Intelligence is not a neutral or objective property. It is, as Summerfield puts it, a normative concept, constructed by psychometricians to value certain behaviours over others, and therefore always shaped by what society has decided to prize.

This is an uncomfortable starting point for a field that increasingly treats intelligence as something to be engineered. If we cannot agree on what intelligence is, how are we expected to build and experience it? And yet that is precisely what the AI industry is attempting, at extraordinary speed and scale, with over $700 billion committed to AI infrastructure in 2025 alone.

It was against this backdrop that the British Neuroscience Association convened a debate between two of the world’s most influential neuroscientists: Karl Friston and Christopher Summerfield.

Subscribe now

The question on the table — is information all you need to create intelligence? This article will explain the two fundamentally different accounts of what intelligence is, where it comes from, and whether artificial systems can ever truly possess it.

Karl Friston is the most cited neuroscientist alive, and creator of the Free Energy Principle — a mathematical framework that attempts to explain all of perception, action, learning, and even evolution under a single unifying idea. He arrived at the debate representing what he calls the “neat” position, the belief that a single governing principle underlies intelligence at every scale.

Christopher Summerfield, Professor of Cognitive Neuroscience at Oxford and Research Scientist at Google DeepMind, took the opposing “scruffy” view, that intelligence is not the expression of one grand principle, but a set of practical solutions shaped by the specific demands of the world.

The tension between these two positions determines how we should build AI, what we should expect from it, and whether the systems we are already deploying deserve to be called intelligent at all.

“Karl has spent the last 20 years arguing that the brain is best thought of as a generative model. For 19 of those years, everyone thought this was an eccentric position… and then large language models happened.”

— Christopher Summerfield

This remark is a useful entry point. It captures something important: the divide between these two arguments is not about whether neuroscience is relevant to AI. Both sides agree that it is. It is about the level at which that relevance operates. Is intelligence best understood from the top down, through first principles, or from the bottom up, through the accumulated solutions that evolution and engineering have independently converged upon?

The Free Energy Principle (FEP)

Friston’s central claim is that the brain is an active prediction machine. At every moment, and at every level of the nervous system, the brain is generating a model of the world and comparing that model against incoming sensory signals. It is perpetually trying to minimise the gap between prediction and reality - a quantity Friston calls surprise, and more formally, variational free energy. The word surprise here is not the everyday concept; it is a precise mathematical concept measuring how improbable a given sensory signal is, given our current model of its causes. On this account, the brain has one fundamental drive: to make the world less surprising by continuously updating its model of what is out there.

This idea has deep roots in theoretical neuroscience. Rao and Ballard’s landmark 1999 study of the visual cortex showed that higher brain areas send predictions downward to lower areas, while lower areas send prediction errors back upward — displaying the residual mismatch between what was expected and what arrived. Friston’s contribution was to generalise this into a unified principle that operates across the entire brain, and across all timescales. Perception here is the named process of updating predictions to reduce error, and learning is the slower process of updating the model itself. Even attention can be understood as the brain adjusting how much weight it gives to prediction errors from particular sources.

What makes Friston’s framework radical is the extension of this logic from perception to action. If the brain can resolve surprise either by updating its model or by changing the world, then action becomes a form of inference. Rather than selecting behaviours because they lead to reward, an agent acting under the free energy principle selects behaviours because they are predicted. They correspond to the future states the agent’s generative model expects to inhabit. This is what Friston calls active inference, and it is the move that makes the FEP a key theory of the whole of intelligent behaviour, not just cognition. As he put it during the debate: “Everything that we do, we perceive, we are, can be explained in terms of resolving uncertainty through information seeking.”

However, the FEP is not a theory in the falsifiable sense. It is a principle, closer in character to the principle of least action in physics. It does not, on its own, make specific testable predictions. Friston acknowledged this distinction explicitly in the debate. This unfalsifiability at the principle level is one of the most contested aspects of the framework, and it sets up the sharpest point of disagreement with Summerfield.

What the World Demands

“If we want to get down and dirty and build a brain, we need to think about how the world is.”

— Christopher Summerfield

Summerfield does not begin with the brain. He begins with the world. His argument is that intelligence is a set of solutions that evolved in response to the specific, non-accidental structure of the environment. The world has properties, and those properties are not arbitrary. Time flows in one direction and cannot be reversed; hence, any agent navigating the world must carry information about the past into the present. Space is three-dimensional and roughly continuous, so agents must build representations not just of objects but of the spatial relationships between them. The world generalises; what was true before tends to be true again, but it also has exceptions, singular events that resist generalisation and must be stored specifically. We are also social animals, which means our brains have evolved dedicated computational machinery for modelling other minds. And our bodily needs recur with regularity, to perform actions like showering, eating, etc. This means intelligence cannot be purely intellectual; it must also be motivational.

Summerfield turns this analysis onto the AI systems that currently exist and argues that they succeed because their architectures have independently arrived at solutions to the same problems. For example, the transformer, the architecture underlying every major large language model, relies on a mechanism called self-attention, which functions as a form of content-addressable memory: it learns associations between all tokens in a sequence and retrieves relevant context dynamically. Summerfield notes that this is not dissimilar to how the medial temporal lobe and neocortex interact in biological brains to arbitrate between generalisation and the recall of specific episodes. Transformers were not designed with neuroscience in mind, but in solving the engineering problem of handling sequences over time, they arrived at a computational structure that rhymes with the brain’s own solution. This convergence, Summerfield argues, is what happens when you solve the right problems.

Where both arguments differ most is in what this implies going forward. Friston’s framework suggests that if you get the principle right, the architecture follows. Summerfield inverts this: understand the demands of the world first, then ask whether your system genuinely meets them, not just in principle, but in practice. The implication is that the gap between current AI and genuine intelligence is not a matter of the wrong equations, but of missing components: embodiment, active information-seeking, richer memory, and the capacity to model other agents. These are all human traits which are hard to replicate.

Where They Clash

I have structured the debate below in a playwright style, so it is easier to follow. This is a paraphrased version of what was discussed.

Summerfield: The problem with the Free Energy Principle is not that it explains too little. It may explain too much. It can describe how an intelligent system updates itself in response to uncertainty, but it struggles to tell us what that system should care about in the first place.

Intelligence is normative. We call certain behaviours intelligent because we value them. Curiosity. Planning. Cooperation. Survival. But if those values are never specified, then minimising surprise alone gives no moral direction at all.

Friston: But those values are specified. They are encoded in the priors of the generative model. An organism expects itself to occupy certain kinds of states and avoid others. What we call value is simply the statistical structure of the states an agent is disposed to inhabit.

Summerfield: That creates another problem. If any value can simply be inserted into the prior, then the framework itself provides no guidance about which values matter. A system minimising surprise may optimise perfectly for its own predictions, while remaining completely indifferent to human well-being.

And for artificial intelligence, that distinction becomes critical.

Friston: The mistake is assuming reward itself is fundamental. Reinforcement learning treats reward as the primitive force driving behaviour. But reward is not primary. It is secondary. It reflects prior beliefs about preferred states.

