WEBVTT

1
00:00:00.000 --> 00:00:02.750
In this lecture, Dominik Lukeš shows researchers and

2
00:00:02.880 --> 00:00:06.190
professional services staff what AI agents can do

3
00:00:06.240 --> 00:00:07.870
in research workflows.

4
00:00:07.920 --> 00:00:09.870
He asks the room when they last used

5
00:00:09.920 --> 00:00:10.838
ChatGPT.

6
00:00:10.910 --> 00:00:13.340
His own answer is that he no longer

7
00:00:13.390 --> 00:00:15.495
uses it. He uses Codex.

8
00:00:15.570 --> 00:00:18.800
He still uses the same large language models,

9
00:00:18.850 --> 00:00:21.360
but a chatbot is no longer his main

10
00:00:21.410 --> 00:00:22.720
interface to them.

11
00:00:22.770 --> 00:00:25.280
Codex is always open on his computer, and

12
00:00:25.330 --> 00:00:27.930
a Mac Mini at home runs an always-on

13
00:00:28.050 --> 00:00:29.792
assistant called Claw.

14
00:00:29.870 --> 00:00:32.620
Dean W. Ball predicted in January that by

15
00:00:32.670 --> 00:00:34.300
the end of this year the least important

16
00:00:34.350 --> 00:00:35.420
thing you will be able to do with

17
00:00:35.470 --> 00:00:38.620
frontier AI systems will be getting chatbots to

18
00:00:38.670 --> 00:00:41.980
answer questions. He says that prediction has

19
00:00:42.030 --> 00:00:43.300
already come true.

20
00:00:43.370 --> 00:00:46.680
When ChatGPT was first released, AI was

21
00:00:46.730 --> 00:00:49.960
an oracle: it answered questions, summarised

22
00:00:50.010 --> 00:00:51.281
and translated.

23
00:00:51.331 --> 00:00:54.160
Then came vibe coding, where AI writes whole

24
00:00:54.210 --> 00:00:57.200
apps. People used Cursor, Lovable and

25
00:00:57.250 --> 00:01:00.560
Google AI Studio to make scripts, dashboards and

26
00:01:00.610 --> 00:01:01.641
simulations.

27
00:01:01.720 --> 00:01:04.950
Then AI learned to use tools as well

28
00:01:05.000 --> 00:01:07.430
as make them, and he says that changes

29
00:01:07.480 --> 00:01:10.710
everything. It works with files and utilities,

30
00:01:10.760 --> 00:01:13.670
manages data and replicates analyses.

31
00:01:13.720 --> 00:01:16.150
His advice is to help AI discover the

32
00:01:16.200 --> 00:01:18.310
right answers the way you would have done

33
00:01:18.360 --> 00:01:18.986
before.

34
00:01:19.060 --> 00:01:21.650
His own example is a Git repository that

35
00:01:21.700 --> 00:01:24.930
had become too big, and GitHub stopped working.

36
00:01:24.980 --> 00:01:27.650
In December he spent hours pasting screenshots into

37
00:01:27.700 --> 00:01:30.169
ChatGPT without fixing it.

38
00:01:30.220 --> 00:01:32.090
Over Christmas he gave the problem to an

39
00:01:32.140 --> 00:01:35.450
agent, which solved it in about fifteen minutes

40
00:01:35.500 --> 00:01:36.931
because it could keep running.

41
00:01:37.010 --> 00:01:40.090
After giving workshops at the University of York

42
00:01:40.210 --> 00:01:42.400
in July, he had to claim the costs

43
00:01:42.450 --> 00:01:45.200
of a four-day trip, using a Word document

44
00:01:45.250 --> 00:01:46.979
and an Excel spreadsheet.

45
00:01:47.030 --> 00:01:49.220
He told his agent to find the receipts

46
00:01:49.270 --> 00:01:50.377
in his Gmail.

47
00:01:50.450 --> 00:01:52.880
The agent checked how much could be claimed,

48
00:01:52.930 --> 00:01:55.360
filled in the forms, added his signature and

49
00:01:55.410 --> 00:01:58.339
looked in his private accounts for other receipts.

50
00:01:58.410 --> 00:02:00.920
It put a draft email in Outlook with

51
00:02:00.970 --> 00:02:02.440
the receipts attached.

52
00:02:02.490 --> 00:02:04.708
Two weeks later he had the money.

53
00:02:04.780 --> 00:02:06.810
Most people still think of AI as a

54
00:02:06.860 --> 00:02:09.130
chatbot, and so give up many

55
00:02:09.180 --> 00:02:12.410
capabilities. What changed in twenty twenty-six is

56
00:02:12.460 --> 00:02:14.837
better models and better harnesses.

57
00:02:14.890 --> 00:02:18.040
Models now plan, choose tools, interpret results and

58
00:02:18.090 --> 00:02:18.795
revise.

59
00:02:18.870 --> 00:02:21.700
The speed of development is enormous.

