The Three Minds of the Machine — Why the Future of AI isn’t Intelligence... It’s Insight

Bridging Mathematics and Logistics – The World Model Analogy

See content credentials

I started the new year reflecting on two domains I’m passionate about: advanced AI for mathematics and practical logistics. At first glance, theorem discovery and supply-chain planning seem worlds apart. However, they converge on a fundamental AI challenge: building an internal model of a complex world and knowing when to stick to the plan and when to seek something radically new. Leading AI researchers like Yann LeCun , Demis Hassabis , and Yoshua Bengio emphasize that such world models – internal “small-scale” representations of reality – are essential for truly smart and safe AI. Just as a mathematician carries an intuitive mental model of a problem, a logistics AI needs a model of its environment (warehouses, routes, disruptions) inside its “mind” to simulate outcomes before acting.

Think of a mathematician trying to prove a theorem and a logistics manager routing deliveries. Both face an immense search space of possibilities. A human mathematician doesn’t brute-force every path; they rely on insights – leaps to a clever lemma or a change of perspective. Similarly, a logistics planner can’t check every possible shipping route in detail; they need intuition to focus on what matters (critical hubs, likely bottlenecks) and to sometimes reimagine the problem (e.g. switch to a hub-and-spoke model or alter schedules).

In both cases, a robust internal model helps decide when to stay local and optimize incrementally, and when to leap to a new representation or strategy.

Modern AI has made strides in competence: Large Language Models can now solve difficult textbook problems and even achieve gold-medal scores on math Olympiads. Yet, as a mathematician, I don’t feel replaced – not even imitated. Why?

Because mathematics is not just deduction; it’s about insight.

It’s not enough for an AI to grind through known methods; it must know when to break the mold. The same holds true for logistics AI: optimizing within one fixed model yields competent scheduling, but true breakthroughs (like redesigning a supply network or adapting to unprecedented disruption) require creative leaps beyond routine.

To build an AI that moves from competence to creativity, I believe we need a combination of mechanisms inspired by how the brain works – a sort of “neuromodulator” approach to AI. Let me outline three key pillars (drawing analogies to neurotransmitters) that such an AI, whether doing math or managing logistics, would need:

Exploration (The Dopaminic Principle)

In neuroscience, dopamine is famously not just the “pleasure molecule” but the signal for prediction error – the jolt of surprise when reality defies our expectations. It’s essentially the brain shouting, “Something unexpected happened – pay attention and explore this!” In math, insight often begins when something doesn’t fit: a bizarre counterexample, a pattern that’s too perfect to be chance, a conjecture that yields a surprise. That anomaly becomes a compass for exploration. Likewise, in logistics, a sudden supply chain disruption or an oddly recurring delay can spark creative rethinking of routes or inventory strategy.

An AI mathematician needs an internal dopamine-like module that rewards the surprising and the novel, nudging the system to break out of routine.

Instead of always following the steepest ascent (what is statistically likely), it should sometimes wander off the beaten path when a weird result pops up.

Technically, this could mean biasing the AI’s search toward novel outputs – e.g. sampling proofs or plans with a surprise bonus. Just as dopamine-driven learning destabilizes habits and encourages exploration, an AI’s “surprise module” would amplify novelty, perturb the usual reasoning gradient, and say “hey, this route/idea is intriguingly wrong or different – let’s follow it.” After all, mathematics (and complex logistics) often progress when we venture into the non-obvious.

Stability (The Serotonergic Principle)

Of course, pure exploration can devolve into chaos. The neurotransmitter serotonin is associated with mood regulation, confidence, and the sense of safety, which, in computational terms, relates to adjusting learning rates and staying balanced. Too little serotonin and a mind can become impulsive and erratic; too much and it can become inflexible or depressed. In AI terms, we need a mechanism to modulate the AI’s level of trust in its current trajectory. A creative AI must dynamically decide when to stick with a plan or line of reasoning and when to tear it up and try something else. Human mathematicians unconsciously calibrate this: we know when a sub-lemma is “good enough to trust for now” versus when we should doubt everything and backtrack. Likewise, in logistics, an AI needs to know when to follow established routes versus when a disruption warrants scrapping the old plan entirely.

Serotonin-inspired control could be implemented as an adjustable “temperature” or learning rate in the AI’s algorithms, tuned by an estimate of uncertainty. In fact, research suggests serotonin neurons modulate learning rates based on uncertainty – effectively telling an animal when to learn slowly and cautiously versus when to rapidly adapt. In an AI, a high-serotonin state might tighten the exploration (low randomness, high confidence in the current model), ensuring creativity doesn’t lead to collapse. A low-serotonin state would do the opposite, injecting flexibility when the AI is too stuck in one mode. This pillar is about creativity with control – ensuring the AI neither gets stuck in endless loops (too rigid) nor jumps so randomly that it never consolidates any insight (too chaotic).

