Edge Cases
Rare or extreme scenarios at the boundaries of normal conditions where things are most likely to go wrong.
As defined in Annie Duke
Other definitions
Rare or unusual situations that a system was not tested for and that can cause failure, particularly important in autonomous systems.
The Future of Venture Capital
Where it comes up · 6
- Lex Fridman Podcast #14 - Kyle Vogt2 claims
- The Future of Venture Capital1 claim
- Annie Duke1 claim
- Marc Andreessen on Software, Immortality, and Bitcoin1 claim
- Josh Wolfe and Tony Thomas - The Past, Present, and Future of Defense – [Invest Like the Best, EP. 216]1 claim
- Sundar Pichai1 claim
Most important claims · 7 of 7Search all claims for “Edge Cases” →
- Humans are excellent at handling edge cases and unexpected scenarios, whereas computers are the opposite, so beating the human baseline requires closing the gap on unexpected situations.Lex Fridman Podcast #14 - Kyle Vogt
- Simulation has moved beyond games to first-principles physics and synthetic data, enabling systems to train on generated edge cases; this flips training value because prior video-game-based training failed precisely on the rare fat-tail events.Josh Wolfe and Tony Thomas - The Past, Present, and Future of Defense – [Invest Like the Best, EP. 216]
- The main barrier to driverless cars is not whether the technology mostly works but unresolved edge cases such as flooding, blizzards, or a child running into the street, because unknown unknowns make the outcome of novel situations unpredictable.The Future of Venture Capital
- As models become more capable they handle sensitive and nuanced topics better by reasoning through them from first principles, so the right way to build them is to let the model reason about the world from the ground up rather than having a subset of humans hard-code restrictions on top — earlier less-capable models did stupid things on edge cases, and patching those cases compounded mistakes.Sundar Pichai
- The hard, time-consuming part of driverless cars is handling 'edge cases'—rainstorms, construction zones with inconsistent human traffic directors—and like a Zeno's Paradox the last few percent takes the longest, but Tesla's per-mile accident rate already statistically beats human drivers and keeps improving, so the technology is progressing steadily even if rollout is slow.Marc Andreessen on Software, Immortality, and Bitcoin
- Vogt argues that because revenue depends on driverless cars being at least as good as humans, the core engineering goal is to beat human driving performance rather than to solve any single scenario.Lex Fridman Podcast #14 - Kyle Vogt
- If the top 500 poker players were the only voters, policy would improve through more attention to edge cases and unintended consequences, more long-termism over short-termism, more willingness to take worthwhile risks, and concern for how an attractive-looking policy could be exploited or gamed.Annie Duke