WEBVTT

1
00:00:00.000 --> 00:00:02.750
Dominik Lukeš gave this lecture to Oxford

2
00:00:03.090 --> 00:00:05.960
students the day after a session on

3
00:00:06.000 --> 00:00:09.210
study methods. He told them they are

4
00:00:09.260 --> 00:00:12.290
at Oxford to change their brain and to

5
00:00:12.340 --> 00:00:13.420
change the world.

6
00:00:13.470 --> 00:00:15.640
He said AI is a good use when

7
00:00:15.690 --> 00:00:18.310
it makes the world better and their brain

8
00:00:18.360 --> 00:00:18.757
better.

9
00:00:18.810 --> 00:00:21.410
You cannot build muscle by sending somebody else

10
00:00:21.460 --> 00:00:21.992
to the gym.

11
00:00:22.060 --> 00:00:24.680
Students still do the heavy lifting of study,

12
00:00:24.730 --> 00:00:26.680
but with heavy weights it helps to have

13
00:00:26.730 --> 00:00:29.850
a spotter. AI can be one, and it

14
00:00:29.900 --> 00:00:30.831
is optional.

15
00:00:30.881 --> 00:00:33.580
He gave three rules: do not say you

16
00:00:33.630 --> 00:00:36.180
did something when AI did it, do not

17
00:00:36.230 --> 00:00:38.680
say you worked unaided when you had help,

18
00:00:38.730 --> 00:00:40.850
and do not use AI when told not

19
00:00:40.900 --> 00:00:41.114
to.

20
00:00:41.164 --> 00:00:43.530
Whether a use is help or a shortcut

21
00:00:43.580 --> 00:00:45.133
depends on your stage of learning.

22
00:00:45.183 --> 00:00:47.410
He set out five roles for AI.

23
00:00:47.680 --> 00:00:50.890
The intern or assistant takes over time-consuming

24
00:00:50.940 --> 00:00:52.730
and distracting tasks.

25
00:00:52.850 --> 00:00:56.060
The tutor reduces cognitive load by chunking

26
00:00:56.110 --> 00:00:56.875
and repetition.

27
00:00:56.930 --> 00:00:59.830
The consultant gives feedback on ideas.

28
00:00:59.880 --> 00:01:02.910
The translator makes understanding easier,

29
00:01:02.960 --> 00:01:04.750
academic language included.

30
00:01:04.800 --> 00:01:07.580
The toolmaker helps you make tools for yourself

31
00:01:07.630 --> 00:01:08.161
and others.

32
00:01:08.211 --> 00:01:11.690
ChatGPT, Gemini and NotebookLM

33
00:01:11.880 --> 00:01:14.990
run on a large language model, which

34
00:01:15.040 --> 00:01:17.942
takes in text and produces text.

35
00:01:17.992 --> 00:01:20.790
It has no hands, but the text can

36
00:01:20.840 --> 00:01:24.130
be code: when ChatGPT searches the web,

37
00:01:24.180 --> 00:01:27.050
it sends code to a computer asking it

38
00:01:27.100 --> 00:01:28.182
to run the search.

39
00:01:28.232 --> 00:01:31.080
The intern has many skills, not all

40
00:01:31.130 --> 00:01:33.250
reliable, so it needs supervision.

41
00:01:33.300 --> 00:01:36.560
Dominik pasted a book cover into ChatGPT

42
00:01:37.140 --> 00:01:39.590
and asked for background on the book, its

43
00:01:39.640 --> 00:01:41.956
author, reviews and similar arguments.

44
00:01:42.010 --> 00:01:44.310
One link was broken, but the item was

45
00:01:44.360 --> 00:01:47.310
real. The same request in Claude also made

46
00:01:47.360 --> 00:01:48.805
its searches in the background.

47
00:01:48.855 --> 00:01:51.800
Dominik gave ChatGPT a long description of his

48
00:01:51.850 --> 00:01:54.050
week, asked it to break the week into

49
00:01:54.100 --> 00:01:56.454
something manageable, then asked for a calendar file.

50
00:01:56.504 --> 00:01:58.500
It imported straight into his calendar.

51
00:01:58.550 --> 00:02:00.670
Work like this can come down from two

52
00:02:00.720 --> 00:02:03.670
hours to thirty minutes, not to one

53
00:02:03.720 --> 00:02:06.930
minute, and ChatGPT's contribution should never be a

54
00:02:06.980 --> 00:02:07.541
hundred per cent.

55
00:02:07.610 --> 00:02:10.730
Dominik asked ChatGPT, as a peer

56
00:02:10.780 --> 00:02:13.740
reviewer, to critique a tweet he planned to

57
00:02:13.790 --> 00:02:16.900
post. It said the tweet risked being read

58
00:02:16.950 --> 00:02:18.492
as anti-rational.

59
00:02:18.542 --> 00:02:20.330
He then asked it to expand on Paul

60
00:02:20.380 --> 00:02:23.670
Feyerabend, and ChatGPT warns that it can make

61
00:02:23.720 --> 00:02:26.678
mistakes, so check important info.

