AI Research Could Split Into Tasks Thousands of AIs Could Do at Once
Like a giant math project with thousands of checkable jobs, AI research has many separable jobs where success can be checked.

Dwarkesh Podcast · The avalanche that hit math is coming for AI developmentA great novel probably cannot be written much faster by adding more writers, but building better artificial intelligence might be different, argues AI commentator Dwarkesh Patel.
The reason is structure. A novel must hold together as one voice, with each chapter depending on the rest. Much AI research breaks into separate jobs that can be tried and judged on their own.
A clear example is the effort to translate the proof of Fermat's Last Theorem, a famous math claim settled in the 1990s, into a form computers can check step by step.
That project contained thousands of modular subtasks. Different contributors could complete pieces independently, and each could check locally whether their piece worked.
AI development looks closer to that project than to novel writing, in Patel's view. Labs already divide the work among dozens of teams, each improving a different part of the system and testing whether its change helped.
One team cleans and selects the massive text supply used for initial training. Another builds artificial practice environments where models learn by trial and reward. A third writes a better attention kernel, a small piece of computer code that helps a model focus efficiently on the relevant parts of what it reads.
In Patel's hypothetical future, a swarm of automated researchers could attack each stream separately, trying a thousand approaches and keeping the winners. That kind of persistent, independent trial and error, he argues, tends to speed up this sort of work a lot.
As Patel puts it, "having 100,000 collaborators doesn't mean that you'll go 100,000 times faster, probably not even 100 times faster."
Patel acknowledges steeply diminishing returns from simply adding researchers, but argues that with so many testable parts, a vastly larger pool of automated workers would still plausibly make progress at least a couple of times faster, especially combined with faster researchers.
Questions people ask
What makes a research task verifiable?
It means there is a quick, local test of whether a change helped, such as whether a proof step checks out or whether a model scores higher on a test. Researchers do not have to wait for the entire system to be finished to know they made progress. That makes it possible for many workers to try different ideas at the same time.
Why doesn't adding more researchers speed things up proportionally?
Many jobs overlap, require coordination, and compete for the same limited resources. In AI work, experiments also need computing power, communication takes time, and many ideas duplicate each other. Patel's point is that returns diminish steeply as more researchers are added.
What is an attention kernel?
It is low-level software that carries out attention, the process by which a model decides which earlier words or data to emphasize. A more efficient version can make training or answers faster or cheaper. Because its performance can be measured directly, it is a natural target for independent experimentation.

