In 1905, Einstein began with a puzzle about a magnet and a conductor. The effect depended only on how the two moved relative to each other, yet physics described it differently depending on which one moved [1]. To resolve it, he had to ask something more basic: what it means for two distant events to happen at the same time [1, Sec. 1].
Breakthroughs in science often look like this. Someone finds a new way to frame the problem, and the old difficulty dissolves [2, Ch. 10]. These moments come from creativity, and they are rarely examined directly. For much of the last century, philosophers of science set them aside: Hans Reichenbach called this the context of discovery, and left it to psychology [3, pp. 6–7].
Through most of the history of science, the moment a problem gave way happened once, inside one person. What science kept was what worked, though most attempts fail [4].
Put a thousand agents on the same hard problem, and a few of them may solve it. The rest will not. Runs like these could leave something new: the rare success beside the many failures, in records we can read. Inside these runs are the moments that create progress, a jump in performance or a step in science.
Lightfast is a research lab that wants to study those moments: what creates them, and whether that can become a science.
To invent is to choose
Our thesis is that what sets the successful runs apart is what they choose to try. People who have made great work, in art and in science, describe it the same way. In a lecture in London in 1885, James McNeill Whistler said nature holds the elements of every picture, as the keyboard holds the notes of all music, "but the artist is born to pick, and choose, and group with science, these elements" [5]. Henri Poincaré, writing about his own discoveries in mathematics, said almost every new combination of known ideas is worthless: "Inventer, c'est discerner, c'est choisir," to invent is to discern, to choose [6, p. 48].
Often the choosing is what to leave out. Sculpture, Michelangelo wrote to Benedetto Varchi, is made "per forza di levare," by taking away [7]. The sculptor sees the shape in the block and exposes it. In music, much of what tells one style from another is the space it leaves between sounds. A jazz saying often credited to Miles Davis, "it's the notes you don't play that matter," has never been found in his words [8]. In his own words, what struck him about the pianist Ahmad Jamal was "his concept of space, his lightness of touch, his understatement" [9, p. 178].
We call the sense of what to choose, and what to leave out, taste. It is not only artistic. For a scientist, it is how they think and work during an experiment: how to frame the problem, which question to ask, which of ten promising ideas to try first, when a result is good. And it is trained. Rick Rubin writes that studying great work is meant to "calibrate our internal meter for greatness. So we can better make the thousands of choices that might ultimately lead to our own great work" [10, p. 50].
Models have less of it. In a study with more than a hundred NLP researchers, reviewers rated ideas from a language model as more novel than ideas written by experts [11]. But the model was poor at evaluating its own ideas [11], and when the ideas were carried out, the model's lost far more of their scores than the experts' did [12].
Less, chosen well
A model's taste is trained by what it learns from, and most of what it learns from records finished results: the published paper, the final code, the painting on the wall. It rarely records the choices that led there, the ideas tried and dropped, or why one was worth pursuing. We believe this is part of why models have less taste. Runs like the ones above could keep that record.
Choosing what a model learns from already changes what it can do, and often less, chosen well, does more. A large model tuned on 1,000 carefully selected examples was preferred to the same base model tuned on 52,000; its authors called the paper "Less is more for alignment" [13]. In mathematical reasoning, 800 selected problems took a model from 16.5% to 63.3% on the AIME 2024 competition benchmark, past a model trained on 114,000 [14].
These results have limits. Each one draws out a skill the model already learned from far more data. When a small filtered set is repeated long enough, unfiltered data overtakes it [15], and at a million examples most selection methods do little better than random [16]. We have found no study that tests whether less, chosen well, produces new ideas rather than better answers.
Lenses to see the shape within
What a person learns from can change what they are able to see. Leonardo da Vinci thought painting was a science. Painting, he wrote, "taught figuration to geometry", and it teaches the makers of perspective, the astronomers, the machine-builders and the engineers [17, p. 256], [a]. He studied the body because he had to paint it [18]. Experience, he wrote, is "the common mother of all the sciences and arts" [19, Sec. 18].
Vasari, writing the lives of these artists in 1568, had one word for the source of architecture, sculpture and painting: disegno, the idea formed in the mind before the hand shows it [20, Ch. XV]. In 1613 the Florentine academy named for it elected Galileo, entered as "Matematico" [21]. Four years before, he and Thomas Harriot had each turned a telescope on the Moon. Harriot saw only spots [22]. Galileo saw mountains and craters [23]; on one historian's reading, his training in light and shade let him see them [22].
We know what these people made. We know far less about the moments when they saw differently. Those moments belong to no single field, so studying them serves the sciences and the arts at once. If we learn what creates them, breakthroughs could come more often, and models could frame problems themselves: in mathematics and the sciences, in engineering, and in the images, objects and buildings people make.
What we will find out
Our first runs will test our thesis: whether what sets the successful runs apart is what they choose to try, and whether less, chosen well, leads a model to new ideas rather than better answers. We will read them for the moments themselves: what happens just before a hard problem gives way to a new framing.
We would like to hear from teams working at the edge of what their models can do and asking what comes next, whether at a frontier lab, in robotics, or anywhere the next step is not obvious. And from anyone who holds data they want to reach those teams. Write to us.