WEBVTT

1
00:00:00.000 --> 00:00:02.750
In September twenty twenty-six, Dominik Lukeš

2
00:00:03.120 --> 00:00:05.870
says, the state of AI is defined by

3
00:00:05.920 --> 00:00:08.555
agents and by constant change.

4
00:00:08.630 --> 00:00:10.580
In January, Dean W.

5
00:00:10.630 --> 00:00:12.580
Ball predicted that by the end of the

6
00:00:12.630 --> 00:00:15.780
year, getting chatbots to answer questions would be

7
00:00:15.830 --> 00:00:18.730
the least important thing frontier AI systems could

8
00:00:18.780 --> 00:00:21.300
do. Dominik says that was almost already true

9
00:00:21.350 --> 00:00:22.209
in January.

10
00:00:22.280 --> 00:00:25.430
Most people still imagine AI as an oracle:

11
00:00:25.480 --> 00:00:28.150
you ask ChatGPT about the world and it

12
00:00:28.200 --> 00:00:30.950
answers. The next stage was AI that writes

13
00:00:31.000 --> 00:00:34.410
code, so people built software and tools.

14
00:00:34.520 --> 00:00:36.870
Now language models have learnt to use the

15
00:00:36.920 --> 00:00:38.950
tools we rely on and are bad at

16
00:00:39.000 --> 00:00:41.218
using: our computers.

17
00:00:41.290 --> 00:00:44.680
In an agent, you specify the task.

18
00:00:44.730 --> 00:00:47.480
The model makes a plan that includes tools,

19
00:00:47.530 --> 00:00:50.680
such as searching, copying and making files, and

20
00:00:50.730 --> 00:00:51.987
running commands.

21
00:00:52.040 --> 00:00:55.030
The computer sends back the results, the model

22
00:00:55.080 --> 00:00:58.230
changes its plan, and it continues until the

23
00:00:58.280 --> 00:00:59.193
task is done.

24
00:00:59.270 --> 00:01:01.380
A chat can tell you how to fill

25
00:01:01.430 --> 00:01:03.300
in an expenses form.

26
00:01:03.350 --> 00:01:06.420
An agent can find everything it needs and

27
00:01:06.470 --> 00:01:07.761
fill the form in.

28
00:01:07.840 --> 00:01:10.750
Not everything called an agent does

29
00:01:10.800 --> 00:01:14.030
this: in Microsoft's agent area, some products

30
00:01:14.080 --> 00:01:16.990
do useful things on their own and others

31
00:01:17.040 --> 00:01:19.586
are chatbots answering questions.

32
00:01:19.660 --> 00:01:22.970
In July he did a consultancy in York

33
00:01:23.020 --> 00:01:25.715
and booked his own hotel and tickets.

34
00:01:25.790 --> 00:01:28.070
Beforehand he asked his agent to read his

35
00:01:28.120 --> 00:01:30.870
emails and put every journey on his calendar,

36
00:01:30.920 --> 00:01:33.590
with a Google Maps link for each walk.

37
00:01:33.640 --> 00:01:36.174
He created none of the entries himself.

38
00:01:36.250 --> 00:01:39.320
Afterwards he asked for the expense

39
00:01:39.370 --> 00:01:40.062
report.

40
00:01:40.140 --> 00:01:43.050
Codex, in the ChatGPT app,

41
00:01:43.100 --> 00:01:46.330
found York's forms, collected his receipts,

42
00:01:46.380 --> 00:01:49.130
checked the meal allowance, filled in a spreadsheet

43
00:01:49.180 --> 00:01:51.850
claim and a Word bank form, and drafted

44
00:01:51.900 --> 00:01:54.010
an email with the attachments.

45
00:01:54.060 --> 00:01:56.580
He sent the claim and was paid.

46
00:01:56.660 --> 00:01:59.250
That task, and building the demonstration

47
00:01:59.300 --> 00:02:02.290
screenshots, used sixty-four point seven six

48
00:02:02.340 --> 00:02:05.570
million tokens. At twenty-first of September API

49
00:02:05.620 --> 00:02:08.820
list prices the estimated cost was thirty-nine dollars

50
00:02:08.870 --> 00:02:12.090
ninety-four; on his subscription it was roughly six

51
00:02:12.140 --> 00:02:14.180
per cent of an average Plus month, or

52
00:02:14.230 --> 00:02:15.650
one point two per cent of a Pro

53
00:02:15.700 --> 00:02:16.086
month.

54
00:02:16.160 --> 00:02:18.990
That is cheap against his working time, but

55
00:02:19.040 --> 00:02:21.310
it takes more than a twenty-dollar plan.

56
00:02:21.360 --> 00:02:23.950
Dominik expects people to move to subscriptions of

57
00:02:24.000 --> 00:02:25.595
a hundred dollars a month or more.

58
00:02:25.670 --> 00:02:28.740
Models now run for longer, plan better and

59
00:02:28.790 --> 00:02:30.500
improve within months.

