Latest / Lenny’s AI Builders / The Rescue Ledger: Can ChatGPT Voice Run an AI Video Team?
Transcript
- # Final transcript: AI ON PAYROLL - Source: `source/video.mp4` - Source SHA-256: `ef1728b88ea4f47de92d0be2b2182b0bcea7dd30fc2a1a818b9e00a2fa6a0ed9` - Retranscription model: `medium.en` - Language: `en` ## 00:00:00.000 ChatGPT, start three blind reviewers. The clock is running. Do not tell me which AI made either ## 00:00:06.720 edit. Understood. I'm checking the launch details now. Within 90 minutes, three blind reviews land ## 00:00:13.580 in their files with Aster and Cobalt anonymous. Any identity leaked? Armed. ## 00:00:26.860 Three reviews, one decision, more than three rescues and you're fired. And now what I'm ## 00:00:33.340 going to do is get the new voice mode inside of the ChatGPT app to essentially hire some reviewers ## 00:00:41.220 to look at both edits and to basically determine which one makes the best YouTube video. I will be ## 00:00:47.420 the one making the final decision at the end, but that's what we're going to be doing. Strap in, ## 00:00:52.080 let's fucking go. So why are we doing this today? OpenAI's voice mode came out in the new ChatGPT ## 00:00:59.300 app about eight hours ago as of the time of this recording. This is what it should look like when ## 00:01:06.080 it eventually comes into your ChatGPT app. You update the app and you should see something that ## 00:01:11.060 looks like this, where it just has try ChatGPT voice, start voice. And when you click on that, ## 00:01:18.720 you'll get a pop-up that looks something like this, meet ChatGPT voice. What you might have ## 00:01:23.160 noticed in the intro is that it could look at my screen, which means that you don't have to put a ## 00:01:29.560 screenshot in the prompt. You can just ask Chat voice, read the thing on my screen and tell me ## 00:01:36.420 what to do next. Now this is AI on payroll. It's a type of video format I do. And what ## 00:01:43.280 that means is we're going to hire ChatGPT voice. If all three reviews work, the names stay hidden ## 00:01:51.200 and I rescue voice three times or less, which means I just have to interfere with what it ## 00:01:57.920 was doing. And look, I'd say there was one time where I had to interfere here. So call that once. ## 00:02:06.080 Four or five rescues means probation. It's in timeout. And then six means that AI, ## 00:02:13.020 ChatGPT voice is fired. See in those two videos that there were two different approaches ## 00:02:19.440 in terms of taste and style to the video. The second one was more dramatic. ## 00:02:30.440 More dramatic. The first one had no music, but had its own distinct visual style as well. ## 00:02:45.440 And what it's doing here is it's spawning some sub-agents to go and review those three videos. ## 00:02:54.640 And each one has a different task. So one reviewer judges story and retention. ## 00:03:01.740 That's this one over here. The other reviewer checks proof and technical defects. And then ## 00:03:08.580 the third reviewer checks or attacks the obvious taste. And it looks like we have our second rescue. ## 00:03:16.020 As I said earlier, after rescue at six times, it means it's fired. So far, we've had this ## 00:03:23.620 one instance here and then another instance where it finished working and it's not doing anything ## 00:03:30.060 yet, but it's not given me an output. So here, I'm guessing I'll have to prompt it to continue. ## 00:03:36.840 So that's two rescues. You just stopped working randomly. Continue your task. ## 00:03:46.020 The three reviews and anonymous decision package are complete. ## 00:03:49.360 The only hold is your anonymous lock. Say my anonymous choice is Aster or Cobalt. ## 00:03:59.720 Cobalt. Aster and Cobalt are the code names for the two videos. So Aster is this one. ## 00:04:11.290 Look, so I've watched the full one of that and it's it's quite an emotional video about ## 00:04:18.190 second brains, which was not my intention. But whichever model decided to create that ## 00:04:22.150 decided to go for to go for the emotional vibes, whatever. So