OpenAI Just Gave Codex a Superpower & More AI News You Can Use

Welcome to another week in AI. There's a lot of things that were released. I think honestly, most of them do not matter. Look, I know this can be a lot, but these are all the stories we looked at this week. And some of it sounds really fancy, but doesn't work as you would expect it to. Whereas other stuff is just not practical to a consumer who's just trying to get things done. And that's why this Friday we'll be focusing on the stories that make a difference. I think there's two things here really. First one is OpenAI Codex Record and Replay, where you can just do something, have AI record you while you do it, and then it can repeat that. And this one actually works. We've seen that idea before, plus some chat GPT upgrades. That's what we're going to talk about today. This episode of AI News You Can Use, where we basically round up the stuff that matters in the AI space. And yeah, as per usual, I'm on the go in the Slovakian mountains. As you can see, this is real. Let's get into it. Starting with Codex. The only unfortunate thing about this story is that it's Mac only. And you do need a subscription. But I think this is going to be widely spread in a few months now, probably in the free plans too. I mean, this is premium functionality. The breakdown of it goes like this. Inside of Codex, you enable this functionality. And when you hit record, you have 30 minutes to record any workflow on your computer. It records your microphone and it records what you're doing. And as you do it, it creates what they call a skill, which is a bit more in-depth than Claude skills. And then you can reuse that skill to redo the workflow that you recorded at any time. You could even schedule it out. Now, some of you might be saying, hey, Igor, but this has been a thing before. Different providers, maybe most notably Anthropic's Claude, have this. If you use the Anthropic Claude Chrome extension that I have right here, there's the exact functionality in there where you hit this button, teach Claude, you enable the microphone, and yeah, then you can do things and then Claude repeats them while considering the context from your mic. Yes, it's not that reliable though. This has been around for a while and it sounds so good in theory, but for me in practice and the team that tested it too, the one that came out in Codex this week is just more reliable than the Claude one. This might change in the future, but right now my initial take after testing Codex and testing Claude is Codex is so much more reliable that it actually makes me want to use it. Whereas Claude is kind of just under that threshold. You can make up your own mind, you know, at the end of the day, I'm just here to show you stuff and you can try it yourself. But what we're going to be doing now, we're going to be doing this workflow in Codex. Again, Mac only, and you need at least a $20 plan to do this. But if you have that, this is what it looks like. You head into Codex desktop application and here on the plugins, it's not one of these plugins in here, but you need to go to create on the top right. And then you go to this new option up here called record skill. Now, this might be geofenced too. I know, ugh, paid plan, Mac, and only certain regions. But hey, we'd like to feature that on this channel, the stuff that is brand new and that actually works really well for everybody else. Hey, there's the Claude one that you can use right now. But yeah, back to Codex. So what we did in our testing is we tried a few examples. And I thought one of the most interesting ones is just something manual that you would do on a computer, something like uploading files to a service, like we do with some of our Shorts every week, and then moving them from one folder to the next. So this is what we're gonna do here. We have one folder with Shorts over here, okay? These are finished edited Shorts from our editor ready to be added. Then we want to upload them to the YouTube channel through the browser. Simple as that, you know, click upload and then add them. And then as a Short has been uploaded, we want to move it from one folder to another one, which is basically an archive of uploaded Shorts. It's really a 3-step process, but with the upload we just have to make sure it's a draft. So here's what's going to happen. We're just going to navigate up here and say record a skill. And as we select that, we get the choice of in which project we want to do it. I'm just gonna say don't work in a project. We're just gonna record a general skill with no extra context. And now when I hit enter, it's gonna start, well, recording. It runs for 30 minutes and I can just do something. So I'm gonna do exactly what I said I was going to do. I'm going to go into the desktop. You can see the Shorts right here. I'm going to upload upload one of them to YouTube. And then additionally, you can actually comment over it. So whatever I say into the microphone as I record this will be considered as context. So in practice, when we were trying this the first time, essentially what you do is you comment on every little step just like you would to a coworker, right? You would be saying things like, hey, now we're going to take the files from this folder, the order doesn't really matter. And we're going to go into YouTube Studio up here, go to this upload button, and then add one of the Shorts. As soon as the Short is uploading, we will switch it to be a draft. And then once the short is done uploading, we will take the file on the desktop, verify if it's the same file name with YouTube Studio, and then we will move the file from one folder into the next. And that concludes our workflow. Everything you just said there will be transcribed and included as context for the skill that will be repeating this. Pretty neat. So I can tell you in practice, this took about 10 minutes in total, and we edited it down