WEBVTT

1
00:00:00.000 --> 00:00:02.550
The AI industry has three layers:

2
00:00:02.610 --> 00:00:05.161
models, chatbots and apps.

3
00:00:05.450 --> 00:00:08.480
Dominik Lukeš opened his Oxford lecture with that

4
00:00:08.540 --> 00:00:09.063
map.

5
00:00:09.250 --> 00:00:12.120
Models come in general and specialised kinds.

6
00:00:12.170 --> 00:00:14.570
Chatbots are either made by the labs or

7
00:00:14.620 --> 00:00:17.340
combine several models, and apps are either built

8
00:00:17.390 --> 00:00:19.548
around AI or have had it added.

9
00:00:19.600 --> 00:00:22.230
Underneath, every one of them is a model

10
00:00:22.290 --> 00:00:24.930
plus orchestration, which is ordinary

11
00:00:24.980 --> 00:00:25.916
software.

12
00:00:26.170 --> 00:00:29.400
The model is the source of intelligence,

13
00:00:29.450 --> 00:00:31.700
and you can only control it with a

14
00:00:31.750 --> 00:00:32.570
prompt.

15
00:00:33.200 --> 00:00:36.100
In a chatbot you prompt and the model

16
00:00:36.150 --> 00:00:37.019
responds.

17
00:00:37.200 --> 00:00:39.650
In an app you click, the app sends

18
00:00:39.700 --> 00:00:41.890
a prompt, the model sends a result and

19
00:00:41.940 --> 00:00:43.140
the app parses it.

20
00:00:43.190 --> 00:00:45.660
Progress comes from three sources:

21
00:00:46.720 --> 00:00:49.850
better models, new orchestration that takes advantage of

22
00:00:49.900 --> 00:00:52.540
them, and new use cases that people think

23
00:00:52.590 --> 00:00:54.331
of once they see what is possible.

24
00:00:54.381 --> 00:00:56.469
Three established leaders come first.

25
00:00:56.660 --> 00:00:59.670
OpenAI started this wave with ChatGPT,

26
00:01:00.450 --> 00:01:02.870
and now has the Codex coding agent, the

27
00:01:02.920 --> 00:01:05.150
GPT series of models, and audio and image

28
00:01:05.200 --> 00:01:05.960
models.

29
00:01:06.820 --> 00:01:09.840
Anthropic makes Claude and Claude Code and focuses

30
00:01:09.890 --> 00:01:11.940
on coding models, with no image or audio

31
00:01:11.990 --> 00:01:12.338
models.

32
00:01:12.390 --> 00:01:14.180
Google started the path to today's models with

33
00:01:14.230 --> 00:01:17.240
the twenty seventeen paper Attention Is All You

34
00:01:17.290 --> 00:01:17.928
Need.

35
00:01:18.140 --> 00:01:20.840
Dominik says its Gemini models are now behind

36
00:01:20.890 --> 00:01:23.300
the frontier in many ways, though it has

37
00:01:23.350 --> 00:01:26.520
the most complete offering of the three: audio,

38
00:01:26.570 --> 00:01:29.280
video and image models, Gemini

39
00:01:29.330 --> 00:01:32.730
Notebooks, the Antigravity coding agent,

40
00:01:32.780 --> 00:01:35.142
AI Studio and its own suite.

41
00:01:35.192 --> 00:01:38.000
A second group are returning after falling off

42
00:01:38.050 --> 00:01:39.142
people's radar.

43
00:01:39.300 --> 00:01:42.030
Meta, known for its open Llama models, now

44
00:01:42.080 --> 00:01:44.820
has the Muse models, a Muse assistant and

45
00:01:44.870 --> 00:01:46.134
Muse Code.

46
00:01:46.410 --> 00:01:49.440
SpaceX AI makes the Grok models,

47
00:01:49.490 --> 00:01:51.523
competent but not the best.

48
00:01:51.610 --> 00:01:54.500
Microsoft's MAI models are decent for voice and

49
00:01:54.550 --> 00:01:56.990
images, and its Copilot products make it, in

50
00:01:57.040 --> 00:01:58.737
Dominik's view, a follower.

51
00:01:58.790 --> 00:02:00.507
China has its own group.

52
00:02:00.560 --> 00:02:03.820
DeepSeek surprised everyone with a high-quality open

53
00:02:03.870 --> 00:02:05.890
model at the turn of twenty twenty-five, and

54
00:02:05.940 --> 00:02:09.200
Alibaba's Qwen, Z.ai, Xiaomi,

55
00:02:09.250 --> 00:02:12.360
Kimi from Moonshot and Minimax have

56
00:02:12.410 --> 00:02:13.393
followed.

