ChatGPT Codex Voice + browser + Sites: an expert’s AI workflow | Nick Baumann (OpenAI)

Nick Baumann is on the Developer Experience team at OpenAI, where he spends his days building with, testing, and communicating the capabilities of ChatGPT Codex and ChatGPT Work. In this episode, Nick walks me through several features that have launched or evolved recently: the new voice interface with its screen-reading orb, the Heartbeats automation system in ChatGPT Work on mobile, the live ChatGPT Sites deployment feature, and his personal use case for AI-assisted UGC video editing. What you’ll learn: How two-person voice chat worksHow Heartbeats workHow to build and deploy a live website with ChatGPT SitesHow to delegate a flight search, hotel booking, and expense report to Codex in a single voice conversation without opening a single app manuallyWhy ChatGPT Work on mobile is the most underutilized AI workflow for people already using the ChatGPT appHow to use a custom UGC Video plugin to feed 50 raw clips into ChatGPT, let it pull transcripts, pick the best takes, and assemble a finished vertical video overnight— Brought to you by: Bolt.new https://bolt.new/partner/howiai—Turn your idea into a real product Hyperagent https://hyperagent.com/howiai—Deploy fleets of agents that handle real work — In this episode, we cover: (00:00) Introduction to Nick Baumann (02:56) What’s new in Codex (05:40) ChatGPT Work and Heartbeats (06:40) Live Codex voice demo (13:25) Latency vs. intelligence (14:36) Quick recap (15:04) Voice on mobile and the ChatGPT Sites workflow (21:24) Live UGC video demo (32:30) How I AI website results (34:04) Lightning round and final thoughts — Tools referenced: • ChatGPT Codex: https://chatgpt.com/codex • ChatGPT Sites: https://chatgpt.site https://chatgpt.site/ — Where to find Nick Baumann: LinkedIn: linkedin.com/in/nick--baumann http://linkedin.com/in/nick--baumann — Where to find Claire Vo: ChatPRD: https://www.chatprd.ai/ Website: https://clairevo.com/ LinkedIn: https://www.linkedin.com/in/clairevo/ X: https://x.com/clairevo — Production and marketing by https://penname.co/. For inquiries about sponsoring the podcast, email [email protected].

• 0:00Speaker 00

  If we go back to, you know, 2022, 2023, around the release of ChatGPT, I think a lot of people felt the AI then. A chatbot that I say things to and intelligently says things back to me. I think with the advent of coding agents and like the coding agent harness in particular, now this AI is going out and like reading files and creating files, like running commands. And this is crazy. I sell it one thing and then it figures out all the other things on its own.

• 0:22Speaker 01

  I haven't gone deep with voice. I want to see it from the pro. How should we be using it? What's cool about it?

• 0:27Speaker 00

  So I'm just going to trigger voice. I've got a hotkey for it.

• 0:30Speaker 02

  I'm listening. Go ahead.

• 0:32Speaker 00

  Okay, there should be some Amazon receipt for a mic I just bought recently. It's for DX uses. I'm hoping you can help me expense it. Could you spin up another thread to handle taking care of that expense report for me?

• 0:42Speaker 02

  The expense task is running. It will find the Amazon receipt and stage the Nirvan report, but not submit it until you confirm. I'll let you know when it's ready to review.

• 0:51Speaker 01

  If anybody has been lucky enough to have an assistant, this is how it gets done. And it is just like such a nice delegation experience to be like, hey, can you take care of this? Hey, can you fix this? Tell me what's going on.

• 1:02Speaker 00

  Typically, voice is great. You can like talk back and forth, but it's usually with a less strong model that is less capable. Whereas this, it's fully able to delegate and manage these fully five, six sole threads on its own, which is great because then you can have it do things on your behalf.

• 1:17Speaker 01

  Hot take question, what do you think matters more on voice experience latency or intelligence?

• 1:25Speaker 01

  Welcome back to How I AI. I'm Claire Voe, product leader and AI obsessive here on a mission to help you build better with these new tools. Today, we have Nick Baumann at OpenAI, and he's going to show us some of the advanced use cases of ChatGPT Codex and ChatGPT Work, including how you can talk to your computer to book your flights, use ChatGPT sites to build websites you can share with anybody or nobody, and use my favorite workflow, which is editing creator content in Codex. Let's get to it. This episode is brought to you by Bolt New, the AI app builder for people who have ideas and want to ship them. Most AI tools spit out code that looks great in a demo

• 2:09Speaker 01

  and falls apart the second you try to do anything real with it. Or they lock you into their own platform with no real way out. Bolt is different. You describe what you want to build, a startup MVP, a landing page, an internal tool, a side project, and Bolt generates production-ready code in minutes. Connect Stripe or any other MCP, hook up your domain and deploy it live. Founders are using Bolt to build businesses doing real revenue. Product managers are shipping prototypes their teams actually use. Designers and marketers are launching campaigns without waiting in line. Anyone can build, engineering can ship, everyone wins. You just need an idea and a weekend.

