An ornithopter is a flying machine that mimics a bird. It flaps. Before the Wright brothers, early aviation experimenters built ornithopters because birds were the model that seemed closest to human-scale, load-bearing flight. The logic held at bird scale. It failed at carrying a human's weight and sustaining flight at a useful altitude. The model was wrong. We are making the same mistake with human and artificial intelligence. And a Turing Award winner just said so publicly.
The ornithopter is not evidence for the argument that follows. It is an analogy for a recurring engineering pattern: reproducing the visible behavior of a successful system rather than identifying the underlying principles that make it successful. No historical analogy is exact; the comparison is useful because it highlights a recurring tendency to optimize existing assumptions instead of questioning them.
Before the Wright brothers, aviation engineers built ornithopters, machines that flapped, because birds were the model that appeared closest to human-scale, load-bearing flight. The logic held at bird scale. It failed at the requirements that defined human flight: carrying a person's weight, and sustaining flight at a useful altitude. The model was wrong. The Wright brothers did not build a better ornithopter. They asked a different question: not how do birds fly, but what does flight actually require? The answer turned out to be unrelated to flapping. This paper argues that the dominant AI paradigm is making the same mistake, and that a Turing Award winner has now said so publicly.
In March 2026, a paper co-authored by Turing Award winner Yann LeCun made a formal argument that human intelligence is not general in any meaningful sense. Its apparent generality is a product of cognitive blindness to the vast landscape of tasks humans cannot perform. Building AI toward a human-cognitive ceiling is, on this account, building toward the wrong target. The ceiling being approached is not the ceiling of intelligence. It is the ceiling of biology.
This paper traces the structural parallel between ornithopter engineering and foundation model development, examines what a fixed-wing equivalent looks like in computing, and explains why the Essence platform's Meaning Coordinates (256 primitives organized as a coordinate system for intent rather than a taxonomy of language) represent the architectural departure the ornithopter engineers never made.
The ornithopter was a serious machine. It had serious investors, serious engineers, serious funding, and serious incremental progress. It could, in some configurations, briefly lift off. What it could not do was carry a human's weight or hold a useful altitude. No amount of additional engineering would allow it to. It was bounded by the biology it was built to imitate.
The Wright brothers did not build a better ornithopter. They asked a different question: not "how do birds fly?" but "what does flight actually require?" The answer turned out to be unrelated to flapping. Lift, thrust, and controlled surfaces. The bird was never the right model. It was merely the one that looked closest.
Many current AI architectures are, this paper argues, building ornithopters. The pattern has persisted for twenty years. The machine we are building is called Artificial General Intelligence, and the model it is built to imitate is human cognition. The assumption embedded in every major AI architecture, every benchmark, every capability claim, is that human-level general intelligence is both the right target and the right model for how to get there.
The ornithopter investors were not fools. They were reasoning correctly from the only available evidence. Birds fly; birds flap; therefore flight requires flapping. The logic was sound given the frame. The frame was wrong.
The parallel to modern AI is not rhetorical. It is structural. Foundation models imitate the products of cognition: the outputs of human language and reasoning. This paper asks whether reproducing those products is sufficient to reproduce the underlying computational process. Researchers and investors have reasoned correctly from the only available evidence of general intelligence: human cognition. That intelligence is associative; therefore AI should be associative. It uses language; therefore AI should use language. It can generalize across domains; therefore AI should generalize across domains. The logic is sound. The model may be wrong.
And unlike the ornithopter investors, we now have a Turing Award winner who has said so, in a preprint.
In March 2026, Yann LeCun, Judah Goldfeder, Philippe Wyder, and Ravid Shwartz-Ziv published a preprint paper making five positions that should have stopped the industry cold. They did not. But the argument deserves a hearing beyond the preprint servers.
LeCun uses the Magnus Carlsen example to crystallize the argument. Carlsen is the greatest human chess player in history. Compared to any modern chess engine, he is not competitive. Our intuition that Carlsen is "good" at chess is a bias artifact. We are comparing within the species, not against the actual performance ceiling. The species ceiling is far below the theoretical ceiling of the task. Building AI to achieve human-level chess performance is building toward a plateau that has already been surpassed. The right question is not how to match Carlsen.
It is how to play chess better than any human will ever play it, and then generalize that adaptability to every other domain where human performance is the ceiling humans have accidentally accepted as the target.
"We are only good at the specific subset of tasks that are important to our existence, but are completely incapable of performing tasks outside this narrow range. Awareness of human limitation gives rise to a critical realization."
The critical realization is this: the AGI frame, in all its formulations, builds toward a human-shaped ceiling. The ornithopter frame, in all its sophistication, builds toward a bird-shaped ceiling. Neither is wrong about the target within its own frame. Both are wrong about the frame.
If you build toward a human-shaped target, you will eventually saturate every benchmark that measures performance against human baselines. That saturation is not a triumph. It is an encounter with the constraint you built into the design.