57
00:02:13.560 --> 00:02:16.050
Most of their models are open, which means

58
00:02:16.100 --> 00:02:17.820
that in theory people can run them on

59
00:02:17.870 --> 00:02:18.738
their own machines.

60
00:02:18.790 --> 00:02:21.360
These labs now share one pattern.

61
00:02:21.410 --> 00:02:23.890
Each has a model family, a chatbot and

62
00:02:23.940 --> 00:02:27.260
a coding agent, free use with limits, and

63
00:02:27.310 --> 00:02:30.003
a subscription plan with credits.

64
00:02:30.210 --> 00:02:32.600
Dominik points out that there is virtually no

65
00:02:32.650 --> 00:02:34.463
unlimited use left.

66
00:02:34.520 --> 00:02:36.318
It is pay as you go.

67
00:02:36.370 --> 00:02:38.722
Many other companies build on these models.

68
00:02:38.772 --> 00:02:41.210
Notion began as a note-taking tool and is

69
00:02:41.260 --> 00:02:43.236
now putting AI first.

70
00:02:43.410 --> 00:02:46.620
Lovable, which did not exist before twenty twenty-four,

71
00:02:46.670 --> 00:02:49.710
builds web apps and websites from a request,

72
00:02:49.760 --> 00:02:51.747
and Gamma makes slides.

73
00:02:51.820 --> 00:02:54.780
Figma and Canva do design, and Manus, from

74
00:02:54.830 --> 00:02:56.699
China, helps with research.

75
00:02:56.750 --> 00:02:58.914
Established companies add models as features.

76
00:02:58.964 --> 00:03:02.270
Adobe, Salesforce, Cloudflare, Amazon, Apple and Oracle are

77
00:03:02.320 --> 00:03:03.058
doing it.

78
00:03:03.220 --> 00:03:06.220
Amazon summarises customer reviews, and Canvas, the learning

79
00:03:06.270 --> 00:03:09.000
platform Oxford uses, now has AI features.

80
00:03:09.080 --> 00:03:11.914
Everyone is adding AI, in every category.

81
00:03:11.964 --> 00:03:13.890
To see what people use, Dominik turned to

82
00:03:13.940 --> 00:03:16.550
the investment firm Andreessen Horowitz and its list

83
00:03:16.600 --> 00:03:18.255
of the top web apps.

84
00:03:18.500 --> 00:03:21.790
By monthly visits, ChatGPT is first,

85
00:03:21.840 --> 00:03:24.760
then Gemini, then Claude, which gained

86
00:03:24.810 --> 00:03:26.236
popularity this year.

87
00:03:26.330 --> 00:03:29.695
By revenue, OpenAI and Anthropic dominate.

88
00:03:29.830 --> 00:03:32.060
The revenue comes from payment data, not from

89
00:03:32.110 --> 00:03:32.722
the companies.

90
00:03:32.772 --> 00:03:34.338
He sorts the apps into six categories.

91
00:03:34.390 --> 00:03:36.202
Productivity has Notion and Microsoft.

92
00:03:36.252 --> 00:03:39.073
Multimedia has ElevenLabs, Speechify and Runway.

93
00:03:39.240 --> 00:03:41.846
Meeting transcription has Otter.ai and Granola.

94
00:03:42.030 --> 00:03:44.590
Research has Elicit, Consensus, Manus and

95
00:03:44.640 --> 00:03:45.462
Perplexity.

96
00:03:45.512 --> 00:03:48.440
Coding has Lovable, Replit and Cognition, and design

97
00:03:48.500 --> 00:03:50.330
has Figma, Canva and Gamma.

98
00:03:50.380 --> 00:03:53.280
Many of these apps may matter less, though

99
00:03:53.330 --> 00:03:55.354
he does not say they will go away.

100
00:03:55.570 --> 00:03:58.710
Apart from multimedia and transcription, much of

101
00:03:58.760 --> 00:04:01.750
what design, coding, research and productivity

102
00:04:01.800 --> 00:04:04.700
apps do is being built into chatbots.

103
00:04:05.470 --> 00:04:08.530
A chatbot asked for a decent PowerPoint used

104
00:04:08.580 --> 00:04:10.850
to fail, and since the start of this

105
00:04:10.900 --> 00:04:13.560
year it can do it, because models,

106
00:04:13.610 --> 00:04:16.328
orchestration and apps got good enough.

107
00:04:16.540 --> 00:04:18.747
The new category is agents.

108
00:04:18.950 --> 00:04:21.880
Dean Ball wrote in January: "By the end

109
00:04:21.930 --> 00:04:23.700
of this year, the least important thing you

110
00:04:23.750 --> 00:04:25.780
will be able to do with frontier AI

111
00:04:25.830 --> 00:04:28.840
systems will be getting chatbots to answer

112
00:04:28.890 --> 00:04:32.075
questions." Dominik says that has already come true.