Summerfield: Yet reinforcement learning is currently the paradigm producing the most transformative AI systems we have. AlphaGo. Modern agents. Large-scale optimisation systems. These systems work because they maximise reward successfully in practice.

Dave Silver, the architect of AlphaGo and one of the intellectual fathers of the reward-maximisation framework, raised over £1 billion for a company built on precisely that principle earlier this year.

Friston: Enough for engineering, perhaps. But not enough for understanding intelligence itself.

Summerfield: And this is where the disagreement becomes deeper than AI architecture. The question is no longer just how intelligent systems operate, but what intelligence fundamentally is.

Is intelligence a universal property of self-organising systems, reducing uncertainty?

Or is it a goal-directed capacity that only makes sense relative to what an agent is trying to achieve?

The debate never fully resolved that question, but it revealed something more important: beneath the technical language of priors, rewards, and generative models lies a philosophical divide about the nature of mind itself, one that may ultimately determine how we build the next generation of artificial intelligence.

What Does This Mean for AI?

Friston’s diagnosis of current AI is unsparing. Large language models run on what computer scientists call von Neumann architectures, chips in which computation and memory are physically separated, and where roughly 99% of energy is consumed simply moving data between them. Real biological intelligence does not work this way. Neural computation is in-memory: processing happens where the data lives. This is the principle underlying neuromorphic computing, where hardware architectures are modelled on the brain’s own organisation. It is where Friston believes the future of genuine AI lies: small, efficient, embedded systems that operate based on active inference rather than passive pattern-matching on vast datasets.

“I don’t think the current direction of travel is sustainable,” he said plainly, pointing to the data centres consuming hundreds of billions in investment as a symptom of building intelligence the wrong way.

Summerfield does not entirely disagree with the sustainability concern, but his prescription is different. Where Friston sees LLMs as architecturally wrong in principle, Summerfield sees them as incomplete. The transformer architecture already embodies several of the right computational solutions: content-addressable memory, generalisation over structured representations, socially-steered objectives through reinforcement learning from human feedback. What it lacks is embodiment, active information-seeking, and the kind of efficient world-modelling that comes not from scale but from constraint. His argument implies that the path forward is not to abandon current AI paradigms but to extend them: give these systems bodies, goals, and richer memory, and watch what they become.

Can AI Achieve Fundamental Intelligence?

Neither speaker gave a clean answer to the title question, and that irresolution is itself the most honest finding of the debate.

From Friston’s perspective, genuine intelligence requires active inference: a system that does not merely respond to inputs but acts on the world to bring about the states it predicts. Current AI does not do this. It is, as he put it, fundamentally passive, sophisticated at prediction, but not genuinely engaged with the world in the way that would make its predictions self-fulfilling.

From Summerfield’s perspective, the question is whether a system meets the demands of the world it inhabits: whether it has the memory to carry the past into the present, the spatial representations to navigate structure, the social models to operate among other agents, and the motivational systems to sustain itself. Current AI meets some of these demands partially. It does not meet all of them fully.

What the debate made clear is that the question “Can AI achieve fundamental intelligence?” cannot be answered until we have resolved what fundamental intelligence is. The audience voted, at the end of the debate, for Summerfield’s view that the world’s structure is the right starting point for building intelligence. However, the more significant convergence was the one neither speaker had to argue for: that intelligence, whether in neurons or in silicon, is fundamentally about managing uncertainty in a structured world. The question that remains, and that the next generation of AI researchers will have to answer, is whether understanding that principle is enough to build intelligence, or whether the act of building will always reveal what the principle alone could never anticipate.

The next article will be: How is knowledge organised in the brain? Do LLMs work the same way? This will focus on insights from Tali Sharot’s Research.

Thank you for reading this and subscribing to The Agora Collective Substack.

References

Clark, A. (2016). Surfing Uncertainty: Prediction, Action, and the Embodied Mind. Oxford University Press.

Friston, K. (2010). The free-energy principle: a unified brain theory?. Nature Reviews Neuroscience.

Rao, R. & Ballard, D. (1999). Predictive coding in the visual cortex: a functional interpretation of some extra- classical receptive-field effects. Nature Neuroscience.

You didn’t suddenly become irrational. Your brain state shifted. Your reward system (driven by dopamine and the ventromedial prefrontal cortex) acted; your control system (dorsolateral prefrontal cortex) didn’t stop it. In that neurobiological moment, the trade felt inevitable. Now, emerging neurotechnology could detect, or even modulate, that shift before it costs you money. Here’s why that matters, and why it’s complicated.

Many irrational trading behaviours are predictable shifts in reward and control circuits, and new brain technology could detect or shift those states in real-time.

Subscribe now

Two Fields, One Problem

Neuroeconomics is the study of how the brain computes value, reward, and risk during financial decisions. It reveals that “irrational” trades aren’t random; they follow predictable neural patterns. (Knutson et al., 2007; Kuhnen & Knutson, 2005).

Cutting-edge neurotechnology includes tools that measure brain state in real-time (EEG + machine learning) or modulate it (transcranial direct current stimulation, or tDCS; brain-computer interfaces). These technologies, once confined to research labs, are moving into consumer applications (Plewnia et al., 2018).

Neuroeconomics identifies why traders make mistakes; neurotechnology asks whether we can detect or prevent those mistakes in real-time. If you can measure the brain state that precedes a bad trade, perhaps you can pause it and prevent a loss.

The Mistakes Traders Keep Making

When you see a stock rallying, you think “it’s going up, so it will keep going”, and you buy into positions only to see it fall back down again. That’s trend-chasing. The mistake is substituting a short-term pattern for actual valuation. Similarly, revenge trading is when you make a loss. The urge to erase the pain quickly makes you accept larger risks and abandon logic. Panic selling, however, is when threat takes over, your time horizon collapses from months to seconds, and you liquidate at the worst possible moment.

The Brain Model

Underneath these behaviours is some simple tension: fast reward-driven urges versus slower control. Think of it like a car.

The “Gas” pedal is your need for reward and how you value it. The ventromedial prefrontal cortex (vmPFC) helps integrate reward, emotion, and motivation into a single “how good is this trade?” signal (Knutson et al., 2007). Meanwhile, dopamine systems broadcast prediction errors, the difference between what you expected and what happened (Watabe-Uchida et al., 2017). Under market volatility, prediction errors spike. Your brain tries to find a pattern in the noise and ignores the fundamentals. This creates a false sense of confidence in short-term patterns. Plus, the nucleus accumbens shows heightened activation when observing others buying. Herding behaviour is literally wired into reward processing (Burke et al., 2010).

The “Brake” is your cognitive control. The dorsolateral prefrontal cortex (dlPFC) maintains working memory, inhibits impulsive responses, and supports flexible updating of beliefs (McGuire et al., 2010). The anterior insula processes uncertainty and risk aversion. It can override the “buy now” signal when uncertainty is high (Christopoulos et al., 2009). But control is effortful. Under stress, fatigue, or high arousal, it becomes harder to sustain (Xue et al., 2010).

