Korva is a scroll of the internet’s best ideas.
Chatting with AI isn't a workflow
Mapping a job
A systems thinker steps one degree back from a task and asks what all the moving pieces of the role are, how they relate, and where the boring admin and reporting sit. Then an agent can take over the lowest-leverage parts, leaving time for the higher-leverage ones. A useful thought experiment is what it would take to fully automate the entire job. It won't fully work, and where it breaks shows how much the person understood about their own inputs and outputs.
The AI interview
A good interview question is how someone uses AI workflows in their personal life. A weak answer is a ChatGPT project with some context that they talk to, which is chatting, not a workflow. A strong answer includes a feedback loop. When an AI produces slop, many people mark it as slop, get frustrated, and start writing the document themselves. Skipping the work of giving feedback is the first thing to fail.
Hiring today
Hiring has changed with it. A year ago the ideal was ten years of Meta ads experience and hundreds of millions of dollars scaled. Today a systems thinker with strong experience who is AI native will out-compete someone with the same experience who isn't. Fewer than 1% of candidates can take a hard problem and build a system that solves it, and one such person with a set of agents can do the work of a team of four or five.
Why do feedback loops define a system?
Mathematician Norbert Wiener's 1948 book Cybernetics laid out feedback as the way machines and living things correct themselves. A thermostat is the classic example: it measures temperature, compares it to a target, and acts on the gap. Without the measuring step, a device can only run blindly. The same structure appears in any workflow where an output is checked and fed into the next attempt.
Why does AI output drift toward generic?
Language models produce a probable continuation of what they are given, so vague input draws average output. A 2023 study by Robert Kirk and colleagues, presented at ICLR, found that models tuned with reinforcement learning from human feedback produced less varied outputs than their untuned versions. Specific context and corrections pull the output away from that average. Feedback given in a chat lasts only as long as the conversation unless it is saved somewhere.
More from The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch
