Korva is a scroll of the internet’s best ideas.
Spend tokens thinking while you sleep
A perfect memory
Always on, it would look at everything you look at on your computer, listen to every meeting you are in, and read every document you read. A memory like that could bring up the email read six weeks ago, or exactly what happened in a meeting seven and a half weeks ago, right at the moment it feeds into a decision, which feels pretty magical next to a human memory.
Thinking while you sleep
A slider could set how many tokens it spends thinking while you sleep: coming up with useful new ideas, doing whatever work it can, and working out what to do next, so the output is better by the next morning. Someone could drag that slider quite far and be willing to spend a lot for it.
The compute limit
The obstacle is compute. All of that takes a lot of tokens, and if everybody in the world wants to drag the slider pretty far, the amount of compute required is a lot.
How could AI recall a meeting weeks later?
Recording everything is easy; finding the right piece later is the harder step. Systems typically convert transcripts and documents into numerical vectors called embeddings, which place passages with similar meaning close together. When a question arrives, the system searches for the nearest passages and hands them to the model. The approach is known as retrieval-augmented generation, introduced in a 2020 paper from Facebook AI Research.
Why does thinking longer improve AI answers?
Reasoning models write out intermediate steps before answering, and each step gives the next one something to build on. Google researchers described a related technique, chain-of-thought prompting, in 2022 and found that asking large models to show their steps improved results on math word problems. When OpenAI announced its o1 model in September 2024, it reported that accuracy on hard math and coding problems rose as the model was given more compute to think.
Why do human memories fade so quickly?
German psychologist Hermann Ebbinghaus tested himself on lists of nonsense syllables in the 1880s. He found that forgetting was steepest right after learning and then slowed, a pattern now called the forgetting curve. Memory is also rebuilt each time it is recalled rather than played back like a recording. Studies of eyewitness testimony by psychologist Elizabeth Loftus have shown that remembered details can shift between recalls.
More from Invest Like the Best with Patrick O'Shaughnessy
