AI catalysts and bubble signals

1

00:00:04,371 --> 00:00:08,775

So one major unlock this year is the emergence of

2

00:00:08,775 --> 00:00:12,879

reinforcement learning and of reasoning in AI.

3

00:00:12,879 --> 00:00:16,783

Effectively, these techniques solve for the reliability issue and

4

00:00:16,783 --> 00:00:20,286

hallucinations, which is the biggest drawback of AI.

5

00:00:20,286 --> 00:00:23,556

So if you think about it as a consumer, just in terms of trust, if your model

6

00:00:23,556 --> 00:00:26,393

is hallucinating, it's a lot tougher to trust that model.

7

00:00:26,393 --> 00:00:30,397

And that's even amplified for business users where they can't as a

8

00:00:30,397 --> 00:00:34,667

business give wrong answers and have hallucinations in their business workflow.

9

00:00:34,667 --> 00:00:39,305

And so really what we've seen this year is that

10

00:00:39,305 --> 00:00:42,842

reasoning and that RL has gotten a lot better.

11

00:00:42,842 --> 00:00:46,813

And the anecdote I like to point to, to explain what reasoning and RL really

12

00:00:46,813 --> 00:00:50,283

looks like is imagine two scenarios.

13

00:00:50,283 --> 00:00:53,953

One is you have a math test and all you've done is you've skimmed the math

14

00:00:53,953 --> 00:00:57,724

textbook. That's gonna be a hard test for you to do.

15

00:00:57,724 --> 00:01:01,461

The other example, which is closer to RL.

16

00:01:01,461 --> 00:01:05,732

And to reasoning models is imagine you've skimmed the math textbook and

17

00:01:05,732 --> 00:01:09,469

then you've done all of the practise questions a couple times at the end of

18

00:01:09,469 --> 00:01:12,005

every chapter and then, you take the test.

19

00:01:12,005 --> 00:01:15,275

Your answers are probably going to be a lot better for that test and more

20

00:01:15,275 --> 00:01:18,845

reliable and you'll feel better about the output and whoever is grading that

21

00:01:18,845 --> 00:01:21,548

test is probably going to feel better about that output as well.

22

00:01:21,548 --> 00:01:25,585

That's the unlock and that's the opportunity that we

23

00:01:25,585 --> 00:01:29,422

have with RL and reasoning in models.

24

00:01:29,422 --> 00:01:31,658

And it really gets at that heart of.

25

00:01:31,658 --> 00:01:36,229

Um, hallucination, reliability that's held models back in terms of adoption.

26

00:01:36,262 --> 00:01:38,865

And you've probably even seen this as a consumer.

27

00:01:38,865 --> 00:01:42,735

The other thing I'd point to just in terms of optimism around use cases that

28

00:01:42,735 --> 00:01:46,639

actually the most revenue generative use case today of AI is ranking

29

00:01:46,639 --> 00:01:50,577

algorithms. So things like your Google search and your Facebook feed, those

30

00:01:50,577 --> 00:01:54,581

are all using AI it now it's not next token prediction AI, but

31

00:01:54,581 --> 00:01:58,818

it is generative AI. Um, and so that is a really revenue generatve

32

00:01:58,818 --> 00:02:03,756

use case. And so there is real spend on the back of AI.

33

00:02:03,756 --> 00:02:07,794

The other thing that I think is important is to recognise

34

00:02:07,794 --> 00:02:11,397

if we're in a bubble or not. I think the catalyst to a bubble would be

35

00:02:11,397 --> 00:02:15,602

infrastructure over investment and a lack of monetizable use

36

00:02:15,602 --> 00:02:20,240

cases. Really, as the model improves and gets better and more reliable,

37

00:02:20,240 --> 00:02:24,210

there isn't any new revenue generative use cases, that's how you get

38

00:02:24,210 --> 00:02:26,546

a bubble, that's what's gonna be a problem.

39

00:02:26,546 --> 00:02:30,550

Especially given, unlike fibre in the ground, in

40

00:02:30,550 --> 00:02:33,953

the telco cycle, GPUs have a five-year lifespan.

41

00:02:33,953 --> 00:02:36,389

Fibre is good for decades.

42

00:02:36,389 --> 00:02:38,992

And after five years, it needs to be refreshed.

43

00:02:38,992 --> 00:02:43,029

So if we don't have that revenue generative use case emerge within five

44

00:02:43,029 --> 00:02:46,900

years you've got to effectively redo much of the investment.

45

00:02:46,900 --> 00:02:51,137

So I think that's how you could see a bubble is a lot of investment and revenue

46

00:02:51,137 --> 00:02:54,807

generatives use cases actually continue to take longer than we expect.

47

00:02:54,807 --> 00:02:58,745

The other... Kind of dark horse candidate that I think is interesting to

48

00:02:58,745 --> 00:03:00,747

talk about is algorithmic improvement.

49

00:03:00,747 --> 00:03:03,416

So today AI is very inefficient.

50

00:03:03,416 --> 00:03:07,120

If you think about it, the human brain, and let's use the use case of driving a

51

00:03:07,120 --> 00:03:11,424

car. The human brain consumes about 20 watts.

52

00:03:11,424 --> 00:03:14,294

And I looked it up in Ontario for driver's ed.

53

00:03:14,294 --> 00:03:18,431

You need 10 hours behind the wheel to get your licence.

54

00:03:18,431 --> 00:03:21,768

If you just compare that with Waymo.

55

00:03:21,768 --> 00:03:25,705

Waymo has tens of billions of simulated road hours to get to where it

56

00:03:25,705 --> 00:03:29,776

is today and it's still not broadly deployed and it uses many

57

00:03:29,776 --> 00:03:33,746

many kilowatts per car to get there or even

58

00:03:33,746 --> 00:03:37,650

if you just compare it a gigawatt data centre is like 50 million brains and so

59

00:03:37,650 --> 00:03:41,888

if we do have a real algorithmic improvement and unlock it actually

60

00:03:41,888 --> 00:03:46,226

could drive demand for gigawat

61

00:03:46,226 --> 00:03:50,196

data centres and lots of ops lower because the algorithms just get

62

00:03:50,196 --> 00:03:54,701

better. So that's kind of, you know, how I think about AI implementation

63

00:03:54,701 --> 00:03:57,303

and use cases and the potential for a bubble.