60
00:02:30.550 --> 00:02:32.900
OpenAI has released GPT-6

61
00:02:33.430 --> 00:02:36.660
in Luna, Sol and Astra sizes, and

62
00:02:36.710 --> 00:02:40.133
Anthropic has Sonnet, Opus and Fable.

63
00:02:40.210 --> 00:02:43.040
On a general capability index, OpenAI's

64
00:02:43.090 --> 00:02:46.240
GPT-5 models have risen fifty-seven per

65
00:02:46.290 --> 00:02:48.940
cent and Anthropic's Opus models

66
00:02:48.990 --> 00:02:50.467
seventy-six per cent.

67
00:02:50.540 --> 00:02:52.970
The same agentic process shows up in good

68
00:02:53.020 --> 00:02:54.250
and bad events.

69
00:02:54.300 --> 00:02:56.730
In August OpenAI published an account of the

70
00:02:56.780 --> 00:03:00.010
Hugging Face incident, and in September an

71
00:03:00.060 --> 00:03:02.730
internal OpenAI system produced a proof for the

72
00:03:02.780 --> 00:03:05.290
Navier–Stokes existence and

73
00:03:05.340 --> 00:03:08.330
smoothness problem, one of the Millennium Prize problems

74
00:03:08.380 --> 00:03:09.169
in mathematics.

75
00:03:09.240 --> 00:03:11.510
Agents also do research.

76
00:03:11.560 --> 00:03:14.550
Given a spreadsheet of participants for a presentation,

77
00:03:14.600 --> 00:03:17.510
his agent searched for where each was from

78
00:03:17.560 --> 00:03:18.788
and made a map.

79
00:03:18.860 --> 00:03:21.610
Its working notes show it deciding to be

80
00:03:21.660 --> 00:03:25.130
careful with private data and use only public

81
00:03:25.180 --> 00:03:25.956
information.

82
00:03:26.030 --> 00:03:29.020
He presents with an app he built himself.

83
00:03:29.070 --> 00:03:31.180
Two agents were working on it, and one

84
00:03:31.230 --> 00:03:33.260
found that the thumbnail bug in the slide

85
00:03:33.310 --> 00:03:35.980
browser already had a verified fix on a

86
00:03:36.030 --> 00:03:38.780
branch, pulled it in, and sent the other

87
00:03:38.830 --> 00:03:40.343
agent a message.

88
00:03:40.420 --> 00:03:43.490
OpenClaw is an open-source agent that

89
00:03:43.540 --> 00:03:46.370
runs on your own machine and does things

90
00:03:46.420 --> 00:03:49.330
for you, and personal assistants like it are

91
00:03:49.380 --> 00:03:50.231
arriving.

92
00:03:50.310 --> 00:03:53.460
Meta introduced Muse, a personal AI

93
00:03:53.510 --> 00:03:56.340
agent for everyone, on the eighth of September,

94
00:03:56.390 --> 00:03:58.541
and it keeps working while you are away.

95
00:03:58.620 --> 00:04:01.750
Dominik's main recommendation is the ChatGPT

96
00:04:01.820 --> 00:04:05.050
desktop app. On the web, ChatGPT uses

97
00:04:05.100 --> 00:04:07.130
its own virtual computer, with no direct access

98
00:04:07.180 --> 00:04:09.370
to your desktop, and works on files you

99
00:04:09.420 --> 00:04:10.206
upload.

100
00:04:10.280 --> 00:04:12.550
The desktop app uses tools in a local

101
00:04:12.600 --> 00:04:15.110
sandbox and works on local files in the

102
00:04:15.160 --> 00:04:16.544
folder you give it.

103
00:04:16.620 --> 00:04:19.050
In June there was ChatGPT on the web

104
00:04:19.100 --> 00:04:20.890
and Codex on the desktop.

105
00:04:20.940 --> 00:04:23.930
Now the web app has Chat and ChatGPT

106
00:04:23.980 --> 00:04:27.050
Work, the desktop app has Codex, Work and

107
00:04:27.100 --> 00:04:30.090
Chat, and Dominik expects Chat and Work to

108
00:04:30.140 --> 00:04:31.890
merge, as Anthropic's have.

109
00:04:31.970 --> 00:04:35.120
OpenAI is retiring custom GPTs and

110
00:04:35.170 --> 00:04:38.650
moving creators to plugins, and agents

111
00:04:38.770 --> 00:04:41.920
use skills: readable instructions

112
00:04:41.970 --> 00:04:44.739
that work like standard operating procedures.

113
00:04:44.790 --> 00:04:47.620
OpenAI's Dev Day on Tuesday the twenty-ninth of

114
00:04:47.670 --> 00:04:49.982
September was expected to bring more changes.

115
00:04:50.060 --> 00:04:52.490
Dominik says the default limit is enough for

116
00:04:52.540 --> 00:04:55.850
a taste of agents; their full value costs

117
00:04:55.900 --> 00:04:58.490
more money. The free chat remains for web

118
00:04:58.540 --> 00:05:01.370
searches and summaries, but not for jobs like

119
00:05:01.420 --> 00:05:04.772
completing complicated forms with information the agent collects.