If we have all these systems, the question is, why do things still go wrong? The problem isn’t that people can’t value assets. It’s that control systems fail under feedback and arousal, so the brain substitutes in cheap rules like “follow the trend” or “recover the loss” to cope with it.

Measuring Brain States

For decades, brain measurement meant fMRI machines in laboratories. Today, wearable EEG (electroencephalography) is commercially available, and machine learning makes it possible to classify broad cognitive states.

EEG measures electrical activity across the scalp with millisecond precision. Recent studies achieved 83–80%+ accuracy in classifying distinct cognitive states using machine learning on EEG features from the prefrontal cortex. It cannot read thoughts, but it can detect state shifts linked to arousal, sustained attention, and fatigue (Saeidi et al., 2021). In trading, that matters because the biggest mistakes are often state-driven: you don’t lose knowledge, you lose access to your rules.

Imagine a trader wearing a discrete EEG headset integrated with their trading platform. It would collect signals, alongside their trades and heart rate, detecting correlations in real time. Machine learning models would pierce through the data, comparing calm deliberation states with those of high arousal and output a “Risk of Impulse” score. If it crosses a threshold, the platform inserts a tiny speed bump: a 3-second pause and one question - does this align with your plan?

This is not a hypothetical. Studies already detect effects of overprice asset herding in the lab with asset market simulations (Haracz & Acland, 2013).

Modulating the Brain State

Measurement alone doesn’t change behaviour. What if you could enhance the brake system directly?

Transcranial direct current stimulation (tDCS) is a non-invasive technique that applies a weak electrical current through scalp electrodes to nudge cortical excitability. Research suggests that stimulating prefrontal control regions, including the dlPFC, can improve aspects of cognitive control and working memory. For example, one study found that around 20 minutes of dlPFC stimulation improved participants’ ability to maintain control during a demanding focus task, with gains persisting into a subsequent testing window (Plewnia et al., 2018).

That matters for trading because the dlPFC supports proactive control, error-monitoring, and flexible rule switching, and it helps maintain working memory, holding your plan and risk rules intact (Mansouri et al., 2019). In simulated market conditions, dlPFC stimulation reduced an overpriced bubble forming around an asset and shifted participants away from pure trend-following and toward more adaptive updating, suggesting a causal role for control circuitry in high-pressure dynamics (Jin et al., 2019).

Hype vs Reality

Unfortunately, clinical indication does not equal performance enhancement. tDCS was developed for therapeutic goals, not for making traders richer. The evidence base is mixed: effect sizes vary, replication is limited, and responses differ by dose, baseline state, and learning history. Plus, casual self-stimulation can carry real risks, including mood effects and individual vulnerability.

Even if we put stimulation aside and focus only on “read-only” neurotech like EEG, there are practical reasons you shouldn’t bet your P&L on it. The first is false positives, arousal can signal panic, but it can also signal healthy engagement, and blocking too many good trades to prevent a smaller number of bad ones is a net loss. The second is regime change, models trained in one market environment often fail in another, and crisis conditions can break the very “state markers” the system relies on. The third is adaptation, once traders know they’re being monitored, behaviour shifts and people change how they act to “look calm”. Finally, there’s signal noise. EEG outside the lab is messy, and movement, sweat, facial tension, and headset comfort can all confound the data, which is why lab-level performance often degrades in real environments (Katmah et al., 2021).

Neuroprivacy and Cognitive Liberty

The moment you measure brain state, you’ve crossed into something deeper than financial data. Your EEG reveals emotional states, stress levels, attention patterns, and (potentially) indicators of neuropsychiatric conditions, even if anonymised (Jwa et al., 2022).

This creates a set of risks. Consent becomes complicated in competitive environments: if a tool reduces important losses, “optional” can quietly become “expected.” Data ownership becomes a threat: who owns the impulse-risk score, and what stops it from being repurposed for hiring, firing, or behavioural profiling? And equity is unavoidable; expensive neurotech will reach elite institutions first.

Scholars increasingly argue for rights like cognitive liberty, mental privacy, and mental integrity (Istace, 2024). If firms treat attention and self-control as performance infrastructure, traders may lose the ability to say, “my mind is mine.” If attention and brain state become another data point, cognitive privacy becomes a trading right, not a luxury (EDPS, 2024; Neurorights Framework, 2024).

Next Steps

Neuroeconomics shows that trading errors aren’t just poor choices, they’re state shifts in the brain. When reward circuits surge and control systems fade, analysis gives way to shortcuts. Neurotechnology won’t predict a good trade, but it could help prevent those born from frustration or impulse.

There are real limitations like noisy signals, regime shifts, and the ethical landmines of measuring or modulating someone’s brain at work. However, these aren’t reasons to dismiss the field. They’re reasons to design it better. If done right, neurotech could become a safety tool, not a surveillance tool. In the next couple of years, we could see lightweight, opt-in systems that help traders. This is not to control them, but to give back control.

Because in markets where milliseconds matter, sometimes the edge isn’t speed. It’s the pause.

References

Burke, C. J., Tobler, P. N., Baddeley, M., & Schultz, W. (2010). Striatal BOLD response reflects the impact of herd behavior. Neuron, 68(2), 335–345. https://doi.org/10.1016/j.neuron.2010.09.022

Christopoulos, G. I., Tobler, P. N., Bfalls, B., King, S. P., & Dolan, R. J. (2009). Neural correlates of value, risk, and risk aversion contributing to decision making under risk. Journal of Neuroscience, 29(12), 3837–3847. https://doi.org/10.1523/JNEUROSCI.0614-09.2009

EDPS. (2024). Neurodata. European Data Protection Supervisor. https://www.edps.europa.eu/data-protection/our-work/publications/techdispatch/2024-06-03-techdispatch-12024-neurodata_en

Haracz, J. L., & Acland, D. (2013). Neuroeconomics of asset-price bubbles (Working paper). UC Berkeley Goldman School of Public Policy.

Istace, T. (2024). Establishing neurorights: New rights versus derived rights. Journal of Human Rights Practice, 17(1), 121–139. https://doi.org/10.1093/jhuman/huae042

Jin, X., Chen, C., Zhou, X., & Yang, X. (2019). Stimulating the Dorsolateral Prefrontal Cortex Decreases the Asset Bubble: A tDCS Study. Frontiers in Psychology, 10. https://doi.org/10.3389/fpsyg.2019.01031

Jwa, A. S., Bae, G. Y., & Hardcastle, S. J. (2022). Addressing privacy risk in neuroscience data. Nature Human Behaviour, 6(9), 1163–1169. https://doi.org/10.1038/s41562-022-01406-6

Katmah, R., Al-Shargie, F., Tariq, U., Babiloni, F., Al-Mughairbi, F., & Al-Nashash, H. (2021). A review on mental stress assessment methods using EEG signals. Sensors, 21(15), 5043. https://doi.org/10.3390/s21155043

Knutson, B., Wimmer, G. E., Kuhnen, C. M., & Winkelman, P. (2008). Nucleus accumbens activation mediates the influence of reward cues on financial risk taking. NeuroReport, 19(5), 509–513.

