WEBVTT

1
00:00:00.000 --> 00:00:02.750
In this lecture, Dominik Lukeš maps

2
00:00:02.800 --> 00:00:05.134
how AI in research has changed.

3
00:00:05.184 --> 00:00:08.280
When ChatGPT arrived in twenty twenty-three,

4
00:00:08.330 --> 00:00:11.590
everybody was concerned that AI would write or

5
00:00:11.640 --> 00:00:14.560
edit papers, answer questions about

6
00:00:14.610 --> 00:00:17.620
difficult concepts and maybe find literature.

7
00:00:17.670 --> 00:00:20.080
They also worried it would hallucinate and make

8
00:00:20.130 --> 00:00:20.525
errors.

9
00:00:20.580 --> 00:00:23.820
By twenty twenty-six, AI makes independent discoveries and

10
00:00:23.870 --> 00:00:27.050
runs full research workflows on its own.

11
00:00:27.100 --> 00:00:28.920
Much of this has become true only in

12
00:00:28.970 --> 00:00:30.809
the last ten to twelve months.

13
00:00:30.860 --> 00:00:33.430
Computers have taken on research roles since the

14
00:00:33.480 --> 00:00:35.850
theorem solvers of the nineteen fifties.

15
00:00:36.190 --> 00:00:39.200
Now there are whole AI scientist systems that

16
00:00:39.250 --> 00:00:42.340
generate ideas, run the computational work and

17
00:00:42.390 --> 00:00:45.205
write a manuscript in one go.

18
00:00:45.280 --> 00:00:48.140
In early September, OpenAI announced that its models

19
00:00:48.190 --> 00:00:51.020
had solved the Navier–Stokes problem, one of the

20
00:00:51.070 --> 00:00:52.890
Millennium Prize problems.

21
00:00:52.940 --> 00:00:55.940
The run used about ten thousand agents, at

22
00:00:55.990 --> 00:00:58.400
an estimated cost of a hundred thousand pounds

23
00:00:58.450 --> 00:00:58.844
or more.

24
00:00:58.894 --> 00:01:02.230
Autonomous discovery gets the headlines, but it

25
00:01:02.280 --> 00:01:03.920
is still the exception.

26
00:01:03.970 --> 00:01:07.230
The lecture concentrates on assisted discovery,

27
00:01:07.280 --> 00:01:09.780
where researchers use these tools to run more

28
00:01:09.830 --> 00:01:12.320
experiments and gather more kinds of data.

29
00:01:12.370 --> 00:01:14.780
They also end up busier, because they fill

30
00:01:14.830 --> 00:01:17.458
the time AI saves with more work.

31
00:01:17.510 --> 00:01:20.190
For most of twenty twenty-five, the tool everyone

32
00:01:20.240 --> 00:01:23.410
talked about was Deep Research, and by the

33
00:01:23.460 --> 00:01:25.530
end of that year nearly every tool had

34
00:01:25.580 --> 00:01:26.980
it built in.

35
00:01:27.030 --> 00:01:29.144
Agents have changed everything.

36
00:01:29.194 --> 00:01:31.670
At the start of twenty twenty-six, Dean W.

37
00:01:31.720 --> 00:01:33.540
Ball predicted that by the end of the

38
00:01:33.590 --> 00:01:36.850
year, getting chatbots to answer questions would

39
00:01:36.900 --> 00:01:39.740
be the least important thing frontier AI systems

40
00:01:39.790 --> 00:01:42.620
could do. Most people still think of AI

41
00:01:42.670 --> 00:01:43.047
that way.

42
00:01:43.100 --> 00:01:45.670
At first, AI was an oracle.

43
00:01:45.720 --> 00:01:48.770
You asked a chatbot like ChatGPT,

44
00:01:48.860 --> 00:01:51.780
Claude or Gemini, and it answered,

45
00:01:51.830 --> 00:01:53.525
summarised or translated.

46
00:01:53.575 --> 00:01:56.440
Then, in twenty twenty-five, came vibe coding.

47
00:01:56.800 --> 00:01:59.550
AI became a tool maker: it could write

48
00:01:59.600 --> 00:02:02.090
code, so people built their own scripts,

49
00:02:02.140 --> 00:02:05.150
dashboards, simulations and whole

50
00:02:05.200 --> 00:02:08.370
apps in tools like Cursor, Google

51
00:02:08.420 --> 00:02:10.317
AI Studio and Lovable.

52
00:02:10.370 --> 00:02:13.780
Around December twenty twenty-five, AI became a

53
00:02:13.830 --> 00:02:17.180
tool user. Better models could plan, work

54
00:02:17.230 --> 00:02:20.230
with files on your computer, run utilities, manage

55
00:02:20.280 --> 00:02:22.687
your data and replicate analyses.

