Every discovery
starts somewhere else.
A five-minute guide to how progress really happens, what Singularity does about it, and what you can do with it — whether you are sixteen, a first-year student or simply curious.
Ideas travel. Slowly.
Many of the greatest advances in science were not new ideas. They were old ideas arriving in a new place. Pick one and see how long the journey took.
If a method can wait thirty years at the door of the field that needs it, the fastest way to speed up science is not only to invent — it is to open the door sooner.
Three words are all you need.
Everything on this site is built from three ideas.
Method
A tool for thinking or doing: a kind of equation, an algorithm, a model, a measurement technique. Monte Carlo sampling, graph neural networks, Bayesian inference.
Field
A community of researchers working on similar problems, with their own journals, language and habits. Fluid dynamics, neuroscience, number theory. The flagship model follows 252 of them, across all of science.
Crossing
The moment a method enters a field that had never used it. Singularity counts it as real only when at least three papers of that field use the method — one paper can be an accident, three are a trend.
A forecast, then, is one sentence you can check.
“At least three experimental particle physics papers will use Kolmogorov-Arnold networks by the end of 2028.”
One method, one field, one deadline. It will be true or false, and everybody will be able to see which. Open this real forecast →
Can you see the future?
Ten real crossings from the past. For each one you see only what was known at the time — exactly what our model saw. Decide: did it happen within three years? Then see what happened next, even years later, and compare yourself with the model.
What Singularity does, in four moves.
No magic and no crystal ball: a machine that has studied how every method in its archive spread, field by field, for decades.
Read
Millions of titles and abstracts, with their field and date. Only public, open metadata.
Recognise
Find the methods inside the text, among the scientific concepts of a curated vocabulary, with their synonyms and acronyms.
Learn from history
Replay the past year by year and learn which signals came before a method crossed into a new field.
Forecast, then wait
Rank every crossing that has not happened yet, publish the top ones with a date and a fingerprint, and let reality judge.
Your move.
The best moment to learn a method is just before your field discovers it. Here is how to use Singularity to choose a thesis, a project or simply what to study next.
- 1
Open the Radar and pick the field you love.
Medicine, ecology, economics, chemistry, mathematics: 248 subfields of every discipline are there.
- 2
Pick a method from the top of the list.
Those at the top are the strongest bets. Next to each one you see its chance of arriving within three years, measured on fifteen years of history; open it to see how much stronger that is than chance.
- 3
Read three papers where it already works.
Every dossier links real papers from the neighbouring fields. Read how they use it.
- 4
Try it on one small problem of your field.
A dataset, an equation, a simulation. Small is fine: the first three papers are what make a crossing.
- 5
You could be one of the first three.
That is how we write our own studies, each in days. See them →
See the whole of science on one map.
All 163 million works we read, on one map you can zoom like a sky. Tap a region to see what it studies, which methods it uses and who its neighbours are; type a method to see everywhere it already lives; slide back in time to watch science grow.
Open the atlas →Glossary
- Backtest
- Pretending to be in the past. We give the model only the papers published up to, say, 2015, ask it for forecasts, and check them against what really happened in 2016–2018. We did this for every year from 2008 to 2023.
- Untouched crossing
- A field that had not a single paper with the method when the forecast was made. Guessing these is much harder than guessing trends already under way — and much more useful.
- Top 100
- The hundred crossings the model considers most likely, out of millions possible.
- Hit rate
- The share of forecasts that came true. Ours is compared with two simple strategies: betting on the most popular methods, and picking at random.
- Preregistration
- Writing down the rules of a test — what will count as success — before running it, and running it only once. It prevents us from fooling ourselves.
- Fingerprint (SHA-256)
- A 64-character code computed from the exact text of the forecasts. Change one letter and the code changes completely. Publishing it proves the forecasts were not edited afterwards.
- Horizon
- How far ahead a forecast looks. Ours: three years.
- OpenAlex
- An open catalogue of the world's scholarly works, from every discipline. Our flagship model reads 163 million of them — the ones with a title and an abstract it can understand.