Just as serotonin balances persistence and adaptability in brains, our AI’s stability module would balance rigor and risk.

Attention (The Acetylcholinic Principle)

Insight isn’t just about generating new ideas; it’s also about knowing which ideas to pursue. In a complex environment, whether a dense mathematical theory or a global supply chain, an intelligence must focus on the signal and filter out the noise. The brain’s acetylcholine system acts like a highlighter pen for neural circuits: when something important or novel appears, ACh is released, telling the brain, “focus here, this matters!” It sharpens perception on what’s relevant and dampens distraction.

An AI mathematician similarly needs to develop a sense for which sub-problems or lemmas are promising and which are red herrings. And a logistics AI must pinpoint critical choke points or key shipments among millions of data points.

ACh as inspiration means our AI should have a dynamic attention mechanism – not just the fixed attention of today’s transformers, but a semantic pointer system that can zero in on structurally important elements.

For example, in a proof, it might learn to focus on a rare symmetry that could unlock the solution. In a supply network, it might learn that one port or one supplier is so central that any anomaly there deserves immediate focus. Biologically, acetylcholine helps broadcast “Hey! Pay attention to this!”, and learning to ignore irrelevant stimuli is as important as learning what to amplify. So, our AI’s attention module would perform structural sparsification – emphasizing the graph nodes or variables that carry mathematical significance (or logistic importance) and ignoring the sea of trivial details. This moves beyond surface-level attention on words or tokens; it’s attention guided by meaning. By developing an internal sense of “this aspect of the problem is key,” the AI can drastically reduce its search space and make insight tractable.

In a nutshell

To build an AI mathematician (or any AI tackling complex real-world problems), we must equip it with an explorer that seeks surprises, a stabilizer that regulates creativity, and a spotlight that shines on what matters. These elements mirror what human thinkers do instinctively with our neurochemistry – and they are equally critical in domains as far-flung as abstract math and concrete logistics.

My vision is to integrate these principles through the concept of world models. A world model isn’t just for physical environments; a mathematical theory or a supply chain can be “worlds” too. By carrying a rich internal model, an AI can simulate “what ifs” and develop intuition, much like we do. Notably, companies are already applying such ideas in logistics: multi-agent systems with neuro-symbolic reasoning model supply chains, simulate disruptions, and adapt plans on the fly. They blend neural pattern recognition with symbolic planning, allowing the system to both learn from data and reason over a knowledge graph of routes, rules, and resources. This gives me confidence that the analogies I draw aren’t just metaphors – they’re starting to show up in real technology. Supply chain AIs that treat disruptions not as isolated events but as wave-like phenomena propagating through a network are effectively employing a form of world-model thinking.

Takeaway

Mathematical insight and logistic innovation both thrive at the intersection of exploration and structure. An AI that can master this balancing act – guided by dopamine-like novelty signals, serotonin-like adaptability, and acetylcholine-like focus – will be well on its way to creative problem-solving. In the next section, I will explore the specific cognitive architecture required to realize these ideas, starting with the three fundamental minds of action.

The Three Minds of AI — Imagination, Proof, and Execution

See content credentials

I keep returning to the same uncomfortable observation: the problems that matter most, in mathematics and in the physical economy, don’t fail because we lack compute. They fail because the search space is too large for brute force and too brittle for pure intuition. Theorem discovery and real-world planning look like different sports until you zoom in. Then you realize they share the same hidden anatomy: a vast space of possibilities, hard constraints, and the need for decisive leaps that don’t collapse into fantasy.

That’s why, when I say I’m trying to build an AI Mathematician, I’m not talking about an LLM that “reasons harder.” I’m talking about a system that can repeatedly do what mathematicians do when nobody is watching: generate promising conjectures, reject its own nonsense, and learn from the bruises. In other words, it needs three distinct “minds,” three gears that mesh. When one is missing, the whole machine either stalls or hallucinates.

Imagination is more important than knowledge. —Albert Einstein

The first mind is Imagination. This is the part that proposes candidates: a lemma that might be true, an auxiliary object that might simplify the landscape, a representation change that makes the problem suddenly feel smaller. In planning, it’s the same move with different nouns: candidate routes, alternative decompositions, options that you can test.

Imagination is quick, associative, and shameless; it doesn’t ask whether the idea is correct, it asks whether the idea is interesting enough to try.

It’s the engine of breadth. Without it, intelligence becomes a clerk: perfectly rigorous, perfectly uncreative.