56
00:02:22.740 --> 00:02:25.968
People moved from chatbots to desktop agents.

57
00:02:26.020 --> 00:02:29.350
Most people can install one now: Codex, or

58
00:02:29.400 --> 00:02:31.890
Claude Code from Anthropic.

59
00:02:31.940 --> 00:02:35.200
Google also has Antigravity, but he does not

60
00:02:35.250 --> 00:02:37.750
recommend it, because Google is fairly behind at

61
00:02:37.800 --> 00:02:38.428
this stage.

62
00:02:38.480 --> 00:02:40.620
A chatbot can tell you how to fill

63
00:02:40.670 --> 00:02:42.243
in an expenses form.

64
00:02:42.293 --> 00:02:45.480
An agent can find the form, search

65
00:02:45.530 --> 00:02:48.450
your emails and folders for the receipts, add

66
00:02:48.500 --> 00:02:51.680
up the totals, attach the receipts and fill

67
00:02:51.730 --> 00:02:52.780
in the form.

68
00:02:52.830 --> 00:02:55.750
AI still makes errors, but he finds

69
00:02:55.800 --> 00:02:58.211
agents reliable for jobs like this.

70
00:02:58.261 --> 00:03:00.840
An agent works through a loop.

71
00:03:00.890 --> 00:03:03.640
You type a prompt into a desktop app

72
00:03:03.690 --> 00:03:07.120
called the harness, such as Codex or Claude

73
00:03:07.170 --> 00:03:10.180
Code. The harness sends it to the model

74
00:03:10.230 --> 00:03:11.370
on a server.

75
00:03:11.420 --> 00:03:14.170
The model makes a plan and sends

76
00:03:14.220 --> 00:03:15.556
commands back.

77
00:03:15.630 --> 00:03:18.460
The harness passes the commands to your computer,

78
00:03:18.510 --> 00:03:21.010
which runs them and returns the results to

79
00:03:21.060 --> 00:03:21.919
the harness.

80
00:03:21.970 --> 00:03:24.100
The results go back to the model, which

81
00:03:24.150 --> 00:03:26.980
changes its plan, and the loop repeats until

82
00:03:27.030 --> 00:03:28.340
you get the answer.

83
00:03:28.390 --> 00:03:31.220
It can run for minutes, hours, even

84
00:03:31.270 --> 00:03:31.917
days.

85
00:03:31.990 --> 00:03:33.480
Take a literature review.

86
00:03:33.530 --> 00:03:35.520
He used to log into Elicit and let

87
00:03:35.570 --> 00:03:37.214
it find and summarise the papers.

88
00:03:37.264 --> 00:03:40.620
Now his agent sends the searches to Elicit

89
00:03:40.670 --> 00:03:43.840
through its API, and every step is shown

90
00:03:43.890 --> 00:03:44.754
and saved.

91
00:03:44.804 --> 00:03:47.130
For a paper he is writing, the agent

92
00:03:47.180 --> 00:03:49.930
built a collection in his Zotero reference

93
00:03:49.980 --> 00:03:53.320
manager, with subcollections, a note on why each

94
00:03:53.370 --> 00:03:54.869
paper is there, and tags.

95
00:03:54.920 --> 00:03:57.546
Done by hand, this is slow and inconsistent.

96
00:03:57.600 --> 00:04:00.070
It also built him an interactive visualiser of

97
00:04:00.120 --> 00:04:03.380
the references. He clicks a category, sees its

98
00:04:03.430 --> 00:04:06.010
papers, and inspects the evidence for each.

99
00:04:06.060 --> 00:04:08.390
That saves him time for reading and understanding,

100
00:04:08.440 --> 00:04:10.201
without outsourcing the review.

101
00:04:10.251 --> 00:04:12.320
A year ago, he spent half a weekend

102
00:04:12.370 --> 00:04:15.030
using Google AI Studio to pull quotes from

103
00:04:15.080 --> 00:04:17.830
a handful of videos, copying and pasting every

104
00:04:17.880 --> 00:04:18.976
prompt by hand.

105
00:04:19.030 --> 00:04:21.180
The weekend before the lecture, he spent about

106
00:04:21.230 --> 00:04:23.640
the same time on a much more ambitious

107
00:04:23.690 --> 00:04:26.523
project, with a desktop agent.

108
00:04:26.573 --> 00:04:29.010
From two thousand and eight to twenty fifteen

109
00:04:29.060 --> 00:04:31.050
he had a radio show on a community

110
00:04:31.100 --> 00:04:33.340
station in Norwich, and he had kept the

111
00:04:33.390 --> 00:04:34.078
recordings.

