WEBVTT

1
00:00:00.000 --> 00:00:02.590
Dominik Lukeš gave this talk on the thirtieth

2
00:00:02.640 --> 00:00:02.750
of September

3
00:00:03.260 --> 00:00:05.860
twenty twenty-six, to a group concerned with

4
00:00:05.910 --> 00:00:09.320
admissions. He covered what had changed in AI

5
00:00:09.480 --> 00:00:10.254
recently.

6
00:00:10.330 --> 00:00:13.390
He picked five trends from the

7
00:00:13.440 --> 00:00:16.540
many: capability, agents,

8
00:00:16.700 --> 00:00:19.840
voice, price and detection.

9
00:00:19.890 --> 00:00:23.280
The first is that models keep improving,

10
00:00:23.390 --> 00:00:25.324
constantly and rapidly.

11
00:00:25.400 --> 00:00:28.300
A new model version now arrives every few

12
00:00:28.350 --> 00:00:29.028
weeks.

13
00:00:29.100 --> 00:00:32.320
On general benchmarks, not super reliable but an

14
00:00:32.370 --> 00:00:35.500
indication, scores rose by fifty-seven

15
00:00:35.550 --> 00:00:37.370
to seventy-six per cent.

16
00:00:37.420 --> 00:00:40.170
Any assumption about what AI can do today

17
00:00:40.220 --> 00:00:42.689
is likely to be out of date soon.

18
00:00:42.760 --> 00:00:45.590
AI began as an oracle that answers questions

19
00:00:45.640 --> 00:00:47.300
and hallucinates references.

20
00:00:47.350 --> 00:00:49.710
Last year it became a toolmaker that makes

21
00:00:49.760 --> 00:00:52.980
software. This year it is a tool user,

22
00:00:53.100 --> 00:00:56.000
using tools to find things and manage

23
00:00:56.050 --> 00:00:58.672
data. The earlier roles remain.

24
00:00:58.750 --> 00:01:01.420
For his York workshops, he had to pay

25
00:01:01.470 --> 00:01:03.440
for travel and a hotel and claim the

26
00:01:03.490 --> 00:01:04.338
money back.

27
00:01:04.410 --> 00:01:06.620
Codex found the forms in his email, collected

28
00:01:06.670 --> 00:01:08.480
the receipts, filled everything in and left the

29
00:01:08.530 --> 00:01:11.330
claim in his drafts, and the money arrived.

30
00:01:11.400 --> 00:01:14.310
Agents drive the growth in capability, and in

31
00:01:14.360 --> 00:01:16.822
the industry's money and excitement.

32
00:01:16.900 --> 00:01:19.180
Most policies and assumptions still sit in the

33
00:01:19.230 --> 00:01:22.440
chat box: write me a letter, summarise

34
00:01:22.490 --> 00:01:25.400
this. Those tasks are no longer the most

35
00:01:25.450 --> 00:01:27.364
interesting thing AI does.

36
00:01:27.440 --> 00:01:30.340
Agents have become personal assistants.

37
00:01:30.390 --> 00:01:33.140
OpenClaw drew attention in February, then

38
00:01:33.190 --> 00:01:36.560
SpaceXAI launched Grok Bot in August,

39
00:01:36.610 --> 00:01:39.440
Meta launched Muse in September and

40
00:01:39.490 --> 00:01:42.416
OpenAI launched dots the day before the talk.

41
00:01:42.470 --> 00:01:44.900
They are always on in a cloud virtual

42
00:01:44.950 --> 00:01:48.170
computer, monitoring email and working around

43
00:01:48.220 --> 00:01:48.920
the clock.

44
00:01:48.990 --> 00:01:52.440
Without being asked to check anything, dots

45
00:01:52.490 --> 00:01:54.850
asked whether he was sure about the date

46
00:01:54.900 --> 00:01:56.509
on his opening slide.

47
00:01:56.580 --> 00:01:58.400
He was not impressed at first.

48
00:01:58.450 --> 00:01:59.513
Now he was.

49
00:01:59.590 --> 00:02:02.650
Live voice models can now listen until they

50
00:02:02.700 --> 00:02:05.130
are called upon, where taking turns used to

51
00:02:05.180 --> 00:02:06.031
be hard.

52
00:02:06.081 --> 00:02:09.300
Until recently they were less capable, because liveness

53
00:02:09.350 --> 00:02:11.630
was paid for with intelligence.

54
00:02:11.680 --> 00:02:13.696
Now they can call stronger models.

55
00:02:13.770 --> 00:02:16.670
For interviews, someone could have an assistant in

56
00:02:16.720 --> 00:02:18.914
their ear that knows what is going on.

