llm's can jump

July 31, 2026 · 588 words · 3 min read · ml, agents, critic

I was reading this paper ,

very fascinating paper, and a brief read on the history of einstein’s work on relativity irrespective of whether you think LLMs can or cannot intuit relativity.

the premise of the paper was that, when Einstein was working on general relativity there was no pressure to discover it because almost everything pre-existing fit correctly under the Newtonian principles. only several years after Einstein’s work on relativity were there contradictions to Newtonian mechanics discovered by others. during Einstein’s time pretty much the only contradiction was a tiny shift in Mercury’s orbit called the advance of perihelion even for which there were teams working on proving that there is another planet behind the view of the sun.

Despite having so little signal from science at the time, Einstein was able to point at a problem critical to the world a decade from when he started working on it. The authors argue this is fundamentally incompatible with the creativity as data compression theory. because if you had compressed the frontier science at the time there was no way for an LLM to intuit relativity.

The paper suffers from fundamentally from focusing too much on the discovery of relativity as an example and trying to derive a general theory of discovery from it.

Let’s keep Einstein and discovery aside for a moment, in the last 200 years some of the greatest leaps for humanity were,

• Germ theory by Louis Pasteur
• Theory of Evolution by Charles Darwin
• Electromagnetism by Michael Faraday
• Structure of the DNA by watson & Crick
• Transistors
• Nuclear Fission

Were all crucial but almost all of these, came from the individuals diligently going into the world in search of signals in their area of expertise. Famously Charles Darwin did six major expedition before he started working on “on the origin of species”. Faraday worked on electromagetism for 10 years before he found a solution.

And in almost all of these cases, the signals have started to accrue and multiple people jointly started to notice the patterns. Watson and Crick were competing with Rosalind Franklin and Maurice Wilkins for the Double helix theory. Similar stories for Electromagnetism, Nuclear Fission etc.

my argument is that, Inductive and Deductive creativity have been the major engine of human scientific discovery. not Abductive creativity.

without signals to validate discoveries become conjecture until proven otherwise. Similar to Gregor Mendel’s lifelong toil which went ignored until 15 years after his death the principles were rediscovered. How many of similar toils just went to naught because we couldn’t trace them back or, weren’t signals at all?

With LLMs though signal pattern matching can be scaled relative to the available compute,

just in the last six months,

• OpenAI model has proved planar unit distance problem an 80 year open problem. (Mar 2026)
• Claude proved the Jacobian Conjecture wrong an 87 year old problem. (July 2026)

a few of those notoriously difficult enough that the smartest brains of our generation couldn’t arrive at a solution after a decade working on those.

The problem with Abductive creativity that the paper argues is that, there already was not a lot of it to go around in the first place. One Extreme example is not enough. the question should be reframed as, “How long would an LLM have taken to arrive at Relativity, from the time the anomaly was discovered?”. I believe, much faster than humans.