🚑 Why AI Fails in Medical Emergencies

📅 made 2026-10-12 · ⏱ 108 s · ✅ 22% quoted · ✏️ 0% rewritten · ❌ 9% corrected · of 23 lines

The ungrounded maximum-likelihood estimator: right 95% of the time, and the 5% is the one that kills.

❌ What the errata show

The strongest read in the stack: it reached the maximum-likelihood estimator on its own. The two errors are the usual pair — 99.85 % sold as savings, and a limit on the model that we never set.

❌ Errata, line by line

“and, get this, it actually slashes token burn costs by 99.85%.” 01:31Same: rung 2 against a counterfactual, never a cost cut.sells rung 2 (99.85 %) as token-bill savings · settled by C563b
“Ultimately, until generative AI is restricted by strict operational boundaries,” 01:37We place what the model did against what was declared before it ran; the halt is the deployer's to wire.makes the tape a brake: kill switch, failed build, forced boundary · settled by CLAUDE.md TAPE, NOT BRAKE, NOT ENGINE

✅ Said right, quoted as said

“I mean, imagine calling 911 and getting an automated meditation guide.” 00:33
“The AI is basically an ungrounded maximum-likelihood estimator,” 00:46
“Sure, the AI might be statistically right 95% of the time,” 00:59
“but if that remaining 5% represents the fat tail that actually kills people,” 01:03
“isn't being mostly right totally useless?” 01:07

📝 The transcript, word for word with timestamps · mdx seed: “The Automated Meditation Guide” ★★★★★