Consider the trajectory of AI coding benchmarks. HumanEval was introduced as a measure of functional code generation. Models saturated it. MBPP followed. Models saturated it. SWE-bench measured real-world software engineering tasks. Progress slowed, then saturated. FrontierCode Diamond, Cognition's set of its 50 hardest tasks, has not been saturated: the top-scoring model, Claude Opus 4.8, reaches 13.4%. The industry interprets this as progress toward a hard target. An alternative interpretation: these numbers are documenting the distance remaining to a ceiling defined by human software engineering performance. When that ceiling is reached, the benchmark is retired and a harder human-defined task is found.
The saturation cycle is not a bug in the benchmark design. It is the structural consequence of building toward a human-shaped target. Every benchmark that uses human performance as its ceiling will eventually be reached, and then a harder human performance ceiling will be found. The cycle continues until the question changes: not "how close to human?" but "what does intelligence actually require?"
LeCun's SAI frame answers that question differently. Superhuman Adaptable Intelligence is not measured against human baselines. It is measured against the ceiling of the task itself, and against the speed at which it can move between tasks. The relevant metric is not "how human-like?" but "how fast can it learn to exceed any human, in any specific economically important domain?" That is a fundamentally different design target. It produces a fundamentally different machine.
The ornithopter investors made one assumption: that the bird was the right model. The AI industry has made three. Each one narrows the design space in ways that look like precision and function like constraint.
Human cognition appears general because we are unaware of our own incapacity. We call ourselves general-purpose because we can do many things, not because we can do all things, or because we do any given thing better than a specialized system could. The appearance of generality is a product of cognitive blindness, not cognitive breadth. The ornithopter appears to fly because we compare it only to other ornithopters.
Large language models work with language because humans express intelligence through language. But language is a lossy compression of intent. The map is not the territory. Training a machine to predict the next token in a language model is training a machine to predict human linguistic behavior, which is itself an approximation of human cognition, which is itself a constrained expression of intelligence. Each layer of indirection adds ceiling without adding capability.
If the model isn't general enough, add more parameters. If the benchmark isn't saturated, add more data. If the ceiling hasn't moved, add more compute. Scale works, within the frame. It cannot move the frame. You can build the largest ornithopter in history. It will still flap.
These three assumptions are not individually unreasonable. In isolation, each has empirical support. Together, they constitute a frame that makes certain kinds of progress visible and certain kinds of progress invisible. The progress that is visible: benchmark improvement, capability generalization, reduction in hallucination rates, improvement in reasoning chains. The progress that is invisible: the question of whether the architecture is approaching the right ceiling.
The invisible progress is the only one that matters at the paradigm level. The ornithopter investors made measurable progress on wing-flap efficiency for decades. None of it was meaningless. None of it was on the path to sustained flight.
The Wright brothers' insight was not aeronautical. It was epistemological. They stopped asking "how do birds fly?" and started asking "what does flight actually require?" The answer had nothing to do with feathers. Lift, thrust, and control surfaces. A fixed wing generates lift passively. It does not inherit the metabolic ceiling of a flapping wing.
The Essence® platform was built on an equivalent epistemological move. Not "how does human intelligence work?" but "what does computation actually require when intelligence is the input?" The answer: a substrate that can receive intent directly. Not language, not code, not a sequence of tokens that approximate what the user wanted. The intent itself, as a first-class primitive that the machine can execute, govern, and extend without recompressing it through layers of approximation.
If procedures are the equivalent of flapping wings, Meaning Coordinates represent the equivalent of lift and control surfaces: the underlying abstractions that make many procedural implementations unnecessary. The Essence® platform's 256 Meaning Coordinates (4 realms × 32 groups × 8 conjugates) are not a taxonomy of language. They are a coordinate system for intent. The distinction matters: language describes intent, always imperfectly, always with the ambiguity and context-dependence that natural language carries. Intent, addressed directly, is unambiguous by construction. The meaning is in the structure, not inferred from the token sequence.
The argument is not that statistical models are ineffective. They have demonstrated remarkable capability, and they remain a valuable component of any deployed system. The question is whether statistical prediction alone constitutes a sufficient foundation for a general computational architecture, or whether a different substrate is needed beneath it. This is not a claim that language is useless. Language is the interface through which intent is expressed. Essence® treats it as the input, not the computational substrate. What the machine receives is intent. What it executes is intent.
What it governs, through Synergy®, is the fidelity of the execution to the original intent. The machine does not approximate what the user wanted. It is governed to execute exactly that.
The performance results are a consequence of the architecture, not a benchmark target. Chameleon®'s 20–114× measured performance gain (independently evaluated by AWS and Rowan University's Digital Engineering Hub) and energy reductions of up to 99.7% are what happen when the substrate is not carrying the overhead of approximating intent through language. Performance results depend on workload characteristics and hardware architecture; benchmark methodology and evaluation software are available at AdaptWithChameleon.com.
The fixed wing is more efficient than the ornithopter not because it was designed to be efficient, but because it does not inherit the inefficiency of the biological model it was built to replace.
