WEBVTT

1
00:00:00.000 --> 00:00:02.550
In this lecture, Dominik Lukeš explains how

2
00:00:02.730 --> 00:00:05.580
desktop AI agents can help with study and

3
00:00:05.630 --> 00:00:06.576
research.

4
00:00:06.700 --> 00:00:09.730
He starts from why people study at Oxford:

5
00:00:09.780 --> 00:00:12.550
to change their brain, to have an impact

6
00:00:12.600 --> 00:00:14.220
on the world, and to prove to the

7
00:00:14.270 --> 00:00:16.500
authorities that they have changed their brain enough

8
00:00:16.550 --> 00:00:17.247
to do it.

9
00:00:17.370 --> 00:00:20.050
He thinks AI is fine to use for

10
00:00:20.100 --> 00:00:22.780
that, but nobody builds muscle by asking someone

11
00:00:22.830 --> 00:00:24.331
else to go to the gym.

12
00:00:24.410 --> 00:00:27.440
Students still do the heavy lifting, and

13
00:00:27.490 --> 00:00:29.405
AI can be a spotter.

14
00:00:29.530 --> 00:00:32.210
For exams and theses the rules differ, so

15
00:00:32.260 --> 00:00:35.550
students should ask, and never claim AI's work

16
00:00:35.600 --> 00:00:38.890
as their own or use AI when told

17
00:00:38.940 --> 00:00:39.654
not to.

18
00:00:39.780 --> 00:00:42.630
Every AI tool can make

19
00:00:42.680 --> 00:00:45.980
mistakes, and the user is responsible for

20
00:00:46.030 --> 00:00:47.350
what they do with it.

21
00:00:47.480 --> 00:00:50.770
People make mistakes too: a study of

22
00:00:50.820 --> 00:00:54.020
data files published with articles from twenty fourteen

23
00:00:54.070 --> 00:00:57.370
to twenty twenty found problems in about thirty-one

24
00:00:57.420 --> 00:00:59.255
per cent of articles.

25
00:00:59.380 --> 00:01:01.620
At his own computer he never logs in

26
00:01:01.670 --> 00:01:03.698
to a chatbot on the web.

27
00:01:03.830 --> 00:01:06.850
He uses Codex or Claude Code,

28
00:01:06.900 --> 00:01:09.400
and his Codex use has grown from a

29
00:01:09.450 --> 00:01:11.690
thin line a year ago to the whole

30
00:01:11.740 --> 00:01:12.514
chart.

31
00:01:12.640 --> 00:01:15.320
For the last hundred and eighty-four days he

32
00:01:15.370 --> 00:01:17.486
has used one every day.

33
00:01:17.610 --> 00:01:20.290
With desktop agents he has collected years of

34
00:01:20.340 --> 00:01:23.170
his writing on a new website, and catalogued

35
00:01:23.220 --> 00:01:26.360
two hundred and seven hours of recordings from

36
00:01:26.410 --> 00:01:28.650
his time as a volunteer at a community

37
00:01:28.700 --> 00:01:31.280
radio station in Norwich, with

38
00:01:31.330 --> 00:01:34.510
snippets and a record of every interview.

39
00:01:34.640 --> 00:01:37.840
After recent IT security news he asked

40
00:01:37.890 --> 00:01:40.830
Codex to audit his accounts and the

41
00:01:40.880 --> 00:01:43.861
software installed across his machines.

42
00:01:43.990 --> 00:01:46.930
It mapped each machine, what was running on

43
00:01:46.980 --> 00:01:49.920
it and its security controls, and then

44
00:01:49.970 --> 00:01:52.470
mapped how people could get in from the

45
00:01:52.520 --> 00:01:53.127
internet.

46
00:01:53.250 --> 00:01:56.200
AI was first an oracle that

47
00:01:56.250 --> 00:01:59.190
answers questions, which is what people have used

48
00:01:59.230 --> 00:02:00.504
chatbots for.

49
00:02:00.554 --> 00:02:03.580
Then people saw that AI can write software

50
00:02:03.630 --> 00:02:05.136
and make tools.

51
00:02:05.220 --> 00:02:07.466
Now it can use tools too.

52
00:02:07.590 --> 00:02:09.740
A desktop agent is roughly a version of

53
00:02:09.790 --> 00:02:12.730
ChatGPT installed on your computer, where it can

54
00:02:12.780 --> 00:02:14.290
work with your files.

55
00:02:14.420 --> 00:02:17.800
In January twenty twenty-six, Dean W.

56
00:02:17.850 --> 00:02:19.730
Ball said that by the end of the

57
00:02:19.780 --> 00:02:21.930
year the least important thing you would be

58
00:02:21.980 --> 00:02:24.310
able to do with frontier AI systems would

59
00:02:24.360 --> 00:02:27.690
be getting chatbots to answer questions.