• 2:51Speaker 01

  Check it out at bolt.nu/HowIAI. Nick, welcome to How I AI.

• 2:59Speaker 00

  Hello, Claire. Thanks for having me.

• 3:00Speaker 01

  I am going to make you laugh because I might be the number one Codex fan, fanboy, tech pro, fanboy in the world. I am like constantly telling people, wait, but have you, have you tried Codex? Have you tried Codex yet? It is just my daily driver, so I'm so excited to have you on because even though I use it all the time, and for some things that you and I think both think are valuable, I have not touched most of the new features that have come out in the last week. It's really hard to keep up. And so I'm really psyched that you're here to just show us some of the new stuff and maybe how you use Codex both for work and life. So tell us, what are the things in Codex you feel like people don't really know about or are brand new

• 3:44Speaker 01

  that you think are completely going to change how, how we all work?

• 3:47Speaker 00

  Yeah, I feel like, you know, with yesterday's release of voice, the ChatGPT Codex app, it's kind of like the amalgamation of all these primitives we've been like putting together kind of quietly. And now we're super useful. So a lot of people don't know that the ChatGPT app, you can ask it to create threads. You could ask it to message your existing threads. And these threads actually talk to each other, too. And it's like not super kind of forward in the app, but these are all capabilities. And now we've added this voice orchestration layer, really, where now you can, you know, with the hotkey, you can trigger this, this orb that like pops on your screen. And no matter where you are in your computer, you can talk to it, it'll talk back to you. And it can see your screen, it can see what you're working on.

• 4:30Speaker 00

  And because we have those primitives of see your existing threads, create new ones, talk to them, it can like basically manage your entire ChatGPT app for you.

• 4:40Speaker 01

  Yeah, one of the things that I don't want people to miss before we even get into like the fancy voice stuff is I really do feel like most people do not know the meta capabilities of Codex and even like we won't say, but like the, the other one. In that like its ability to not only spin off subagents, but actually start completely new threads, fork threads, look through old data, look through its own memory. It's, it's a really good mechanism to let the AI manage context and decide when forking off different tasks makes a lot of sense. And so I've been doing that a lot, but I actually had to discover that capability sort of organically when one day I was sitting there and it was probably five, six was like,

• 5:23Speaker 01

  I'm going to kick off a new thread. And I was like, you're going to do what? Like, what now? And so I do, I do feel one of the challenges with these core kind of platforms like Codexes and ChatGPT is people just don't know how to discover the features. And they're so rich. And so do you have any, you know, like, we'll talk about voice, but are there any other like sort of edge features of Codex or ChatGPT that you feel like people underuse, not just this, this thread one?

• 5:52Speaker 00

  The ChatGPT Work experience, you know, people are talking about, you know, should we be using ChatGPT Work or Codex? I think ChatGPT Work on the web, specifically mobile, and that's where I use it mostly, is so insanely powerful. It's one, it's, it's essentially like the Codex harness, which has this incredible compaction. You have all your plugins. So if you've like already connected Gmail and Gcal, like whatever you use, it's already there. And then you have it all accessible on your phone. So you can like create these automations, also it has heartbeats, and it's already there. It's like, I mean, a few times a day, I'll just get like a notification saying, hey, like I got this email about this thing you bought. Are you sure you bought that? Or like, hey, this package is arriving. Make, make sure you grab that. And I'm not watching my email or my finances,

• 6:36Speaker 00

  but that's just kind of being monitored for me. And that's, that's really magical experience.

• 6:41Speaker 01

  Okay, I love it. All right. So let's get into ChatGPT Voice and what you, what you use it for. Cause I, it came out yesterday. I've been super busy and around humans. We're going to talk about that a little bit later. So I haven't gone deep with voice. I want to see it from the pro. How should we be using it? What's cool about it?

• 7:00Speaker 00

  Yeah, I, I think it's really magical. So I'm just going to trigger voice. I've got a hotkey for it.

• 7:06Speaker 02

  I'm listening. Go ahead.

• 7:07Speaker 00

  I'm hoping you can help me handle a few logistical things right now. Can you, can you see the, the offsite I've got pulled up in Slack?

• 7:16Speaker 02

  I'll take a look. Let me check that Slack view. Yep, it's the DX team offsite in Paris, August 26 through 28. You're due in Tuesday with the program starting Wednesday morning and wrapping Friday afternoon. Flights and hotel are supposed to be booked in Nirvan this week.

• 7:35Speaker 00

  All right. So I muted myself so that Codex doesn't hear me right now. But what happened there is it can already see my screen. You can see that I've got Slack opened up. And what it did is it took an app shot. And so an app shot is something that you can do or whatever window you have open, you just hit Command, Command. And it essentially takes like a smart screenshot loaded with context. I can use computer use or browser use on it instantly. But, you know, kind of a meta feature, Codex Voice already has that built in. So I'm hoping you can help me book my flights and a hotel for the Paris offsite. I'm also thinking about sticking around through the weekend and maybe a couple days after. Could you maybe like check my calendar, see what my availability is like and see if it's, if it's feasible for me to stick around?