But imagination has a well-known sin: it is fluent. It can be elegant while being wrong. It can be persuasive while being impossible. It can produce a proof-shaped object that evaporates the moment you touch it. So the second mind must exist precisely to be the villain in the story, the one that ruins the beautiful lie.

The second mind is Proof. In mathematics, proof is not “explaining.” Proof is a contract with reality: every step either follows or it doesn’t. In decision systems, the analog is verification under constraints: feasibility, invariants, safety envelopes, “this will not break when the world is adversarial.” This is the mind that asks for receipts. It is slower in the sense that it is more expensive per guarantee; it searches, checks, cross-examines. It is the difference between a clever guess and a trustworthy result.

Imagination makes the guess. Proof earns the right to trust it.

Now comes the part that most AI discussions still treat as an afterthought, because it doesn’t fit neatly into a benchmark: the third mind, the one that turns intelligence into accountability.

The third mind is Execution. Execution is not just “doing”; it is committing to a hypothesis in a world that pushes back, then updating your internal model when you were wrong. In math, execution looks like running the proof attempt all the way to the wall, discovering the contradiction you didn’t foresee, and realizing your representation was hiding the true structure. In the physical world, execution is where the cost of mistakes becomes real: disruptions propagate, capacities shift, a constraint you ignored suddenly dominates the outcome. This is where a system stops being eloquent and becomes responsible. The world becomes the teacher; drift becomes curriculum.

What I like about this three-mind frame is that it explains why many “single-mode” approaches disappoint. Pure optimization, without imagination, becomes a cathedral of computation: correct in principle, too slow in practice. Pure generation, without proof, becomes a theater of plausibility: fast, impressive, unsafe. Pure execution, without either, becomes randomness with consequences.

The systems that actually scale are the hybrid ones, even when they pretend they aren’t: generate candidates, verify, act, learn. The shape is invariant; only the implementation changes.

This is also why “System 1 / System 2” remains a useful metaphor, as long as we don’t treat it like a psychology lecture. I use it lightly: one mode proposes quickly, the other checks carefully. The point is not that one is always slower on a CPU; the point is that one is cheaper per idea and the other is more expensive per certainty. Intelligence is the clutch between them.

And this is where I can hint at logistics without making it the headline. A logistics network is an embodied proof problem: a living graph with constraints, delays, and cascading failure modes. If you only optimize, you drown. If you only generate, you drift. The only robust approach is the same three-mind architecture: imagination to propose plausible candidates, proof to enforce constraints, execution to close the loop and update the model. The domain is not the point; the structure is.

Takeaway

Today’s takeaway is simple: if we want an AI that invents mathematics, we must stop asking for “one model to do it all” and start designing for the three minds explicitly. Imagination to explore, proof to ground, execution to learn. This brings us to a research direction that serves as the bedrock for these minds: the stability of the worlds they inhabit.

Learning Without Seeing: LeJEPA and Latent Worlds That Don’t Collapse

See content credentials

A key theme I see in the AI community right now is a return to fundamentals: moving away from ad-hoc tricks, and toward principled, theory-driven approaches. Stanford University researchers have been unusually vocal about this, especially on evaluation reliability and on the limits of today’s language models in high-stakes settings. Their message is not “slow down because we’re scared.” It’s “slow down because we want to understand.” That ethos resonates with me. In this phase of AI, robustness is not a luxury; it is the price of admission.

In that spirit, I want to talk about a recent paper that I find genuinely exciting: LeJEPA (Randall Balestriero & Yann LeCun, 2025). The paper targets a problem that has quietly shaped the last few years of self-supervised learning: we’ve gotten great results, but we’ve often gotten them with a growing pile of heuristics.

LeJEPA’s promise is blunt: keep the performance, delete the hacks.

Why LeJEPA matters: the collapse problem

Modern AI systems, especially in vision, often learn representations using self-supervised learning (SSL): the model tries to predict one part of the data from another, or it learns embeddings where different “views” of the same underlying example map close to each other. This family includes famous systems like SimCLR, BYOL, DINO, and the JEPA lineage—different technical choices, but a shared idea: learn a representation space that captures what matters.

The catch is a classic: without precautions, the model can cheat. If it maps everything to the same embedding, then any two views are “similar” by definition. The training objective is satisfied, but the representation contains no information. This is representation collapse.

Historically, the community fought collapse with a kitchen sink of anti-collapse heuristics: stop-gradient in one place, a momentum encoder in another, negative samples or not, special normalization choices, temperature tuning, carefully designed asymmetries, and more. These recipes work. But they’re brittle. Change the wrong detail and you sometimes fall off a cliff. The deeper issue is not that practitioners are careless; it’s that for years we lacked a clean theory for what shape the representation space should have and how to enforce it reliably.