60
00:02:21.750 --> 00:02:23.620
The model he used for the Git problem

61
00:02:23.670 --> 00:02:26.660
at Christmas now feels slow and

62
00:02:26.710 --> 00:02:27.326
dumb.

63
00:02:27.400 --> 00:02:30.230
OpenAI has released four versions since

64
00:02:30.280 --> 00:02:33.660
January, and Anthropic's Opus went from

65
00:02:33.710 --> 00:02:37.030
four point five in November to Opus five

66
00:02:37.080 --> 00:02:37.940
in July.

67
00:02:38.010 --> 00:02:40.020
Cost has changed too.

68
00:02:40.090 --> 00:02:43.000
In twenty twenty-three OpenAI chose twenty dollars a

69
00:02:43.050 --> 00:02:45.240
month for a subscription, because it could not

70
00:02:45.290 --> 00:02:47.105
afford to serve its users free.

71
00:02:47.180 --> 00:02:49.770
On twenty-ninth September it added a new Pro

72
00:02:49.820 --> 00:02:52.880
option in three usage tiers, at one hundred,

73
00:02:52.930 --> 00:02:55.370
two hundred and five hundred dollars a month,

74
00:02:55.420 --> 00:02:58.354
and even the top tiers have usage limits.

75
00:02:58.430 --> 00:03:01.330
Agents allow more ambitious projects and a

76
00:03:01.380 --> 00:03:03.059
broader range of methods.

77
00:03:03.130 --> 00:03:06.440
Moving from SPSS to R made analysis

78
00:03:06.490 --> 00:03:09.427
more reproducible but demanded much more learning.

79
00:03:09.480 --> 00:03:11.750
An agent that can use R makes any

80
00:03:11.800 --> 00:03:13.150
relevant feature available.

81
00:03:13.230 --> 00:03:16.220
He wanted to send pictures in one HTML

82
00:03:16.430 --> 00:03:19.260
file, and told Codex to try different

83
00:03:19.310 --> 00:03:22.412
compression settings and show them side by side.

84
00:03:22.490 --> 00:03:25.800
Codex used local tools on his computer and

85
00:03:25.850 --> 00:03:29.080
had the comparison in about four minutes.

86
00:03:29.130 --> 00:03:31.560
He could always have done this, but never

87
00:03:31.610 --> 00:03:32.238
would have.

88
00:03:32.310 --> 00:03:34.420
He also had an agent replicate a published

89
00:03:34.470 --> 00:03:37.540
paper's results from its reported numbers and visualise

90
00:03:37.590 --> 00:03:39.540
them. He had to ask it to check

91
00:03:39.590 --> 00:03:41.168
the parameters and data.

92
00:03:41.220 --> 00:03:43.170
Where a paper has public data, he can

93
00:03:43.220 --> 00:03:45.970
check its numbers, because numbers in published

94
00:03:46.020 --> 00:03:47.831
papers are often wrong.

95
00:03:47.910 --> 00:03:50.580
For a researcher working on Asian languages, he

96
00:03:50.630 --> 00:03:53.740
asked Codex to find corpora in Sketch Engine

97
00:03:53.830 --> 00:03:56.580
and compare terms for justice and fairness in

98
00:03:56.630 --> 00:03:58.428
Urdu and Bengali.

99
00:03:58.480 --> 00:04:00.260
The researcher still has to learn what the

100
00:04:00.310 --> 00:04:03.070
results mean, but no longer Sketch Engine's query

101
00:04:03.120 --> 00:04:03.589
language.

102
00:04:03.660 --> 00:04:05.817
He also builds with agents.

103
00:04:05.890 --> 00:04:08.400
One turned a script into a published website

104
00:04:08.450 --> 00:04:11.360
of historical sources, and he has made himself

105
00:04:11.410 --> 00:04:14.720
desktop apps such as Slidewell, which

106
00:04:14.770 --> 00:04:16.122
searches his slides.

107
00:04:16.200 --> 00:04:19.270
An agent is a model plus a

108
00:04:19.320 --> 00:04:21.110
harness plus a loop.

109
00:04:21.160 --> 00:04:22.790
The model makes a plan and sends a

110
00:04:22.840 --> 00:04:24.950
command to the harness, an app such as

111
00:04:25.000 --> 00:04:26.171
Claude Code.

112
00:04:26.221 --> 00:04:28.970
The harness passes it to the computer, which

113
00:04:29.020 --> 00:04:31.168
runs it and returns the result.

114
00:04:31.220 --> 00:04:33.650
The model revises its plan and the cycle

115
00:04:33.700 --> 00:04:34.245
repeats.

116
00:04:34.320 --> 00:04:37.470
Take the search for last year's evaluation numbers.

117
00:04:37.520 --> 00:04:40.110
A person opens a colleague's handover document, then

118
00:04:40.160 --> 00:04:42.830
a PDF report, then searches the notes, and

119
00:04:42.880 --> 00:04:43.908
finds nothing.

