Jumping Intelligence is an automated AI research company.
We’re building toward AI that can evolve its own research agenda.
Fig. 1 · climber vs. jumperillustrative
01Change the representation
The most useful move changes what you can see.
Remove two opposite corners from a chessboard. Can 31 dominoes, each covering two neighbouring squares, tile the 62 squares that are left?
The question stays the same. Only the representation changes.
dominoes placed
0 / 31
arrangements tried
0
representation
squares
62 squares, 31 dominoes. Looks easy.
↑ stop searching. change how you look at it.
Proven
The answer was hiding in a property the placement strategy ignored. The research challenge is to discover useful representations before someone supplies the trick.
Illustrative animation, not an exhaustive solver, and not a test of model capability.
02Look past the score
The experiment you almost threw away.
An optimizer that keeps only what raises its score can discard findings worth investigating.
Optimizer’s log
Keep: what raises the score. Discard: everything else.
Pick one back out.
Recovered report
Original goal
Scored against the goal
What it opened
History & sources
An illustrated counterfactual. In history, each of these researchers kept the result and investigated it. Experiments and timelines are condensed.
03What we’re building toward
A research process that learns where to go next.
Faster experiments don’t tell you which question to investigate next.
1
Generate
Keep competing hypotheses and representations alive, including directions the current score doesn’t reward.
2
Investigate
Turn surprises into experiments that tell competing explanations apart. A model that loses on average might reveal an advantage when training data is scarce.
3
Carry forward
Keep what works reusable: methods, anomalies and partial solutions to revisit when new tasks appear.
Discoveries can improve the process that makes discoveries.
A discovery can become a better training method, evaluator or search procedure. Carrying it forward changes how the next round of research is done.
Deduction derives consequences and induction tests them. Abduction, forming the explanatory hypothesis in the first place, is “the only logical operation which introduces any new idea.”
Discoveries often start from an “unanticipated, anomalous and strategic datum”: a result nobody expected, that seems inconsistent with existing theory or established facts, and that points somewhere new.
Same method. Same astronomer. Opposite outcomes: a new theory, not a new planet, explained Mercury.
Normal science works inside an accepted frame. Le Verrier’s successful prediction of Neptune made the same move reasonable for Mercury, but Vulcan was never confirmed; in 1915 a new theory, Einstein’s relativity, explained the drift.
Existing traits can acquire new uses, including features that evolved for another role. Feathers likely helped with warmth or display long before they were used for flight.
After weeks on Fuchsian functions, the idea came as he stepped onto a bus on a geology trip: the transformations he had used to define them were the same as those of non-Euclidean geometry.
Koestler called creative connections between previously separate frames “bisociation.” Claude Shannon joined Boole’s logic to telephone relays and founded digital circuit design.