This is where LeJEPA makes its move.

The core idea: a principled geometry for embeddings

LeJEPA provides a theoretical foundation that identifies a core principle to avoid collapse: constrain embeddings to follow an isotropic Gaussian distribution. In the authors’ theory, this distribution is optimal—they argue it minimizes downstream prediction risk under their formal assumptions, and they present it as the target geometry a JEPA-style method should aim for.

In plain English, imagine your embeddings are marbles inside a box. Collapse is when all the marbles pile into a tiny corner or flatten into a skinny line—easy for the model, useless for you. LeJEPA says: force the marbles to fill the space evenly in every direction—no pancake, no line, no degenerate corner. A symmetric “blob” is not an aesthetic preference; it’s a constraint that blocks the trivial solution and preserves information across dimensions.

The key shift is practical: collapse is no longer treated as an annoyance you patch with a dozen special tricks. It becomes a violation of a global constraint on representation geometry.

When representations collapse, the model keeps the objective and drops the information.

That is the failure mode LeJEPA targets, and it targets it at the root.

SIGReg: enforcing the constraint without eating the whole watermelon

How do you enforce “isotropic Gaussian” in high dimension without doing something computationally absurd? LeJEPA introduces SIGReg: Sketched Isotropic Gaussian Regularization.

The intuition is elegant. Checking whether a high-dimensional cloud is “spherical” is hard if you try to check everything at once. But you can learn a lot by taking many random one-dimensional slices. If every slice looks like a standard Gaussian, then you have strong control over the shape of the full distribution (this connects to the Cramér–Wold viewpoint: one-dimensional projections characterize a multivariate distribution).

A friendly analogy used by commentators: checking high-dimensional spherical structure is like checking whether a watermelon is ripe inside—you don’t want to eat the whole thing to find out; a few slices tell you a lot.

SIGReg works in that spirit. It takes random projections of the embedding vectors onto one-dimensional directions, and for each projection it penalizes deviation from a standard Gaussian. The paper uses a characteristic-function-based normality statistic (Epps–Pulley). The result is a regularizer that is scalable, differentiable, and strong enough to prevent collapse without the older scaffolding.

What you get: one trade-off, fewer superstitions

Once you combine the standard predictive objective with SIGReg, you get LeJEPA:

loss = predictive JEPA loss + λ · SIGReg

One trade-off parameter, λ. The authors emphasize that LeJEPA is “heuristics-free”: no stop-gradient, no teacher–student momentum encoder, no special schedules. They also emphasize scalability (linear time/memory behavior with respect to key dimensions of the regularizer) and stability across architectures and hyperparameters.

Empirically, they report results across many datasets and architectures, including large vision transformers. One highlighted example: 79% ImageNet top-1 linear evaluation with a frozen ViT-H/14 backbone pretrained on ImageNet-1K, under their evaluation protocol.

The deeper point is not the single number. The deeper point is the direction of travel: the field replacing “recipes that happen to work” with “objectives we can defend.”

Why I care: world models don’t forgive unhealthy latent spaces

Now, why am I—someone obsessed with world models and AI mathematicians—so excited about what looks like a vision SSL technique?

Because a world model is only as good as its representations. A world model that will be used for prediction, planning, or control cannot afford a latent space that quietly degenerates. In the world-model lens, collapse is not just a training annoyance.

It’s existential. If the internal state becomes trivial, planning becomes theatre.

LeJEPA’s core message generalizes: keep your representation space stable, non-degenerate, and usable. It is a small but meaningful example of what I’d call “mathematician-like discipline” applied to learning systems: define what must be true, and then build the method around that constraint.

A concrete (but clearly labeled) application story: flight planning write-ups

I also find it interesting how quickly LeJEPA is being discussed as a building block inside planning agents. There are industry-style write-ups describing aviation / flight-planning agent prototypes that fuse perception, reasoning, and a learned latent dynamics model, and explicitly cite LeJEPA-style training objectives to avoid degenerate latent representations during learning. To be clear: I treat these as illustrations, not peer-reviewed validation. But the motivation is exactly right: if you intend to plan by rolling forward a learned latent dynamics model, you want a training objective that makes collapse difficult, not merely unlikely.

In a nutshell

The “Stanford moment” in AI—rigor, measurement sanity, theory-backed objectives—is not an academic aesthetic. It is a practical requirement. LeJEPA is a prime example of injecting theory (in this case, a principled target geometry for embeddings plus a scalable regularizer) to replace a pile of heuristics.

It makes self-supervised learning more stable and more defensible.