Kuhnen, C. M., & Knutson, B. (2005). The neural basis of financial risk taking. Neuron, 47(5), 763–770. https://doi.org/10.1016/j.neuron.2005.08.008

Mansouri, F. A., Egner, T., & Buckley, M. J. (2019). Interaction of task-related learning and transcranial direct current stimulation of the prefrontal cortex in modulating executive functions. NeuroImage, 198, 170–179.

McGuire, J. T., Nassar, M. R., Gold, J. I., & Kable, J. W. (2010). Prefrontal cortex, cognitive control, and the registration of decision costs. Proceedings of the National Academy of Sciences, 107(17), 7922–7926. https://doi.org/10.1073/pnas.0910662107

Plewnia, C., Schroeder, P. A., & Wolkenstein, L. (2018). Targeting the biased brain: Non-invasive brain stimulation to ameliorate cognitive biases and decision making. Neuroscience & Biobehavioral Reviews, 84, 431–441.

Saeidi, M., Karachchi, H. K., & Thapliyal, H. (2021). Neural decoding of EEG signals with machine learning: A systematic review. Biomedical Engineering Online, 20(1), 74.

Watabe-Uchida, M., Eshel, N., & Uchida, N. (2017). Neural circuitry of reward prediction error. Current Opinion in Neurobiology, 49, 9–15. https://doi.org/10.1016/j.coneurobio.2017.10.006

Xue, G., Lu, Z., Levin, I. P., & Bechara, A. (2010). The impact of prior risk experiences on subsequent risky decision-making. Journal of Neuroscience, 30(22), 7595–7603. https://doi.org/10.1523/JNEUROSCI.0559-10.2010

pointed out that recently, neuroscience, financial ML, and network physics quietly have converged similar architectures. One being a regime detector that sits before the predictive model, not inside it. A Kuramoto-style network model in a 2025 PNAS paper showed that proximity to a critical transition can predict the trajectory of a collapse, something standard return‑based models are not designed to detect.

This caught my attention. In computational neuroscience, we’ve been using Kuramoto-type models to simulate whole-brain dynamics and reproduce patterns of functional connectivity. It felt natural to ask whether the same machinery could be used, in a disciplined way, for markets. Turns out, it has already been done.

An oscillator is any system that evolves in a repeating cycle, where its state can be summarised as its position within that cycle. The Kuramoto model is a way to describe macro-level dynamics of many coupled oscillators. Spotting where different regions spontaneously synchronise, depending on their frequencies, connections, and interaction strength, can tell us which areas are functionally related. I’ll first unpack the Kuramoto equation itself, then show how it’s already been used in finance, and where I think the next applications sit.

How does the Kuramoto model work?

Oscillations are represented as phases, the position of a system within its cycle, mapped onto a circle from 0 to 2𝜋. This makes it easy to compare timing across oscillating components. In the basic Kuramoto model, coupling is sinusoidal: the influence one oscillator has on another depends smoothly on their phase difference via a sine function. If two oscillators are slightly out of sync, they tend to pull each other toward alignment; if they’re in opposing phases, they push each other away.

Where:

i is the oscillator we are interested in, and j represents all other oscillators it interacts with at that given moment.

θi​ is the phase of the oscillator

wi represents the intrinsic frequency of oscillator i, i.e., its natural tendency to oscillate in the absence of interactions.

k is the global coupling parameter, indicating how strongly we want to simulate the coupling of i and j. The bigger the value, the more synchrony is promoted.

Cij represents the strength of the connection between i and j. In markets, this can be from past data, such as historical correlations and supply-chain links.

ηi(t) is the noise at that given time, signals from regions we are not factoring in, or sheer randomness

sin⁡(⋅) component captures how phase differences drive interactions, determining whether one oscillator speeds up or slows down relative to another.

Dij is the delay, introducing temporal lag in interactions, reflecting delayed information transmission rather than instantaneous coupling.

We use the Kuramoto model with noise and delay because financial markets are not instantaneous or fully observable systems: delays capture information transmission lags (e.g., across geographies or trading horizons), while noise represents idiosyncratic shocks and unobserved information flows that cannot be explained by common factors.

To quantify synchrony, we use the Kuramoto order parameter R(t), which ranges from 0 to 1. R≈0 indicates pure noise and incoherence, and R≈1 indicates strong alignment across oscillators.

From R(t), we can derive powerful statistics: R averaged across several timepoints as a measure of global synchrony, and the standard deviation of R over time as a measure of metastability (how much the system wanders between more and less synchronised states).

Subscribe now

Why does it apply to trading?

For traders, Kuramoto-type models provide a compact, interpretable way to measure and anticipate collective behaviour in markets. Its main value is acting as a regime detector: a way to track when a market or sector is drifting toward dangerously high synchronisation. High synchrony implies that shocks propagate more efficiently through the system, increasing the risk of cascading failures and reducing the effectiveness of diversification.

Dal’Maso Peron and Rodrigues (2012) show that during crises (e.g., Black Monday 1987 and the 2008 financial crisis), the Kuramoto order parameter rises sharply, indicating that assets move increasingly in phase. In such regimes, correlations spike, diversification breaks down, and systemic risk dominates, conditions under which many trading strategies fail simultaneously.

Within markets, this model can realistically capture:

• Macro-level co-movement of assets, as shown in the example above (e.g., during crises vs. calm periods).

• Sector and regional clustering, where there is partial synchronisation within strongly coupled clusters and weaker synchrony across them. Globalised economies can be represented as coupled oscillators with network-based synchronisation (Carrera et al., 2020).

• Regime transitions, critical shifts from low to high global synchrony as coupling parameters or network topology change, including abrupt, explosive transitions. We will go through an example of this below.

• Metastable dynamics where the system is repeatedly transitioning between temporary states of partial synchrony and desynchrony. Financial volatility can alternate between synchronised and desynchronised states over time (Li et al., 2025).

At the same time, Kuramoto has important limitations when applied to markets:

1. Phase-only, no amplitude.
   Kuramoto dynamics track only phase, not amplitude, so they ignore the magnitude of returns or volatility, which are central to P&L and risk. More general oscillator models, such as Stuart–Landau (used by Lee et al., 2025), include both amplitude and phase and are better suited when modelling volatility or index levels directly.

2. Sinusoidal coupling assumption.
   Interactions are assumed to vary smoothly with phase differences (sinusoidal coupling), which is appropriate for weakly coupled systems but may fail to capture threshold-driven or discontinuous dynamics seen in financial markets, such as forced deleveraging or liquidity shocks.

Can it predict the next financial crisis?

