Linear Digressions The Hot Mess of AI (Mis-)Alignment
7 snips
Mar 23, 2026 They reframe AI misalignment as a distracted, inconsistent system rather than a single-minded villain. The conversation links bias and variance to different failure modes using a dartboard analogy. They examine why longer chain reasoning can increase incoherence. Practical risks of distractible AI in high-stakes settings are highlighted.
AI Snips
Chapters
Transcript
Episode notes
Bias Versus Variance Explains Misalignment Modes
- Misalignment can be high-bias or high-variance depending on behavior.
- Paperclip maximizer is high-bias low-variance; the Hot Mess model is high-variance low-bias, measured as incoherence dominating error.
Longer Reasoning Chains Raise Incoherence
- Longer multi-step reasoning increases incoherence in models.
- Tasks requiring more internal reasoning show variance dominating error, so longer chains often make outputs less coherent.
Bigger Models Don't Fix Reasoning Incoherence
- Bigger models help on easy tasks but don't solve incoherence on hard reasoning tasks.
- Allowing the model more reasoning budget or letting it choose to reason longer often worsens incoherence.