Deconstructing the Kuhnian Revolution

In 1962, Thomas Kuhn published The Structure of Scientific Revolutions, a monograph that arguably did more to change our understanding of “discovery” than any other work in the 20th century. Before Kuhn, the prevailing view of science was one of “logical empiricism”—a belief that science was a steady, linear accumulation of facts, building brick-by-brick toward an ultimate truth.

Kuhn dismantled this cumulative view. Instead, he proposed a morphology of scientific development characterized by long periods of dogmatic stability interrupted by violent, radical upheavals. This “Kuhnian thinking” suggests that science does not march toward a fixed finish line; rather, it evolves through a cyclical process of paradigms, crises, and revolutions.

To understand this structure is to understand the history of human inquiry—and, crucially, the current trajectory of Artificial Intelligence.

1. Normal Science: The Cumulative Framework

Kuhn introduces the concept of “Normal Science” to describe research firmly anchored in a shared paradigm. A paradigm is more than just a theory; it is a worldview. It is a set of “internalized” ideas, tools, and standards that a scientific community accepts as the foundation of their practice.

• Mopping-Up Operations: During this phase, scientists are not trying to invent new theories or challenge the status quo. Instead, they are engaged in what Kuhn calls “mopping-up operations.” They are solving puzzles defined by the paradigm, aiming to increase the precision of existing knowledge.
• The Rigidity of Belief: Progress here is cumulative and highly efficient because the community agrees on what questions are worth asking. Dissent is discouraged; the focus is on “filling the gaps” of the current framework.

Contextual Example:

Consider the era of AI research prior to 2017. The dominant paradigm was Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks. The community had “internalized” the belief that language was inherently sequential—meaning a machine must read a sentence one word at a time, from left to right, to understand it. For years, “normal science” in AI consisted of making incremental improvements to this sequential framework, never questioning the sequential premise itself.

2. Anomalies and the State of Crisis

Normal science is successful, but it is not perfect. Eventually, the rigid framework encounters anomalies—violations of expectation that the paradigm cannot explain.

• From Irritant to Crisis: Initially, scientists ignore these anomalies or devise ad-hoc fixes to explain them away. However, when anomalies persist and “systematically thwart the puzzle-solving efforts,” the field enters a state of crisis.
• The Loosening of Rules: A crisis is a period of profound professional insecurity. The rules of normal science loosen. Researchers begin to argue; they express discontent; and crucially, they become willing to try “speculative theories” that would have been considered heretical during the period of normal science.

Contextual Example:

In the world of NLP (Natural Language Processing), the anomaly was the bottleneck of sequential processing. No matter how much researchers optimized RNNs (normal science), the models could not handle massive datasets or remember the context of long paragraphs. The sequential nature of the paradigm—once considered its greatest strength—became the very thing holding it back. The field realized that the old map could no longer guide them.

3. Scientific Revolutions and Incommensurability

The crisis is resolved only by a Scientific Revolution—a non-cumulative event where the old paradigm is replaced by an entirely new one. Kuhn likens this to a “Gestalt Switch” (like looking at an image that flips between a duck and a rabbit). Once you see the new pattern, you cannot un-see it.

• Incommensurability: Kuhn argues that the new paradigm is incommensurable with the old. This is a difficult concept: it means the two paradigms are not just different; they are incompatible. They utilize different vocabularies, different methodologies, and value different things. You cannot measure the new paradigm using the yardstick of the old.
• Destruction as Creation: The transition is not an addition to knowledge; it is a reconstruction of the field from new fundamentals. The old paradigm is not “fixed”; it is abandoned.

Contextual Example:

The introduction of the Transformer architecture (Google, 2017) was a textbook Kuhnian revolution. It didn’t “improve” sequential processing; it rejected it entirely. It proposed that Attention Is All You Need—that a model could process an entire sentence simultaneously (parallel processing) rather than sequentially. This was incommensurable with the old way; it changed the very definition of how a computer “reads.”

4. Progress as Evolutionary Competence

Perhaps Kuhn’s most controversial contribution is his rejection of teleology—the idea that science is moving toward a specific, pre-destined goal called “The Truth.”

• Evolutionary, Not Teleological: Kuhn argues we should view scientific progress like biological evolution. Evolution has no “goal.” A generic species does not evolve toward a “perfect ideal”; it evolves away from primitive beginnings.
• Problem-Solving Competence: Therefore, a new paradigm is not selected because it is arguably “truer” in a metaphysical sense. It is selected because it is a more “powerful organizer of successful inquiry.” It possesses a higher competence in solving the specific puzzles that the community currently faces.

Analogy for Understanding: The Darwinian Shift

To visualize Kuhnian progress, look to the evolution of a species.

A species does not evolve to become a “Perfect Animal.” It evolves to become better adapted to its specific environment.

• Normal Science is a species thriving in a stable environment.
• An anomaly/crisis is a sudden change in climate or the introduction of a predator. The old traits no longer work.
• A revolution is a genetic mutation that offers a survival advantage in this new world.

When a species evolves, we do not say it is “closer to the truth of biology.” We say it is more competent at surviving. Similarly, we moved from RNNs to Transformers not because Transformers are the “True Mind,” but because they are vastly more competent at surviving the environment of massive internet-scale data.