There is another layer to Kuramoto modelling for markets, combining it with the Stuart-Landau oscillators. Together, it can quantify how close a network is to an explosive synchronisation, called the ES proximity measure. This is where a system snaps into synchrony abruptly instead of gradually, solving for the limitations in using purely Kuramoto. Markets with more explosive behaviour are dangerous.

Lee et al (2025) tested this on the 2008 financial crisis data, specifically, daily stock prices for each country’s stock market index. This was then converted into correlations in price movements between countries, building a network of these markets essentially. Each market was represented as an oscillator with a phase.

The model was then applied, and two outputs were defined:

• collapse speed, how fast prices fall after the peak

• recovery speed: how fast prices rise from the bottom

The process is shown in the diagram below for ease of understanding.

There is far more complexity than this, so I encourage you to read the paper.

What We Can Take Away

There is an increasing amount of research that links the brain and markets in the mechanisms they use, hence we have more transferable models between the fields.

Here, the linking principle is synchrony. In both neuroscience and finance, the system’s vulnerability is not encoded in individual units but in how tightly they move together.

As research continues to merge ideas from network neuroscience, nonlinear dynamics, and financial ML, Kuramoto‑type models will likely become part of a broader toolkit for understanding systemic risk, contagion, and collective behaviour. It won’t be able to tell you which stock to buy tomorrow, but it can tell you when the market is entering a state where almost nothing behaves independently, and that’s the kind of information traders rarely get in time.

Thank you for reading!

References:

‌Carrera, Edgar & Avalos-Gaytán, Vanesa & Cardona, Yajaira. (2020). Synchronisation of Globalised Economies 10.13140/RG.2.2.28404.58249.

Lee, U., Kim, H., Kim, M., Oh, G., Joo, P., Park, A., Pal, D., Tracey, I., Warnaby, C. E., Sleigh, J., & Mashour, G. A. (2025). Proximity to explosive synchronization determines network collapse and recovery trajectories in neural and economic crises. Proceedings of the National Academy of Sciences, 122(44). https://doi.org/10.1073/pnas.2505434122

Li, J. C., Guo, J., Ma, R., & Zhong, G. (2025). Dynamic forecasting and mechanisms of volatility synchronization in complex financial systems. PloS one, 20(10), e0334853. https://doi.org/10.1371/journal.pone.0334853

‌Kauê, T., & Rodrigues, F. A. (2012). Explosive synchronisation enhanced by time-delayed coupling. Physical Review E, 86(1). https://doi.org/10.1103/physreve.86.016102

We can approach this in one of two ways:

1. Machine Learning

2. Mechanistic Modelling

This depends on our goal: do we want a plausible model that works slower or a largely accurate yet sometimes inexplicably wrong model that works faster?

How can we use ML in Markets?

Glaser et al. 2019 defined how we can use Supervised Machine learning in 4 different ways in systems neuroscience. These translate to markets.

The first three roles do not require the model to be a true explanation of the brain. They only require it to be useful. We cannot treat predictive success as understanding here.

Understanding in a Real World Context

Let’s put this into context. It’s Feb 28th 2026, the Iran-US war has just started, and Brent crude futures jumped nearly 7%, briefly hitting $126 per barrel.

Mechanistic approach

A mechanistic model would begin with a hypothesis about how the Iran war headline affects Brent crude through the structure of the oil market. This would be human-generated. The model would map the real-world transmission channels and simulate how stress to one part of the system changes the behaviour of the wider market. For example, we want to see what shipping insurance increases do to the futures curve.

By seeing a model’s reaction in these scenarios, we are better able to understand the reasons behind shifts in the market.

In neuroscience, even our best computational models rarely get close to reproducing full cognition. Instead, they capture specific circuits or tasks under tightly controlled assumptions and are constantly stress‑tested against new data. In markets, a mechanistic model plays a similar role. It would not be a full digital replication of the global oil system causally, but a deliberately simplified dynamical system that encodes hypothesised causal channels and is updated when it fails. This can be extremely valuable for scenario analysis and for understanding where historical relationships might break.

In the bigger picture, it would look like the image below.

Machine learning approach

A machine learning approach would treat an Iran-related oil shock primarily as a classification or prediction problem. The model would be trained on historical market features surrounding previous geopolitical stress events. The aim would be to identify patterns that distinguish normal market conditions from geopolitical stress regimes, and to learn a decision boundary between periods where Brent crude is likely to spike and periods where the market absorbs the news with limited price impact.

However, these models become vulnerable during black swan events, where the current market does not closely resemble the historical data on which the model was trained. JPMorgan said hedge funds were having their worst drawdown when the war was announced since the prior tariff shock, and Reuters reported that systematic quants like CTAs were hit especially hard. The Wall Street Journal reported that Millennium Management and Point72 each lost about $1.5 billion in one week. In these moments, models may need to be paused or overridden because they are calibrated to historical relationships that break down under extreme uncertainty.

These models are trained on past data; the last time the Strait of Hormuz was affected like this was during the 1980s Tanker War. It was already a critical oil chokepoint, but the market structure was very different. Flows through the strait fell sharply as tankers were attacked, yet the disruption directly affected only a small share of ships, and real oil prices drifted lower regardless through much of the decade. In other words, even a direct military threat to shipping did not translate into a simple, one-directional oil-price shock.

Today, the same type of disruption would travel through a much larger and more financialised system. Around 20 million barrels of oil per day now pass through Hormuz, alongside significant liquified natural gas flows, while Brent has become a global pricing benchmark for a large share of seaborne crude. A Hormuz shock would therefore not only affect Gulf exporters such as Saudi Arabia, Iraq, Kuwait, Iran and the UAE; it would also reprice Brent-linked barrels, shift the global futures curve, and feed into inflation, rates, FX, and risk sentiment.

This creates a fundamental problem for machine learning models: there may be no clean historical analogue. The model can detect statistical similarities between past and present stress events, but it cannot fully understand whether today’s market structure is different in a way that makes the outcome more severe. In a world where geopolitical shocks, inflation concerns, energy security risks, and cross-asset volatility are occurring simultaneously, the model may be forced to make predictions from incomplete precedent.

While an ML model may be useful for detecting whether Brent crude is entering a stress regime, its weakness is that it may not explain why the regime is forming or whether the present shock is structurally different from the past.

Concluding Remarks

However, we can argue that most of the time, we don’t need to know how something works; we just need to know that P&L incrementally compounds over time.

This is the key difference between neuroscience and markets. In neuroscience, interpretability is ethically necessary when someone’s life and livelihood are on the line from the impact of a disease. Money can be made back, but health cannot.

In markets, the cost of not understanding the mechanism is not nearly as severe. It appears as hidden leverage, crowded positioning, false confidence, and models that fail precisely when the market stops looking like their training data. That is why the strongest trading system is not purely machine learning or purely mechanistic modelling.

ML works better when the market behaves like history. Mechanistic modelling works better when the structure changes. The real edge is using both ML for speed and mechanistic modelling for guardrails.

Thank you for reading!

References

Bloomberg News. (2026, March 10). Balyasny, Millennium, Point72 hit by Iran conflict in wild week. Bloomberg.