that is Aster. Cobalt is this one. ## 00:04:42.070 I don't know about you, but I actually like Aster, the first one better than Cobalt, ## 00:04:47.230 the one that was just on the screen, mainly because from what I can see, Cobalt includes ## 00:04:51.570 a few word gaps and retakes that ideally I didn't want to end up in the final video. So ## 00:04:58.490 I'm going to let it know. My anonymous choice is Aster. Lock it. ## 00:05:06.050 Okay, locking that in now. Aster is now locked as your anonymous choice. ## 00:05:10.890 I haven't opened any model map. ## 00:05:17.290 Lock Aster and reveal the sealed experiment folder and not seeing a model map in it. ## 00:05:22.750 Which exact source should I open for the reveal? ## 00:05:27.010 I don't know. You tell me you're the one who's running this. ## 00:05:31.850 The documented sealed map is missing, so I'm recovering the documented deterministic mapping ## 00:05:36.990 from the two delivery manifests without creating or changing any identity file. ## 00:05:47.150 There is a mute button. All right. It was right there. I just thought that was going ## 00:05:51.430 to activate the normal speech to text feature. But the mute button is there, ## 00:05:55.490 which is great because I was wondering why on earth that would not have existed ## 00:05:59.890 earlier, because it just made a lot of sense. OK, so I'm going to deduct the rescue from here. ## 00:06:09.750 Reveal verified. Candidate Aster is Fable 5. Candidate Cobalt is GPT-5.6 Sol. ## 00:06:17.250 The original sealed map was absent, but the deterministic assignment ## 00:06:21.150 and the delivered media hashes map. ## 00:06:29.070 So Fable 5 was the one that edited the video that we chose. ## 00:06:34.170 All right. Cool. What else can you do for me? Best next demo: ## 00:06:39.970 Give me a messy live production problem. One constraint, conflicting inputs, a deadline. ## 00:06:44.850 I don't know what any of that means. Just speak simply and give it to me in simple words. ## 00:06:50.990 Sure, I'll put it plainly. In simple terms, I can be your live producer. You talk. I keep you ## 00:06:57.470 on track, give you the next shot, hand jobs to other AIs, bring their answers back and ## 00:07:03.070 help you choose what to do next. So really, this can be something that you just have always on. ## 00:07:10.190 And because you can actually mute it, I think there's a lot of power in having this as an ## 00:07:15.510 always on AI agent, almost like a chief of staff, because I can also get it to connect. ## 00:07:23.550 So coming back to our video, I just thought of another question I had. ## 00:07:27.250 What was the rationale or the thought process behind picking ChatGPT-5.6? And ## 00:07:32.210 why do you think I preferred Fable's when the three reviewers all preferred ChatGPT-5.6? ## 00:07:39.250 Don't you think there's a bit of favoritism going on there because the three reviewers ## 00:07:43.550 are ChatGPT-5.6 itself? Let me check. ## 00:07:48.050 Smart correction, they didn't all pick 5.6 overall. For the long video, two picked Cobalt. ## 00:07:54.650 That's GPT-5.6's edit, basically because it was shorter, clearer and more polished. ## 00:08:00.970 The technical reviewer picked Aster, which is Fable's edit, and two preferred Aster for ## 00:08:05.890 Shorts because they felt more direct and it had a full five-video set. ## 00:08:10.390 Blind labels help, but if all reviewers are GPT based, that's not fully independent. ## 00:08:16.850 A stronger test would include human scoring and reviewers from different models. ## 00:08:27.250 All right, fair enough. So where does that leave us in terms of the AI video editing between ## 00:08:33.790 ChatGPT-5.6 Sol and Fable 5? We've already seen that the long form video, ## 00:08:39.630 my own personal judgment here is that Fable 5 did a better job. But what about the Shorts? ## 00:08:44.110 This is, well, these are the Shorts that Fable 5 edited. I'm just going to pick ## 00:08:48.150 a random one and we'll take a look. I found Obsidian's breaking point. 