From an epoch of measure and thought to integrated perceptual intelligence

The word "scientist" is less than two centuries old.

William Whewell proposed it in 1834, in a review of Mary Somerville's book on the connections between the physical sciences. Until then, people who studied nature were natural philosophers, or men of science. The new word named a person with a profession: someone whose work was to produce reliable knowledge about the world, in institutions built for that purpose.

What grew around that person is the architecture we still live in. Instruments measure. People interpret. Journals, reviewers and replications decide what counts. Other people, often in other institutions, act on the result. Each step is separated from the next, and each separation does a job. Specialisation lets one person go deep. Independent checking catches the errors of the first observer. Publication lets knowledge outlive the people who produced it.

The cost of the arrangement is time. Between something happening in the world and a reliable statement about it entering the shared record, there are instruments to build, funds to win, data to collect, papers to write and results to reproduce. Some of those steps take seconds. Others take decades. Much of the apparatus of science can be read as a response to that distance: a way of keeping claims honest when observation is slow, expensive and done by a few.

The division was never complete. Experimenters have always thought with their apparatus, instrument makers and theorists have shaped each other's work, and some of the most productive science has happened where observing and reasoning were done by the same hands at the same bench. What the epoch separated was less the activities than their scale: the larger and more consequential the question, the more the work was spread across people, places and years.

We call this the epoch of measure and thought. It is defined less by any instrument or theory than by a division of labour between the act of measuring, the act of reasoning about what was measured, and the act of doing something about it.

A civilisation can hold enormous quantities of measurement and still be blind to what its instruments were never designed to notice.

The founding text of information theory is explicit about what it leaves out. Claude Shannon wrote that "these semantic aspects of communication are irrelevant to the engineering problem." His theory says how much can be sent through a noisy channel. It says nothing about what is worth sending. That restraint is why it works, and it is also the clearest statement of the epoch's habit: decide in advance what matters, then measure that well.

Later work made the choice of relevance explicit. The information bottleneck method of Tishby, Pereira and Bialek compresses observations while keeping what bears on a chosen variable. Everything else can go. The method is precise, and the precision depends on someone having named the variable first. Kolchinsky and Wolpert have proposed a definition of semantic information as the part of a system's information about its environment that matters for its continued existence. It is a proposal rather than a settled measure, but it points at the gap Shannon left open on purpose: between information that is present and information that means something to someone.

Most data infrastructure sits on the near side of that gap. Signals are sampled at rates chosen for yesterday's questions, averaged into summaries, sorted into categories and stored. What was discarded on the way cannot be recovered later, however good the analysis becomes. Our article Noise is signal works through what a measurement can contain and where recovery stops.

The way a sensor reports the world is itself a design choice, and it is changing. Compressed sensing showed that, when a signal has the right structure, it can be rebuilt exactly from far fewer measurements than classical rules require. Event cameras report changes in brightness pixel by pixel, as they happen, instead of whole frames on a clock. Unlimited sampling folds a signal that would saturate a conventional converter and recovers it afterwards. Each changes the grammar in which the world is allowed to speak, and each makes visible something the older grammar hid.

Perceiving is something a system does in the world, not a picture it receives from it.

In 1991 Rodney Brooks built robots that behaved intelligently without a central model of their surroundings, by coupling sensing directly to action. O'Regan and Noë later argued that seeing is a skill of exploring, governed by how sensations change as we move. Karl Friston's free-energy principle offers one formal account in which perception and action are two sides of the same inference. All three positions are debated, and none is a consensus. Together they mark a break with the picture of a detached thinker reasoning over a finished representation.

A second question follows: where does a perceiving system end? Clark and Chalmers argued that cognitive processes can extend into the tools and surroundings a person relies on. The most careful empirical case for something like this comes from an airline cockpit.

In "How a cockpit remembers its speeds", Edwin Hutchins studied how a crew flies an approach. The right speeds depend on the aircraft's weight and flap setting, and they change through the descent. No pilot holds them all in memory under workload. Instead, the crew looks the speeds up in a booklet of cards, sets small markers called speed bugs around the airspeed indicator, and confirms the settings aloud with each other. During the approach, remembering a number becomes seeing where the needle sits against a bug. Hutchins' conclusion was that the unit doing the remembering is the cockpit: two people, an instrument, a booklet and a procedure.

An integrated intelligence would be a crowd, not a single mind.

In 1978 Premack and Woodruff asked whether chimpanzees attribute mental states to others, and named the capacity a theory of mind. It has since become a formal object. Baker, Saxe and Tenenbaum modelled how people infer someone's goals by treating their actions as the output of a plan, and inverting it. Interactive planning frameworks let artificial agents model not only the world but the beliefs of other agents, including their beliefs about each other. Machine learning systems now learn to predict other agents' behaviour, and large language models have been tested against people on theory-of-mind tasks. Scoring well on such tasks says something about prediction. It does not settle whether anything is understood or experienced.