113
00:04:32.130 --> 00:04:35.220
Each leader now pairs its web chatbot with

114
00:04:35.270 --> 00:04:38.040
a desktop agent: ChatGPT with

115
00:04:38.090 --> 00:04:41.360
Codex, Claude with Claude Code, and

116
00:04:41.410 --> 00:04:43.263
Gemini with Antigravity.

117
00:04:43.313 --> 00:04:45.910
The desktop version uses tools in a local

118
00:04:45.960 --> 00:04:48.321
sandbox and works on local files.

119
00:04:48.371 --> 00:04:50.880
In March Jensen Huang said: "Every company in

120
00:04:50.930 --> 00:04:53.360
the world today needs to have an OpenClaw

121
00:04:53.410 --> 00:04:56.392
strategy, an agentic system strategy.

122
00:04:56.450 --> 00:04:59.750
This is the new computer." OpenClaw is

123
00:04:59.810 --> 00:05:02.310
an always-on assistant that runs on your own

124
00:05:02.360 --> 00:05:05.190
computer and talks to you on WhatsApp.

125
00:05:05.880 --> 00:05:08.790
Microsoft announced Autopilot on the twenty-ninth of

126
00:05:08.840 --> 00:05:11.750
June, SpaceXAI launched Grok Bot on the

127
00:05:11.800 --> 00:05:14.950
eleventh of August, Meta's Muse followed on the

128
00:05:15.000 --> 00:05:17.830
eighth of September and OpenAI's dots on the

129
00:05:17.880 --> 00:05:19.506
twenty-ninth of September.

130
00:05:19.560 --> 00:05:21.916
Several of these are limited to the US.

131
00:05:21.970 --> 00:05:23.947
Agents change the price.

132
00:05:24.190 --> 00:05:27.650
In twenty twenty-three, ChatGPT Plus cost twenty dollars

133
00:05:27.770 --> 00:05:28.674
a month.

134
00:05:28.800 --> 00:05:31.850
In twenty twenty-six, full agent use costs a

135
00:05:31.900 --> 00:05:34.236
hundred to five hundred dollars a month.

136
00:05:34.520 --> 00:05:36.850
People are paying more as they see the

137
00:05:36.900 --> 00:05:37.112
value.

138
00:05:37.162 --> 00:05:40.090
Paid use is growing from a small base.

139
00:05:40.610 --> 00:05:43.830
In YipitData's US panel, the share of eligible

140
00:05:43.880 --> 00:05:47.180
consumers with a paid personal ChatGPT, Gemini

141
00:05:47.230 --> 00:05:49.890
or Claude subscription rose from two point one

142
00:05:49.940 --> 00:05:52.480
per cent in August twenty twenty-five to four

143
00:05:52.530 --> 00:05:55.440
point five per cent in August twenty

144
00:05:55.490 --> 00:05:56.620
twenty-six.

145
00:05:57.750 --> 00:06:00.170
The top one per cent of payers account

146
00:06:00.220 --> 00:06:03.000
for nineteen point five per cent of spending,

147
00:06:03.050 --> 00:06:05.470
which is more than the bottom half's sixteen

148
00:06:05.520 --> 00:06:06.300
point six.

149
00:06:06.350 --> 00:06:09.370
Dominik recommends a video in which two photographers

150
00:06:09.420 --> 00:06:12.290
use AI to manage their business and build

151
00:06:12.340 --> 00:06:12.993
apps.

152
00:06:13.200 --> 00:06:15.300
One of them pays about a thousand dollars

153
00:06:15.650 --> 00:06:18.830
a month, rotating five two-hundred-dollar

154
00:06:18.880 --> 00:06:21.900
subscriptions, because each one has credit limits and

155
00:06:21.950 --> 00:06:22.701
he runs out.

156
00:06:22.751 --> 00:06:25.740
Dominik expects many of these apps to persist,

157
00:06:25.790 --> 00:06:28.170
but to change in what they do and

158
00:06:28.220 --> 00:06:31.110
how they work, as tools such as Claude

159
00:06:31.160 --> 00:06:33.334
Code and Codex do more.

160
00:06:33.384 --> 00:06:36.370
He is sure agents will become the normal

161
00:06:36.420 --> 00:06:39.610
way of using AI, whether in months

162
00:06:39.660 --> 00:06:40.821
or in a year.

163
00:06:40.880 --> 00:06:43.970
The next sessions cover AI in research, the

164
00:06:44.020 --> 00:06:47.087
model landscape and AI agents.