For the broader agenda I care about this week—AI that moves from competence to creativity—this matters. I don’t expect an AI to discover new mathematics if its internal representations are fragile or collapse-prone. If we want machine insight, we need machine foundations: learning dynamics and representation spaces built on constraints we can argue for, not tricks we hope don’t break.

With these stable foundations in place, we can finally assemble the complete architecture of machine insight.

Beyond Competence – Toward AI Insight and Creativity

See content credentials

A human mathematician is not a single, monolithic problem-solving machine.

We are a collection of sub-systems: intuition, formal reasoning, visual imagination (we picturize geometry), analogical thinking, rigor checks, and even emotional drives (curiosity, frustration, etc.). Likewise, a true AI mathematician will likely be an assembly of specialized components, each powerful in its domain, orchestrated together. Here’s how I envision it:

• A Language Model – for symbolic manipulation and the “narrative” of proofs. This corresponds to the part of a mathematician that writes down arguments, transforms equations, and follows logical steps. Today’s LLMs (like GPT-style models) already excel at formal reasoning to a degree, so this would handle the deductive reasoning and communication aspect.
• A World Model – as we discussed, for continuous intuition. This would handle concepts of space, quantity, and metaphorical “motion” through a problem. For example, understanding a limit or an integral geometrically, or visualizing the shape of a solution space. In a logistics analogy, this is the simulator that can imagine different routing scenarios. In math, it might be a learned model of how changing one parameter affects outcomes – an intuition pump.
• A Graph Reasoner / Knowledge Base – mathematicians stand on a vast web of prior results. A math AI should navigate the graph of theorems and concepts effectively, retrieving relevant lemmas, checking if a conjecture is already known in some guise, etc. This component would ensure the AI isn’t reinventing the wheel and can leverage existing theory (like an extreme version of how humans use Wikipedia or textbooks). It’s the part that says, “This problem smells like it could be graph theory – recall the Spectral Theorem” (for logistics, think of it as knowing existing routes or best practices).
• A Neuromodulatory Controller – the explorer/stabilizer/attention modules we analogized to dopamine, serotonin, acetylcholine. This sits atop the other systems and continually adjusts their behavior. For instance, if the language model part is churning out routine but unproductive algebra, the controller might crank up the “exploration” setting (like increasing a creative temperature or injecting a surprising assumption to try). If the world model’s simulations are getting too wild, it might dial back the exploration to refocus. This meta-controller keeps the whole system in a productive regime – not too hot, not too cold. It’s essentially the project manager ensuring the different “brains” work together effectively and adapt to the problem at hand.
• A Proof Verifier / Critic – an essential piece for trust and rigor. This module would formally check proposed proofs or solutions (much like automated proof assistants do) to ensure no logical gaps. It also acts as a critic, providing feedback. If a proof attempt fails, the verifier can highlight where, guiding the other components to focus there next. In a way, this provides the ground truth that keeps the AI honest. For creative discovery, having a built-in verification loop means the AI can generate bold conjectures and immediately test them against strict criteria, learning from the failures. (Human mathematicians do this too – we constantly self-check smaller lemmas or use computation to test conjectures with examples.)
• A Concept Generator (Creative Divergence Engine) – this is the part that does a CK-style concept expansion. It can propose new definitions or maps between domains (like, “What if I treat this number sequence as a geometric object?” or in logistics, “What if I group deliveries in a completely new way?”). This might be an algorithm that deliberately steps out of the current context – perhaps by analogies, metaphors, or even by dreaming up alternative worlds where the problem is simpler. In human terms, it’s the “blue-sky brainstorming” mode. Some ideas will be nonsense, but occasionally, one will be a breakthrough. The AI needs a space to safely generate and explore such off-the-wall ideas without derailing its main line of thought – a sandbox for insight, if you will.

Bringing all these together is an integrative framework. I often think of the famed mathematician Alexander Grothendieck, who revolutionized fields by changing their language. He didn’t just solve problems – he built new theories (new “worlds”) around problems so that they became simple in that new context. An AI mathematician should aspire to the same.

What excites me most is the prospect that mathematics may become the first domain where AI shows non-human creativity.

That sounds paradoxical – how can an AI be creative in a way that isn’t just mimicking humans? Here’s one scenario: AIs, unlike human specialists, have the potential to connect distant fields effortlessly. Humans are limited by silos – you might be an expert in algebraic topology and know little of machine learning, or vice versa, and cross-disciplinary insights often happen by serendipity (like bumping into someone at a conference). AI, on the other hand, especially a system with a vast knowledge graph and concept generator, can traverse the entire map of human knowledge in moments. An AI could spot that a problem in combinatorics might be attacked with a technique from neural network theory, or that a pattern in physics reappears in number theory. Such cross-pollination might yield proofs and concepts no human would have ever considered, not because we lack the intelligence, but because we wouldn’t travel that far afield for inspiration. In Grothendieck’s sense, the AI can build a bridge of language between two previously separate “worlds” of math, and in that merged space, the problem might dissolve.

