Korva is a scroll of the internet’s best ideas.
Finding old research
Finding which research was done in what year, with what question and what test, used to mean sending an email that interrupts someone, often an expert who had been at Netflix for 20 years and had seen every experiment. Now it can be found almost instantly. A person can form their own view of what is interesting and skip a couple of steps toward deciding whether anything is actionable.
A head start
AI is powerful at distilling information, looking across a broad set of material, and getting to the core insights to start with. It is a head start and not something to rely on exclusively. Business stakeholders in finance, content, and advertising can do the same, then bring back an initial hypothesis before working more deeply with data scientists and engineers.
Checking the results
Quality depends on checking that the results are valid, working with a local data scientist, and confirming that the source-of-truth data was used. Applied this way, AI goes beyond prototyping and coding to general analytical thinking, translating data into action and insight.
Why does organizational knowledge hide in people?
Psychologist Daniel Wegner studied this in the 1980s and called it transactive memory. In studies of couples and small groups, he found that people divide up what to remember and keep track of who knows what. A group's memory then lives partly in its members' heads and partly in their knowledge of one another. When a key member leaves, the group loses both the facts and the map to them.
How does AI find research across documents?
Many search tools turn text into embeddings, which are lists of numbers that capture meaning. Texts with similar meaning land close together, so a question can match a document that uses different words. Google researcher Tomas Mikolov and colleagues popularized the idea for words in 2013 with the word2vec model. Modern systems extend the same approach to whole passages and documents.
Why do language models state things that are wrong?
A language model generates the most likely next piece of text, which is different from looking up a stored fact. A 2025 OpenAI paper by Adam Tauman Kalai and colleagues argued that models guess because standard tests reward a confident guess over an honest admission of uncertainty. Wrong answers produced this way are called hallucinations. Grounding a model in a company's own documents reduces them without removing them.
More from Lenny's Podcast: Product | Growth | Career
