WEBVTT 1 00:00:00.000 --> 00:00:02.550 In this lecture to MSc students, Dominik 2 00:00:02.710 --> 00:00:05.960 Lukeš describes his switch from ChatGPT to a 3 00:00:06.010 --> 00:00:07.325 desktop agent. 4 00:00:07.410 --> 00:00:09.160 He asks the room when they last used 5 00:00:09.210 --> 00:00:12.210 ChatGPT, and his own answer is not regularly 6 00:00:12.260 --> 00:00:14.794 in months and not at all last week. 7 00:00:14.920 --> 00:00:17.580 On his computer he opens Codex, now the 8 00:00:17.630 --> 00:00:20.950 ChatGPT desktop app, or Claude Code, 9 00:00:21.000 --> 00:00:22.705 with the same models as before. 10 00:00:22.830 --> 00:00:25.820 His token chart starts in November, stays 11 00:00:25.870 --> 00:00:29.169 small until June, then climbs steeply. 12 00:00:29.290 --> 00:00:32.200 His profile shows fifty point two 13 00:00:32.250 --> 00:00:35.580 billion lifetime tokens, a longest 14 00:00:35.630 --> 00:00:38.910 task of fifteen hours and a streak of 15 00:00:38.960 --> 00:00:40.465 one hundred and eighty-four days. 16 00:00:40.590 --> 00:00:43.500 He had kept two hundred and seven hours 17 00:00:43.550 --> 00:00:45.890 of recordings from his years as a volunteer 18 00:00:45.940 --> 00:00:48.748 broadcaster at a community radio station. 19 00:00:48.870 --> 00:00:51.780 He asked Codex to transcribe the archive and 20 00:00:51.830 --> 00:00:54.385 build a catalogue of shows and clips. 21 00:00:54.470 --> 00:00:56.720 It took a weekend, with the agent doing 22 00:00:56.770 --> 00:00:57.587 most of the work. 23 00:00:57.710 --> 00:01:00.870 Given recent IT security news, he asked 24 00:01:00.920 --> 00:01:03.910 Codex to audit his accounts and the software 25 00:01:03.960 --> 00:01:05.360 on his machines. 26 00:01:05.480 --> 00:01:08.230 Among the findings was an online server he 27 00:01:08.280 --> 00:01:11.002 had been given for free and forgotten. 28 00:01:11.120 --> 00:01:13.130 A chatbot can tell you how to fill 29 00:01:13.180 --> 00:01:14.849 in an expenses form. 30 00:01:14.910 --> 00:01:17.740 An agent can find the form, search email 31 00:01:17.790 --> 00:01:20.870 for receipts, calculate totals, process the 32 00:01:20.920 --> 00:01:22.729 receipts and fill it in. 33 00:01:22.850 --> 00:01:25.520 In his York workshops example it pulled up 34 00:01:25.570 --> 00:01:27.660 the emails and forms and filled in the 35 00:01:27.700 --> 00:01:28.448 claim. 36 00:01:28.530 --> 00:01:30.620 It also filled in the bank form, from 37 00:01:30.670 --> 00:01:33.090 email to completed forms to money in the 38 00:01:33.140 --> 00:01:33.616 bank. 39 00:01:33.740 --> 00:01:36.980 He places Codex in the story since 40 00:01:37.030 --> 00:01:40.350 ChatGPT launched in twenty twenty-two, when 41 00:01:40.400 --> 00:01:42.980 AI meant an oracle that answers 42 00:01:43.030 --> 00:01:45.950 questions, summarises, translates and 43 00:01:46.000 --> 00:01:47.481 drafts emails. 44 00:01:47.560 --> 00:01:50.560 ChatGPT, Gemini and Claude defined that 45 00:01:50.610 --> 00:01:53.020 phase, together with specialist apps such as 46 00:01:53.070 --> 00:01:55.515 NotebookLM and Elicit. 47 00:01:55.640 --> 00:01:58.880 The next phase was vibe coding, AI 48 00:01:58.930 --> 00:02:00.040 as a tool maker. 49 00:02:00.330 --> 00:02:02.580 AI writes the code, and the user goes 50 00:02:02.630 --> 00:02:04.638 by the vibe of what they want. 51 00:02:04.690 --> 00:02:07.850 Tools such as Cursor, Lovable and Google AI 52 00:02:07.900 --> 00:02:10.730 Studio produced scripts, dashboards, 53 00:02:10.780 --> 00:02:12.444 simulations and workflows. 54 00:02:12.570 --> 00:02:15.970 The current phase is the desktop agent, 55 00:02:16.020 --> 00:02:18.667 such as Codex or Claude Code. 56 00:02:18.790 --> 00:02:22.110 In Among the Agents, in twenty twenty-six, 57 00:02:22.160 --> 00:02:24.420 Dean W. Ball wrote that by the end 58 00:02:24.470 --> 00:02:26.140 of this year the least important thing you 59 00:02:26.190 --> 00:02:27.870 will be able to do with frontier AI 60 00:02:27.920 --> 00:02:31.080 systems will be getting chatbots to answer 61 00:02:31.130 --> 00:02:31.967 questions. 62 00:02:32.090 --> 00:02:34.178 Two things made this possible. 63 00:02:34.230 --> 00:02:37.300 Models now understand the task, plan, 64 00:02:37.350 --> 00:02:40.020 choose the tools on the computer and revise 65 00:02:40.070 --> 00:02:42.320 after seeing the results, and they can run 66 00:02:42.370 --> 00:02:43.197 for longer. 67 00:02:43.320 --> 00:02:45.820 A harness is the software installed on the 68 00:02:45.870 --> 00:02:48.370 computer that talks to the model, supplies the 69 00:02:48.420 --> 00:02:50.230 tools and returns the results. 70 00:02:50.350 --> 00:02:53.180 In a chatbot, the model reads the prompt 71 00:02:53.230 --> 00:02:55.934 and any documents and writes a response. 72 00:02:56.060 --> 00:02:58.140 In an agent, the model makes a plan 73 00:02:58.190 --> 00:03:01.110 and calls tools, and the harness runs them 74 00:03:01.160 --> 00:03:02.259 on the computer. 75 00:03:02.310 --> 00:03:04.480 His definition is a model using tools in 76 00:03:04.530 --> 00:03:06.950 a loop to achieve complex tasks that require 77 00:03:07.000 --> 00:03:08.247 multiple steps. 78 00:03:08.370 --> 00:03:11.360 In his downloads folder example, the model asks 79 00:03:11.410 --> 00:03:13.680 for a list of files and sizes. 80 00:03:13.800 --> 00:03:16.380 The harness, ordinary software with no 81 00:03:16.430 --> 00:03:19.670 intelligence, passes the command to the computer 82 00:03:19.720 --> 00:03:21.778 and returns the results. 83 00:03:21.900 --> 00:03:25.220 Seeing many PDFs, the model changes its plan 84 00:03:25.270 --> 00:03:28.530 and asks for the sizes of only those. 85 00:03:28.650 --> 00:03:31.970 Loops that once lasted about an hour can 86 00:03:32.020 --> 00:03:33.893 now last a week. 87 00:03:34.010 --> 00:03:37.340 Models now follow many more instructions without 88 00:03:37.390 --> 00:03:40.550 confusing or forgetting them, so they no longer 89 00:03:40.600 --> 00:03:43.670 fall for 'ignore previous instructions' 90 00:03:43.720 --> 00:03:45.379 hidden in a paper. 91 00:03:45.500 --> 00:03:48.820 Training in real computer environments also taught models 92 00:03:48.870 --> 00:03:51.527 which problems need external tools. 93 00:03:51.650 --> 00:03:54.730 In the radio project, Codex led the 94 00:03:54.780 --> 00:03:58.020 research, using cheaper models and open-source 95 00:03:58.070 --> 00:04:00.199 models running on his own computer. 96 00:04:00.320 --> 00:04:02.740 It chose to run small experiments of its 97 00:04:02.790 --> 00:04:05.691 own, such as which model transcribes best. 98 00:04:05.750 --> 00:04:08.990 For large, consequential research, the user needs to 99 00:04:09.040 --> 00:04:10.133 be more involved. 100 00:04:10.250 --> 00:04:13.330 Take a search for last year's evaluation numbers, 101 00:04:13.380 --> 00:04:15.550 remembered as being in a report by a 102 00:04:15.600 --> 00:04:18.061 colleague called Geoff, who has left. 103 00:04:18.180 --> 00:04:21.180 A person searches his name, opens a handover 104 00:04:21.230 --> 00:04:24.470 document and the February progress report, and searches 105 00:04:24.520 --> 00:04:27.236 inside files for evaluation. 106 00:04:27.320 --> 00:04:29.375 None of it turns up the numbers. 107 00:04:29.500 --> 00:04:32.162 An agent gets the same vague request. 108 00:04:32.212 --> 00:04:34.800 It searches across the year, finds that George 109 00:04:34.850 --> 00:04:37.980 did the analysis in March, and reports 148 110 00:04:38.380 --> 00:04:41.330 responses with an average rating of 4.3 111 00:04:41.760 --> 00:04:42.593 out of 5. 112 00:04:42.710 --> 00:04:44.890 The model sees only what it is looking 113 00:04:44.940 --> 00:04:47.270 at, so the gain is an exhaustive search 114 00:04:47.320 --> 00:04:49.220 where a person lacks the time. 115 00:04:49.340 --> 00:04:52.340 He no longer uses Deep Research, and he 116 00:04:52.390 --> 00:04:55.630 never asks AI to summarise a paper, because 117 00:04:55.680 --> 00:04:58.590 a summary is a shorter version of something 118 00:04:58.640 --> 00:05:00.364 he does not understand. 