WEBVTT

1
00:00:02.030 --> 00:00:04.070
If it can answer your question,

2
00:00:04.570 --> 00:00:07.310
that does not mean it has permission to

3
00:00:07.310 --> 00:00:08.330
roam the building.

4
00:00:09.450 --> 00:00:12.450
An answer is not an access badge.

5
00:00:13.750 --> 00:00:14.990
Start with the model.

6
00:00:15.470 --> 00:00:18.170
It takes input and produces output using

7
00:00:18.170 --> 00:00:20.250
patterns learned during training.

8
00:00:20.950 --> 00:00:23.170
Ask it to draft a plan and it

9
00:00:23.170 --> 00:00:24.110
gives you a response.

10
00:00:24.990 --> 00:00:26.710
Now the agent.

11
00:00:27.570 --> 00:00:29.150
People use that word differently.

12
00:00:29.150 --> 00:00:33.970
Here we mean a larger software system that uses a

13
00:00:33.970 --> 00:00:35.910
model to work toward a goal.

14
00:00:36.770 --> 00:00:39.970
It can choose a step, use a tool, check the

15
00:00:39.970 --> 00:00:42.490
result and decide what to do next.

16
00:00:43.510 --> 00:00:45.550
The model helps guide those decisions.

17
00:00:46.470 --> 00:00:48.570
The surrounding software runs the tools.

18
00:00:49.530 --> 00:00:52.950
Think of the model as a brilliant consultant

19
00:00:52.950 --> 00:00:54.470
in a locked room.

20
00:00:54.470 --> 00:00:58.790
Building an agent is like adding a work order, a

21
00:00:58.790 --> 00:01:01.470
toolbox and an access badge.

22
00:01:02.530 --> 00:01:05.230
Which seems like an excellent time to read

23
00:01:05.230 --> 00:01:06.490
the change ticket.

24
00:01:07.330 --> 00:01:08.510
But remember,

25
00:01:09.050 --> 00:01:10.150
that is an analogy.

26
00:01:10.730 --> 00:01:13.670
The model did not grow ambitions or a

27
00:01:13.670 --> 00:01:14.850
secret career plan.

28
00:01:15.490 --> 00:01:17.090
People built the system.

29
00:01:17.330 --> 00:01:19.450
They chose the tools and granted the access.

30
00:01:19.450 --> 00:01:23.550
They also need to decide when it asks for approval

31
00:01:23.550 --> 00:01:24.990
and when it stops.

32
00:01:26.030 --> 00:01:28.710
At work, a model might draft a plan

33
00:01:28.710 --> 00:01:30.350
for a software update.

34
00:01:31.110 --> 00:01:35.450
An agent might also check the system, request approval,

35
00:01:36.010 --> 00:01:38.570
run the update and inspect the result.

36
00:01:39.250 --> 00:01:42.190
That depends on which tools and permissions

37
00:01:42.190 --> 00:01:44.510
the surrounding system provides.

38
00:01:45.990 --> 00:01:47.930
So when somebody says,

39
00:01:48.590 --> 00:01:50.490
are we buying an AI agent?

40
00:01:51.090 --> 00:01:51.670
ask:

41
00:01:52.690 --> 00:01:55.690
Can it only answer or can it act?

42
00:01:56.470 --> 00:01:58.190
Which tools can it use?

43
00:01:58.990 --> 00:02:00.630
What can it access?

44
00:02:01.630 --> 00:02:03.810
How many steps can it take?

45
00:02:04.890 --> 00:02:06.690
Where does a human approve?

46
00:02:07.890 --> 00:02:09.150
And what stops it?

47
00:02:11.180 --> 00:02:13.300
A model may be brilliant.

48
00:02:14.100 --> 00:02:17.360
The system around it determines the blast radius.

49
00:02:18.080 --> 00:02:20.040
How far a mistake can reach.

50
00:02:21.060 --> 00:02:24.120
Next signal, what is a token?

51
00:02:24.980 --> 00:02:28.020
And why are we spending them like arcade coins?

52
00:02:29.060 --> 00:02:30.340
Until then,

53
00:02:30.920 --> 00:02:32.460
understand the machine,

54
00:02:33.160 --> 00:02:37.460
question the story and please read the change

55
00:02:37.460 --> 00:02:40.160
ticket before energizing.