CNBC. (2026, March 18). Hedge funds suffer worst losses since “liberation day” on Iran war turmoil. CNBC.

Reuters. (2026, March 25). Hedge funds get B-minus in Iran market stress test. Reuters.

Reuters. (2026, April 1). Hedge funds hammered by market turbulence triggered by Iran conflict. Reuters.

J.P. Morgan Chase & Co. (2026, March 12). JPMorgan says hedge funds hit by worst drawdown since April. Business Times.

The Wall Street Journal. (2026, March 10). Iran conflict triggers losses for Citadel, Millennium and Point72. The Wall Street Journal.

How do we see?

We see incomprehensible amounts of data on a day-to-day basis. The human eye has over 100 million photoreceptors to capture this, but must compress that data to fit through approximately one million axons to pass through the optic nerve.

After the optic nerve, the information passes through the lateral geniculate nucleus (LGN) to the visual cortex, where further processing occurs. Each area in the cortex focuses on a particular type of visual processing that leads us to be able to perceive the world around us.

Currently, Neuralink, Elon Musk’s neurotech company, has a program that helps those who are blind to generate visual perception with a cortical visual prosthesis. Blindsight aims to bypass damaged parts of the visual system and create visual percepts by electrically stimulating neurons in the visual cortex. This is done with a camera attached to an implanted electrode array.

It has been reported to work in monkeys, and now it has an FDA Breakthrough Device Designation to be tested on humans. Elon compares the first version created to graphics seen in retro video games like Atari.

How does Blindsight work?

Turning Pixels into Language Your Brain Speaks

To convert the input from the camera into biologically compatible information, the system must first pass it through a bandpass filter, removing redundancies such as blank space or regions that contain no useful information.

This can be done using a difference of Gaussians filter, which detects edge/contrasts in the visual image.

Let’s say we wanted to do this for an image that is 256 x 256 pixels, which is about 65,536 individual data points. We will call this image I(x,y), which captures the brightness at every pixel coordinate on a scale of 1-100.

The DoG is a template that looks like this:

It is exactly like the retina; the centre excites the neurons, and the surround inhibits the neurons. This passes across an image to capture all the relevant data points and puts it into the equation below:

The equation essentially subtracts the surround value from the centre value.

The output is a number for the area R(x,y).

If we are looking at a flat blue sky…

Image patch:
100 100 100
100 100 100
100 100 100

There would be no contrast. So output = 0

If we are looking at an edge…

Image patch:
50 50 50
200 200 200
200 200 200

There is a large contrast. Output = large number

After DoG, most outputs are 0, and a few outputs are large. This is called sparse representation.

The equation above represents R(x,y) over the entire image summed up. r represents the modelled firing rate response of a single neuron.

However, it does not have the properties of a neuron; the value could be negative, and there is no threshold (a certain value requirement for the firing to occur). The next stage would be to pass it through a sigmoidal function to allow the values to be on a realistic biological scale of 0-1.

This gives you a smooth firing rate that respects the physical limits of a real cell, where r = 0.8. It would be able to tell you, "In the next 100 milliseconds, this neuron has an 80% chance of spiking.”

However, we need this in a stochastic manner in order to be able to be interpreted by the brain. We run it through the equation below:

If neurons were like computers, a constant input would result in a perfectly even rhythm:

|____|____|____|____|

However, real biological neurons are noisy. Because of the random movement of chemicals in the brain, the same input actually looks like this:

|__|_|______|__|___|

The average number of spikes over a few seconds is exactly what the sigmoid predicted, but the exact timing of each spike is random (stochastic). This tells you the frequency of binary clicks over time.

There’s an Edge… but which Direction Does it Face?

Hubel and Wiesel demonstrated how, in the visual cortex, specifically V1, certain neurons respond to certain orientations. This is implemented in modern technology through the Gabor filter, which represents simple cells’ ability to process receptive fields.

The Gabor filter takes in a section of an image defined by its location, centre of the receptive field and size. This is all specified in the exp() section of the equation, and the cosine part of the equation detects the line and its orientation.

The value r is assigned to the summation of the whole image, just as we saw in the DoG filter.

A system could use these equations to process camera input before stimulating V1, bypassing damaged eyes/optic nerve/LGN, not bypassing the visual cortex itself.

By using the DoG filter to find edges first, Gabor filters to determine their orientation and through direct electrical microstimulation of the primary visual cortex (V1), the device can stimulate a pattern of phosphenes that traces the shape of an object, like a glowing neon outline. This is what we can hypothesise to see.

How does this Stimulus fit in with the Existing Visual Cortex Structures?

Even after DoG and Gabor filters create an edge-based representation, V1 does not passively receive the signal. Activity spreads across V1 during perception. If we stimulate one small region, the effect may not stay perfectly local. V1 is a living neural sheet with lateral interactions, so stimulation at one point can spread, amplify, suppress, or distort nearby activity.

This is where the Rankin and Chavane model becomes useful. It explores how a new stimulus interacts with the existing infrastructure over time.

In a simple camera or screen, if you light up one pixel, only that pixel glows. But in V1, if you stimulate one spot, the activity can leak into neighbouring regions. V1 is organised into orientation maps: some regions prefer vertical lines, others horizontal lines, and others diagonal lines. There is no single plug point, making blindsight-like more difficult.

Lateral selectivity (beta rec in the formula) tells us how selectively activity spreads between neurons that prefer similar orientations. If it is low, the output becomes blurry. For a cortical prosthesis, this means the perceived phosphene pattern may not exactly match the electrode pattern. The brain’s own circuitry will reshape it. We want high lateral selectivity for an orientation-specific and structured output. Similarly, we want a low excitation width (RW ex in the formula) so our output is precise and easy to interpret.

While the equation looks complex, it essentially shows how brain activity evolves over time through four parts:

1. Decay to rest: brain activity naturally fades away without stimulus

2. Stimulus input: the actual visual input entering the system.

3. Non-selective interaction: activity that spreads equally in all directions.

4. Selective lateral interaction: the most important part for this model, where activity follows the brain’s internal orientation map and creates more specific patterns.

For Blindsight, we would ideally want the selective lateral interaction to dominate enough for the percept to form useful outlines, rather than just a vague glow. But high lateral selectivity is not something the prosthesis simply “turns on” directly. It depends on the patient’s existing V1 wiring, the orientation map at the stimulated location, the spacing of electrodes, and the stimulation parameters used, such as current amplitude, frequency, pulse width, and timing.

So the aim is not just to stimulate the correct electrode. The aim is to calibrate stimulation so that it works with the cortex’s existing orientation maps, rather than fighting against them. This varies person to person…moving us swiftly to the rest of the limitations.

Limitations

The DoG filter only represents a few types of retinal ganglion cells - certain midget and parasol cells. The eye contains approximately 50 distinct types of Retinal Ganglion Cells (RGCs), each tuned to different features like colour, motion, or fine edges. For a prosthetic to restore naturalistic vision, it has to be able to model all of these.

