Korva is a scroll of the internet’s best ideas.
Some work is never finished
Tasks with an end
Some workflows are sufficiency-based: you do the task and it's done. Updating a CRM is the example, since there isn't much room to get better at it. For those, asking whether a model can do it or not makes sense.
Tasks with no ceiling
A second class shouldn't be thought of in binary terms at all. Legal arguments and, to an extent, medical advice can always be better, so a percentage of tasks completed is the wrong framing, and the better one is continuous, uncapped rewards.
Security as the case
Cybersecurity is the clearest case. It is competitive, a cat-and-mouse game between offensive and defensive capabilities, so it works like an AlphaGo-type situation with uncapped rewards and a field that is constantly moving. There will never be a 90 percent line in security, because the goalposts always move. Demand for data on cyber defense is growing very rapidly, and the data types are uncapped, evolving, and adversarial.
Why does cybersecurity never reach a final score?
Attackers adapt to every defense, so the target keeps moving. Biologist Leigh Van Valen proposed the Red Queen hypothesis in 1973 after finding that the chance of a species going extinct stayed roughly constant over time, suggesting species must keep evolving just to survive. The name comes from Lewis Carroll's Through the Looking-Glass, where the Red Queen must run to stay in place. Security researchers borrowed the image for the arms race between attackers and defenders.
How did AlphaGo get uncapped rewards?
AlphaGo Zero, reported by DeepMind in 2017, learned Go purely by playing against itself. Each version faced an opponent exactly as strong as itself, so the challenge grew as it improved and had no fixed ceiling. After three days of training it beat the earlier version that had defeated Lee Sedol, 100 games to 0.
How are open-ended answers graded at scale?
Standardized tests such as the AP exams use trained readers who score essays against a written rubric, awarding points along a scale. Readers calibrate by scoring sample essays together before grading, which keeps scores consistent across thousands of papers. Rubrics break quality into parts such as argument, evidence, and organization. Machine learning projects borrow the same structure, with graders scoring model answers against a list of criteria.
More from 20VC with Harry Stebbings
