WEBVTT 00:00:12.700 --> 00:00:16.591 An agent made this explainer from one paper, in about five minutes: 00:00:16.902 --> 00:00:20.351 the script, the slides, the captions, and the video. 00:00:20.781 --> 00:00:23.247 Every sentence points to the page it came from. 00:00:23.638 --> 00:00:26.825 And the paid voice? The agent never saw the key. 00:00:29.200 --> 00:00:32.250 AI for Research Efficiency. 00:00:32.300 --> 00:00:35.610 Episode 12: Make a narrated explainer of your paper. 00:00:38.700 --> 00:00:41.946 Last time, you mapped a research field with OpenAlex. 00:00:42.485 --> 00:00:47.035 This time, one paper becomes a short narrated video, for a lab's page, 00:00:47.085 --> 00:00:52.254 a class, or a grant's broader impacts. One real run, replayed. 00:00:55.700 --> 00:00:57.382 The folder: the paper, 00:00:57.512 --> 00:01:00.326 an instruction file, and three empty folders, 00:01:00.476 --> 00:01:03.420 for the work, the voice, and the finished video. 00:01:04.130 --> 00:01:07.398 The paper is a 2025 review 00:01:07.448 --> 00:01:11.347 of water management in rice, licensed CC BY: 00:01:11.397 --> 00:01:14.984 anyone may adapt it, with credit. It's the review behind episodes 3 and 4. 00:01:15.834 --> 00:01:17.837 The instruction file sets the rules. 00:01:18.248 --> 00:01:22.806 About 150 words, one sentence per line, each ending with its page. 00:01:23.237 --> 00:01:26.682 Only what the paper says, and every number as printed. 00:01:27.512 --> 00:01:32.501 And one rule for the voice: the agent has no key, and must not look for one. 00:01:32.933 --> 00:01:35.417 It writes a request, and waits for the audio. 00:01:38.900 --> 00:01:42.385 The prompt asks for the script first, checked against the paper; 00:01:42.776 --> 00:01:45.921 then the voice request, the slides, and the video. 00:01:46.611 --> 00:01:51.394 In its first 10 seconds, the agent splits the paper into one text file per page, 00:01:51.605 --> 00:01:53.906 so each sentence can carry its page. 00:01:54.825 --> 00:01:57.287 35 seconds in, a first draft: 00:01:57.597 --> 00:02:01.660 215 words, too long for a minute. It cuts. 00:02:02.570 --> 00:02:03.572 Two more drafts. 00:02:03.983 --> 00:02:06.619 Then it checks each figure against the page it cites, 00:02:06.710 --> 00:02:08.533 and writes down every check. 00:02:09.423 --> 00:02:12.470 At 82 seconds: 159 words 00:02:12.520 --> 00:02:15.466 in 13 sentences, each with its page. 00:02:17.300 --> 00:02:22.620 Methane fell by 31 to 62%: page 1, in the abstract. 00:02:23.737 --> 00:02:25.760 Both gases together, 33.6 to 56.2% lower: 00:02:25.810 --> 00:02:29.836 page 3. 00:02:30.907 --> 00:02:34.813 Safe AWD raised yield by 8.9%, over a 5-year experiment: 00:02:34.863 --> 00:02:37.546 page 5. 00:02:38.956 --> 00:02:43.220 Page 5 says one thing more: the experiment was at a single site. 00:02:43.591 --> 00:02:46.994 The first draft said so; cutting for length took it out. 00:02:48.270 --> 00:02:50.792 The numbers are right, but the context is thinner. 00:02:51.223 --> 00:02:53.095 That's the check only you can make: 00:02:53.445 --> 00:02:56.347 does each sentence still mean what the paper means? 00:02:58.100 --> 00:02:59.001 Now the voice. 00:02:59.471 --> 00:03:03.825 The agent writes a request: the text, a stock voice, and the model. 00:03:04.215 --> 00:03:05.196 Then it waits. 00:03:05.961 --> 00:03:09.406 You make the paid request yourself, outside the agent's session, 00:03:09.536 --> 00:03:14.863 with the key from your own .env file. The companion page has a short script for it. 00:03:15.699 --> 00:03:21.038 First an estimate: 966 characters, about 107 credits. 00:03:22.151 --> 00:03:24.859 About 20 seconds later, the audio is back, 00:03:25.070 --> 00:03:30.124 with the start and end of every character. It cost 106 credits. 00:03:30.843 --> 00:03:34.954 The agent never saw the key. It saw two files appear. 00:03:38.900 --> 00:03:41.985 Meanwhile, the agent drew 12 slides in Python: 00:03:42.356 --> 00:03:45.221 one idea each, the key number large. 00:03:45.942 --> 00:03:48.589 It looked at them all at once, and fixed what it saw: 00:03:49.041 --> 00:03:54.015 a clipped number, a drawing over a title, and text running off two slides. 00:03:54.926 --> 00:04:00.477 The character timings become captions: a sentence at a time, at most two lines. 00:04:01.348 --> 00:04:06.462 Then ffmpeg puts it together: each slide while its sentences are spoken, 00:04:06.713 --> 00:04:09.159 the voice, and the captions burned in. 00:04:10.050 --> 00:04:14.555 It checks the result: the length matches the voice, every slide appears, 00:04:14.785 --> 00:04:19.070 and the captions can be read. It made their boxes darker until they could. 00:04:19.756 --> 00:04:22.966 5 minutes and 7 seconds, start to finish. 00:04:26.900 --> 00:04:31.545 Its report is candid: the video runs 89 seconds, not 60. 00:04:31.956 --> 00:04:36.620 This voice reads slowly, and the paper's address alone takes 12 seconds. 00:04:37.566 --> 00:04:40.855 And it names what it could not check: it cannot listen. 00:04:41.266 --> 00:04:45.557 So listen yourself; every number should sound the way the paper prints it. 00:04:46.567 --> 00:04:49.790 Before you share it: each sentence against its page, 00:04:49.960 --> 00:04:52.993 the voice against the numbers, the credit to the paper, 00:04:53.283 --> 00:04:56.105 and a line that says the voice is AI. 00:04:58.100 --> 00:04:59.441 About the voice service. 00:04:59.871 --> 00:05:02.583 ElevenLabs' free plan can't be used commercially, 00:05:02.753 --> 00:05:07.516 and anything you publish from it must have elevenlabs.io in its title. 00:05:08.181 --> 00:05:10.364 Paid plans include a commercial license, 00:05:10.454 --> 00:05:14.259 and what you make during a subscription stays licensed after it ends. 00:05:14.970 --> 00:05:17.434 Never clone someone's voice without their consent, 00:05:17.644 --> 00:05:20.067 and never hide that a voice is AI. 00:05:21.073 --> 00:05:25.966 This series is narrated with ElevenLabs, on a paid plan we use and value. 00:05:26.277 --> 00:05:30.046 A free voice that runs on your own computer is on the companion page. 00:05:31.700 --> 00:05:34.111 Here is what the run made, in full. 00:07:07.700 --> 00:07:11.867 Try it with your own paper: copy the instruction file from the companion page, 00:07:12.118 --> 00:07:13.450 ask for the script first, 00:07:13.701 --> 00:07:17.547 and check every sentence against its page before you make the voice. 00:07:22.100 --> 00:07:26.103 So: pin every sentence to its page, keep the key with you, 00:07:26.433 --> 00:07:30.136 listen before you share, and say the voice is AI. 00:07:31.500 --> 00:07:34.078 Next: how this series is made.