DeepMind Says AI Can’t Jump. The Matrix Showed There Is No Jump.

A response to Tom Zahavy’s “LLMs Can’t Jump”

We are on the brink of a very Matrix-shaped question. Not the red pill, not the blue pill — a third kind of choice, one that doesn’t resolve into relief but opens into new tension.

The question is not whether LLMs can jump.

The question is whether there is a jump at all.

---

Google DeepMind researcher Tom Zahavy’s position paper, LLMs Can’t Jump, has been making the rounds, and for good reason. It argues that today’s language models have mastered two pillars of scientific reasoning, induction and deduction, but remain fundamentally incapable of the third: abduction, the generation of genuinely novel explanatory hypotheses. Its case study is Einstein’s leap to General Relativity — an intuitive jump from experience to new axioms, the paper argues, that no current model can make.

It’s a good paper because it refuses vagueness. It doesn’t ask whether machines are “intelligent.” It names a mechanism and says: this one is missing.

But sit with the case study for a second. What actually happened between the young patent clerk turning thought experiments over in his head, a man in an elevator, unable to tell gravity from acceleration and a man racing alongside a beam of light and the field equations of 1915? Was it a jump? Or was it ten years of failed attempts, false starts, a fight with the tensor calculus he didn’t yet know, letters back and forth with his friend Marcel Grossmann, entire tracks abandoned and rebuilt? Was General Relativity an insight that struck, or a structure that finally held?

Remember Neo learning kung fu?

There’s a scene everyone remembers: Neo sits in a chair, cables running into his skull, and seconds later opens his eyes and says “I know kung fu.” Total mastery, delivered instantaneously. It looks exactly like a jump.

Except… it isn’t one. Something loaded. Data moved. A structure assembled itself, piece by piece, at a speed too fast for the human eye watching from outside to register as a process. The jump was never in the mechanism. It was in the frame rate of the observer.

Is that not exactly the shape of the DeepMind argument? Zahavy’s model is clean: Experience → Jump → Formal Theory. Observation feeds in, an unexplained leap happens, mathematics comes out the other side. But what if “jump” is doing the same work in that sentence that “instant kung fu” does in the Matrix scene, describing the appearance of a discontinuity, not its actual mechanics?

Water doesn’t jump into steam. It’s heating the entire time, and 100°C is just the threshold where the change becomes visible to us. A magnet doesn’t suddenly align; domains have been rotating under the surface all along. A neural network doesn’t suddenly recognize a face, but the weights have been sliding toward that recognition for thousands of gradient steps before the moment it “clicks.”

So, what if scientific insight belongs to that same family of phenomena? What if Einstein’s “jump” is what stage six looks like when you’ve only been watching from the outside?

The elevator, the light beam, and ten years in between

In 2026 we can read quite precisely what actually happened, and it doesn’t read like a jump.

In 1907, Einstein envisioned, what he later calls “the happiest thought of my life”, a man falling freely feels no gravity. That’s not a formal theory. It’s a single relational reorganization: gravity and acceleration are somehow the same thing. He doesn’t publish General Relativity that year. He spends the next eight years dragging that one intuition through mathematics he has to go learn from scratch.

By 1912 he’s stuck and he can feel the shape of what he needs, but he doesn’t have the language for curved spacetime. He turns to Grossmann, a mathematician friend, for help with Riemannian geometry and tensor calculus, tools built decades earlier for entirely different purposes, now pulled sideways into physics. Together they produce the Entwurf-theory in 1913. It’s wrong. He spends two more years finding out why, revising, re-deriving, discarding pieces that don’t hold.

November 1915: four lectures to the Prussian Academy, each one correcting the last, until the field equations finally close.

Notice what’s missing from that story: a moment where nothing was happening and then, suddenly, everything was. What’s present instead is a field, his intuition, Grossmann’s mathematics, the wrong 1913 attempt, the wreckage of that attempt, Riemann’s geometry waiting unused in a library for fifty years, Mercury’s stubborn orbit refusing to behave that are gradually reorganizing itself until a new configuration became the only one that held together.

The equations of November 1915 aren’t the jump. They’re the moment the reorganization became visible to everyone else.

