The Missing Link

The field made one assumption at the start: that thinking is the manipulation of symbols. It was never abandoned, only implemented better. Expert systems, neural networks, language models, now world models. Seventy years of progress built on the same premise.

Once intelligence is defined as symbol manipulation, every new architecture inherits the same assumption. Progress improves the implementation, not the premise. Language models therefore operate on descriptions of the world rather than on the world itself. Hallucination, missing grounding and ever-increasing training costs are not independent problems waiting to be solved. They are different expressions of the same assumption.

Nature solved intelligence differently. A cell, a flock and a body organise, adapt and repair without language, rules or plans. A soap film finds a minimal surface without geometry, optimisation or a blueprint. Order is not represented first and executed afterwards. It emerges directly from local interactions.

link takes that literally. Instead of representing the world through symbols, every element is coupled only to its local surroundings. There are no symbols, no rewards and no predefined rules. Structure is not programmed. It settles through coupled gradients and local stress reduction. Language models are not replaced. They remain the interface to people. link becomes the computational layer underneath.

The principle is deliberately tested where no domain knowledge can be hidden inside the system, and the first experiments have run. A coupled architecture identified tumour cells without ever being told what a tumour cell is, developed herding behaviour without language or behavioural rules, and found an efficient wing profile without aerodynamic equations. One mechanism across unrelated domains, which makes the mechanism matter more than any of the applications.

If this holds, capability comes from coupling to the world rather than from model scale. Many of the problems modern AI is trying to solve cease to be engineering problems and become artefacts of an unnecessary assumption. That is what the test is designed to settle.

The test. Four people, 24 months, modest compute. Success is defined in advance: a coupled system that remains stable in a rich, unpredictable environment and adapts to situations it has never encountered before. The question is simple: does the principle carry from simulated environments into the real world, or does it not?

For hundreds of millions of years, sunset meant darkness. Today the days are dimmer and the nights are brighter, flattening one of the strongest environmental signals biology has ever relied on. Medicine usually asks what modern life added. This asks what it removed.

Artificial light at night is associated with depression at a 3.29% increase in risk per lux of indoor exposure across 560,219 participants, with allergic disease at an odds ratio of 1.88, and with breast cancer and multimorbidity across nearly every major disease category. These conditions are treated as separate because they affect different organs and belong to different medical specialties. They may instead represent different expressions of the same upstream disruption.

The proposed translation mechanism is the mast cell. It sits at the intersection of immune, nervous, endocrine and vascular signalling and is regulated by melatonin through MT1 and MT2 during the biological night. Without a strong circadian amplitude there is no complete nightly reset. Natural experiments including blindness, shift work and polar environments repeatedly point in the same direction.

If that mechanism is correct, pharmacology should converge before diagnosis does. It does. Thirteen drugs developed for depression, diabetes, ADHD, asthma, urticaria, fibromyalgia and mastocytosis improve conditions outside their registered indications despite acting through unrelated primary targets. One property consistently reappears: mast-cell stabilisation. Montelukast provides the counterexample. It acts downstream without stabilising mast cells, fails to show the same cross-syndrome effects and carries an FDA black-box warning for depression and suicidal ideation.

Read backwards, the hypothesis reaches psychiatry itself. A 2022 umbrella review covering more than 115,000 participants found no consistent evidence that depression is explained by reduced serotonin. Imipramine binds the H1 receptor more strongly than classical antihistamines, a finding published in 1978 and largely ignored. Omalizumab improves mood despite having no monoaminergic target at all. The drugs may have been right while the mechanism inferred from them was wrong.

The same framework explains why several diseases disproportionately affect women. Estrogen lowers the activation threshold of mast cells while androgen raises it. Endometriosis, lipedema and hypermobile Ehlers-Danlos syndrome cluster far beyond chance despite belonging to different specialties. The shared mechanism may have been hidden by diagnostic boundaries rather than biological ones.

The hypothesis is deliberately written to be falsified. Its predictions are stated before testing. If mast-cell stabilisation provides no benefit in biologically selected patients, the hypothesis fails.

Existing bright-light therapy cohorts already contain the critical experiment. Mood, circadian amplitude, actigraphy and stored biosamples have been collected and paid for. The question is whether biological recovery follows restoration of circadian amplitude, and whether mast-cell biomarkers change before symptoms improve. If not, the hypothesis fails before a new trial is ever started.

Electricity has become digital everywhere except inside the distribution board. Every circuit is still planned in advance, connected by hand, verified manually and documented afterwards. The intelligence is expected to exist in the installer rather than in the system itself.

That assumption made sense when protection devices could only be electromechanical. It no longer does. Semiconductor switching, current sensing and embedded control have all become commodity technologies. The missing component is not hardware but architecture.

A distribution board should not require anyone to decide in advance where every circuit belongs. Every port should be generic. The board should identify what has been connected, determine its role, configure itself and refuse to energise until every connection has been verified. Documentation becomes a by-product instead of a task.

The opportunity appears because standards have finally caught up. Semiconductor-based protection inside distribution boards is now recognised by international standards. The limiting factor is no longer regulation. It is that nobody has built the complete system.

The problem becomes largest where skilled electricians are scarce. Millions of homes are built every year in markets where electrical expertise cannot scale with demand. Reducing installation from craftsmanship to verification changes both cost and safety at the same time.

The hypothesis is straightforward to falsify. If a board cannot reliably identify, configure and verify itself under real installation conditions, the concept fails regardless of the commercial opportunity.

Build one working board. Generic ports, automatic identification, internal routing and generated documentation. Install it in one apartment and compare installation time, documentation quality and commissioning against a conventional distribution board. If the system cannot outperform today’s process, there is no product.

The bottleneck of the energy transition has moved. Large transmission networks are no longer the dominant constraint. Congestion is increasingly created inside the low-voltage distribution grid, where millions of decentralised generators and consumers connect but almost nothing can be controlled. In the second quarter of 2025, 49% of all redispatch measures in Germany originated there, against 23% in 2023.

The industry is responding by building larger batteries. That solves the wrong problem. A battery connected to the transmission grid cannot relieve congestion inside a neighbourhood because the electricity still has to travel through the same overloaded cable. The limitation is geographical before it is electrical.

Grid operators therefore do not primarily lack storage. They lack individually addressable nodes at known physical locations. Every controllable action begins with knowing exactly where it can happen. In Germany they are obliged under §14 EnWG to keep the distribution grid stable, and they have nothing to do it with.

The cheapest way to build that control layer is not through expensive home energy systems but through a plug-in device any household can own outright for €199, at any of the roughly 50 million connection points in the country rather than the six million that already have a solar system or a heat pump. The battery is simply the mechanism that occupies the socket. The strategic asset is the address.

Once millions of addresses exist, the hardware becomes secondary. Electricity suppliers, aggregators and distribution grid operators can all steer the same physical network through software. Whoever controls the optimisation creates the value. The installed base becomes the platform.

One thing is worth naming. The market that would pay for this at scale exists in law and not yet in practice: the obligation under §14c is suspended until the regulator issues specifications, and there is no timetable for them. The bet is that the control layer has to be in the ground before that market opens, because afterwards there is no time to install it.

The hypothesis is easy to falsify. Real grid topology and real load profiles should demonstrate whether geographically distributed low-cost nodes remove congestion more efficiently than centralised storage. If they do not, the concept fails before any hardware is built.

Model one real distribution grid using existing topology and load data. Simulate several thousand addressable nodes and compare congestion relief against conventional storage investments. If distributed control does not outperform central storage where congestion actually occurs, the project ends before the first prototype is built.