• 8:20Speaker 02

  I'll check your calendar against the offsite dates and then look at travel options.

• 8:24Speaker 01

  Is this using the connectors?

• 8:27Speaker 00

  Good question. Yeah. So this is using the plugins underneath the hood right now. So it's not having to actually go through.

• 8:31Speaker 02

  Clear after the offsite through Monday, August 31st. A Tuesday, September 1st return would conflict with a late afternoon meeting. The simplest plan is returning Monday, August 31st. Do you want me to plan around that?

• 8:45Speaker 00

  You know what? I'm thinking I might just take PTO on that Monday and Tuesday. I'm flying out of SFO. So why don't you just go ahead and use Nirvan, spin up another task to figure out the travel, pick a hotel that's reasonable and within our company policy. And then when it's ready, let me know and I'll, I'll click book now.

• 9:05Speaker 02

  Sure, I'll sort that out. I'll find options for flying out of SFO, arriving Tuesday before the Wednesday start, and returning after Tuesday, September 1st. Then bring you choices that fit policy for you to book.

• 9:20Speaker 01

  So is this chat going to be in? Oh, sorry. I, I am talking to, I'm talking to his Codex. Invite Claire to the Paris offsite. Put her up in the ritz. Get her in business class.

• 9:34Speaker 00

  And honestly, we don't, we don't really have a Paris offsite scheduled yet, but so much of our DX team is French that I'm hoping that I can like, you know, start seeding the, the ideas here.

• 9:44Speaker 02

  That task is running now. Comparing flights from SFO and a practical hotel for August 25th through September 2nd. I'll share the suggested itinerary when it's ready.

• 9:54Speaker 00

  Cool. So we've got this task already running. And make sure it uses my existing Chrome browser for this one.

• 10:01Speaker 02

  I'll check that. I've updated the travel task to use your existing Chrome session and keep it to preparing options, not booking.

• 10:11Speaker 00

  Awesome. So right now it just, it just sent a message back to that task. It's already running. But I think while we've got that going, I'm going to ask it to do another thing for me. While I've got you, I've got an email, I think. There should be some Amazon receipt for a, a mic I just bought recently. It's for DX uses. I'm hoping you can help me expense it. Could you spin up another thread to handle taking care of that expense report for me?

• 10:35Speaker 02

  Sure, I'll take a look. I have started that expense task and it'll use your existing Chrome session. The expense task is running. It will find the Amazon receipt and stage the Navan report, but not submit it until you confirm. I'll let you know when it's ready to review.

• 10:52Speaker 00

  Awesome. So those are both running in the background. I've muted our voice so we can talk without getting interrupted. But we can already see that in Chrome, this task is using the browser. It's figuring out our flight options. It'll figure out the hotel. Frankly, this is something I just enjoy not having to do. It's nice to just have this handled in the background for me.

• 11:13Speaker 01

  What I like about voice is I do think sometimes people with AI really struggle still with like blank chat window syndrome, which is like they look at it and they're like, what am I going to ask? What am I going to ask this brilliant AI to do? And so what I love about first just like voice transcript to text, which a lot of people have been using is it is what our beloved multi-time guest, Hillary calls the yapers API, which is like the best, highest bandwidth way to communicate with an LLM is just to yap out loud and just context them. But what I love about this like voice to voice interaction is if anybody has been lucky enough to have an assistant, like this is how it gets it gets done.

• 11:57Speaker 01

  And it is just like such a nice delegation experience to be like, hey, can you take care of this? Hey, can you fix this? Tell me what's going on, blah, blah, blah, blah, blah. And so I'm guessing the hypothesis here with this voice kind of experience is there's going to be like better discovery of use cases and sort of like more sprawling use cases, because people can sort of delegate in a more natural way as opposed to having to like sit there and think through how to instruct kind of in text the models. Have you found that like there are specific things that you reach for voice with that you don't? That you don't like type with your human fingers? Are you a full voice pill? Like, where are we?

• 12:36Speaker 00

  Yeah, a few thoughts there. And one, I think it's like this iteration itself is like kind of a new primitive in that typically voice is great. You can like talk back and forth, but it's usually with a less strong model that is less capable. Whereas this, it's fully able to delegate and manage these, you know, fully, you know, five, six soul threads on its own, which is great because then you can have it do things on your behalf. Other thoughts there? I find the like the thinking process in voice is a lot better. When it comes to, you know, I could dictate for, you know, two minutes, send off a long paragraph, get back a couple of long paragraphs and read those.

• 13:20Speaker 00

  And it's really hard for me to get in flow. I think of like planning in voice to be really, really effective.

