Why your boss isn't worried about AI – "can't you just turn it off?"
17 hours ago
- #Software Misconceptions
- #Expert-Novice Gap
- #AI Safety
- Public understanding of software bugs is misapplied to AI, leading to incorrect assumptions.
- AI vulnerabilities stem from training data issues, not code mistakes, making them harder to diagnose and fix.
- Unlike regular software, AI bugs cannot be pinpointed or guaranteed to stay fixed after retraining.
- AI behavior is unpredictable with slight input changes, unlike deterministic regular software.
- AI capabilities and behaviors cannot be fully specified or guaranteed before training.
- Experts and novices in AI have unspoken, differing assumptions, causing communication gaps.
- AI safety requires recognizing that AI cannot simply be 'patched' like regular software.