For most of history, the bottleneck on human knowledge has not been raw intelligence. It has been the cost of doing the work. Running an experiment. Reading the prior literature. Translating a paper. Holding ten ideas in your head at once long enough to notice the connection between idea three and idea nine. Every step that looks like "thinking" in the abstract turns out, on closer inspection, to involve a great deal of tedious, error-prone labor.
AI is collapsing the cost of that labor toward zero. And when you collapse the cost of an input that sits underneath everything else, you do not get a small improvement. You get an explosion.
The bottleneck was all too human
Ask a working scientist where their time actually goes. Only a small fraction is thinking hard on the problem. It is the months of writing grant proposals, the year of building a group of PhD students. It is reading three hundred papers to find the four that matter. It is waiting for a collaborator in another time zone to respond to an email.
None of this is glamorous, and none of it is what we mean when we celebrate "discovery." But it is where the years go. Strip those years away by compressing the literature review into an afternoon, the apparatus build into a weekend, the cross-disciplinary translation into a single conversation, and the rate of discovery does not increase linearly. It explodes.
Compounding is the whole story
A 10% improvement in the speed of doing science, applied once, is uninteresting. The same 10% improvement, applied at every stage of every workflow, every day, for a decade, is civilizational. This is the lesson of every previous tooling revolution, such as the printing press or the computer. AI is a tooling revolution that touches more steps of the knowledge-production pipeline than any of them.
What is new this time is that the tool itself can reason. Earlier tools amplified specific narrow steps: indexing a corpus, transmitting a message, performing arithmetic. AI amplifies the messy steps in the middle, that previously required a human to sit in a chair and concentrate. Once those are amplified, the loops that used to take months instead take hours.
What "amazingly fast" actually looks like
Three concrete shifts are already underway:
The literature stops being a moat. A graduate student in 2026 can synthesize the relevant prior work in a field they have never touched, in an afternoon, at a level that would have taken a postdoc a year in 2015. The "tax" that interdisciplinary work used to pay (learning the jargon, the conventions, the unwritten rules) has dropped by an order of magnitude. Fields that should have merged decades ago finally will.
The experiment loop tightens. Wet labs, simulation pipelines, codebases: anywhere there is a tight design / test / measure loop, AI shortens each turn. The number of hypotheses a single researcher can take seriously in a year goes up by a factor we have not seen since the introduction of the computer itself.
Tacit knowledge becomes legible. The knowledge that used to live only in the heads of senior practitioners, the "I know it when I see it" of a great engineer or a great clinician, can now be elicited, written down, and queried. The bottleneck of expert attention, which has gated whole industries, begins to dissolve.
The pessimist's objection, and why it misses
A common reply: but most of what AI produces is mediocre. The literature reviews are shallow, the code is buggy, the summaries are confidently wrong. This is true. It is also beside the point.
The relevant question is not whether the median AI output is excellent today. That doesn't matter. First, the models are significantly improving every year. Second, the median human doesn't know how to use the models effectively, so we wouldn't expect their output to be excellent anyway. Finally, the median output is sufficient to unblock the next step. A mediocre first draft, produced in practically no time, is infinitely more valuable than a perfect draft you do not have. A flawed simulation you can correct beats a perfect simulation you never ran. The explosion does not require AI to be better than the best human given years of time. It requires AI to be good enough, while being cheaper and faster than the human, so that the AI when run in cycles can produce output that is pushing the frontier quickly and efficiently.
What this means for the future of humanity
Paul Romer, the Nobel prize winning economist, showed that investments in new ideas and technology are among the strongest way to drive economic growth. If we want to improve GDP per capita for all of humanity, improving everyone's lives substantially, we need to generate new significant ideas. That's one of the great promises of AI and of the coming knowledge excplosion: by discovering important new ideas, we will make a significant improvement to GDP per capita. Everyone's lives should be orders of magnitude richer in 2050 than they are now in 2026, assuming AI alignment doesn't go off teh rails (and as an optimist, I don't believe it will).
The knowledge explosion is not some sci-fi far future phenomenon. It is one of the most profound events of the coming years. Get excited.