Korva is a scroll of the internet’s best ideas.
More AI workers than people
Ten times a year
Picture OpenAI with ten million AI laborers this year, a hundred million the next year, and a billion the year after. Even if compute scaling slows, it takes only a few more years before a single company has more labor equivalents than there are people on Earth, which is plausible by the end of this decade.
With or without self-improvement
The tenfold growth holds without any recursive self-improvement, for a fixed level of capability such as a competent remote worker, software engineer, or researcher. With recursive self-improvement, the population could grow a hundredfold or a thousandfold a year, or intelligence could rise while the population stays flat, or some mix of the two.
Where the minds sit
Today this matters little because AIs can't yet do full jobs or work as autonomously as people. The trend still points to most of the world's working minds sitting inside two labs that consume more and more of the world's compute. If those AIs are misaligned, most of the world's minds are misaligned. Even if they are aligned, very few companies hold a great deal of control over the future.
Why does the compute needed per capability keep falling?
Better algorithms, cleaner data, and improved training recipes let a model reach the same skill with less math. A 2020 study by OpenAI found that the compute needed to match AlexNet-level image recognition fell 44-fold between 2012 and 2019. The same pattern has shown up in language models, where smaller recent models match the output of much larger older ones.
How could an AI improve itself?
Mathematician I. J. Good described the idea in 1965, imagining a machine that designs better machines, each better than the last. Today the loop would run through AI writing code, designing chips, running experiments, and building training data for its successors. Each step shortens the time to the next model, which is why the idea is called an intelligence explosion.
More from Dwarkesh Podcast