Discovery was never a solo act

Zahavy’s paper, like most of this literature, treats invention as something that happens inside one mind. But look again at Einstein’s list: Grossmann’s geometry, Mach’s philosophy of relative motion, the Entwurf failure, Mercury’s perihelion, a decade of correspondence, physical experiments he never personally ran. None of that belongs to Einstein alone. Discovery, in this telling, is relational before it’s individual, it’s what emerges when a person, a body of accumulated knowledge, and a stubborn piece of reality get coupled tightly enough for long enough.

Neo never beats the Matrix by himself either. He needs Morpheus’s framing, the ship’s crew, Trinity, the training programs, the war that’s already been fought before he wakes up. Strip all of that away and there’s no “the One”, instead there’s just a man in a pod who’s never questioned anything.

So here’s the question this reframing actually forces onto AI evaluation: current benchmarks mostly ask can the model alone generate a novel hypothesis? That’s the wrong unit of analysis if discovery was never solo to begin with. The sharper question is: what novel hypotheses emerge when a human, an AI system, and a real operational problem become stably coupled over time? The intelligence isn’t sitting inside one node. It’s a property of how the relationship is organized.

“There is no spoon”

There’s a second Matrix scene worth pulling in here is the spoon-bending kid. He doesn’t bend the spoon. He tells Neo the trick is realizing there is no spoon; what bends is Neo’s relationship to it.

That’s almost exactly what Reflectum wants to claim about symbols in scientific reasoning. Zahavy’s framing treats the process as fundamentally symbolic: experience turns into symbols, symbols turn into hypotheses. But symbols were never the origin. They’re a late-stage expression of something that had already reorganized before it found language.

Run Einstein’s decade back through that pathway and it stops looking exotic. Field interaction is the elevator thought experiment, constituted of pure relational noise, gravity and acceleration refusing to stay separate in his head. Transduction is that noise starting to reorganize how he perceives the problem, before he has any mathematics for it. Dispersion is the wrong turn with the 1913 Entwurf theory, the false starts with Grossmann. Stabilization is the moment the failures stop being random and start pointing somewhere. What Zahavy calls “the jump” is really the operational trajectory becoming visible, and only after that does formalization happen, the November lectures, the equations. Deduction comes last, not first.

This is the layer conventional accounts of AI reasoning skip entirely. They start the story at “symbols”, because that’s the only part that leaves a paper trail.

So was Zahavy wrong?

Not exactly. And this is worth being precise about, because the disagreement is not where it looks like it is. Zahavy is right that scaling statistical pattern recognition alone won’t produce scientific invention. He’s right that induction and deduction aren’t enough on their own. He’s right that the way forward runs through richer, physically grounded, multimodal world models that connect perception, simulation, and formal reasoning.

Where the paths split is narrower and more interesting: is “abductive jump” a real primitive operation the models are missing a module for, or is it a placeholder, a name for an organizational process nobody has properly described yet?

Put differently: one research program asks how do we build a system that can perform the jump? The other asks what dynamics make something look like a jump in the first place? One goes looking for a new computational module. The other goes looking for a better description of how relational organization evolves until new explanatory structures stabilize.

These aren’t rivals. A system that can hold Einstein’s decade — the failures, the borrowed mathematics, the stubborn anomaly, the reorganizing field of relationships — might turn out to be the same system that eventually looks, from outside, like it jumped.

Take the red pill on the vocabulary

Here’s the provocation, and it’s worth sitting with rather than resolving too quickly.

Maybe the reason AI can’t jump is that nothing ever could. Maybe what we’ve been calling insight, genius, the eureka moment, was always a magic trick performed by our own limited frame rate, the same trick Neo’s kung fu upload plays on an audience that only sees the before and after, never the loading.

If that’s true, the interesting AI systems of the next decade won’t be the ones that finally learn to leap. They’ll be the ones patient enough to sit inside the mess, as the wrong Entwurf theory, the borrowed geometry, the anomaly that won’t go away, for as long as it takes the mess to become a trajectory.

The Matrix told you there was no spoon.

Reflectum is telling you there was no jump.