WEBVTT

1
00:00:00.000 --> 00:00:01.560
You think you understand AI.

2
00:00:01.780 --> 00:00:04.540
Almost nobody does, and the
reason isn't intelligence.

3
00:00:04.900 --> 00:00:05.540
It's your brain.

4
00:00:06.920 --> 00:00:09.740
Your brain runs a shortcut:
easy to read feels true.

5
00:00:09.900 --> 00:00:14.140
That's processing fluency, and a language
model produces the most fluent text ever

6
00:00:14.160 --> 00:00:17.800
written. So your truth detector
fires before you check a single fact.

7
00:00:18.920 --> 00:00:20.800
A language model doesn't look up facts.

8
00:00:21.100 --> 00:00:26.480
It ranks the next token, the most plausible
continuation built from patterns in human text.

9
00:00:27.300 --> 00:00:31.900
To do that, it compresses: trillions of
words squeezed into a few hundred billion

10
00:00:31.920 --> 00:00:35.400
numbers. Compression keeps
patterns and discards what's rare,

11
00:00:35.640 --> 00:00:37.880
and truth usually lives in the rare parts.

12
00:00:39.160 --> 00:00:40.580
It can't tell you what it threw away.

13
00:00:40.860 --> 00:00:44.520
No model has a window into its own
ignorance, so when it doesn't know,

14
00:00:44.560 --> 00:00:47.880
it doesn't hesitate. It hands you the
most plausible sentence available:

15
00:00:48.260 --> 00:00:50.050
confident, fluent, wrong.

16
00:00:51.280 --> 00:00:52.360
Now, the deeper problem.

17
00:00:52.580 --> 00:00:55.440
We say the model wants,
thinks, understands, lies.

18
00:00:55.460 --> 00:00:57.620
Each of those words smuggles in a human.

19
00:00:57.640 --> 00:00:59.600
Your brain evolved to read minds.

20
00:00:59.620 --> 00:01:02.920
It finds intention in a
shadow, in a face on Mars,

21
00:01:02.940 --> 00:01:04.340
in a matrix of numbers.

22
00:01:04.360 --> 00:01:09.720
So we get two ways to be wrong: treating
a statistical engine as an oracle,

23
00:01:09.740 --> 00:01:13.460
or calling it “just autocomplete” and
dismissing what it can actually do.

24
00:01:13.480 --> 00:01:16.400
Same mistake: explaining a
new thing with an old word.

25
00:01:16.420 --> 00:01:19.100
Drop the question, “Does it understand?”

26
00:01:19.120 --> 00:01:21.460
Ask instead: What does it do reliably?

27
00:01:21.480 --> 00:01:23.580
Where does it fail? And how would I know?

28
00:01:23.600 --> 00:01:26.860
Read it like an overconfident colleague
who never says “I don't know.”

29
00:01:27.320 --> 00:01:29.320
Verify everything. Keep
your name on the work.

30
00:01:30.520 --> 00:01:34.600
That's the real skill now: not using AI,
but knowing exactly what you're holding.

31
00:01:35.180 --> 00:01:36.620
Subscribe, and next time we go deeper.
