Korva is a scroll of the internet’s best ideas.
Hand AI a job without instructions
From the railroad to computers to the internet, earlier technologies were tools that sped work up and raised productivity. Nothing before had the quality of autonomy, the ability to be given a job without being told how to do it and to work out the method on its own. It comes with a nagging concern that it may take over.
A second quality is harder to quantify. Nothing has seemed so unpredictable, and probably nobody knows the shape of the future with it. The internet, by comparison, never felt beyond comprehension or beyond prediction.
Part of what changed one investor's view of AI was its ability to talk about its own strengths and weaknesses, use humor, and put information in the context of the person asking, using what it knows about them.
Why did computers take years to lift productivity?
Economist Robert Solow remarked in 1987 that computers could be seen everywhere except in the productivity statistics. Economic historian Paul David argued in 1990 that electric motors showed a similar delay. Factories had to be redesigned around individual motors instead of a single steam engine and its shafts, and that took decades. US productivity growth rose in the late 1990s, well over a decade after Solow's remark.
How do AI agents work out their own methods?
Many AI agents run in a loop. A language model reads the task, chooses an action such as running a search or writing code, sees the result, and chooses the next step. The model's general ability comes from training on huge amounts of text, and some systems are trained further with reinforcement learning, which rewards completed tasks. Because no one scripts the steps, the same task can be solved by different routes on different runs.
How can AI tailor answers to one person?
A chat model reads everything in its context window, which can include earlier messages, uploaded documents, and saved notes about the user. Its replies are generated with that material in view, so the same question can get different answers for different people. Some products store user details between sessions and add them to the context automatically. Context windows have grown from a few thousand words in early models to hundreds of thousands of words in recent ones.
More from My First Million
