1
00:00:06,000 --> 00:00:11,119
What's the difference between generative AI and agentic AI?

2
00:00:11,119 --> 00:00:16,399
Well, they're two distinct approaches to artificial intelligence, and I think we're all familiar

3
00:00:16,399 --> 00:00:21,920
with generative AI, things like chatbots and image generators and the like.

4
00:00:21,920 --> 00:00:28,639
And they are really fundamentally reactive systems.

5
00:00:28,639 --> 00:00:36,040
They wait for you to do something, specifically, they wait for you to prompt them.

6
00:00:36,040 --> 00:00:44,759
And once you prompt them, their job is to generate some kind of content based upon what

7
00:00:44,759 --> 00:00:46,520
you provided in the prompt.

8
00:00:46,520 --> 00:00:50,080
And they're using patterns they learned during training.

9
00:00:50,639 --> 00:00:58,560
Now, the things that it can generate, well, that might be some text, or it might be an

10
00:00:58,560 --> 00:01:06,720
image, or it might be a piece of code, or it might be some audio.

11
00:01:06,720 --> 00:01:10,720
These are all sorts of things that we can generate with generative AI.

12
00:01:10,720 --> 00:01:14,160
And they're essentially sophisticated pattern matching machines.

13
00:01:14,160 --> 00:01:19,319
They've learned the statistical relationships between words and between pixels and between

14
00:01:19,319 --> 00:01:20,319
sound waves.

15
00:01:20,319 --> 00:01:22,760
And they've learned that from massive data sets.

16
00:01:22,760 --> 00:01:28,160
So when you provide a prompt, gen AI predicts what should come next based on its training.

17
00:01:28,160 --> 00:01:31,319
But its work does end at generation.

18
00:01:31,319 --> 00:01:35,480
It doesn't take further steps without your input.

19
00:01:35,480 --> 00:01:42,319
Now agentic AI systems, by contrast, those are not reactive.

20
00:01:42,319 --> 00:01:47,120
They are proactive systems.

21
00:01:47,120 --> 00:01:52,879
Now like generative AI, they often start with a user prompt.

22
00:01:52,879 --> 00:02:00,959
But that prompt is then used to pursue goals through a series of actions.

23
00:02:00,959 --> 00:02:04,839
And an agentic system basically goes through a bit of a lifecycle.

24
00:02:04,839 --> 00:02:12,039
So the way this works is it kind of, first of all, perceives its environment, if you

25
00:02:12,039 --> 00:02:13,039
like.

26
00:02:13,039 --> 00:02:19,399
And once it's done that, it can decide an action to take.

27
00:02:19,399 --> 00:02:25,039
Once it's decided that action, it can then execute that action.

28
00:02:25,039 --> 00:02:31,199
And then once that action has been executed, it can kind of learn from the output.

29
00:02:31,199 --> 00:02:38,000
And then round and round we go, all with minimal human intervention.

30
00:02:38,000 --> 00:02:44,279
Now both of these AI approaches often share a common foundation.

31
00:02:44,279 --> 00:02:50,759
And that common foundation is large language models, or LLMs.

32
00:02:50,759 --> 00:02:55,919
LLMs serve as the backbone for chatbots.

33
00:02:55,919 --> 00:02:59,559
And yet there's actually other tools that are used for some of these other generative

34
00:02:59,559 --> 00:03:03,479
things, diffusion models typically for images and audio.

35
00:03:03,479 --> 00:03:05,800
But for chatbots, we use LLMs.

36
00:03:05,800 --> 00:03:11,479
And LLMs also provide the reasoning engine that powers agentic systems.

37
00:03:11,479 --> 00:03:17,399
But before we go any deeper into that, let's talk about some real world applications and

38
00:03:17,399 --> 00:03:18,919
use cases.

39
00:03:18,919 --> 00:03:22,720
Now maybe this doesn't put me in the best of lights, but I don't think I'm the only

40
00:03:22,720 --> 00:03:32,559
one using generative AI to help with the task of content creation, and especially creative

41
00:03:32,559 --> 00:03:35,020
content creation.

42
00:03:35,020 --> 00:03:40,979
Now before work this morning, and this is completely true, I used a chatbot to help

43
00:03:40,979 --> 00:03:45,740
write the next chapter of my Nelson DeMille fan fiction novel.

44
00:03:45,740 --> 00:03:50,460
And right now you're probably thinking how profoundly cool and absolutely non-nerdy this

45
00:03:50,460 --> 00:03:51,779
guy is.

46
00:03:51,779 --> 00:03:57,100
But for many of us, gen AI does help with daily tasks.