62
00:02:26.730 --> 00:02:30.010
AI as consultant is informal

63
00:02:30.140 --> 00:02:31.760
and always there.

64
00:02:31.810 --> 00:02:34.260
It can be a people pleaser, and it

65
00:02:34.310 --> 00:02:36.430
can be wrong, as can the other students

66
00:02:36.480 --> 00:02:39.100
at Oxford. AI is another part of the

67
00:02:39.150 --> 00:02:41.535
mix and does not replace them.

68
00:02:41.585 --> 00:02:43.370
People make errors too.

69
00:02:43.420 --> 00:02:46.680
A 2021 study found gene-name errors in

70
00:02:46.730 --> 00:02:49.210
thirty point nine per cent of articles with

71
00:02:49.260 --> 00:02:52.210
supplementary Excel gene lists, and

72
00:02:52.260 --> 00:02:55.390
doctors sometimes operate on the wrong limb.

73
00:02:55.440 --> 00:02:58.130
The fix is built-in verification.

74
00:02:58.180 --> 00:03:01.300
Surgical sites are marked, and in a pharmacy

75
00:03:01.350 --> 00:03:03.070
somebody else checks the prescription.

76
00:03:03.120 --> 00:03:05.570
Dominik asks of any task whether it has

77
00:03:05.620 --> 00:03:07.861
verification, and adds it if not.

78
00:03:07.911 --> 00:03:10.360
The tutor never tires and can walk you

79
00:03:10.410 --> 00:03:11.519
through any issue.

80
00:03:11.570 --> 00:03:14.030
Dominik asked a model to summarise a difficult

81
00:03:14.080 --> 00:03:16.700
passage by Butler; the result was only

82
00:03:16.750 --> 00:03:20.040
shorter, and he could repeat it without understanding.

83
00:03:20.090 --> 00:03:21.790
The model should take him to where he

84
00:03:21.840 --> 00:03:23.688
can write that summary himself.

85
00:03:23.740 --> 00:03:25.770
So he asks for bullets, which show the

86
00:03:25.820 --> 00:03:28.860
structure, then for the propositions the text

87
00:03:28.910 --> 00:03:31.327
makes, such as X causes Y.

88
00:03:31.380 --> 00:03:33.830
A table puts the verbs and relations in

89
00:03:33.880 --> 00:03:36.660
view. Then come definitions and

90
00:03:36.710 --> 00:03:38.468
multiple-choice sense checks.

91
00:03:38.520 --> 00:03:41.310
ChatGPT can supply practical examples of an

92
00:03:41.360 --> 00:03:44.470
abstract idea, since lecturers never give enough, and

93
00:03:44.520 --> 00:03:47.060
find others who agreed or disagreed with the

94
00:03:47.110 --> 00:03:47.332
author.

95
00:03:47.382 --> 00:03:50.060
The last step is the point: write your

96
00:03:50.110 --> 00:03:52.286
own understanding and ask the model for feedback.

97
00:03:52.340 --> 00:03:55.450
As a translator, AI speaks almost any

98
00:03:55.500 --> 00:03:58.700
language and translates styles, modalities

99
00:03:58.750 --> 00:03:59.590
and ideas.

100
00:03:59.640 --> 00:04:02.110
Participants said academic papers are like a foreign

101
00:04:02.160 --> 00:04:05.217
language, and AI is good with academic language.

102
00:04:05.270 --> 00:04:07.880
Dominik turned lyrics about the passage into a

103
00:04:07.930 --> 00:04:09.480
song with Suno.

104
00:04:09.600 --> 00:04:12.550
A song does not teach by itself, but

105
00:04:12.600 --> 00:04:14.486
it keeps you near your studies for longer.

106
00:04:14.540 --> 00:04:17.720
NotebookLM turns a text into a mind map,

107
00:04:17.770 --> 00:04:19.794
or a podcast of two voices.

108
00:04:19.844 --> 00:04:22.130
The toolmaker is the role with the most

109
00:04:22.180 --> 00:04:24.630
impact for good or bad, and it is

110
00:04:24.680 --> 00:04:26.798
driving much of the recent progress.

111
00:04:26.850 --> 00:04:29.760
In Google AI Studio he described a website

112
00:04:29.810 --> 00:04:32.850
for exercising Czech, which he used to teach,

113
00:04:32.900 --> 00:04:34.007
and it built one.

114
00:04:34.060 --> 00:04:36.850
The danger is a time sink: you can

115
00:04:36.900 --> 00:04:38.270
do too much of it.

116
00:04:38.320 --> 00:04:41.020
Dominik built a search for his fourteen thousand

117
00:04:41.070 --> 00:04:44.190
highlights and a replacement for PowerPoint.

118
00:04:44.240 --> 00:04:47.530
He calls this his version of AI psychosis

119
00:04:47.580 --> 00:04:48.877
and does not recommend it.