• 13:26Speaker 01

  Hot, hot take question. What do you think matters more on voice experience latency or intelligence? If you could only pick one, I'm going to make you pick one.

• 13:35Speaker 00

  I would say when I use ChatGPT, I put it on a high intelligence voice. I think there's like a middle ground. I think if there are good tools it can delegate to, then I care more about latency. So I guess I'll say latency.

• 13:53Speaker 01

  Yeah, I think I think it's the I think it's what we're going to be talking. And this is the Clara Vau prediction. I think it's we're going to talk more and more about latency in the in the second half of the year, because I just think it's like almost the thing to differentiate on right now is like how real time can these experiences really be? I see a lot of people abandon great, great AI workflows because of a spinner or because of a delay. And so like the more you can close that gap, the more I think people can discover cool things and actually pull through and get some stuff done.

• 14:22Speaker 00

  Yeah, I feel like so many product assumptions in design and AI tooling is like built around this like there's going to be latency. There's going to be delay. And the more we shrink that, we can kind of drop those assumptions and new products will, you know, will emerge.

• 14:36Speaker 01

  Amazing. Okay, so, so what we've seen here, just to reiterate for people who are maybe not watching or listening, is you can hotkey spin up voice. It can manage these like very intelligent five, six threads. You can delegate tasks. It can be hooked up to plugins. It can use your browser. It can use app shots. It can use all these things. And it's just like a very high bandwidth, low friction way to talk to super intelligence that can get nice stuff done. And you mentioned that a little bit earlier that you're also loving like mobile and this combination of like dictation and managing through mobile. Can you tell me how like being able to do some of this on your phone has changed your workflow?

• 15:20Speaker 01

  Maybe like give me an example of that.

• 15:22Speaker 00

  So I put together this site. And so this is just a aggregation of all these really cool tweets that, you know, both people in like the Codex and ChatGPT community and also people who work at OpenAI have shared with how they're using Codex ChatGPT. This was done just with with ChatGPT, like the ChatGPT app where I had it use its own in-app browser, go through Twitter for like a couple of hours and then embed these links and kind of add a little bit of description to them. And now what I do is I've got a thread on my work phone that is essentially managing this site. And when I see something I like, I just drop in that link. I say add it and it does.

• 16:05Speaker 00

  We also have this prompts tab and I do the same thing there. If I have like an idea for how somebody could be maybe he's like newer to like ChatGPT work, I'll just like dictate my phone, explain like, hey, there's this prompt for like having it manage your inbox and your finances. Can you add this to the site as well? And in terms of actually editing the site, deploying it, that can all happen from ChatGPT work on mobile.

• 16:31Speaker 01

  And I think what people are not seeing here, which again, so many new features coming out that I think people aren't taking advantage of is the domain here is chatgpt.site. So you can now deploy these like artifacts, these sites live. So can you just walk us through a little bit how that works? Or maybe we can do one, just so people can like understand the power of these sites and how you might use them or what their limitations are even?

• 16:58Speaker 00

  Yeah, I think like for like 10 second description for the technical audience, it supports a SQL database. It has S3 storage for files and even has environmental variables you can put there. For the non-technical audience, it's essentially a website that you can sort things on, which is great. And you can also control who sees it. You can, you can filter by email. So it can be private, it can be fully public. That's up to you. But what we could do if you're interested is we could try building a site live.

• 17:29Speaker 01

  Let's do it. We love to build live.

• 17:31Speaker 00

  Yeah. All right. So let's jump into it. So I've got an idea for a site. I'm here right now with Clara Vau. She hosts the How I AI podcast. And we're hoping that we can build a site that takes all the best tips from the videos on her YouTube channel and makes it so you can like easily filter through them and have quick links to like different timestamps on her videos.

• 17:58Speaker 01

  Make it match the How I AI branding. We use good old AI blurp all purple and black and white. And let's have it, if possible, categorized by tool. So like what AI tool is being used in the workflow and also function. So if it's for designers, if it's for engineers, it's for personal productivity, any of those things. Otherwise, I'm excited to see what you come up with.

• 18:26Speaker 00

  Awesome. That sounds great. And go ahead and deploy it privately. And then we'll take a look at it when it's ready and then you can deploy it publicly from there if it if it looks good.

• 18:35Speaker 01

  While this is running, let me make you make you laugh because this two-human chat to ChatGPT I did recently with my mother. And she was like stuck on an administrative task and she was like getting all frazzled about. She's like, oh, I'm going to have to look up all this stuff and blah, blah, blah, blah, blah, blah. And I was on the phone with her and I was like, Mom, just hold on one second. And I opened up ChatGPT on my desktop and I turned on voice. And then I was like, can you just please tell me what you need to do? And I turned the speakerphone off and I put it up to the computer and had her like babble kind of like we did. And I asked her a couple of questions. And then I was like, thanks. Pressed enter and I was like, I did it for you.