for you. But the beautiful thing is you take these 10 minutes once, and now I can type slash and find the skill that I recorded. And when I hit enter, it's going to execute the same skill. And look at that right here, it's taking the second video file, it's uploading it to YouTube Studio, and voilà, it's in a new folder. Now you could go more complex, you could do 5-step, 10-step workflows. I think at a certain point, it's probably going to start falling apart. Even with this one that is like 3 steps, we found that by running it a few times, well, the first 1 or 2 times, you're going to run into things like, hey, can I have permissions to this? And are you sure I can use the browser and things like that? So I would recommend running the skill a few times as after you recorded it. But then once you're over those hurdles, there's just what I mentioned in the beginning, which is that slight unreliability. It's not 100% reliable like a human who has been doing the work for a few years. I mean, sure, humans are never 100%, but you know what I mean in terms of, hey, move the file from here to here. If something goes wrong, a human self-corrects or he finds a replacement for the work that needs to be done, whereas AI kind of just gets stuck and that's that. And it did happen even in our testing. Like right here where it copied an unrelated screenshot into Shorts published. And then it's asking, may I undo that copy? And then you kind of have to tell it, yeah, you can undo your mistake. That's okay, Codex. I think in practice with these things, it's always the same. The idealistic version of it sounds so good. It's like, oh my God, I'm going to do this once and never again. And then in practice, it kind of does it 8 out of 10 times, 9 out of 10 times. Still a huge lift, but you just don't build this like deep trust with the workflow or with the tool here. But I guess what I'm saying is that Codex is just more reliable than all the ones we've tried before. And that's a win. We just haven't reached this point where you like record 50 of these and then you have to work like 1 day a week instead of 5. It's more like you record these and instead of spending, I don't know, 10 minutes on the task, you're just going to check if it ran correctly and then you can run multiple in parallel and kind of more like a human overseer than a human worker. I think that's just the reality today. But nevertheless, it's really exciting to just know that, hey, you can, you know, hit record and then semi-automate things in such a simple manner. I mean, let's be real, clicking and talking, everybody can do that. And I think that's the beauty of this. The other story that I wanted to cover this week is some of the minor ChatGPT updates. There's a bunch, but honestly, I think most of them are like, okay, the image upload on mobile is a bit faster now, great. But I think the one that I thought was interesting as a regular user is this ability to pin chats. Look, you can now pin both chats and projects from this menu. So you can kind of create a favorites list up here. So you could Just pin something and voilà, your pinned chats include that. I think this is really neat. Another one is for all the football or soccer fans. ChatGPT made a football product, which hey, I don't know about you, but during the World Cup, I become a soccer fan. Over here, if you go to chatgpt.com/football, you will see this interface where you can look at all the standings and really easily learn more. Obviously, it's chat. Maybe there's sites out there that do it really well, but just having this inside of ChatGPT, super clean, seeing exactly what games are coming up with all the standings in one place. Is pretty great. And hey, look, look, look, I know soccer fans out there, you're going to have your sites where it's like, hey, I have that. What are you talking about, Igor? This is not unique. I just mean for everybody else who might not have this or wants an even cleaner view that, I don't know, comes without ads. It's right there in chat. You can kind of follow up and ask things about the groups, about the rules, about whatever. I think this is best experience for people who are new to soccer. And then maybe even some more experienced people might enjoy this. So I wanted to show that to you. And by the way, as I live there, Go Portugal. Oh, I guess I should speak about one thing that we tried. We thought this was going to be the lead story. It was Claude Tag, where you can just add Claude and bring it into any conversation. We actually tried this a bunch with the team and gave up on it about 24 hours or something. It's just not as reliable as we want it to be. Now, I can't help but think that maybe we're doing something wrong, but we literally had all our technical moderators try it and talking back and forth. And sometimes it just completely lost the context and thread of what was actually going on and gave nonsensical answers. I thought that idea was really interesting and got really hyped on Twitter, but practically, we don't love it. If you have a different opinion, or if I might be missing something, let me know. Okay, and those are the stories. Honestly, as of the rest, you might know this, but a few months ago, I used to cover everything that came out. And something about that just didn't feel right. I was like, I don't want to represent stories that I do not genuinely care about or want to recommend in these videos. The frame that I always keep in mind for the YouTube channel is, hey, just me and you getting coffee or dinner and you asking, hey, what's new in AI that matters? And then, you know, I tell you what I think that matters. That's the point of the channel. But if you want to dig in Hey, here's the overview of our research database this week. But that's pretty much the stories from this week. I hope you found something useful or inspiring for yourself. My name is Igor Pogany and I will see you very soon.