Language models hit wall on scientific breakthroughs; world models may offer escape route

Tom Zahavy's analysis rests on a classical philosophical framework from Charles Sanders Peirce, who divided all reasoning into three modes. Deduction derives guaranteed conclusions from fixed rules, like running a program that produces provably correct output. Induction spots patterns in data by observing many instances—the classic example being observing a thousand white swans and generalizing that all swans are white. Abduction is the creative leap: inventing a cause to explain a surprising phenomenon.

Zahavy identifies two levels of abduction. Ordinary abduction picks the most plausible explanation from a set of known candidates, much as a doctor matches symptoms to a disease—a task language models can perform. The harder version, which Zahavy calls "manipulative abduction," invents a cause for which no linguistic template exists yet. This, he argues, is the real bottleneck of scientific invention, and machines cannot achieve it. The evidence supporting his claim is striking: language models already excel at statistical pattern recognition, and they're rapidly conquering formal derivation. Systems like AlphaProof, Gemini, and GPT-5 now achieve gold-level scores on International Mathematical Olympiad problems. Zahavy even concedes that a language model could probably derive general relativity if given Einstein's assumptions as a starting point. But formulating those assumptions in the first place—making the manipulative leap to reach them—remains the bottleneck.

To illustrate why machines struggle with this leap, Zahavy points to how AI models typically learn: by comparing their predictions to reality and adjusting based on the error, the gap between prediction and outcome. Without a detectable error, there's nothing for the system to work with. This was precisely Einstein's situation. When Einstein was working, Newton's physics had been confirmed with extreme precision. The only known anomaly was a tiny shift in Mercury's orbit, which had been attributed to a hypothetical hidden planet called Vulcan. An optimization-driven AI, following the logic of error reduction, would have had no reason to overthrow physics; it would have done what the astronomers of the era did—invented an extra planet to account for the small discrepancy rather than rethinking space and time. The data confirming Einstein's theory, such as Eddington's measurement of light deflection, didn't arrive until years after the theory was formulated.

So where did the insight come from? Zahavy points to Einstein's "happiest thought": the freely falling observer who no longer feels gravity. This insight arose from embodied simulation—Einstein mentally playing through a physical sensation. He imagined a physicist inside an accelerating elevator in space and concluded that acceleration and gravity are indistinguishable from the inside. Zahavy draws a parallel to Archimedes, who discovered his buoyancy principle not through calculation but, as the story goes, through the physical feeling of water rising as he stepped into a bathtub. In both cases, a foundational principle emerged that didn't yet exist in the language of the time.

Language models lack exactly this sensory grounding. Zahavy compares them to John Searle's "Chinese Room," a thought experiment where a person shuffles Chinese characters according to a rulebook without understanding a single word. Language models shuffle the symbols of physics in much the same way, without access to the physical experience that gives those symbols meaning. Current automation systems like Sakana's AI Scientist and DeepMind's AlphaEvolve can recombine existing concepts and optimize brilliantly, yet both require either a clear error signal they can shrink step by step or operate only within existing conceptual frameworks. Neither can make the leap into an entirely new framework of thought.

As a possible way forward, Zahavy points to physically consistent world models. He distinguishes between video generators like Veo, which simply predict the most likely next frame—falling objects fall not because the model understands gravity but because falling is the most common continuation in the training data—and action-controllable world models like Genie. The latter let an agent actively intervene in a simulation and run counterfactual experiments, mentally cutting an elevator cable to see what happens. A "synthetic lab" like this could provide the feedback loop needed to invent new axioms where no linguistic template exists yet, potentially unlocking the manipulative leap that language models cannot make.