• 19:18Speaker 01

  So I really do feel like people are underusing like two-person voice chat. You can do it on your phone. You can use speakerphone. Like it is this really nice, again, like high bandwidth way to get requirements into the system. And it doesn't just have to be you alone.

• 19:35Speaker 00

  Yeah, I've had times where, you know, I need context with somebody I work with. And I'm like, please just like ramble dictate into Slack and send me like garbage. I don't care. And they're like, no, let's just like set up a meeting. And I'm like, no, don't, don't do that. I don't need that. I don't need this to be like pretty or anything. Just, just tell me that's fine.

• 19:58Speaker 01

  So are we at this meeting could have been a voice note like level?

• 20:02Speaker 00

  Yeah, I think so. I mean, all it's more like. Give me information that unstructured, however it is, that my agent can decipher, and that's good enough for me.

• 20:12Speaker 01

  This episode is brought to you by Hyperagent, the platform for deploying always-on agents that actually run your business. With Hyperagent, you build agents in the cloud and deploy them where your work already happens, like Slack, Telegram, or email. An agent will scan your inbox and draft replies to vendor follow-ups, another monitors competitors and spins up rich ad kits and landing pages, a third notices a deal going cold in Salesforce and writes the save email with full account context. These aren't chatbots waiting for a perfect prompt. They're proactive, learning your preferences, retaining your playbooks, and getting better with every run. One user built four agents to run an outbound sales

• 20:55Speaker 01

  pipeline, prospecting, outreach, follow-ups, CRM updates, all in a single afternoon. No local setup, no VPS bills, no fragile permissions on your laptop, just powerful agents with full control over skills, tools, and guardrails. Hyperagent was built by the team behind Air Table, and How I AI listeners get $1,000 in free inference to start building. Claim yours at hyperagent.com/HowIAI. Okay, so this is gonna go spin up. We're gonna let it kind of like whirl a little bit. Maybe we'll come back to what we see. But there's one other use case that both you and I really like in Codex that I thought you could show, and particularly how you

• 21:40Speaker 01

  prompt it, which is Codex for video editing. And for those that don't know it, I do a lot of this for the podcast. I also do a lot of this just for like general work stuff. I'm cutting a lot of videos. So tell me like how you came to this use case, why you feel it so useful, and then how easy is it to prompt, kind of like video editing?

• 22:04Speaker 00

  You know, I do a little bit of content as part of the DX team, and you know, we're trying to reach people that might not be inside like the Twitterverse, which you know, there's a lot of us there, but there's more of us not there. And I think explaining, you know, how you can use these tools in really relatable ways, where you're explaining to like a camera in like 45 seconds is great. And so I started just kind of like going to a park, filming myself, filming my camera, getting a smattering of little clips. And so this is from, like this is from yesterday. I went to the park. These are like, most of these are takes that we're not gonna use. But what I'll do is I'm literally just gonna

• 22:49Speaker 00

  drag these into ChatGPT work. And then, so what I've done, and I'm gonna share this, but I've created a plugin called UGC Video. Actually, let me like get the one that I use for, I made a version that has like open I branding, but you can actually see it here. And so this basically, after going through like a few runs of this, what I've noticed is that just having like some guidance around like, hey, make sure that we're not entering like the safety zone with Instagram ads, or let's do these formats. We want nine by 16, also four by five. That's great. But what I'll generally do, and I'll stick tape this. So we've got a bunch of clips here. I recorded a video for the record and

• 23:34Speaker 00

  replay feature. It starts with me just explaining that I'm apartment hunting, and that's kind of a pain. And then I walk through and show the feature in real life. I show myself actually apartment hunting, and I show how Codex can do it for me. And then I have a closer about how this skills that's been made for me can be put on a heartbeat, and I can have Codex just search for apartments in the background. Can you go through these clips, first like pull the transcripts, find the best takes, and then kind of piece this together into a UGC-style video? So that's running. A few things that, to kind of explain how

• 24:19Speaker 00

  this works. So I've only submitted like 20 some odd like video clips. And so what Codex is gonna do is gonna start by processing those clips to get just the transcripts. This will help it understand like the story, what we're talking about, and even like my light description will help it understand that. And then what we'll do then, actually it's gonna ask you some questions. Sure, we'll do organic vertical. Let's do complete and clean. Sure. So first get the transcripts, and then it's gonna go through the takes and actually like process images from them to understand what I'm showing. And then it'll choose like the best takes. And when I'm recording, I'll say things like, cut,

• 25:03Speaker 00

  or that's a bad take, or that's a good take. And it's able to like use that as guidance to know like, you know, what's actually useful from these.