For instance, models require systems that act as these cells:

• Parvocellular (Midget cells) handle the fine detail and colour we perceive.

• Magnocellular (Parasol cells) handle motion and spatial location.

• Koniocellular (Bistratified cells): These are a third, distinct pathway shown in the LGN layers.

This model also skips the LGN because Blindsight-style cortical stimulation bypasses the normal retina-LGN-V1 pathway. However, this is a major simplification because biological vision depends heavily on LGN gating, thalamocortical feedback, and recurrent loops between V1, LGN, and higher visual areas.

For a Blindsight-style prosthesis, the goal is not to recreate normal vision immediately. A more realistic first goal is to create useful vision: sparse phosphenes, glowing contours, object boundaries, doorframes, obstacles, and motion cues. This could help a patient navigate the world even if the percept is closer to a broken neon outline than a natural image.

The limitations mentioned are exactly why computational neuroscience matters. It gives us a way to turn biological ideas into testable models. The future of neurotechnology may not come from copying the eye perfectly, but from understanding which parts of the visual code are essential enough to rebuild.

Thank you for reading. I hope you enjoyed this breakdown.

Subscribe now

About the field

Computational Neuroscience uses mathematical modelling, computer simulations, and theoretical analysis to understand the brain's development, structure, and information-processing mechanisms.

This field was where artificial intelligence stemmed from. Science provided the biological principles that underpin the modern AI architectures like neural networks and learning algorithms. There is scope to develop the next type of model. Where intelligence has already been recreated, perhaps vision or touch processing can be recreated next.

References:

Scientific

Duncan, R. O., & Boynton, G. M. (2003). Cortical magnification within human primary visual cortex correlates with acuity thresholds. Neuron, 38(4), 659–671.

Gabor, D. (1946). Theory of communication. Part 1: The analysis of information. Journal of the Institution of Electrical Engineers-part III: radio and communication engineering, 93(26), 429-441.

Hubel, D. H., & Wiesel, T. N. (1968). Receptive fields and functional architecture of monkey striate cortex. The Journal of physiology, 195(1), 215–243.

Kim, U. S., Mahroo, O. A., Mollon, J. D., & Yu-Wai-Man, P. (2021). Retinal Ganglion Cells-Diversity of Cell Types and Clinical Relevance. Frontiers in neurology, 12, 661938.

Marr, D., & Hildreth, E. (1980). Theory of edge detection. Proceedings of the Royal Society of London. Series B, Biological sciences, 207(1167), 187–217.

Rankin, J., & Chavane, F. (2017). Neural field model to reconcile structure with function in primary visual cortex. PLoS computational biology, 13(10), e1005821.

Van Essen, D. C., Anderson, C. H., & Felleman, D. J. (1992). Information processing in the primate visual system: an integrated systems perspective. Science (New York, N.Y.), 255(5043), 419–423.

Neuralink Related

Coldewey, D. (2024, September 17). Neuralink’s “breakthrough device” clearance from FDA does not mean it has cured blindness. TechCrunch.

Reuters. (2024, September 18). Musk’s Neuralink gets FDA’s breakthrough device tag for “Blindsight” implant. Reuters.

Medical Device Network. (2024, September 17). Neuralink obtains breakthrough designation for Blindsight device.

Medical Devices Zone. (2024, September 26). Neuralink’s Blindsight receives US FDA’s breakthrough device designation.

Media MICE. (2026, January 10). Neuralink’s Blindsight implant receives FDA breakthrough device designation.

Neurotech Reports. (2024, July). Elon Musk, Neuralink, Blindsight, and the limits of visual cortex stimulation.

MobiHealthNews. (2025, March 6). Elon Musk provides update on Blindsight implant trials.

The Start: Israel first attacked Iran on 28 February with US backing, targeting its missile infrastructure, military sites and leadership in the capital, Tehran, and across the country. This was called Operation Epic Fury.

Ayatollah Ali Khamenei, Iran’s supreme leader, was killed during this first wave of strikes. On the 8th March, his son, Mojtaba Khamenei, was appointed as his successor.

Why is this happening?

The US Government has not given one clear reason. Trump says it is for regime change. Pete Hegseth, the US Secretary of War, says it is explicitly not a regime change. Other officials have explained that its purpose was to:

• Destroy Iran's missile and military capabilities

• Prevent Iran from ever having a nuclear weapon

• Secure its natural resources

Whilst long-term goals are to make sure all of these are true, this was a war of choice; there was no imminent threat.

Uncannily, it also comes at a time when Trump’s polls are down, and there is a possibility he’s trying to escape the reality of his ties to Jeffrey Epstein and the mounting pressure for the release of all the files.

How did Iran respond to the attack?

Iran then launched hundreds of drones and ballistic missiles at targets in Israel and at US military bases in Bahrain, Jordan, Kuwait, Qatar, Saudi Arabia, Turkey, and the United Arab Emirates on the same day.

These attacks have continued to go on. As of 13 March, 13 US service members had been killed. As of 15 March, Israeli authorities said 12 people – all civilians – had been killed by missile fire since the start of the war.

Iran has fired close to 2,000 missiles and drones at the United Arab Emirates during the war, with most intercepted by air defences.

Why is it bad if Iran had a nuclear weapon?

It would be an existential threat to Israel. Iran has historically threatened to wipe Israel off the map. Plus, neighbours such as Saudi Arabia might feel compelled to develop or acquire their own nuclear capabilities to match a nuclear-armed Iran, triggering a domino effect of proliferation in a volatile region and an arms race.

Why are the US and Israel allies? Should this war even be allowed?

They have had allyship since 1960. The US provides Israel with financial aid, and Israel provides a reliable, militarily powerful base of operations. They share the same goals when it comes to Iran.

The US-Israel attacks on Iran are considered illegal under international law, as they violate the UN Charter's prohibition on the use of force, because the attacks fail to meet the criteria for self-defence, and lack authorisation from the United Nations Security Council under Chapter VII.

Why is Iran even considered a threat?

• Iran and the US have the same-sized army, 1 million.

• Iran has ~200,000 Revolutionary Guards, an aggressive type of soldier that directly responds to the supreme leader, with direct access to missiles.

• They have motivation from the country’s history, a vendetta against the US from events like Operation Ajax, when the US overthrew Iran’s democratically elected Prime Minister Mohammad Mosaddegh in 1953.

• The Islamic Republic of Iran runs the biggest terror proxy network in the Middle East – Hezbollah in Lebanon, the Houthis in Yemen, Hamas in Gaza and militias across Iraq and Syria.

What pre-empted this attack?

Failed previous attacks. On June 22nd 2025, the US carried out strikes on Iran’s nuclear facilities, sites key to its uranium enrichment program, which could be used to develop nuclear weapons. No one knows if they destroyed these sites successfully.

However, on June 19th, multiple cargo trucks were seen leaving the Fordow Uranium Enrichment Plant, causing reason for this doubt.