60
00:02:27.820 --> 00:02:30.550
He thinks that has already come true.

61
00:02:30.680 --> 00:02:33.100
In a chat, AI can tell you how

62
00:02:33.150 --> 00:02:35.603
to fill in an expenses form.

63
00:02:35.730 --> 00:02:38.760
An agent can find the form,

64
00:02:38.810 --> 00:02:42.010
search your email for the receipts, calculate

65
00:02:42.060 --> 00:02:44.718
the totals and fill the form in.

66
00:02:44.840 --> 00:02:47.960
He also had Codex put his train receipts

67
00:02:48.010 --> 00:02:50.377
and workshop schedule on his calendar.

68
00:02:50.500 --> 00:02:53.440
The models are much better: they understand the

69
00:02:53.500 --> 00:02:56.080
task, plan, choose tools,

70
00:02:56.130 --> 00:02:59.476
interpret results and revise their plans.

71
00:02:59.600 --> 00:03:01.840
The desktop agents, called

72
00:03:01.890 --> 00:03:05.180
harnesses, supply the tools and manage the

73
00:03:05.230 --> 00:03:06.466
exchange with the model.

74
00:03:06.590 --> 00:03:09.880
Given a downloaded paper, an agent

75
00:03:09.940 --> 00:03:13.022
can build a replication he can explore.

76
00:03:13.100 --> 00:03:15.410
He did this with a paper on energy

77
00:03:15.480 --> 00:03:16.956
systems in a building.

78
00:03:17.080 --> 00:03:20.110
He finds that more useful than a summary,

79
00:03:20.160 --> 00:03:22.007
which is just shorter.

80
00:03:22.130 --> 00:03:24.720
He wanted to email a set of pictures,

81
00:03:24.770 --> 00:03:26.425
but they were too large.

82
00:03:26.550 --> 00:03:29.320
Codex tried different compression levels and built a

83
00:03:29.370 --> 00:03:32.660
side-by-side comparison and a table, with the

84
00:03:32.710 --> 00:03:35.390
best trade-off at about ninety-three per cent

85
00:03:35.440 --> 00:03:36.346
saving.

86
00:03:36.470 --> 00:03:38.620
He asked for the visual comparison so that

87
00:03:38.670 --> 00:03:40.792
he did not have to trust Codex's word.

88
00:03:40.920 --> 00:03:43.950
He does not believe AI tools steal

89
00:03:44.000 --> 00:03:44.675
data.

90
00:03:44.800 --> 00:03:47.210
It is a question of trust, as with

91
00:03:47.260 --> 00:03:50.030
OneDrive, SharePoint, Google Drive or

92
00:03:50.080 --> 00:03:53.020
Dropbox: each provider says what it

93
00:03:53.070 --> 00:03:56.100
promises, and users read that and

94
00:03:56.150 --> 00:03:56.945
decide.

95
00:03:57.070 --> 00:04:00.280
An agent is a model using

96
00:04:00.330 --> 00:04:03.530
tools in a loop to achieve complex tasks

97
00:04:03.580 --> 00:04:05.500
that take several steps.

98
00:04:05.630 --> 00:04:08.568
A chatbot takes a prompt and answers.

99
00:04:08.620 --> 00:04:11.030
An agent starts the same way, then the

100
00:04:11.080 --> 00:04:13.670
model makes a plan and calls tools, such

101
00:04:13.720 --> 00:04:15.780
as a search or a command on the

102
00:04:15.830 --> 00:04:16.533
computer.

103
00:04:16.660 --> 00:04:19.860
The harness, a normal application on the

104
00:04:19.910 --> 00:04:22.680
desktop, passes those commands to the

105
00:04:22.730 --> 00:04:23.513
computer.

106
00:04:23.640 --> 00:04:26.400
The computer runs them and returns the

107
00:04:26.450 --> 00:04:29.750
results, the model revises its plan, and

108
00:04:29.800 --> 00:04:32.740
the loop continues until the result is

109
00:04:32.790 --> 00:04:33.486
ready.

110
00:04:33.610 --> 00:04:36.730
The user decides which folders the agent can

111
00:04:36.780 --> 00:04:39.421
reach and what data goes to the model.

112
00:04:39.550 --> 00:04:42.660
In an invented office example, someone needs

113
00:04:42.710 --> 00:04:45.920
last year's evaluation numbers and thinks

114
00:04:45.970 --> 00:04:48.910
Geoff, who has since left, did them last

115
00:04:48.960 --> 00:04:49.842
February.

116
00:04:49.970 --> 00:04:52.650
Searching by hand means trying names, dates and

117
00:04:52.700 --> 00:04:54.447
file contents in turn.