• 25:12Speaker 01

  You're gonna give me so many ideas, because I have like deeply abandoned my once-great TikTok, because I guess like UGC is hard, like short form, short form is hard. And I just have such a high bar, that I'm constantly like recording these clips and being like that was garbage, and then I get frustrated and I walk away. But you've given me this idea, which is I could just like keep recording until like there's one that feels good, and then just dump it all. I don't even have to remember which one is good. Dump it all, give some voice notes about how to edit, and then, and then kind of assemble it. And then I love this process of just having the model transcribe, look at the video, come

• 25:51Speaker 01

  up with like good, good cuts, and put it together. I use it a lot for our trailers. So anytime I do a very long talk, like a 30 to 60 minute talk, or I do a podcast that's like 60 minutes, I dump it in, and I'm like, give me a 60-second hype video with the funniest parts of the talk, and like give me a little bit of buffer for clipping. And then it clips that really well. It actually has like pretty good funny taste, which is nice. And so I do that. And then I also do, like I teach like a three-hour-long workshop every quarter on how to like transform your engineering organization to be kind of like AI engineers. It's very, very long. It's like three-hour Zoom clip. And I also do

• 26:33Speaker 01

  that to like clip chapters, clip shorts, do all this. It's just like all this tedium of video editing. It's quite, quite good at. And then just tell us like, what does this plugin actually do?

• 26:46Speaker 00

  Yeah, so the plugin is really just based on like all the steering I've done, really this week of like making these videos, whether it's, you know, describing the kind of formats I need, or, you know, that you don't want to have text in certain areas because it'll be cut off, or natural kind of breakpoints, or how to do captions. It's, you know, literally, I just, I had this massive thread where I did a few of these videos. And there was so much essentially like bespoke knowledge in this process built up, that I was like, hey, can you make a plugin out of this? And there's actually a plugin creator skill in the ChatGPT app, so it's pretty easy. And then I did a few tweaks, and now I just more reliably get, I guess, the happy path that I want.

• 27:31Speaker 01

  And can you show us an app? I know you did one of these for yesterday. Can you show us like what, what the kind of output is that you can get?

• 27:38Speaker 00

  Yeah. So here was my starting prompt. Again, I just dumped in a bunch of clips. And one thing I want to know, you talked about like the tedious stuff.

• 27:47Speaker 01

  Yep.

• 27:48Speaker 00

  Even like little things like going through and deleting bad takes, or finding like the right segments. I feel like I'm so willing just to hand that over.

• 27:59Speaker 01

  Yep.

• 27:59Speaker 00

  And just get that off my plate. I don't want to think about it. So yeah, I started a pretty similar prompt that I just, I kicked off with a little bit before. At one point, and actually I was in an Uber, I said this on my phone. I remembered, I was like, oh yeah, I need, I need four by five also. So I sent that in as like a steer. And I was like, all right, got it. Updating the, updating the delivery matrix. And then this is actually where, you know, I was like, why aren't you asking me questions? So I had to update the plugin at the same time. And this is what the output looks like. So this is actually a few, few different like, you know, back and forth steers. Here, we'll actually haven't like watched this full one, but we'll see how it looks.

• 28:44Speaker 00

  It's insane trying to find apartment in SF or New York City right now. And it's impossible to get to all the listings before they're gone. So what's funny, I don't know if you can see that, it's, it even added blurs to these addresses here.

• 28:55Speaker 01

  Real smart. No, I can see it.

• 28:57Speaker 00

  Yeah. Actually, I did like a couple of these for, like that we're releasing on Instagram next week. And there's like sensitive, like release material information on my Slack that I like bowled up in my ChatGPT work. And it did this like insanely granular work of like blurring out literally lines, and then following it while I'm scrolling. And it does all this like verification where it, it has these blurs, it checks its own work. And then it like does that over and over again until it's good. And then it gives you like, you know, a good result.

• 29:28Speaker 01

  I probably six months ago coded a redactomatic for the How AI podcast because I had so many people being like, whoops, I left like my API key open in that like one piece of code that we were talking about. And so we actually like built one. I need to try it again with Codex because it's probably I, the one challenge we have is the blur is following kind of like screen shares, but.

• 29:54Speaker 00

  Yeah, and actually that gets back to the like, I don't my, I don't even want to deal with the tedious work. Like not only, actually, it's a nonstarter that I'm going to make blurs automatically, do that manually. That's just, that's not going to happen.

• 30:08Speaker 01

  No.

• 30:08Speaker 00

  What's even more annoying though, is having to think about, all right, well, how do I generate outputs that don't have confidential information? And how do I still make it seem genuine? That's also a pain that I just, I don't think about. I just, I leave it to, leave it to the chat GPT app.

• 30:21Speaker 01

  I'm going to give one other benefit for the content creators out there for this like AI editing of content, which is it's real embarrassing to edit your own content. It's like really embarrassing to look at yourself and watch all these videos of yourself and be like, oh yeah, I was like real cute and funny in that one and that one I am not like, I make no sense as a human. I like to offload like the critique of myself and the critique of my performance off to another model. So I don't have to look at my face so much as well, especially as like a solo creator, solo founder, somebody who's like working on their own work. So I do like having a third, you know, like a neutral third party do these cuts. So I don't have to sit there and like, listen to my own videos all day.

