and concepts. A solve flow guides someone through a problem and stores work traces. Calibrate, Learn, and Verify use a generative model of hidden states, state factors, observations, priced edges, and policies. Think-aloud probes and Proof of Work traces make parts of exploration reviewable. Workspace agents can help prepare maps and flows for authors to review. The model is not a readout of a mind. Distance from a cheapest policy is not a certificate of understanding.

Research horizon

The larger aim is an autopilot for knowledge building: adaptation around the person so that acquiring difficult knowledge costs less attention and less conscious management of the path, and so that people can inspect, question, and use work that machines can already produce. Biochemical and neural routes, if ever explored, would be research under the constraints below, not a present offering.

What would count as progress

We would ask evaluation questions, rather than announce results we do not have. Against appropriate baselines: does understanding retain. Does it transfer to unfamiliar problems. Does reasoning quality improve on inspection of the work. Does subjective workload fall. Time saved or a higher score, taken alone, does not establish knowledge acquisition. Effort, in these questions, means cognitive demand and reported workload. It is not metabolic energy.

Existing website copy is a description of intent and of current software. It is not independent validation.

If neural or biochemical research were ever pursued

It would require consent, protection of sensitive data, ethical oversight, and applicable regulation. Workplace use would have to confront the danger of coercion: a person should not have to submit neural or biochemical measurement as a condition of hiring or advancement.

07 · SOURCES

References

Published research, theoretical influences, and historical papers that this page draws on. They are context. They are not evidence that Uncertain Systems works.

Dresler, M., Sandberg, A., Bublitz, C., Ohla, K., Trenado, C., Mroczko-Wąsowicz, A., Kühn, S., & Repantis, D. (2019). Hacking the Brain: Dimensions of Cognitive Enhancement. ACS Chemical Neuroscience, 10(3), 1137–1148.
Mode-of-action taxonomy (biochemical, physical, and behavioral) and further dimensions: domain, person, time scale, side effects, availability, and social acceptance.
Cinel, C., Valeriani, D., & Poli, R. (2019). Neurotechnologies for Human Cognitive Augmentation: Current State of the Art and Future Prospects. Frontiers in Human Neuroscience, 13:13.
Recording versus influencing brain activity, invasive and non-invasive methods, and a 2019 forecast. Vocabulary and context, not a current product-readiness assessment.
Licklider, J. C. R. (1960). Man-Computer Symbiosis. IRE Transactions on Human Factors in Electronics, HFE-1, 4–11.
Historical frame for people and computers adapting together in formulative thinking. An influence on the symbiotic vocabulary, not evidence of measured co-adaptation today.
Friston, K., Rigoli, F., Ognibene, D., Mathys, C., Fitzgerald, T., & Pezzulo, G. (2015). Active inference and epistemic value. Cognitive Neuroscience, 6(4), 187–214.
Definitional paper: epistemic versus pragmatic value of action. Forage for information, rather than immediately chase rewards.
Friston, K., FitzGerald, T., Rigoli, F., Schwartenbeck, P., & Pezzulo, G. (2017). Active Inference: A Process Theory. Neural Computation, 29(1), 1–49.
Process-level account of policies that minimize expected free energy, including information-seeking.
Schwartenbeck, P., FitzGerald, T., Dolan, R. J., & Friston, K. (2013). Exploration, novelty, surprise, and free energy minimization. Frontiers in Psychology, 4, 710.
How exploration, novelty, and surprise reduce uncertainty under free-energy minimization.
Parr, T., Pezzulo, G., & Friston, K. J. (2022). Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. MIT Press.
Readable book-length frame for active inference and the Free Energy Principle.
Friston, K. (2010). The free-energy principle: a unified brain theory?. Nature Reviews Neuroscience, 11(2), 127–138.
The Free Energy Principle overview that the adjacent science section already names as an influence.

Uncertain Systems