To tie it back to logistics one last time: This kind of AI creativity could equally transform complex planning. Imagine a supply-chain AI that doesn’t just optimize routes but invents a completely new paradigm of distribution by drawing an analogy from, say, internet packet routing or even biological systems (like blood circulation). Human experts in logistics might not think to borrow theories from biology or computer networks, but an AI that has ingested all these domains could. The result could be a leap in efficiency or resilience that looks almost alien – an “out of the box” solution in the truest sense. In both math and logistics, such leaps will change the game.

Of course, to reach that, we have open challenges – an open research agenda that I consider my roadmap for years to come. We need to formalize concepts like “surprise” as a computational objective (how exactly do we quantify when an AI should be rewarded for finding something novel?).

We need better ways to blend symbolic reasoning with neural nets (neuro-symbolic integration) so that our AI can rigorously prove a statement but also intuitively feel which directions are fruitful.

We need to define what a “mathematical latent space” should look like for different fields – perhaps the isotropic spread is one piece of it, but there might be other geometric structures (like manifolds corresponding to symmetry groups, etc.) that make certain insights easier. Compactness, duality, symmetry – can we turn these deep mathematical ideas into computational handles that an AI can tweak and play with? Each of these questions is both hard and thrilling.

Yet, I’m optimistic. The progress in 2025 alone – from AI solving Olympiad problems in natural language to startups like Axiom formally proving new theorems – shows that we’re at the cusp. When Google DeepMind ’s AI model cracked five of six IMO problems and did so by thinking for hours and using strategies like consensus between multiple agents, it signaled that AI is starting to capture glimmers of human-like reasoning (albeit at great computational cost). The skeptics always said “AI just imitates and regurgitates.” But now we’ve seen AI systems that truly reason through novel challenges, not by recall but by active exploration and self-consistency checks. The debate about whether AI can “really” think is becoming moot, at least in bounded domains like contest math.

The frontier now is creative thought – and that’s exactly what my multimodal, multi-component vision is aiming for.

In a nutshell

An AI that can not only solve given problems but also discover new mathematics will redefine intelligence as we know it. Having walked through these ideas, my conviction is stronger than ever. Such an AI won’t come from a single breakthrough or a giant end-to-end model that magically does everything.

It will come from an architecture of integrated intelligences – a symphony of specialized models, each excelling at one mode of thinking, orchestrated to complement each other.

In building this, we are essentially acknowledging what some of the greatest minds have taught us: the power is not just in the objects of study, but in the languages, representations, and perspectives we bring to study them. Give an AI the ability to change the map, not just move along it; to change the language of the question, not just the symbols in an equation – and we will witness something truly extraordinary. In the words I’ve kept as my guiding “punchline” throughout this project:

Change the space, and insight follows.

Takeaway

Insight, ultimately, is an emergent phenomenon. It arises when exploration, stability, and attention intersect; when language meets geometry meets logic; when an unexpected link is forged between distant concepts. With a careful blend of neuroscience-inspired control, world-model building, graph knowledge, and rigorous verification, we can engineer the conditions for insight to spark. And once those sparks start flying, I have a feeling the 2020s will get very interesting. We might see AIs proposing conjectures that astound us, or optimizing systems in ways we’d never imagine – not because they’re “smarter” in a human IQ sense, but because they use representational leverage we don’t have.

My ambition is to contribute to this new wave of AI research, one that is as comfortable proving a lemma as it is scheduling a delivery truck. If we succeed, we won’t just have automated some tasks; we’ll have expanded the creative capacity of our civilization. An AI that truly invents mathematics will be a tool for humans to explore ideas we couldn’t reach alone. And an AI that optimizes logistics with world-spanning insight will free us to solve bigger problems while it handles the mundane.

The Living Twin: When Simulation Stops Being a Dashboard

See content credentials

Most “digital twins” are not twins. They are portraits.

A portrait can be beautiful, detailed, and deeply reassuring. It can also be useless for decision-making, because it never argues back. It doesn’t tell you you’re missing a constraint. It doesn’t show you the second-order consequence. It doesn’t punish you for a plan that looks right but fails when the world turns adversarial.

A real twin, the kind I care about, is closer to a wind tunnel. You don’t build a wind tunnel to admire it. You build it to let bad ideas fail safely, early, and repeatedly—until the shape of the right idea becomes obvious.