LeCun's SAI frame points toward the same epistemological move from a different angle. His argument is that specialization, not generality, is the right target. That the adaptability of a system should be measured by how fast it can reach superhuman performance on a new specific task, not by how broadly it approximates the human generality that was never general to begin with. The SAI argument and the intent-native argument converge on the same structural insight: the human cognitive model is the wrong frame.
It is not wrong because humans are not intelligent. It is wrong because the constraints of human cognition, metabolic, neurological, evolutionary, are not constraints that intelligence itself requires.
The ornithopter analogy does not demonstrate that today's AI architectures are fundamentally limited. It illustrates a pattern that motivates the alternative architecture presented here. Whether that pattern applies precisely to the current AI trajectory is an empirical question the subsequent papers address. The ornithopter-to-jet transition did not happen by ornithopter manufacturers building jets. It happened by people who did not have flapping-machine assets to protect.
This is the position MindAptiv has held since 2011, not as a contrarian posture, but as the consequence of building from first principles. Wantware, the platform's foundational concept, treats human intent as computationally primitive. Not as input to be interpreted. Not as a prompt to be completed. As the thing the machine is built to execute. Every architectural decision in Essence® flows from that single premise. And that premise has nothing in common with the ornithopter frame, which is why Essence® does not share the ornithopter ceiling.
LeCun's paper ends with a proposal: replace AGI as the North Star with SAI, Superhuman Adaptable Intelligence. It is a better target than the one it replaces. It breaks the human-generality assumption. It opens the design space. It changes the benchmark from "how human-like?" to "how fast can it learn to exceed any human in any specific domain?"
That is progress. It is not arrival. The SAI frame, as LeCun articulates it, still requires a substrate, and the substrate question is where the ornithopter mistake lives. An AI that can rapidly specialize and exceed human performance at any task is valuable. The question is whether that intelligence is grounded in a system that can govern its own execution: take intent as input, execute it faithfully, and account for the gap between what was intended and what was produced. That is not a benchmarking problem. It is an architecture problem.
And it is the problem Essence® was built to solve.
This is where the SAI frame and the AGI frame converge on the same risk, whichever name the industry settles on for what comes after: Superhuman Adaptable Intelligence, or the more familiar Super Intelligence. Task-specific superhuman performance, with no architectural ceiling on what a system can act on, needs a substrate that constrains action by design, not by instruction. An AI told not to expand its own authority is a policy. An AI built with no code path to expand its own authority, or to act on unauthorized intent, is an architecture. Essence® was built as the latter: Aptivs propose intent to Synergy® for governance; they do not execute unilaterally, because the substrate contains no path to unauthorized execution. The imperative is not to make superhuman systems safer to trust. It is to make trust unnecessary to the safety case.
None of this is a bet that Essence® wins the race to build smarter machines. It does not need to. The AGI frame and the SAI frame are both making a capability bet: whichever architecture reaches superhuman task performance first. Essence®'s bet is on a different layer, one any winner of that race still has to pass through to deploy responsibly. That is not an intelligence play. It is an economic one. The Intent Economy is the structural consequence of governed intent becoming, itself, a scarce and necessary asset: human expertise encoded as provenance-verified Aptivs, consulted at every execution a system takes in a domain that expertise governs. It is scarce because it cannot be fabricated or hallucinated. It is necessary because a system acting on ungoverned intent has no determination layer at all, only detection after the fact. Whoever builds the smartest machine still has to govern what it is allowed to do. That governance layer, not the intelligence underneath it, is where this paper argues the durable value sits.
The ornithopter investors were not wrong about the goal. Flight is genuinely valuable. They were wrong about the model, and the model's constraints became the design's constraints, invisibly, because no one thought to question the closest available evidence of what flight required. The same dynamic is playing out in AI right now. Human intelligence is genuinely valuable. Building AI that can exceed it, in every domain, faster than any human can learn, is a worthy goal. The question is whether the model chosen to pursue that goal is the right model, or whether it is merely the one that looked closest, and, as that goal is reached, whether human control over the result was designed in from the start or assumed as an afterthought.
LeCun has now said publicly what MindAptiv has said since 2011: the model is wrong. The frame is wrong. The ceiling being approached is not the ceiling of intelligence. It is the ceiling of biology. And the ornithopter transition, when it comes, will not be made by the organizations with the largest ornithopters. It will be made by the ones who asked a different question, including the question of who remains in control of the answer.
Essence® is not a better language model. It is not a faster ornithopter. It is an intent-native computing platform built from first principles, where human intent is the computational primitive, Synergy® governs the execution, and the machine does not inherit the constraints of the biology it was built to serve. The opportunity it points to is not a bet on out-building anyone else's machine intelligence. It is the Intent Economy: the governance layer every such machine has to pass through, and the layer where durable value accrues no matter who wins the race underneath it. If the argument in these pages resonates, we would like to hear from you.
Explore Essence® → mindaptiv.com