Where many agents observe the same world, a useful model of another agent has to go further than its beliefs. It includes what the other can observe and what it cannot, how reliable its instruments are, what its reports are for, and which observation would change its mind. Interactive planning frameworks already represent some of this formally. The point here is that the value of anyone else's report depends on a model of how that report was produced. Information theory, game theory and epistemology meet at that point.

Game theory adds the uncomfortable part. Aumann showed that, under strong assumptions about shared priors and common knowledge, rational agents cannot agree to disagree. Morris and Shin showed that more public information can make a coordinating group worse off, because people weigh what others will believe over what is true. Bergemann and Morris treat the design of what each agent gets to see as a lever that shapes outcomes without dictating any action. Information infrastructure, on this view, is never neutral plumbing. It is a form of governance.

This rules out the picture of one planetary brain. Real coordination happens through many overlapping centres, as Elinor Ostrom showed for the governance of shared resources, with incomplete views, private interests, access rules and disputes. Burton and colleagues describe how large language models can strengthen collective intelligence and also flatten it into shared errors. Any integration worth having would be plural, permissioned and contested.

Early versions already exist, and every one of them is partial.

Earth-system digital twins aim to combine observations, models and the consequences of human decisions in one continuously updated representation of the planet; Hazeleger and colleagues argue they must be built with and for people, not only for forecasts. Aurora, a foundation model for the Earth system, transfers across forecasting tasks that used to need separate specialised models. Europe's Destination Earth programme links observation, simulation and decision support across national and disciplinary lines. Computational social science, as Lazer and colleagues described it in 2009, made the traces of daily life into a new instrument for observing society, with the privacy and representativeness problems that come with it.

Laboratories are moving in the same direction. Self-driving labs plan experiments, run them with robots and choose the next one from the results. They matter here as one small, visible case of a wider change: the loop from observation to model to action to correction getting shorter, and running inside one system.

What would turn these pieces into a perceptual system rather than a set of connected instruments? Our proposal is that three things would have to become routine. Observations travel with their provenance: where, how, by what and for what they were made. Each participant keeps a model of the others: what they can and cannot see, how stale their view is, how their reports relate to what they describe. And participants change what and where they measure in response to the gaps in the joint picture. The participants stay separate, with their own interests and permissions. What becomes shared is less a single model than the relationships through which observations flow and are weighed. On this view, the relationships between perceivers become part of perception.

The next step is harder than speed. A perceptual system in the full sense can change what it observes, and how, when its current categories stop working. It can decide that a variable it never measured matters, move or redesign a sensor, or abandon a representation that kept fitting the data while explaining the wrong thing. That capacity is still rare in machines and in institutions. It is the one this essay is about.

Our forecast: within about a century, science may stop being a separate category of human activity, or the word may come to mean something very different.

The argument runs as follows. Science stands apart because empirical checking has been slow, costly and specialised, so it had to be done by particular people in particular places. If observation, interpretation, experiment, action and correction become continuously coupled across ordinary infrastructure, empirical checking stops being a separate activity. It becomes a property of how systems run, in the way that literacy stopped being a profession and became a condition of everyday life.

What would go is the category, not what the category protects. Evidence, doubt, criticism and the unknown would remain, and they would need protecting in new ways. Discovery would remain wherever there are new places, scales and combinations to encounter. Creativity would remain, since it was never only the production of testable claims. Exploration would remain, and might speed up: the faster a civilisation can reconcile new observations with its models, the sooner it can turn to the next thing it does not understand.

The latency will not reach zero. Some observations take time by nature. Information from far away cannot arrive instantly. Two models can remain indistinguishable until the right experiment becomes possible. Zero latency is a direction of travel, not a destination. That makes the distinction between keeping the accessible world continuously checked and exploring the world not yet accessible more important, not less.

Put in one sentence, the transition would take us from a civilisation that builds knowledge about the world to one that develops intelligence in continuous relationship with it. The change would be less in how much is known than in how knowing happens.

Continuous perception is not truth. A system that is always updating can also be always confirming itself.

Accurate prediction can coexist with the wrong causal story. Successful control can coexist with ignorance of side effects. A system that chooses what it observes can learn to observe only what agrees with it. The more capable it is, the harder that becomes to see from inside.

There is already a small example. In 2023 a widely reported autonomous laboratory announced dozens of newly synthesised inorganic materials. In 2026 its authors published a correction in Nature, clarifying that the targets were new to their platform rather than necessarily new to science, and revising which syntheses could be counted as successes, after independent researchers questioned the identification of the products. The loop worked fast. The checking still came from outside it.

So the transition, if it comes, needs things the epoch of measure and thought built well and that integration could erode: independent scrutiny, records of where each observation came from, the right of others to disagree and the means to show it. It also needs answers to problems the old arrangement never fully solved: who may observe whom, who owns what is observed, and how a plural world coordinates without one view absorbing the rest.

UNDAMENTAL works on the preconditions, which are physical before they are computational.

A system can only perceive what reaches it. That depends on what is measured, at what fidelity, how it is kept, and whether it stays connected to the physical context it came from. Our work is to make the physical world more addressable: capture designed for machine intelligence, run alongside the systems organisations already have, that keeps more of what might later matter, can be reconfigured when the questions change, and stays answerable to the world it describes.