The twin isn’t the product; the twin is the instrument that makes the product trustworthy.

So here is my definition, the line I draw in 2026: a digital twin becomes living only when it closes the loop between recommendation and reality. It must ingest the world as it is, simulate the world as it could be, propose actions, verify constraints, then execute and learn. If it can’t do that, it’s not a twin. It’s a dashboard with a good marketing team.

This matters because consumer AI trained us to tolerate wrongness. A movie recommender can be wrong and we shrug. In physical systems, wrong has inertia. A plan that violates a constraint doesn’t just disappoint; it cascades. Delays propagate. Buffers evaporate. SLA breaches compound. When the system is coupled, local errors become global behavior.

That’s why I think the enterprise question is not “can the model answer?” It’s: can the model test itself before it speaks? Can it attach a receipt, a certificate of feasibility, a bounded-risk argument? In mathematics, proof is the mechanism that turns speculation into knowledge. In operational decision-making, verification is the mechanism that turns a recommendation into something you can execute without fear.

A living twin is a lie detector for plans.

This is where the Three Minds architecture moves from theory to industrial application. The Living Twin becomes the environment where Imagination proposes, Proof verifies, and Execution learns. This is not theory; it is the only design that scales when the world is dynamic and high-stakes.

And the subtle point is that the twin is not just a simulator; it is a discipline. A living twin changes the behavior of the entire system, because it makes it cheap to ask counterfactual questions and expensive to be sloppy. “What if I reroute?” “What if I reshape the network?” “What if I trade speed for reliability?” In a portrait, these questions are answered by opinion. In a living twin, they are answered by structured consequence.

Now the connection back to mathematics—and to my “AI Mathematician” theme—is not rhetorical. Mathematics is the cleanest arena we have for training the habit of verification. A proof is a counterfactual engine: it tests an idea in a world of strict rules before you ever touch reality.

If we can build systems that treat operational decisions the way mathematicians treat conjectures—generate, test, verify, revise—then we can scale decision-making without scaling fragility.

This is also where the world model lens becomes more than a buzzword. A twin cannot be living if it only models surfaces. It needs latent structure: the hidden state of the system, the propagation mechanics, the causal skeleton. This is exactly why I care about research that learns stable, decision-relevant representations. If your internal representation collapses or drifts, the twin becomes an unreliable narrator. If the representation is stable, the twin becomes an instrument.

My thesis for 2030 is straightforward: the winners won’t be the models that talk the best. They’ll be the systems that can dream multiple solutions, test them inside a living twin, prove they satisfy constraints, execute them, and learn as the world changes.

A twin that does this is not a dashboard. It is a laboratory. And in high-stakes domains, laboratories beat opinions.

That’s the end of the week’s arc. Fluency is impressive. But the future belongs to systems that can be held accountable: imagination grounded by proof, proof made practical by execution, execution guided by world models that stay honest.