118
00:04:54.570 --> 00:04:57.430
The agent searched the whole year, compared a

119
00:04:57.480 --> 00:05:00.510
draft and a presentation, found a hundred and

120
00:05:00.560 --> 00:05:03.090
forty-eight responses with an average rating of four

121
00:05:03.140 --> 00:05:05.780
point three out of five, and showed that

122
00:05:05.830 --> 00:05:09.084
George did the analysis, not Geoff.

123
00:05:09.210 --> 00:05:12.240
For literature searches, an agent can replace

124
00:05:12.290 --> 00:05:14.790
logging in to a website for Deep

125
00:05:14.840 --> 00:05:15.992
Research.

126
00:05:16.120 --> 00:05:19.060
It runs an Elicit literature survey through its

127
00:05:19.110 --> 00:05:21.960
API and files the papers in a new

128
00:05:22.010 --> 00:05:23.857
Zotero collection.

129
00:05:23.980 --> 00:05:27.190
In a real project every reference came back

130
00:05:27.240 --> 00:05:29.560
tagged, with a note on why it was

131
00:05:29.610 --> 00:05:30.411
chosen.

132
00:05:30.540 --> 00:05:32.860
He asked for a visualisation of how the

133
00:05:32.910 --> 00:05:34.839
references fit together.

134
00:05:34.970 --> 00:05:37.910
It gives the reason each paper is recommended,

135
00:05:37.960 --> 00:05:40.720
and lets him inspect the evidence, click through

136
00:05:40.770 --> 00:05:44.197
to related papers and check the actual paper.

137
00:05:44.320 --> 00:05:47.090
He finds this more useful than a literature

138
00:05:47.140 --> 00:05:50.080
review written for him that he does not

139
00:05:50.130 --> 00:05:51.640
fully understand.

140
00:05:51.770 --> 00:05:54.238
Agents need context.

141
00:05:54.320 --> 00:05:56.730
A model knows only what it learned in

142
00:05:56.780 --> 00:05:59.100
training and what is in front of it:

143
00:05:59.150 --> 00:06:02.127
the prompt, chat history and files.

144
00:06:02.250 --> 00:06:04.400
It does not keep learning as you use

145
00:06:04.450 --> 00:06:04.806
it.

146
00:06:04.930 --> 00:06:07.350
One example from Claude added up to about

147
00:06:07.400 --> 00:06:10.340
a hundred and seventeen thousand tokens, like a

148
00:06:10.390 --> 00:06:11.613
short novel.

149
00:06:11.740 --> 00:06:14.940
Files in folders give an agent

150
00:06:14.990 --> 00:06:16.234
that context.

151
00:06:16.310 --> 00:06:19.650
He recommends a file called AGENTS.md

152
00:06:20.180 --> 00:06:22.330
in every folder you work in with an

153
00:06:22.380 --> 00:06:25.240
agent, a text file saying what is and

154
00:06:25.290 --> 00:06:26.762
is not allowed there.

155
00:06:26.890 --> 00:06:29.300
The harness reads it before doing anything in

156
00:06:29.350 --> 00:06:31.500
that folder, which is why it must have

157
00:06:31.550 --> 00:06:32.348
that name.

158
00:06:32.480 --> 00:06:35.590
A skill is a folder of instruction files

159
00:06:35.640 --> 00:06:38.850
for one task, such as making PowerPoint

160
00:06:38.900 --> 00:06:39.462
files.

161
00:06:39.512 --> 00:06:41.660
It holds the main instructions and links to

162
00:06:41.710 --> 00:06:43.860
other files, and the agent looks those up

163
00:06:43.910 --> 00:06:45.620
when it needs them, as you would from

164
00:06:45.670 --> 00:06:46.391
a manual.

165
00:06:46.520 --> 00:06:49.630
The harness lists the available skills in each

166
00:06:49.680 --> 00:06:50.395
prompt.

167
00:06:50.520 --> 00:06:53.810
Agents keep everything on disk, so the user

168
00:06:53.860 --> 00:06:56.190
can ask for a report on the steps

169
00:06:56.240 --> 00:06:58.435
taken, linked to evidence.

170
00:06:58.560 --> 00:07:01.500
He recommends using a smarter model rather than

171
00:07:01.550 --> 00:07:04.400
a smaller one because it is more reliable,

172
00:07:04.450 --> 00:07:07.089
and always checking what the agent does.

173
00:07:07.220 --> 00:07:10.330
He advises small steps rather

174
00:07:10.380 --> 00:07:12.267
than trying everything at once.

175
00:07:12.320 --> 00:07:14.730
He sees this as the direction for AI

176
00:07:14.780 --> 00:07:17.460
and for better research, and says everyone is

177
00:07:17.510 --> 00:07:19.437
at the start of the journey.

178
00:07:19.560 --> 00:07:21.980
A student he spoke to recently had built

179
00:07:22.030 --> 00:07:24.880
a flashcard app for recall because none of

180
00:07:24.930 --> 00:07:27.330
the existing ones worked the way they wanted.