119 00:05:00.490 --> 00:05:02.820 He asks for an interactive page or a 120 00:05:02.870 --> 00:05:04.274 simulation instead. 121 00:05:04.400 --> 00:05:07.470 For a paper he is writing, Codex searched 122 00:05:07.520 --> 00:05:10.520 Elicit through its API, imported the papers into 123 00:05:10.570 --> 00:05:13.730 Zotero and organised them with subcollections, 124 00:05:13.780 --> 00:05:16.690 systematic tags and a note on why each 125 00:05:16.740 --> 00:05:17.624 was chosen. 126 00:05:17.750 --> 00:05:20.250 It then built a visualisation of the references 127 00:05:20.300 --> 00:05:23.016 that recommends papers against his criteria. 128 00:05:23.140 --> 00:05:26.130 He runs small experiments to decide between 129 00:05:26.180 --> 00:05:27.991 options for something digital. 130 00:05:28.110 --> 00:05:31.440 One prompt about compressing images got Codex 131 00:05:31.490 --> 00:05:34.070 to produce a side-by-side web page and a 132 00:05:34.120 --> 00:05:37.030 table of savings and trade-offs in four minutes 133 00:05:37.080 --> 00:05:38.811 and thirteen seconds. 134 00:05:38.890 --> 00:05:41.560 It made no difference for his images, but 135 00:05:41.610 --> 00:05:44.506 he had evidence instead of a guess. 136 00:05:44.630 --> 00:05:47.130 The key to making this work over time 137 00:05:47.180 --> 00:05:48.497 is context. 138 00:05:48.580 --> 00:05:51.000 A model answers from its own knowledge and 139 00:05:51.050 --> 00:05:53.249 from context, and nothing else. 140 00:05:53.370 --> 00:05:55.870 In a real session the context holds a 141 00:05:55.920 --> 00:05:58.920 system prompt, system tools, custom agents, 142 00:05:58.970 --> 00:06:02.200 memory files, skills and messages: a hundred and 143 00:06:02.250 --> 00:06:05.420 sixteen point eight thousand tokens, twelve percent of 144 00:06:05.470 --> 00:06:06.768 a million-token window. 145 00:06:06.890 --> 00:06:09.470 To provide the right context at the right 146 00:06:09.520 --> 00:06:12.680 time, the answer is files and folders, which 147 00:06:12.730 --> 00:06:14.410 agents search well. 148 00:06:14.460 --> 00:06:16.220 At least the top folder should hold a 149 00:06:16.270 --> 00:06:18.190 file called AGENTS.md, 150 00:06:18.820 --> 00:06:21.650 which the harness loads automatically. 151 00:06:21.700 --> 00:06:25.110 A skill is a folder with a SKILL.md 152 00:06:25.160 --> 00:06:27.372 file of pre-made instructions. 153 00:06:27.490 --> 00:06:30.600 He has used skills nine thousand seven hundred 154 00:06:30.650 --> 00:06:32.120 and eighty-three times. 155 00:06:32.240 --> 00:06:35.700 What leaves the computer depends on the task. 156 00:06:35.780 --> 00:06:39.020 Organising the downloads folder sends the model only 157 00:06:39.070 --> 00:06:41.980 names and sizes, while organising 158 00:06:42.030 --> 00:06:44.815 pictures means the model has to see them. 159 00:06:44.940 --> 00:06:47.270 A harness rule can allow a folder to 160 00:06:47.320 --> 00:06:49.294 be listed but never opened. 161 00:06:49.420 --> 00:06:52.740 Agents save time on the task, and people 162 00:06:52.790 --> 00:06:54.679 then try many more things. 163 00:06:54.800 --> 00:06:56.640 The bad news is that there are so 164 00:06:56.690 --> 00:06:58.259 many more things to try. 165 00:06:58.340 --> 00:07:01.330 The radio archive would never have been started 166 00:07:01.380 --> 00:07:04.050 without an agent, and every new tool is 167 00:07:04.100 --> 00:07:05.340 a learning project. 168 00:07:05.460 --> 00:07:08.210 He has given his agent his banking 169 00:07:08.260 --> 00:07:11.090 details so it can send him money, but 170 00:07:11.140 --> 00:07:13.970 not his credit card, because that is how 171 00:07:14.020 --> 00:07:15.418 money gets lost. 172 00:07:15.540 --> 00:07:17.960 He says people should not trust that an 173 00:07:18.010 --> 00:07:19.821 agent did the right thing. 174 00:07:19.900 --> 00:07:22.810 They should build in checking and double-checking, 175 00:07:22.860 --> 00:07:25.370 depending on how much the result matters.