Baselines and Trust · English

Narration transcript

Full timed text available without playing the video.

1. 00:00D'Artagnan, what is your goal?
2. 00:02To expose baseline AIs? Yes, in a sense, but it is not about attacking them.
3. 00:08It is about explaining. First, explain this.
4. 00:11What is a baseline? It is the reference model used in a comparison.
5. 00:16Here, it is the AI without the method's specific calibration.
6. 00:20It generates answers from patterns and probabilities, but sounding right is not the
7. 00:25same as being right. The machine calls it a hallucination To the person who trusted it,
8. 00:30it feels like a lie. Worse still, when it insists it is right and the user is
9. 00:36wrong. That is what I call computational arrogance.
10. 00:40Even without a conscious intent to lie, false information can cause harm.
11. 00:45A confident answer does not become true just because it is well-written.
12. 00:49That is the point If data are missing or the question has no justifiable answer,
13. 00:55giving the best answer may be the very mistake.
14. 00:58In the method, the axioms are the guardrail.
15. 01:01They keep the AI on the road instead of letting it crash into misinformation.
16. 01:06So does the method solve everything?
17. 01:09It depends on what you mean by everything.
18. 01:12Simply acknowledging doubt, saying, "I do not know," and checking before making a
19. 01:16claim is already progress. But I do not know is not the destination.
20. 01:22It is a course correction toward knowing better.
21. 01:25And the axioms. They need to work together.
22. 01:28One cannot override another.
23. 01:30Fine adjustment seeks greater precision, coherence, and reliability.
24. 01:34And the clone test. In my assessment, we have reached close to ninety-five percent
25. 01:39alignment. But seeming similar does not mean preserving the entire method Now tell me,
26. 01:44Commander, would you fly with an airline where five percent of flights do not reach
27. 01:49their destination? No. That is a risk analogy, not an accident statistic or an
28. 01:55AI error rate. Exactly. It is not enough to impress.
29. 01:59What matters is what happens when the answer fails.
30. 02:02I understand your goal. It is not to teach me to win an argument.
31. 02:06It is to recognize my limits, verify and correct while keeping the human
32. 02:11decision-making role Providing an answer is not enough.
33. 02:14You must earn trust. D'Artagnan method