120
00:04:48.930 --> 00:04:52.280
Dean W. Ball wrote in January twenty twenty-six:

121
00:04:52.330 --> 00:04:54.290
"By the end of this year, the least

122
00:04:54.340 --> 00:04:55.700
important thing you will be able to do

123
00:04:55.750 --> 00:04:58.460
with frontier AI systems will be getting

124
00:04:58.510 --> 00:05:01.720
chatbots to answer questions."

125
00:05:02.090 --> 00:05:05.380
Dominik said answering questions is very helpful but

126
00:05:05.430 --> 00:05:07.306
perhaps the least important thing.

127
00:05:07.360 --> 00:05:10.550
Dominik said ChatGPT is no longer his

128
00:05:10.600 --> 00:05:12.720
main way into a language model.

129
00:05:12.770 --> 00:05:15.260
He uses Codex, and had used it for

130
00:05:15.310 --> 00:05:17.646
one hundred and thirty-nine days without a break.

131
00:05:17.700 --> 00:05:18.930
It runs on more or less the same

132
00:05:18.980 --> 00:05:20.435
models as ChatGPT.

133
00:05:20.500 --> 00:05:22.910
Most people think of AI as an oracle

134
00:05:22.960 --> 00:05:25.250
that answers questions, translates and

135
00:05:25.300 --> 00:05:28.080
summarises. The next step was a

136
00:05:28.130 --> 00:05:30.345
toolmaker that makes apps.

137
00:05:30.400 --> 00:05:32.930
The big step is a tool user that

138
00:05:32.980 --> 00:05:35.940
can configure a computer, do version control, check

139
00:05:35.990 --> 00:05:38.655
email, save notes and deploy a website.

140
00:05:38.710 --> 00:05:42.070
In a chatbot, the model answers a prompt.

141
00:05:42.120 --> 00:05:45.080
In an agent, the harness, a

142
00:05:45.130 --> 00:05:48.000
desktop app, sends the question to the model,

143
00:05:48.050 --> 00:05:50.525
which makes a plan and sends commands back.

144
00:05:50.580 --> 00:05:53.450
The harness passes them to the computer and

145
00:05:53.500 --> 00:05:56.780
returns the results, and the model changes its

146
00:05:56.830 --> 00:05:58.790
plan before it responds.

147
00:05:58.840 --> 00:06:02.120
The harness is ordinary software with no

148
00:06:02.170 --> 00:06:03.101
intelligence.

149
00:06:03.151 --> 00:06:06.110
As of September twenty twenty-six, the desktop

150
00:06:06.160 --> 00:06:09.450
agents he names are Codex, with

151
00:06:09.500 --> 00:06:12.200
Claude Code and Google Antigravity as

152
00:06:12.250 --> 00:06:12.995
alternatives.

153
00:06:13.050 --> 00:06:15.050
They work on the files on your own

154
00:06:15.100 --> 00:06:17.880
computer, so they can organise a messy folder

155
00:06:17.930 --> 00:06:18.560
of files.

156
00:06:18.610 --> 00:06:21.810
His suggested first task is the downloads folder.

157
00:06:21.860 --> 00:06:24.980
Codex visualised its file types and sizes

158
00:06:25.030 --> 00:06:27.820
and found three point two gigabytes of zip

159
00:06:27.870 --> 00:06:28.300
files.

160
00:06:28.350 --> 00:06:30.230
When you get a paper, he said, do

161
00:06:30.280 --> 00:06:31.394
not ask for a summary.

162
00:06:31.444 --> 00:06:34.320
Ask the agent to download the paper's

163
00:06:34.370 --> 00:06:37.830
data and rerun one analysis,

164
00:06:37.960 --> 00:06:40.500
step by step, to see whether it gets

165
00:06:40.550 --> 00:06:41.660
the same result.

166
00:06:41.710 --> 00:06:43.830
It will not work for everything, but it

167
00:06:43.880 --> 00:06:45.898
brings you closer to doing research.

168
00:06:45.950 --> 00:06:49.230
The recent change is that companies made models

169
00:06:49.280 --> 00:06:51.900
better at using tools and following

170
00:06:51.950 --> 00:06:53.560
instructions about them.

171
00:06:53.610 --> 00:06:56.480
They gave models a real computer environment and

172
00:06:56.530 --> 00:06:59.570
tasks, and the models tried, failed and got

173
00:06:59.620 --> 00:07:02.116
feedback, which is reinforcement learning.

174
00:07:02.170 --> 00:07:04.364
This is new in twenty twenty-six.

175
00:07:04.414 --> 00:07:07.600
Dominik called the lecture a menu of options.

176
00:07:07.650 --> 00:07:10.270
It taught that these things exist, not how

177
00:07:10.320 --> 00:07:11.046
to do them.

178
00:07:11.100 --> 00:07:13.720
Learning how is a project in itself, and

179
00:07:13.770 --> 00:07:16.140
the first attempt may not save time.

180
00:07:16.190 --> 00:07:18.750
He advised against trying everything at once.