That is a modest part of the transition described here, and an early one. It is also the part where choices made now are hardest to reverse, because what is never observed cannot be recovered later.

1. Aumann, R. J. Agreeing to disagree. Annals of Statistics 4(6), 1236–1239 (1976). DOI ↗
2. Baker, C. L., Saxe, R. and Tenenbaum, J. B. Action understanding as inverse planning. Cognition 113(3), 329–349 (2009). DOI ↗
3. Bergemann, D. and Morris, S. Information design, Bayesian persuasion, and Bayes correlated equilibrium. American Economic Review 106(5), 586–591 (2016). DOI ↗
4. Bhandari, A., Krahmer, F. and Raskar, R. On unlimited sampling. International Conference on Sampling Theory and Applications (2017). arXiv:1707.06340. arXiv ↗
5. Bodnar, C. et al. A foundation model for the Earth system. Nature 641, 1180–1187 (2025). DOI ↗
6. Brooks, R. A. Intelligence without representation. Artificial Intelligence 47, 139–159 (1991). DOI ↗
7. Burton, J. W. et al. How large language models can reshape collective intelligence. Nature Human Behaviour 8, 1643–1655 (2024). DOI ↗
8. Candès, E., Romberg, J. and Tao, T. Robust uncertainty principles: exact signal reconstruction from highly incomplete frequency information. IEEE Transactions on Information Theory 52(2), 489–509 (2006). DOI ↗
9. Clark, A. and Chalmers, D. The extended mind. Analysis 58(1), 7–19 (1998). DOI ↗
10. European Commission. Destination Earth. Programme documentation (2024). Link ↗
11. Friston, K. The free-energy principle: a unified brain theory? Nature Reviews Neuroscience 11, 127–138 (2010). DOI ↗
12. Gallego, G. et al. Event-based vision: a survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 154–180 (2022). DOI ↗
13. Gmytrasiewicz, P. J. and Doshi, P. A framework for sequential planning in multi-agent settings. Journal of Artificial Intelligence Research 24, 49–79 (2005). DOI ↗
14. Hazeleger, W. et al. Digital twins of the Earth with and for humans. Communications Earth & Environment 5, 463 (2024). DOI ↗
15. Hutchins, E. How a cockpit remembers its speeds. Cognitive Science 19(3), 265–288 (1995). DOI ↗
16. Kolchinsky, A. and Wolpert, D. H. Semantic information, autonomous agency and non-equilibrium statistical physics. Interface Focus 8(6), 20180041 (2018). DOI ↗
17. Lazer, D. et al. Computational social science. Science 323, 721–723 (2009). DOI ↗
18. Levin, M. The multiscale wisdom of the body: collective intelligence as a tractable interface for next-generation biomedicine. BioEssays, e202400196 (2024). DOI ↗
19. Morris, S. and Shin, H. S. Social value of public information. American Economic Review 92(5), 1521–1534 (2002). DOI ↗
20. O'Regan, J. K. and Noë, A. A sensorimotor account of vision and visual consciousness. Behavioral and Brain Sciences 24(5), 939–973 (2001). DOI ↗
21. Ostrom, E. Polycentric systems for coping with collective action and global environmental change. Global Environmental Change 20(4), 550–557 (2010). DOI ↗
22. Premack, D. and Woodruff, G. Does the chimpanzee have a theory of mind? Behavioral and Brain Sciences 1(4), 515–526 (1978). DOI ↗
23. Rabinowitz, N. C. et al. Machine theory of mind. Proceedings of the 35th International Conference on Machine Learning, PMLR 80, 4218–4227 (2018). Link ↗
24. Ross, S. "Scientist": the story of a word. Annals of Science 18(2), 65–85 (1962). DOI ↗
25. Shannon, C. E. A mathematical theory of communication. Bell System Technical Journal 27, 379–423 and 623–656 (1948). DOI ↗
26. Strachan, J. W. A. et al. Testing theory of mind in large language models and humans. Nature Human Behaviour 8, 1285–1295 (2024). DOI ↗
27. Szathmáry, E. and Maynard Smith, J. The major evolutionary transitions. Nature 374, 227–232 (1995). DOI ↗
28. Szymanski, N. J. et al. An autonomous laboratory for the accelerated synthesis of inorganic materials. Nature 624, 86–91 (2023), first published as "…of novel materials"; author correction, Nature 650, E1 (2026). DOI ↗
29. Tishby, N., Pereira, F. C. and Bialek, W. The information bottleneck method. Proceedings of the 37th Allerton Conference (1999). arXiv ↗
30. Whewell, W. Review of On the Connexion of the Physical Sciences by Mrs Somerville. Quarterly Review 51, 54–68 (1834).

The forecast in section 6 is UNDAMENTAL's own synthesis. Levels four and five of Fig. 3.1 extend Hutchins' study. The observers in Fig. 5.1 are invented.

← All research