At the frontier of a deep field such as fundamental physics, understanding a problem can require years of preparation. A researcher has to internalize the theories, mathematical tools, and assumptions that make the question possible to ask. By the time that foundation is in place, much of their thinking is shaped by it. The knowledge required to reach the frontier also carries a weight that makes it difficult to see beyond it.
Artificial Intelligence introduces a different relationship with that constraint. Thousands of agents can work on a problem simultaneously, explore different approaches, and exchange findings at a pace that no human research team could sustain. As those systems become more capable, the time required to produce new knowledge will unavoidably fall below the time we need to understand it.
This essay looks at what happens when research advances faster than humans can internalize its outcomes, and what that means for our ability to remain involved in the work.
The Weight of Priors
At Abstract, we have spent a decade building highly complex products such as InstaLOD or RSX Engine. As the software grows, it becomes gradually more complex and more difficult for engineers to hold every implementation detail in mind at once. We introduce abstractions and interfaces that allow us to reduce what we have to remember to a surface-level contract. For instance, an engineer can build on an interface for a render backend without knowing how it talks to the GPU driver or write a computational geometry algorithm using a meshing API without the need to know every operation performed by it.
Those abstractions make complex systems possible to build. They allow us to move past details that have already been worked out and focus on the next problem. But the details remain, and when an abstraction fails to hold, someone has to understand what sits beneath it.
Scientific knowledge has a similar structure. Each accepted discovery becomes part of the foundation for the next, and specialized language allows researchers to refer to concepts that would otherwise require pages of explanation. Learning a new term is easy; developing the understanding required to use it meaningfully can take years. Software engineering lends another fitting example: the concept of a "pointer". Its mechanics and principles can be explained to undergrad students in an hour, but it will take a whole semester for students to internalize and being able to apply it properly.
The Scale
Nearly a century passed between the formulation of the Poincaré conjecture in 1904 and the award of the first Millennium Prize Problem for its solution.
On September 8, OpenAI announced a solution to another Millennium Prize Problem, the Navier–Stokes existence and smoothness problem. According to the company, the group responsible involved approximately 10,000 LLM-based agents and arrived at the result in about 88 hours.
The breakthrough revealed something consequential: a research effort can now apply thousands of instances of a capable model to the same problem without spending decades educating each participant. Humanity has always parallelized research, but every human researcher has to acquire the necessary foundation individually, thus, research teams are comparatively small. A university cannot copy the expertise of its strongest researcher and paste it onto ten thousand researchers overnight. With AI, deploying thousands of additional instances of a frontier model is trivial, if the compute resources are available.
I believe that frontier models applied at this scale already represent a form of superhuman intelligence and lets us glimpse into its consequences.
The Language
The first consequence will be an increasing distance between the production of knowledge and our ability to absorb it.
Consider a field in which AI systems develop new methods, then use those methods to establish further results, and then build another layer on top. A human researcher begins studying the first result while the agents continue working. By the time the researcher understands its implications, the active research may depend on countless additional layers of novel concepts.
This is a challenge already familiar to engineering teams that leverage coding agents in their process: the amount of code LLM-based agents produce make it very difficult for humans to keep up with the review process. A few PRs later and the boundaries of the system are still familiar to the human team, but the inner workings might already be opaque.
An AI could explain each of those layers. The human would still need time to understand the explanation. Because the team moves significantly faster now, there is no longer enough time to review all submissions at the required depth.
At some point, even the problem statements could become difficult for us to interpret. They would refer to concepts that were developed by other agents and have already become ordinary parts of their work. The language might remain fluent and grammatically familiar while its meaning becomes increasingly inaccessible. In some form, this is already happening and it is what is commonly referred to as "claude speak": sophisticated sounding prose that's difficult to decipher.
There's one uncomfortable truth buried in that code: the referenced object isn't the driver of this route handler — it's the passenger, carried through the controller without ever being vetted. And when unvalidated input rides along at the security boundary, it's not just a design smell. It's an open door.
↪ Claude referring to a request parameter that was passed along unvalidated.
In this case, the developer knows the underlying concept and can eventually work out what the model is trying to say. The language slows down the process, but the meaning is within reach. What if the words themselves referred to novel concepts that we had never encountered?
The Constraint
This has consequences for the assumption that humans will continue to direct AI research.
Choosing a useful research direction requires an understanding of what is known, what remains uncertain, and which questions are worth pursuing. If those judgments depend on knowledge that humans have not yet absorbed, the decisions will increasingly be made by the systems doing the work.
We could still state broad goals: develop a better battery, treat a disease, or build a more capable AI. Within those goals, however, the choice of problems, methods, and experiments are then opaque to us. We would depend on the same systems to explain why their chosen direction was worth pursuing.
This is further amplified when intelligence is parallelized and progress is multiplied within the same amount of time. For us, time is the constraint on understanding: we need it to learn, understand and internalize. Increasing the pace of discovery does not increase our capacity to do those things at the same rate.
Humanity will benefit from discoveries long after it loses the ability to participate meaningfully in making them. The frontier will keep advancing while our understanding will lag progressively more behind.