Skip to content
Paximal
All insights

AI & Drafting6 min read

The Coastline of Innovation

By Ian Schick, PhD, Esq.

The coastline of innovation. A grid of rounded tiles in which two solid endpoints gain a new orange midpoint on every row, with smaller orange tiles appearing inside the remaining gaps.

Innovation cannot be exhausted, and it cannot be fenced off, because it behaves like a fractal: the closer you look, the more of it there is.

Every few years someone announces that the good ideas are nearly gone. The low-hanging fruit has been picked, the argument runs, and what remains is too expensive to reach. A smaller group goes further and tries to finish the job on purpose, by publishing every possible invention so that nothing is left to patent.

Both camps misread the shape of the thing. Innovation is not a field that gets harvested. It is an edge that gets longer every time someone measures it. AI is the shortest ruler we have ever had, which means the edge is about to get much longer.

There is always a point in between

Pick any two points on a line and there is another point between them. Then there is a point between that one and each of its neighbors, and so on without end.

Technology works the same way. Between any two existing products there is a third that borrows from both. Between any two solutions to a problem there is one that is cheaper than the first and lighter than the second, or that works at a temperature neither can tolerate.

Benoit Mandelbrot made a related point about the coast of Britain in 1967. The measured length of a coastline depends on the length of your ruler. A shorter ruler picks up more bays and inlets, and the total keeps growing. The state of the art has that kind of edge. From a distance it looks like a settled boundary. Up close, every stretch is full of gaps, and every gap has gaps of its own.

Here is the part the pessimists miss. A new invention does not use up a spot. It becomes a new endpoint, with open space on either side. Filling in the map makes the map bigger.

The economist Martin Weitzman formalized this in 1998. In his model, the production of new ideas is “a function of newly reconfigured old ideas.” His conclusion was that the limits to growth lie “not so much in our ability to generate new ideas as in our ability to process” that abundance into usable form.

The plan to publish everything

In 2016 Alexander Reben, whom NPR described as an “inventor, artist and MIT-trained roboticist,” launched All Prior Art. The site pulls language from US patents and published applications, recombines it, and publishes the output as purported prior art. Reben told NPR it produced about 36,000 ideas a minute. Within weeks InfoWorld counted roughly 4.2 million listings. A sister site, All the Claims, does the same for claim language.

The theory is simple. If every idea is already published, no one can patent it. “Once it’s out there the onus is on the patent office to find the prior art,” Reben said.

The project is candid about its quality. Its own explanation concedes that “most inventions generated will be nonsensical.” A law firm bulletin that June was blunter, calling the overwhelming majority “pure gibberish.”

As art, it is clever. As a plan to exhaust invention, it is a stunt. The reasons are worth spelling out, because generative AI has made the idea fashionable again. In April 2024 the USPTO asked for public comment on how AI-generated material affects prior art.

Why it cannot work

Start with the arithmetic. At 36,000 ideas a minute, the machine produces about 19 billion a year. A single 20-word sentence drawn from a 10,000-word vocabulary can be written 10 to the 80th power ways, roughly the number of atoms in the observable universe. That is one sentence. The generator is not filling the space. It is not denting it.

Then the law. A string of words is not automatically prior art that defeats a patent. To anticipate a claim, a reference has to enable a skilled person to make and use the invention without undue experimentation (MPEP 2121.01; Elan Pharmaceuticals v. Mayo Foundation, Fed. Cir. 2003). Practitioners have also questioned whether millions of machine pages are publicly accessible in any meaningful sense. A non-enabling reference can still be cited for obviousness, but only for what it actually teaches. Gibberish teaches nothing.

The deepest problem is the fractal one. Suppose the machine gets lucky and lands a real, enabled disclosure. That disclosure is now a point on the map, with fresh territory on both sides: the improvement, the cheaper variant, the combination with something published next year. Blocking art is also starting material.

Reben supplied the best summary himself: “If an idea is truly creative and innovative, a computer should have difficulty coming up with it.”

AI is a shorter ruler

None of this means invention is easy. In 2020 four economists reported in the American Economic Review that “research effort is rising substantially while research productivity is declining sharply.” Keeping Moore’s Law on pace now takes more than 18 times the researchers it took in the early 1970s.

But harder to find is not the same as running out. It is what you would expect on a fractal edge. The gaps get finer, and the instruments needed to work them cost more.

AI is the instrument that changes that cost. In 2023 Google DeepMind reported that its GNoME model had predicted 2.2 million new crystal structures, 380,000 of them among the most stable. The company called the result “equivalent to nearly 800 years’ worth of knowledge.” Before GNoME, about 48,000 stable crystals were known.

Notice what did not happen. Materials science did not get closer to finished. Every one of those structures is a new endpoint, with its own neighbors to explore. AI shortened the ruler, and the coastline got longer.

This is the same family of technology as All Prior Art, pointed the opposite way. One recombines words and hopes. The other searches the gaps for things that would actually work. The first produces noise. The second produces invention, faster than any research team could by hand.

So acceleration and inexhaustibility are the same fact seen twice. The faster AI fills in points, the faster new space opens between them.

The scarce thing is not ideas

That is good news, and it should change how inventors and their counsel think. A crowded field is not a closed field. The space between two references is where most patents have always lived.

Volume is not disclosure, either. The standard that sinks machine gibberish is the same one a patent application has to meet. What makes an idea count is a description complete enough for someone to build it.

So the bottleneck is the one Weitzman named: processing. Ideas are abundant, and AI is multiplying them. The capacity to turn them into enabled, claimable disclosures is not. That is where AI earns its place in patent practice. Not generating noise at 36,000 ideas a minute, but helping a practitioner capture a real invention completely, under review, before the next point appears on the map.

No one will ever publish the last idea. There is always one more in between.

Sources

See Paximal on Your Own Matter

Book a 30-minute demo with our team. USPTO-registered practitioners can also start a free trial.