120
00:04:43.980 --> 00:04:46.410
The agent searches by subject across the whole

121
00:04:46.460 --> 00:04:49.150
year, and finds the figures, one hundred and

122
00:04:49.200 --> 00:04:52.590
forty-eight responses averaging 4.3 out of 5,

123
00:04:52.780 --> 00:04:55.370
in a draft report and a May presentation

124
00:04:55.420 --> 00:04:57.376
rather than the files he remembered.

125
00:04:57.450 --> 00:05:00.360
The model sees more than the prompt.

126
00:05:00.410 --> 00:05:03.590
In his Claude Code screenshot, about thirty-eight

127
00:05:03.640 --> 00:05:06.760
thousand tokens of instructions travel with each prompt,

128
00:05:06.810 --> 00:05:09.994
beside seventy-nine thousand tokens of messages.

129
00:05:10.044 --> 00:05:12.710
With a desktop agent he can inspect everything

130
00:05:12.760 --> 00:05:15.590
that was sent, including every tool call.

131
00:05:15.640 --> 00:05:18.870
Web chat apps send such instructions too but

132
00:05:18.920 --> 00:05:19.857
never show them.

133
00:05:19.930 --> 00:05:22.520
With a desktop agent, files stay in your

134
00:05:22.570 --> 00:05:25.640
own folders, and work can cover many local

135
00:05:25.690 --> 00:05:28.760
files. You must manage the file structure and

136
00:05:28.810 --> 00:05:31.720
decide which folders the agent can use.

137
00:05:31.770 --> 00:05:34.680
A local folder does not mean a

138
00:05:34.730 --> 00:05:35.666
local model.

139
00:05:35.740 --> 00:05:38.730
Analysing a large data set sends little of

140
00:05:38.780 --> 00:05:41.210
the source data to the cloud model, while

141
00:05:41.260 --> 00:05:44.105
summarising a document sends its text.

142
00:05:44.155 --> 00:05:46.420
A local model can describe an image on

143
00:05:46.470 --> 00:05:49.060
your computer, so the cloud model receives only

144
00:05:49.110 --> 00:05:50.942
a caption, not the picture.

145
00:05:51.020 --> 00:05:53.770
For sensitive data, the prompt can say that

146
00:05:53.820 --> 00:05:55.370
data must never leave.

147
00:05:55.420 --> 00:05:57.850
The harness can put a sandbox around the

148
00:05:57.900 --> 00:06:00.410
agent, with folders that are never sent and

149
00:06:00.460 --> 00:06:02.330
a list of allowed commands.

150
00:06:02.380 --> 00:06:04.330
Because the model can interact only through the

151
00:06:04.380 --> 00:06:07.121
harness, it cannot get around the sandbox.

152
00:06:07.200 --> 00:06:09.870
Asked about risk, he says he trusts his

153
00:06:09.920 --> 00:06:12.590
agents more than himself, though they still make

154
00:06:12.640 --> 00:06:13.490
mistakes.

155
00:06:13.540 --> 00:06:15.650
For safety he asks what he would ask

156
00:06:15.700 --> 00:06:18.210
of a researcher with access to the data,

157
00:06:18.260 --> 00:06:20.450
how easy is a mistake and with what

158
00:06:20.500 --> 00:06:23.650
consequences, across three areas: safe

159
00:06:23.700 --> 00:06:26.610
use of the computer, secure tools and data

160
00:06:26.660 --> 00:06:27.370
protection.

161
00:06:27.440 --> 00:06:29.950
Asked whether a project is on track, an

162
00:06:30.000 --> 00:06:32.990
agent reads the proposal and methods, joins the

163
00:06:33.040 --> 00:06:35.950
participant files by ID and writes a repeatable

164
00:06:36.000 --> 00:06:36.698
script.

165
00:06:36.750 --> 00:06:39.660
A good-looking total of ninety-seven point five per

166
00:06:39.710 --> 00:06:42.700
cent hides one group at thirty-five per cent,

167
00:06:42.750 --> 00:06:45.180
so it still needs checks, and often the

168
00:06:45.230 --> 00:06:47.180
check is the user, who knows what the

169
00:06:47.230 --> 00:06:48.419
work should look like.

170
00:06:48.490 --> 00:06:51.320
He tells the agent to record each step

171
00:06:51.370 --> 00:06:52.520
in a file.

172
00:06:52.570 --> 00:06:55.160
That leaves a text record of the conversation

173
00:06:55.210 --> 00:06:58.040
and the agent's notes that people and other

174
00:06:58.090 --> 00:06:59.658
agents can read.

175
00:06:59.730 --> 00:07:02.640
Rerunning the saved script on the same inputs

176
00:07:02.690 --> 00:07:06.000
reproduces the same counts, though that does not

177
00:07:06.050 --> 00:07:07.867
show the method is right.

178
00:07:07.940 --> 00:07:10.450
He closes by saying that the task is

179
00:07:10.500 --> 00:07:13.810
getting used to thinking differently and automatically

180
00:07:13.860 --> 00:07:15.286
reaching for an agent.