112
00:04:34.130 --> 00:04:37.120
On Friday they sat on an external hard

113
00:04:37.170 --> 00:04:40.090
drive, and he was not sure what was

114
00:04:40.140 --> 00:04:40.661
in them.

115
00:04:40.711 --> 00:04:43.280
By Sunday he had published a website with

116
00:04:43.330 --> 00:04:45.740
clips to play, a list of every programme,

117
00:04:45.790 --> 00:04:47.860
a calendar and a list of all the

118
00:04:47.910 --> 00:04:48.428
guests.

119
00:04:48.480 --> 00:04:50.960
Codex acted as the main researcher.

120
00:04:51.010 --> 00:04:53.930
It transcribed the recordings with a speech recognition

121
00:04:53.980 --> 00:04:56.570
model, tried a second one for comparison, and

122
00:04:56.620 --> 00:04:59.100
used another local model to extract the guests

123
00:04:59.150 --> 00:05:00.985
and what each interview was about.

124
00:05:01.035 --> 00:05:03.080
It chose many tools on its own.

125
00:05:03.130 --> 00:05:05.540
It knew there was music in the recordings

126
00:05:05.590 --> 00:05:08.930
and used a music identification model without

127
00:05:08.980 --> 00:05:10.012
being asked.

128
00:05:10.062 --> 00:05:12.560
Everything it did was saved on his drive,

129
00:05:12.610 --> 00:05:13.856
so he took no notes.

130
00:05:13.910 --> 00:05:15.720
He asked the agent what the work would

131
00:05:15.770 --> 00:05:17.360
have taken without AI.

132
00:05:17.640 --> 00:05:20.650
It estimated three to six hundred

133
00:05:20.700 --> 00:05:23.870
hours, or nine to twenty person-weeks,

134
00:05:23.920 --> 00:05:26.166
costing tens of thousands of pounds.

135
00:05:26.220 --> 00:05:28.540
Work that would have taken a project team

136
00:05:28.590 --> 00:05:30.580
weeks can be done by one person in

137
00:05:30.630 --> 00:05:31.470
a few days.

138
00:05:31.520 --> 00:05:33.880
His part was mostly checking in and giving

139
00:05:33.930 --> 00:05:34.485
directions.

140
00:05:34.540 --> 00:05:37.650
The agent acts as research coordinator, and the

141
00:05:37.700 --> 00:05:39.919
researcher as principal investigator.

142
00:05:39.970 --> 00:05:42.030
This works beyond archives.

143
00:05:42.080 --> 00:05:45.260
Researchers can run more interviews or record more

144
00:05:45.310 --> 00:05:47.970
in the lab, and keep the audio, video

145
00:05:48.020 --> 00:05:50.651
and pictures for models to analyse.

146
00:05:50.701 --> 00:05:53.860
Research projects will change: who is on the

147
00:05:53.910 --> 00:05:57.090
team, what aims are realistic, and what it

148
00:05:57.140 --> 00:06:00.490
costs. He used a subscription of two hundred

149
00:06:00.540 --> 00:06:01.554
pounds a month.

150
00:06:01.604 --> 00:06:04.390
If every researcher in a team uses agents,

151
00:06:04.440 --> 00:06:06.590
each may need two to three thousand pounds

152
00:06:06.640 --> 00:06:07.178
a year.

153
00:06:07.230 --> 00:06:10.410
Agents make it quick to produce a paper,

154
00:06:10.460 --> 00:06:13.630
and journals and peer review are overwhelmed

155
00:06:13.680 --> 00:06:14.577
with submissions.

156
00:06:14.630 --> 00:06:17.690
When the machine learning journal Transactions on Machine

157
00:06:17.740 --> 00:06:20.740
Learning Research interviewed ten authors chosen at

158
00:06:20.790 --> 00:06:23.970
random about their submissions, three could not answer

159
00:06:24.020 --> 00:06:26.295
basic questions about their own papers.

160
00:06:26.350 --> 00:06:29.600
All of this is happening very fast, and

161
00:06:29.650 --> 00:06:32.400
most of it was not true a year

162
00:06:32.450 --> 00:06:32.739
ago.

163
00:06:32.790 --> 00:06:35.910
Young researchers use methods that did

164
00:06:35.960 --> 00:06:39.310
not exist until now, so their supervisors do

165
00:06:39.360 --> 00:06:40.916
not know how to supervise them.

166
00:06:40.990 --> 00:06:44.250
Trainers who know what is current are scarce,

167
00:06:44.300 --> 00:06:46.903
which is why he is running more workshops.