Korva is a scroll of the internet’s best ideas.
AI needs more context on you
Too much to read
Nobody can read every post on an internal Slack or every story a customer tells about where ChatGPT worked or failed them. Research papers take a lot of mental energy. An agent constantly trying to help could take in far more context than a person has time or energy for, and bring it to bear on a big decision.
What AI can hold
Product work has correctly focused on model intelligence, but it has thought much less about giving a model more context than any person could hold. No one can read tens of thousands of pages in a few seconds and use them accurately, and AI can.
A working example
Here is one working example: since 2018, every note and highlight from every book David Senra has read goes into a database, along with the transcripts of his Founders episodes. Searching it for what Bob Noyce said about something, or what Rockefeller did, turns up notes from books he read seven years ago and no longer remembers.
Why can't models just hold everything?
A language model reads a fixed window of text, measured in tokens, and cannot see anything outside it. In the original transformer design from Vaswani and colleagues in 2017, self-attention compares every token with every other token. The computation therefore grows with the square of the input length. Longer windows have become practical through engineering advances, but the cost of reading more still rises.
How does searching notes by meaning work?
Text can be converted into a long list of numbers called an embedding, positioned so that passages with similar meaning land close together. A search turns the question into the same kind of numbers and looks for the nearest passages. This finds a note about a topic even when it shares no words with the query. Pairing this kind of retrieval with a language model was described in a 2020 paper by Patrick Lewis and colleagues at Facebook AI Research.
Why do people forget books they read?
Hermann Ebbinghaus mapped the forgetting curve in 1885 by memorizing lists of nonsense syllables and testing himself over time. He found that memory drops steeply in the first hours and days and then levels off. Meaningful material is retained longer than nonsense, but it fades the same way without review. Notes and highlights act as an outside record that holds what the memory loses.
More from David Senra