• 31:07Speaker 00

  Yeah, I know. It's, it's, it's great. Back to like the, it like, it definitely lowers like the bar, like the barrier to entry for stuff like this. If I just need to like go to a park, you know, I recorded like three videos. I saw they recorded the three videos over the course of like an hour and a half a couple days ago. And then I uploaded like 50 or 60 clips.

• 31:28Speaker 01

  Yep.

• 31:29Speaker 00

  And I dictated like what the different narratives for those three videos were. And because I have a plugin and a task that already understands what I want.

• 31:37Speaker 01

  Yep.

• 31:37Speaker 00

  You know, it took a while. I, I went to bed and I came back to it, but I had, I had, had what I needed and actually shipped one to our Instagram a couple days ago.

• 31:45Speaker 01

  Okay. I have one more question on this and then maybe we'll see if the, if the website is up. Do you need Sol Ultra for this? Cause that's what I see below.

• 31:55Speaker 00

  Yeah, I would. In my experience, there's just so many, as much as editing just for speed, there's so many little like micro tasks involved in checking all these frames and checking these transcripts. And there's a lot of parallel processing that can happen and a lot of like parallel judgment, if you will. And what Sol Ultra is, is it's basically this framework for multi-agent, for five, six Sol. And so I find that it's just more efficient and I get better outputs when I use Sol Ultra.

• 32:29Speaker 01

  Love it. Okay, cool. Let's see if the How AI website is, is up and running. Look at that. Black and purple and white, just like I said.

• 32:45Speaker 01

  We also have AI identify which thumbnails are going to do the best this, by the way, this episode, just so you know, it's all look, no hands. We had some problems. Oh, I see what it is. It's the use cases. So each episode is showing up multiple times. That was a very good episode about Codex browser use where I let it shop for me for my Hawaiian vacation.

• 33:06Speaker 01

  And it did quite, quite a lovely, lovely job.

• 33:09Speaker 00

  If you, if you want to get really fancy, you could probably have it have different thumbnails for each clip. We could even have like these image gen to make like a, you know, a singular thumbnail.

• 33:19Speaker 01

  Love it. Amazing. Okay. There's good old Alex over and over and over again. And then we're in it three of me. It just built it. Awesome.

• 33:28Speaker 00

  Yeah.

• 33:28Speaker 01

  And then this can be published, just like remind me how it, how it gets published, who I can share it with, how that all works.

• 33:33Speaker 00

  Yeah. So let's say, and let's say I wanted to like share this with you. I can say, you know, share this. And so now you can, you can put in, you know, whoever you really want to. It can be totally publicly available. You could add individual, individual email addresses.

• 33:50Speaker 01

  Yep.

• 33:50Speaker 00

  And it'll share with those folks.

• 33:52Speaker 01

  Amazing. And then how do you log in?

• 33:55Speaker 00

  It's logging with Chat GPT.

• 33:57Speaker 01

  Great. Easy peasy. I do have a Chat GPT account underneath that email address. So we are, we are good to go. Just to recap, we did a couple of use cases. So we talked a little bit about voice, high bandwidth, yappers API. Just basically like talking to a computer empowered assistant that can just go do a bunch of your work for you with access to browser and other things, including just being able to screenshot whatever's on the screen and have context. We did a, an, a chat GPT site, which I did not know. It has its own computer and memory and browser and all that stuff. So that's fun fact for Claire.

• 34:37Speaker 01

  Where you can have it actually go off and make a website, which you can share publicly or share with specific email addresses, which you just showed. And then again, my favorite use case, which is take a bunch of clips for UGC, dump them in, make it the model's problem, including doing redaction and blurring, which is for any of my fellow YouTubers out there, a, a huge and very annoying issue. And I'm very happy personally to have seen this, this workflow and get it solved. Nick, a couple of questions, lightning round questions for you. We will get you back to all your, all your voice chatting. Question number one is on that topic. In this world where we're all loving our Chat GPT voice experience, are, is everybody in the OpenAI office

• 35:20Speaker 01

  just like mumbling to themselves and to their computer? Like, how are we going to intersect all this voice capability with the realities of like existing around humans that find all this chitter chatter annoying? Like, how does it actually work?

• 35:36Speaker 00

  I think it's like a, like an almost genuinely uncomfortable question. Because when I'm at home, I exclusively dictate. And when I'm at my desk around other people, I will sometimes whisper, but I'm more likely to like use my keyboard. And, you know, we have all these meeting rooms and these like little pods people hop into. And I see people like hopping in there for not calls or often these days. Because they're, they're clearly going in there to like use dictation or to use voice and Codex. And clearly people are adapting to this as like a better experience. But frankly, we, you know, I don't know if the answer is we all walk around like with those masks around us that like shield our voice

• 36:20Speaker 00

  anybody but the agent, but it's, it's not ideal right now. But you can, but we can tell that there's, there's a better, there's a good version and that's using voice. But it's not ideal for our current, you know, office makeup.

