Expand what intelligence can discover and accomplish.
We want to build software that extends the reach of intelligence: what it can find, what it can understand, and what it can do in the world.
That is a direction, not a claim that we have arrived. The products will change as we learn. The ambition reaches beyond any one of them.
Nothing beyond criticism. Everything open to improvement.
A good idea is worth taking seriously. It is also worth trying to break.
Being wrong is unavoidable. Refusing to correct an error is a choice.
Five windows into progress.
The experiments below are invitations to explore. Each opens onto a field where people have changed what we know, and where much remains unexplained.
How much can we learn from light?
Look at the sky, then look more closely. Light carries clues about worlds we may never reach. An observation can change our picture of the universe.
- AncientLights on a turning sphereStars as fixed points on a crystal shell around the Earth.
- 1584Bruno: the stars are sunsA bold guess, with no way yet to test it.
- 1838Bessel measures a star's distance61 Cygni shifts as the Earth orbits: stars are far, real places.
- 1920The Milky Way is everythingThe Great Debate. Hubble's 1924 measurement of Andromeda ended it: other galaxies exist.
- 1929A static, eternal universeHubble finds galaxies receding in proportion to distance. Space is expanding.
- 1965A beginning, still glowingPenzias and Wilson hear the afterglow of the early universe as noise in an antenna.
- OpenWhat is dark energy?Since 1998 we know the expansion is speeding up. We do not know why.
The farther a galaxy, the faster it recedes. One line that turned the night sky into a history.
Auguste Comte: we will never knowwhat the stars are made of.
Twenty-four years later, Kirchhoff and Bunsen read the elements of the Sun from the dark lines in its light. The unknowable became a measurement.
What is the world made of?
An explanation meets an experiment. Something unexpected happens. The picture has to change. Even the familiar can conceal a deeper problem.
- 450 BCEarth, air, fire, waterFour elements, mixed in proportions. It explained everything, so it predicted nothing.
- 1808Dalton's solid atomsElements combine in fixed ratios. Atoms are real, and indivisible.
- 1904Thomson's plum puddingThe electron is found; the atom is a soft positive ball dotted with charges.
- 1911Rutherford's nucleusAlmost all the mass sits in a tiny core. But orbiting electrons should spiral in.
- 1925Quantum mechanicsMatter as waves of possibility. It powers every chip and laser.
- OpenQuantum gravityOur two best theories, of the very small and the very large, disagree with each other.
Energy comes in packets. Planck introduced them to fix one stubborn graph, and they rewrote physics.
Rutherford: as if you fired a shell at tissue paperand it passed throughand it came back and hit you.
The particles that bounced back broke the plum pudding. The anomaly was the discovery.
How does complexity emerge and change?
Variation, inheritance and selection connect the smallest changes to the diversity of life. Explaining that process opens further questions, including how it began.
- AncientSpecies are fixedEach kind made once, unchanging, in its place on a ladder.
- 1809Lamarck: use and disuseTraits acquired in life are passed on. Close, but the mechanism was wrong.
- 1859Darwin and WallaceNatural selection. Right, but nobody knew how traits were carried.
- 1866Mendel's units of inheritanceGenes, from pea plants. Ignored for 34 years, rediscovered in 1900.
- 1953The double helixFranklin's X-ray images; Watson and Crick's model. Inheritance is a code, copied by pairing.
- OpenHow did life begin?We can read and now edit the code (CRISPR, 2012). We cannot yet say how chemistry first became biology.
Without selection, gene frequencies hold still. Every departure from this line is evolution caught in the act.
Lord Kelvin, from the cooling of the Earth:too young for Darwin's evolution.
He was rigorous and wrong. Radioactivity, unknown in his day, keeps the Earth warm. One field corrected another.
How far can an idea extend our abilities?
A machine turns understanding into something we can use. Building it exposes new limits and new problems. Explanation and invention push each other forward.
- AncientMuscle, wind and waterPower limited to what grows, blows or flows.
- 1712Newcomen's engineSteam pumps water from mines, burning a mountain of coal to do it.
- 1769Watt's separate condenserThe same work for about a quarter of the coal. The Industrial Revolution runs on it.
- 1831Faraday's inductionMotion makes electricity, and electricity makes motion. Power can travel down a wire.
- TodayMotors, batteries, robotsLithium batteries (Nobel Prize, 2019) and motors put machines on wheels, in homes and on factory floors.
- OpenAbundant clean energyFusion, and superconductors at room temperature. Both allowed by physics, neither yet done.
No heat engine can do better. It holds for a steam engine, a car and a power station alike.
Engines had worked for a century, and nobody could say why.
Sadi Carnot, aged 28, explained it: work comes from heat falling between two temperatures. His idea became entropy. Practice led; explanation made it boundless.
What can a machine do?
Simple rules can produce extraordinary consequences. Computing gives us a way to explore them, and raises questions about knowledge, creativity and intelligence itself.
- 1930Every question can be settledHilbert: mathematics would be complete, consistent, decidable.
- 1931–36Gödel and Turing prove limitsSome truths cannot be proved; some questions no machine can decide. From that proof came the universal computer.
- 1958The perceptron will learn anythingIn 1969 Minsky and Papert showed what one layer cannot do. Funding froze.
- 1986BackpropagationMany-layered networks can learn, but computers were too slow to prove it.
- 2012–17Deep learning, then transformersChips caught up. Networks learned to see in 2012; the transformer, in 2017, led to today's language models.
- OpenCan machines create new explanations?General and super intelligence are still conjectures. Whether they can be built, and trusted, is the open question of our time.
Information, measured. Its form matches the entropy of heat engines, and every digital message since has been built on it.
Hilbert: “We must know. We will know.”Everything is decidable.
At the same conference, a day earlier, the young Gödel had announced his proof that this was impossible. The refutation came first, and opened the field.
Progress does not stay inside its categories.
A discovery in one field can become an instrument in another. A better explanation can become a new capability. What once seemed unrelated can become the foundation of something new.
That possibility is why we build.
Unfragile
Our first product. Not the limit of our ambition.
Unfragile's current focus is tool search and discovery for agents.
Our ambition for it is broader: a neutral search layer where humans, agents and robots can find trusted answers.
The current work is a beginning. We intend to develop it toward that larger purpose, while remaining willing to change direction as we learn.
Explore Unfragile →This page is an explanation of what we intend to build. It is open to criticism too.
We expect to learn. We expect to change our minds. We intend to make those changes visible, with reasons rather than rewritten history.
There is more to discover.