57
00:02:18.990 --> 00:02:21.040
A synthetic person who takes part in an

58
00:02:21.090 --> 00:02:23.520
interview and shows their room live is not

59
00:02:23.570 --> 00:02:26.710
available yet. What is possible badly now

60
00:02:26.760 --> 00:02:29.970
will probably be possible well, perhaps soon,

61
00:02:30.020 --> 00:02:31.491
but not for everything.

62
00:02:31.560 --> 00:02:34.640
The twenty-dollar monthly price dates from twenty

63
00:02:34.690 --> 00:02:37.860
twenty-three. The day before the talk OpenAI

64
00:02:38.020 --> 00:02:40.796
added a five-hundred-dollar monthly tier.

65
00:02:40.870 --> 00:02:42.920
If you pay more, you can do more,

66
00:02:42.970 --> 00:02:46.030
so assumptions about equity and equality go back

67
00:02:46.080 --> 00:02:47.459
to the starting point.

68
00:02:47.530 --> 00:02:50.410
Detectors are back, and they are just AI

69
00:02:50.490 --> 00:02:53.700
too. Pangram claims low false positives,

70
00:02:53.750 --> 00:02:56.273
and Substack has built it into its app.

71
00:02:56.350 --> 00:02:59.480
Anecdotes cut both ways: some say Pangram is

72
00:02:59.530 --> 00:03:02.590
perfect, others say it wrongly flags human writing

73
00:03:02.640 --> 00:03:03.973
as AI-written.

74
00:03:04.050 --> 00:03:07.180
Agents make experiments easier, so he ran

75
00:03:07.230 --> 00:03:10.210
one. He took ten of his own texts

76
00:03:10.260 --> 00:03:13.250
from before twenty twenty-three, five in English and

77
00:03:13.300 --> 00:03:16.200
five in Czech, and had AI edit each

78
00:03:16.250 --> 00:03:19.101
for clarity and rewrite it from an outline.

79
00:03:19.151 --> 00:03:22.120
His agent scanned all fifty with Pangram 4

80
00:03:22.260 --> 00:03:23.663
and produced the results table.

81
00:03:23.740 --> 00:03:26.560
On that extremely small data set the

82
00:03:26.610 --> 00:03:29.830
detector was not perfect, but it gave no

83
00:03:29.880 --> 00:03:32.630
false positives. That will not hold for

84
00:03:32.680 --> 00:03:35.890
everything: results differ between Czech and English

85
00:03:35.940 --> 00:03:38.547
and depend on language and context.

86
00:03:38.600 --> 00:03:40.720
He ran the experiment in a few hours

87
00:03:40.770 --> 00:03:43.990
one afternoon. Such experiments were once quite

88
00:03:44.040 --> 00:03:44.655
costly.

89
00:03:44.730 --> 00:03:47.710
A score on its own cannot settle

90
00:03:47.760 --> 00:03:50.510
authorship, he said, but the tool is worth

91
00:03:50.560 --> 00:03:53.000
testing on existing data, which agents now make

92
00:03:53.050 --> 00:03:55.790
cheap. False positives are worse than false

93
00:03:55.840 --> 00:03:56.597
negatives.

94
00:03:56.670 --> 00:03:58.870
He proposed assembling a data set of past

95
00:03:58.920 --> 00:04:02.060
submissions to measure the false-positive rate:

96
00:04:02.110 --> 00:04:05.066
experimentation, not implementation.

97
00:04:05.140 --> 00:04:08.120
Everything about capability is now a dated

98
00:04:08.170 --> 00:04:10.840
observation, and most policies still assume a

99
00:04:10.890 --> 00:04:13.870
chatbot. Voice is different, and

100
00:04:13.920 --> 00:04:16.830
rising prices mean people have different capabilities

101
00:04:16.880 --> 00:04:19.086
depending on how much they can pay.

102
00:04:19.160 --> 00:04:21.720
The assumption that something is safe from AI

103
00:04:21.800 --> 00:04:23.662
is not a safe assumption.

104
00:04:23.740 --> 00:04:26.560
His answer was to raise the standards.

105
00:04:26.610 --> 00:04:28.820
A researcher told him that AI is now

106
00:04:28.870 --> 00:04:31.070
his research assistant, so the question is what

107
00:04:31.120 --> 00:04:33.710
a clever undergraduate or postgraduate should be able

108
00:04:33.760 --> 00:04:34.983
to do to contribute.

109
00:04:35.060 --> 00:04:37.030
The group should start thinking about the new

110
00:04:37.080 --> 00:04:38.820
baseline for competence.

111
00:04:38.870 --> 00:04:40.186
He had no answer yet.