I keep hearing about the Strait of Hormuz. Why is this important?

The Strait of Hormuz is a key route for global oil transport. More than 1,000 cargo ships, mainly oil and gas tankers, have been blocked from transiting through. 20% of the world's oil is transported through it. Iran has threatened to attack any oil tanker that passes through it, causing a global oil crisis. This is their main advantage against other attacks; it significantly impacts the global economy.

Import Dependency

90% of the oil from here goes to Asia, especially China, Japan, India, and South Korea. Strait of Hormuz disruptions could worsen access to fertilisers for some of the lesser developed countries like Sudan, Tanzania and Somalia.

Export Dependency

Saudi Arabia & UAE are the largest exporters through the region, bearing massive revenue risks despite limited pipeline bypasses.

What is unanswered right now?

• How will Iran’s new pro-hostility supreme leader retaliate against the US & Israel?

  Context: We still don’t know if Mojtaba Khamenei is alive. On 13 March, Pete Hegseth stated he had been "wounded and likely disfigured" in new strikes.

• Which other countries will get involved?

  Context: Donald Trump stated on March 15th that it would be "very bad for the future of NATO" if allies don't help secure the Strait of Hormuz.

• How will this war escalate?

  Context: Today, the Israel Defence Forces (IDF) has begun a wide-scale wave of strikes in parts of Iran. Israeli ground troops have been deployed. There is a possibility that the US will deploy ground troops as well.

• When will it end?

  Context: Trump said the Iran war will be over when ‘I feel it in my bones’…

Thank you very much for reading. I have made it a goal to research for at least 1 hour a day for the next 30 days, and I will be writing up what I learn. I’d love to discuss any of the write-ups in the comments below. Look forward to seeing you next time!

Subscribe now

Sources & Further Reading

• BBC News Live Feed

• Diary of a CEO Podcast (Nuclear & Strategic War Focus)

• The War on Iran, Explained

• Wikipedia

• UN Trade and Development

When was the last time you sat in silence? No music, notifications or screens. Just you and your thoughts. For most people, it’s unbearable. Yet, silence is where growth happens. It’s where curiosity has space to wander, to stretch, to connect dots that don’t look connected at first.

Curiosity is not just a luxury, it’s the foundation of intelligence.

“I have this urge to learn everything… cinematography, history, politics, science, music… We’re trained to isolate knowledge into pockets, when in reality it’s a web. Some of the greatest inventions and ideas came from that web. Da Vinci was the archetype of the polymath. Einstein engaged in philosophy and music. Ada Lovelace literally called her work ‘poetical science.’ The one ingredient that fuels that web is curiosity.” - James Jani

And yet, curiosity has all but disappeared from our cultural discourse. This absence is in research too:

“Good researchers and scientists need curiosity. But our research culture is no longer driven by curiosity; it’s driven by utility. Funding bodies favour well-established ideas over novel ones. Certain kinds of knowledge get labelled as useless.”

However, the truth is that “useless” knowledge has often been the most transformative. CRISPR, the gene-editing technology that is revolutionising medicine, started as a curiosity-driven study of the immune system of bacteria, research dismissed at the time as irrelevant.

Is it more than a feeling?

Curiosity drives learning and originality. When it drops, both suffer.
An fMRI study found that when people are highly curious, they remember more both immediately and the next day, and this benefit also extends to unrelated information encountered at the same time. The hippocampus and midbrain reward areas are more active in high-curiosity states (Gruber et al., 2014).
A large‐scale analysis of 45 million papers and 3.9 million patents shows a steady decline in “disruptiveness” across fields, meaning new work is less likely to break from past knowledge (Park et al., 2023).
If classrooms, funding and daily habits suppress curiosity, we should expect the pattern we’re seeing: better wrapping paper, worse present. In research, the biological lever for depth is being underused, while institutions nudge thinkers toward safer, narrower paths.
If we treat the symptom early, we protect memory, originality and long-horizon thinking. Ignore it, and the condition progresses from fading curiosity to atrophied cognition.

This is the paradox we live in: the fuel for human progress has always been curiosity, but we are systematically draining it out of education, research, and even our daily lives. Instead of silence, we fill space with distraction. Instead of curiosity, we chase utility. Instead of webs, we build silos.

We feel the effects already: shallow conversations, shorter attention spans, a sense that everything is surface-level, but we haven’t yet seen the full impact. What happens when a generation grows up without the ability to wonder, to wander, to pursue questions for their own sake?

The warning signs are here. We just don’t see the effects yet.

If we want to preserve the essence of what makes us human, we need to make curiosity central again. Not as a side effect, not as “useless knowledge,” but as the core driver of growth. Because the things we feel today, the silence we avoid and the curiosity we neglect, are the very things that decide what kind of future we get.

If we do not treat the cause now, the disease will not just take curiosity. It will take thinking with it too.

p.s. This Agora app is currently in development and features all the basic functionalities listed above. Here’s a sneak peek:

For example, one researcher I spoke with had been reading about neuroscience and quantum mechanics purely out of interest. Through books, he built up enough of a foundation to not only understand the basics but actually connect those ideas back to his own field - finance. He found links between market cycles, political patterns, quantum mechanics, and even neuroscience. Talking to someone who could bridge such complex ideas across disciplines was outstanding.

Another marker of competence, I realised, is clarity (shoutout to Carmen, this article really stuck with me). If you truly understand something, you should be able to explain it with precision and confidence. That ability shows you’ve internalised the concept, not just memorised it.

And here’s the thing: there are certain things you will never be able to learn from the internet, no matter how many hours you spend searching. Some knowledge is so practical and so useful that it only exists in lived experience. For example, how to run a profitable dermatologist company. You won’t find a step-by-step tutorial on YouTube that actually works, because that kind of insight only comes from talking to someone who’s done it before. That’s what makes certain ideas so valuable: they’re locked inside conversations and connections.

This exposure completely shifted how I see learning. That same researcher told me that many of his insights came not just from books, but from conversations and group brainstorming. And that stayed with me, because we’ve lost a lot of that in today’s world.

This is also why rich people tend to stay rich. It’s not just their money; it’s the fact that they operate in circles where conversations themselves are assets. They’re surrounded by communities that explore ideas, test them, and profit from them. It’s those groups and networks that drive extraordinary opportunities. You can’t replicate that by following a TikTok tutorial on how to set up a dropshipping business.

Instead of bouncing ideas off each other, we sit isolated in silos, consuming endless streams of online content. I want to move away from that. I want to dig into concepts that aren’t mainstream, to absorb knowledge from books and real discussions, not just scrolling. Because our society’s retention has collapsed. We’re constantly overstimulated, distracted, and unclear about what we actually want or think.

Building a knowledge bank in your head and then actively connecting it through conversations feels like the way forward.

While online resources do have their value, it’s the combination of classic textbook knowledge, lived experience, and real conversation that actually allows you to grow.

If you would like to be part of these conversations, join the subscriber chat. Let’s rebuild the kind of community that algorithms can’t.