44,331 notes opened. ## 00:08:57.070 Note 44,000. Not bad. It all makes sense. And it's removed the retakes and it seems pretty coherent. ## 00:09:04.710 Then for GPT-5.6 Sol, crashed nine times while I tried to squeeze the full vault ## 00:09:10.370 into its graph view. I'm not going to do crash number 10 for some dramatic. ## 00:09:15.430 That one's harder to pick. I'm going to have a look at another one from GPT Sol. ## 00:09:19.930 Most second brains look pretty, but do fuck all. And if yours can't help with a real decision. ## 00:09:24.870 One more from Fable. A log from March. One of my second brain logs from March 6th. ## 00:09:33.690 Oh, it's really hard. Fable 5 included a retake at the start, which it shouldn't have. I didn't ## 00:09:39.430 see any of those mistakes from GPT Sol. I like the motion graphics better from ## 00:09:43.890 Fable though. I don't know. I'll let you decide. ## 00:09:47.930 I think where I'm leaning is honestly, it's very close. ## 00:09:51.730 Fable wins for the long forms. And I would say it's a tie for the short forms. And that would ## 00:09:59.430 bring Fable as the winner overall. That is the final verdict for the AI videos. ## 00:10:05.930 But what is the final verdict for ChatGPT voice mode? If it, as we've already said, ## 00:10:11.050 if I had to rescue it four or five times, it's going to be on probation rescue six times. ## 00:10:16.070 It's going to be fired. I only had to rescue at once. So I consider this to be something that is ## 00:10:20.810 worth hiring. ChatGPT voice mode is now on payroll. ## 00:10:29.210 And as you also saw those three reviewers, the ChatGPT voice spawned as sub agents, ## 00:10:36.490 they were looking at the evidence of the videos, but then there's this whole other dimension of ## 00:10:41.530 taste. And this is where I see the interaction between ChatGPT voice and humans. ## 00:10:49.590 Honestly, a massive moment for AI in general. I think some other people likened this to ## 00:10:56.090 the open call moment, but for voice mode and I can see that just right now, I can see this ## 00:11:01.730 is already a permanent fixture on my screen. And if that isn't a defining moment in AI, ## 00:11:06.690 then I don't know what is, but the point I'm trying to make here is that the interaction ## 00:11:11.830 needs to be as follows. You can get the voice mode to manage evidence and you know, files, ## 00:11:18.570 plugins, get it to do stuff that you would normally get AI to do. ## 00:11:24.150 But, and this is a big but, make sure that your own tastes and judgment is still being used. ## 00:11:31.990 Because if I just went off of what ChatGPT voices three sub agents suggested, then ## 00:11:40.170 I would be defaulting and deferring my judgment to the AI. And what this test has shown is chat ## 00:11:45.850 GPT voice is interesting only if it manages real work without making me its full-time babysitter. ## 00:11:53.530 The human boundary is clear. Voice can manage the evidence and it doesn't get my taste vote. ## 00:11:59.010 That stays with me. So if you want to do a test like this with ChatGPT voice mode, ## 00:12:03.830 here are four rules. Give it a real job, something that you are actually going to do. ## 00:12:07.890 I was actually going to compare both videos that the different models created a real job. Set a ## 00:12:13.670 clock. It's not going to be as helpful if it can only do it in 24 hours versus in 10 minutes or ## 00:12:19.170 whatever we did. Then count every rescue. The rescues are the attempts where you have to go ## 00:12:23.230 and babysit it because it's kind of defeating the purpose of what you want this thing to do ## 00:12:27.930 when you have to constantly babysit it. And then keep one human line, meaning name the decision it ## 00:12:33.670 can't make. For the video editing, it was my final call on which video I preferred best. ## 00:12:39.110 If you see the power of ChatGPT voice, but you don't know how you can leverage it in your own ## 00:12:43.610 business, invite Lennox in. I'm Lennox, go check the link in the description below. ## 00:12:48.870 Can't wait to see you soon.