47
00:03:57,100 --> 00:04:02,940
Like let's consider how a YouTuber might use a generative AI system to review scripts and

48
00:04:02,940 --> 00:04:07,940
suggest thumbnail concepts, and maybe even generate background music.

49
00:04:07,940 --> 00:04:12,100
But at each step, there is a human.

50
00:04:12,100 --> 00:04:14,220
There is a human creator.

51
00:04:14,220 --> 00:04:19,940
And that human creator is looking at this generated content, and they are reviewing

52
00:04:19,940 --> 00:04:20,940
it.

53
00:04:20,940 --> 00:04:21,940
Check it's what they want.

54
00:04:21,940 --> 00:04:22,940
Probably isn't.

55
00:04:22,940 --> 00:04:25,579
So then they are refining it as well.

56
00:04:25,579 --> 00:04:30,500
And they are really going through and directing this whole process.

57
00:04:30,500 --> 00:04:35,220
The AI generates possibilities, but the human curates them.

58
00:04:35,220 --> 00:04:42,779
Now agentic AI, that kind of thrives in scenarios that require ongoing management and consist

59
00:04:42,779 --> 00:04:46,220
of multi-step processes.

60
00:04:46,220 --> 00:04:49,839
So not just one thing at a time.

61
00:04:49,839 --> 00:04:53,500
So consider a personal shopping agent.

62
00:04:53,500 --> 00:04:59,420
Given a product to purchase as input, it actively hunts for availability across platforms.

63
00:04:59,420 --> 00:05:04,100
It might monitor price fluctuations, it might handle checkout processes, and it might even

64
00:05:04,100 --> 00:05:05,739
coordinate delivery.

65
00:05:05,739 --> 00:05:11,420
Largely by itself, seeking input only from you, only when it's needed.

66
00:05:11,420 --> 00:05:14,059
But how does it do that?

67
00:05:14,059 --> 00:05:18,420
Well it turns out that the LLMs that are behind much of generative AI can also be used to

68
00:05:18,420 --> 00:05:22,720
provide reasoning capabilities to AI agents.

69
00:05:22,720 --> 00:05:28,880
So this essentially here, we're using gen-AI's ability to kind of think, in inverted commas

70
00:05:28,880 --> 00:05:32,000
there, and it's thinking through problems.

71
00:05:32,000 --> 00:05:33,899
And this has a name.

72
00:05:33,899 --> 00:05:40,440
It's called chain of thought reasoning.

73
00:05:40,440 --> 00:05:44,519
And this is what LLMs are so very good at.

74
00:05:44,519 --> 00:05:51,040
It's a process where the agent basically breaks down a complex task into smaller logical steps.

75
00:05:51,079 --> 00:05:54,000
I like how humans tackle difficult problems as well.

76
00:05:54,000 --> 00:05:55,679
So let's imagine one.

77
00:05:55,679 --> 00:06:01,239
Let's imagine that we want to have an agent that is planning a complex task, like organizing

78
00:06:01,239 --> 00:06:02,380
a conference.

79
00:06:02,380 --> 00:06:08,000
So what it's going to do is it's going to use gen-AI to generate an internal dialogue.

80
00:06:08,000 --> 00:06:09,600
And that dialogue might go something like this.

81
00:06:09,600 --> 00:06:14,480
It might say, first I need to understand the conference requirements, so the size, the

82
00:06:14,480 --> 00:06:16,519
duration, the budget, that sort of thing.

83
00:06:16,519 --> 00:06:20,720
Then I should research available venues matching those parameters.

84
00:06:20,720 --> 00:06:24,600
Then it might think, well, for those venues that meet those requirements, I now need to

85
00:06:24,600 --> 00:06:27,160
check availability and so on.

86
00:06:27,160 --> 00:06:32,880
It's effectively the agent really kind of talking to itself to explore the problem space

87
00:06:32,880 --> 00:06:34,640
before taking action.

88
00:06:34,640 --> 00:06:39,720
Gen-AI is basically the cognitive engine driving an agent's decision making.

89
00:06:39,720 --> 00:06:46,399
Now looking ahead, the most powerful AI systems probably won't be purely generative or purely

90
00:06:46,399 --> 00:06:47,399
agentic.

91
00:06:47,399 --> 00:06:50,359
They're going to be intelligent collaborators.

92
00:06:50,359 --> 00:06:56,119
They'll understand when to explore options through generation and when to commit to causes

93
00:06:56,119 --> 00:07:00,079
of action through agentic action.

94
00:07:00,079 --> 00:07:06,119
Like an agent that would know when to generate the next chapter of fan fiction so it's ready

95
00:07:06,119 --> 00:07:09,880
after, oh I don't know, a video shoot.

96
00:07:09,880 --> 00:07:11,959
Maybe it's ready right now.