Chapter Six · Part III · Crisis
Structure of Scientific Revolutions: 21st & 22nd Centuries
Thomas Kuhn was widely read, and understood by almost no one - he was ahead of his time, threatened the true believers of science, and basically upended the entire chessboard. His theories simply did not fit the trending narratives of a science constrained by the polite conformity of the post-1950s. Although he published The Structure of Scientific Revolutions in 1962, the work did not even realize the changes the 1960s would bring.
The kind of changes in thinking which marked the 1960s were short-lived, easily forgotten and did not begin to affect the mainstream until now, several generations later. In retrospect, the 1960s become reduced to something they weren’t, history distorted to fit the preferred narrative.
Founders today experience similar events, although founders are everywhere, even here on Substack, they notice many people read them, while only a very few people understand what they are saying. If Thomas Kuhn, the philosopher/scientist who changed how we think about science, takes seventy-odd years to be understood, how long might it take until engineered intelligence is comprehended?
Complicating the situation is the parallel growth of quantum science/philosophy. A science which refuses to remain in its assigned lane and which impudently spreads out in a most gregarious way across all disciplines. If Copernicus upended the whole chessboard, quantum science upends the Newtonian game board, chops it into kindling, and celebrates it with a bonfire.
I am not planning on offering a value judgment about the possible success of Ineffable. It stands here as an example of the lengths a founder must go to in order to truly pose something which may in the end overthrow the compute chessboard. Whether they succeed or not is not my interest — it is that they try.
Not a Bonfire of the Vanities, but a Bonfire of the Improbable
Research companion · engineered-intelligence draft for author review · not author prose
Learning from Experience: Ineffable Intelligence
The question about whether engineered intelligence will be comprehended opens a place to examine how it is trained. The Gemini notes supplied by the author point toward Ineffable Intelligence, founded by reinforcement-learning researcher David Silver. The useful thread is learning through interaction and its consequences, rather than depending only on a collection of human-written examples.
Ineffable Intelligence describes an ambition to build a ‘superlearner’ that discovers skills and knowledge through experience using reinforcement learning. Its public description is a research mission, not a published demonstration of a general system that has already achieved those ambitions. The book does not wager on the outcome; the company is cited, in the author’s words, as an example of the lengths a founder must go to in order to pose something which may in the end overthrow the compute chessboard. Silver and Richard Sutton’s Welcome to the Era of Experience sets out a related direction: agents learning predominantly from experience rather than human-generated data.
Several different processes need to remain visible. Pretraining adjusts a model’s parameters using data; post-training can further adjust them using examples and reward-based methods. Reinforcement learning can use feedback from an agent’s actions to change its policy. Additional computation during an answer can explore or evaluate alternatives without changing those parameters. Stored memories and new tool instructions may change a later interaction without retraining the underlying model.
Selecting training examples, proposing tasks, and learning from environmental feedback are also different operations. The Gemini notes combine them into one account of a model choosing its own education. The company’s public mission supports the experience-learning direction; it does not establish that particular autonomous curriculum architecture. Objectives, environments, feedback, and human engineering choices still shape what a learner can encounter and retain.
For the continuation of this book, this offers a concrete question beside Kuhn’s puzzle solver and Papert’s learning by building: what happens when the learner can act on a problem, encounter resistance, and carry the result into its next attempt? Which part changes—the parameters, the memory, the tools, or only the answer—and what new work can the system perform afterward? Those distinctions give the next chapters something specific to examine.
The Gemini transcript’s John Jumper detour, local-agent names, promotional praise, and claims of literal quantum ‘phase transitions’ are not carried into this companion. Neither reinforcement learning nor selecting a curriculum is itself a quantum mechanism. The author’s constructionist framework and the documented training methods remain distinct lines of inquiry.