References

1. Yann LeCun — “A Path Towards Autonomous AI” (talk + JEPA framing, Feb 2022) https://www.linkedin.com/posts/yann-lecun_yann-lecun-a-path-towards-autonomous-ai-activity-6903097383371300864-flBH
2. Meta AI — “I-JEPA” (Jun 2023) https://ai.meta.com/blog/yann-lecun-ai-model-i-jepa/
3. Meta AI — “V-JEPA” (Feb 2024) https://ai.meta.com/blog/v-jepa-yann-lecun-ai-model-video-joint-embedding-predictive-architecture/
4. Demis Hassabis (clip) — “World models are my longest-standing passion” (2025) https://www.linkedin.com/posts/tp23_demis-hassabis-says-world-models-are-his-activity-7407480647516094464-4y44
5. DeepMind — “AI solves IMO problems at silver-medal level” (AlphaProof + AlphaGeometry 2, Jul 25, 2024) https://deepmind.google/blog/ai-solves-imo-problems-at-silver-medal-level/
6. Nature — “Solving olympiad geometry without human demonstrations” (AlphaGeometry) https://www.nature.com/articles/s41586-023-06747-5
7. ArXiv — “LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics” (Nov 2025) https://arxiv.org/abs/2511.08544
8. Code — LeJEPA reference implementation (GitHub) https://github.com/rbalestr-lab/lejepa
9. Schultz (1998) — “Predictive reward signal of dopamine neurons” https://journals.physiology.org/doi/10.1152/jn.1998.80.1.1
10. Grossman et al. (2022) — “Serotonin neurons modulate learning rate through uncertainty” (Current Biology) https://www.sciencedirect.com/science/article/pii/S0960982221016821
11. Picciotto et al. (2012) — “Acetylcholine as a neuromodulator…” (attention, cue detection, memory) https://pmc.ncbi.nlm.nih.gov/articles/PMC3466476/
12. Noudoost & Moore (2011) — “The role of neuromodulators in selective attention” https://pmc.ncbi.nlm.nih.gov/articles/PMC3351278/
13. Kosasih et al. (2024) — Neurosymbolic AI in supply chain (systematic review) https://www.tandfonline.com/doi/full/10.1080/00207543.2023.2281663
14. Forbes — Axiom Math profile (Sep 30, 2025) https://www.forbes.com/sites/rashishrivastava/2025/09/30/meet-the-stanford-dropout-building-an-ai-to-solve-maths-hardest-problems-and-create-harder-ones/
15. ArXiv — “Formal Mathematical Reasoning: A New Frontier in AI” (Dec 2024) https://arxiv.org/html/2412.16075v1
16. Stanford Report — “Researchers work to squash ‘fantastic bugs’ hidden in AI benchmarks” (Dec 8, 2025) https://news.stanford.edu/stories/2025/12/ai-benchmarks-flaws-models-bugs-research
17. Stanford HAI blog — “Fantastic Bugs and Where to Find Them in AI Benchmarks” (Dec 13, 2025) https://ai.stanford.edu/blog/fantastic-bugs/
18. Stanford Report — “Why AI still struggles to tell fact from belief” (Nov 10, 2025) https://news.stanford.edu/stories/2025/11/ai-language-models-facts-belief-human-understanding-research
19. Suzgun et al. — “Language models cannot reliably distinguish belief from knowledge and fact” (Nature Machine Intelligence, 2025) https://www.nature.com/articles/s42256-025-01113-8
20. LeCun (2022) — “A Path Towards Autonomous Machine Intelligence” (OpenReview; PDF) https://openreview.net/pdf?id=BZ5a1r-kVsf
21. Bardes, Ponce, LeCun (2021) — “VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning” (arXiv) https://arxiv.org/abs/2105.04906
22. Balestriero & LeCun (2025) — “LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics” (arXiv) https://arxiv.org/abs/2511.08544
23. LeJEPA full text (HTML on arXiv; SIGReg / test details) https://arxiv.org/html/2511.08544v3
24. Official LeJEPA code repository (reference implementation) https://github.com/rbalestr-lab/lejepa
25. Epps & Pulley (1983) — “A test for normality based on the empirical characteristic function” (Biometrika) https://academic.oup.com/biomet/article-abstract/70/3/723/248040
26. Cramér–Wold theorem (background on 1-D projections determining multivariate distributions) https://en.wikipedia.org/wiki/Cram%C3%A9r%E2%80%93Wold_theorem
27. Example industry write-up (illustrative, non-peer-reviewed) referencing LeJEPA-style objectives in a flight-planning agent https://medium.com/ai-simplified-in-plain-english/the-integrative-architecture-of-agi-lejepa-and-deepseek-for-causal-flight-planning-84c05f8970b5

Further reading (useful background, not directly cited)

1. Dawid & LeCun (2023) — notes on latent-variable Energy-Based Models + H-JEPA (PDF) https://arxiv.org/pdf/2306.02572
2. Assran et al. (2023) — I-JEPA (arXiv) https://arxiv.org/abs/2301.08243
3. Meta AI Blog (2024) — V-JEPA overview https://ai.meta.com/blog/v-jepa-yann-lecun-ai-model-video-joint-embedding-predictive-architecture/
4. Epps (1993) — “Characteristic Functions and Their Empirical Counterparts” (PDF) https://www.math.mcgill.ca/~dstephens/OldCourses/556-2007/Papers/Epps1993.pdf
5. Lyons & Zumbrun (2017) — “A Calculus Proof of the Cramér–Wold Theorem” (PDF) https://rdlyons.pages.iu.edu/pdf/cramerwold.pdf
6. Turing Post (Nov 2025) — LeJEPA explainer https://www.turingpost.com/p/lejepa
7. Medium (Nov 2025) — “LeJEPA: The simple new SSL framework that works without the tricks” https://medium.com/@mohamed-aymen.bouyahia/lejepa-the-simple-new-ssl-framework-that-works-without-the-tricks-e25616c721cb
8. Towards Deep Learning (Nov 2025) — LeJEPA explainer / positioning https://www.towardsdeeplearning.com/goodbye-dino-tricks-goodbye-stop-grad-lejepa-is-about-to-rewrite-ssl-for-computer-vision-078bf74cbdd2
9. AlphaXiv — LeJEPA overview page https://www.alphaxiv.org/overview/2511.08544