• 36:33Speaker 01

  Infrastructure. Yeah. I, I imagine we like press a button and like a tube comes down.

• 36:39Speaker 00

  That would be ideal. Yeah, that's, that's what I'm.

• 36:41Speaker 01

  And if we talk and we unpress our button and the tube, the tube goes up. Okay. We're going to, we're going to figure this out until we get the direct Codex to mind connection, which I'm sure will be a 2027 release. Okay. Second question. If, if people really wanted to like feel the AGI, what are, what are the things that you really think are just huge step changes in the last three months? Maybe like voice aside, what is like the one thing that you would tell people to try or that you're really thinking is, is the new interface? Cause there's so many things that you've shown here.

• 37:16Speaker 00

  In terms of like, in terms of feeling AGI, I think a lot of people like you and me have already felt the AGI. I think if we go back to, you know, 2022, 2023, around the release of Chat GPT.

• 37:28Speaker 01

  Yep.

• 37:28Speaker 00

  I think a lot of people felt the AI then, you know, what, what AI can be became understandable. Like you understand this conception of a chatbot that I say things to and intelligently says, says things back to me. And that's, that's pretty cool. And then I think with the advent of coding agents and like the coding agent harness in particular, I think people in the developer community and like AI tinkerers, they felt this AGI moment where wow, now, now this AI is going out and like reading files and creating files and like running commands. And this is crazy. I tell it to do, I tell it one thing and then it figures out all these other things on its own. And that was a huge AGI moment for people. But I do think there was a certain barrier, like barrier to entry with that.

• 38:11Speaker 00

  Where if you're not an engineer or you're not a tinkerer, you know, there's, this is limited to a small ish group of people. But now, you know, with something like Chat GPT work in the web, I think we're taking a lot of the primitives, like the coding agent harness, like having always on agent. And we're making them not only accessible to the web, but they're accessible to your phone. And so for people that are already using the Chat GPT app, which is a lot, they can just switch over to the work tab. They can ask a question like, hey, can you check my Gmail and also my finances? Which I know are two, you know, plugins that you can install really quickly.

• 38:52Speaker 00

  And just monitor those and let me know if anything stands out or anything has gone awry. And now they suddenly have this agent that's, you know, maybe twice a day checking their email, checking their finances. And I think that like small thing, you know, when they first get a notification, you know, letting them know maybe the package has arrived or there's some weird charge. I think those are like the small moments that people might kind of like, it might be like a foot in the door for a lot of people.

• 39:17Speaker 01

  Okay. Amazing. And then I have one last question I ask every guest and because you speak to AI. I'm really curious here, what, what you say. When I, I saw a little bit of your prompting strategy. It's a little bit of what I do, but when AI is not responding, when it is not doing what you want, when it is not using that plugin, what is your prompting strategy? Do you yell? Do you all caps?

• 39:42Speaker 00

  Well, it depends if I'm sitting at my desk or if I'm in the living room. I guess the, my treatment of, of AI is different.

• 39:54Speaker 00

  It might be all caps in the office. Is I like banging my keyboard?

• 39:58Speaker 01

  Yep.

• 39:58Speaker 00

  At home, alter on dictation, maybe share a few expletives, try to redirect things.

• 40:05Speaker 01

  I'm just like constantly, constantly being like, this is trash. Why? Why are you, why are you this way? I used to be so polite and now my bar is so high. I'm like, you, you can do better. I believe in you.

• 40:20Speaker 00

  I would say lately I feel like I have been less upset. Things are just working better, you know?

• 40:26Speaker 01

  I've been more upset.

• 40:27Speaker 00

  People, people are, are you getting more upset?

• 40:29Speaker 01

  Yeah, because my bar is higher. I'm like, I know you can do this. You're not a dumb dumb. Come on.

• 40:35Speaker 00

  That's interesting. Yeah. All right. Yeah. Treat it, treat it as an equal or superior and you're surprised when it, when it doesn't come through.

• 40:41Speaker 01

  Okay. Well, Nick, this has been super fun. Where can we find you? How, how can we be helpful?

• 40:46Speaker 00

  Yeah. And find me on Twitter, X, my name, Nick Baumann underscore. If you have any questions about Codex, Chat GPT work, anything we're releasing at OpenAI, feel free to reach out to me. DMs are open. Great to, great to be out with you, Claire.

• 41:03Speaker 01

  Yeah. Thanks for being here.

• 41:06Speaker 01

  Thanks so much for watching. If you enjoyed this show, please like and subscribe here on YouTube or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app. Please consider leaving us a rating and review, which will help others find the show. You can see all our episodes and learn more about the show at howiaipod.com. See you next time.