YT YouTube2Text

YouTube2Text → Transcripts → OpenAI Just Merged ChatGPT and Codex. This Changes Everything.

OpenAI Just Merged ChatGPT and Codex. This Changes Everything. — Transcript

by Riley Brown · 9,485 words · 1,302 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00OpenAI just combined Codeex and ChachiBT
0:03into one super app. And they also
0:06released a new model that's nearly as
0:08good as Fable. They also made updates to
0:09their sites feature which turns it into
0:11a full vibe coding platform. And they
0:13made updates to their in-app browser.
0:15And so I have so many questions about
0:17all of this and how this actually works.
0:19So today I'm bringing on Ross Mike and
0:22we're going to break down all the new
0:24models, how the new platform works, and
0:26how he uses Codeex as a developer and
0:29how I use it as a marketer and founder.
0:31You're watching Agent Native. My name's
0:33Riley Brown. Let's dive in. Ross Mike,
0:36here we are. We're in the new Chat GPT
0:39app, which used to be Codeex. OpenAI
0:42decided to merge chat GBT and Codeex and
0:45also introduce three new models, one of
0:47which is very close to Fable 5. I I want
0:52to dive into a lot today, but I first
0:54want to talk about the new platform. Do
0:56you want to open up the work tab real
0:58quick?
0:59>> No, right here. Everyone's favorite tab
1:01on Twitter right now.
1:03>> Yeah. Yeah. So, like I'm just really
1:05curious. I I am still on the codeex tab.
1:08I just don't understand the benefit for
1:10me. Should I just stay on the codex tab?
1:12What do you think?
1:12>> Yeah, stay on the codex tab. It's not
1:14for you. The the work tab like the work
1:15tab is not for us, right? And again,
1:17they are trying to get Think about it.
1:19983
1:22million users. That's more than the
1:24population of America twice. Add Canada
1:27in and it's still more, right? So, you
1:30have these people and most of them are
1:32not using Chad GBT to the fullest. In
1:34fact, a lot of people are starting to
1:36think Claude is the agentic platform and
1:38Chad GBT is just a place where I, you
1:40know, send texts and get questions
1:42asked. Like I've heard a lot of people
1:44uh in my church writing, they'll come up
1:45to me and be like, "Yeah, like Chad GBT
1:47is cool, but you know, with Claude, I
1:49can do stuff." Right? So, there's not
1:51only like a branding issue where people
1:54are just thinking of chat GBT as chat,
1:56but there's also an awareness issue,
1:58right? This drop down specifically
2:00saying for getting work done is to tell
2:03people, "Oh, maybe I could do some
2:05knowledge work." Right? So, if you've
2:06already been using Codex, you could
2:08stick to Codex. Even if you're doing
2:09knowledge work, the work tab isn't for
2:11me and you. I if I was an investor in
2:13OpenAI and I had some shares, I wish I
2:15had some shares. I would be bullish on
2:17this decision because there are millions
2:19of people who have no idea the full
2:21power and capabilities of Codeex and
2:23this is how they're going to get used to
2:24it
2:24>> 100%. Um, okay. So, I want to talk about
2:28the models a little bit. So, have you
2:30So, you've taken this model for a ride.
2:32You've tested the the different models
2:34that were released. What are your
2:36thoughts? Okay. Okay. What are your
2:38thoughts on Actually, first, can we
2:40break down the three different models?
2:42So, there was three models released.
2:43There was GBT, 5.6 Soul,
2:46>> 5.6 Terra, and 5.6 Luna.
2:50>> Uh, Soul is the one that I think most
2:52people care about. I think naturally
2:54they just care about the most powerful
2:55model. What are your thoughts? What are
2:57your thoughts on soul?
2:59>> So, uh, Terra and Luna are probably
3:02going to live in the work tab, right?
3:04This is probably what most normal people
3:05are going to use. It's going to live in
3:06the chat GBT tab cuz you don't really
3:09need super big models for almost 70% of
3:12tasks. But for most of us who are
3:14interested in these longunning tasks,
3:16who maybe build applications with AI,
3:19Soul is the interesting one. Now, one
3:21thing about OpenAI and their model
3:23development is they're very token
3:26efficient and they're cheaper than the
3:28enthropic models, right? Um price isn't
3:30necessarily what you just want to look
3:32at. You want to look at the token
3:33efficiencies. One of the first tests
3:34that I run is when I use a model, I
3:37check how many tool calls it does. And
3:39one thing about Gemini, our fallen foe,
3:42is Gemini, although on an intelligence
3:45index is extremely smart when it comes
3:49to anything agentic, it just it just
3:51spirals into hell, it calls 15, 20
3:53different tools. It like psyches itself.
3:56It itself experiences hyosis, right? So
3:59the one thing that the GBT 5.5 models
4:02really good at was being token efficient
4:04and was good at tool calling. 5.6 6
4:07takes it at a different realm. Now 5.6
4:10in my opinion is not a fable competitor.
4:13GPT6 will be a fable competitor. This is
4:15just a version bump. But even in the
4:17version bump, I genuinely think it is
4:19very very comparable for what you get
4:22for your dollar. But not only that, in
4:24terms of sheer power, right, its ability
4:27to do really run long running tasks. One
4:30issue I had with 5.5, um, it was very
4:33needy, right? like you give it a full
4:35plan, you tell it to execute on the
4:36plan, and it'll be like, "Oh, I did
4:38phase one. Um, should I continue on
4:41doing phase two?" It's like, "I gave you
4:42a whole freaking plan. Do the entire
4:44thing, right? Come back to me once it's
4:45it's done." 5.6 is a lot more eager.
4:48It's a lot more faster. Um, and it's
4:50token efficient as well.
4:53But it's not a Fable equivalent. Like, I
4:56I I definitely think the Fable
4:58equivalent will be a six, right? And
5:00this is why they did a version bump of
5:025.6 6 versus GPT6. But all in all,
5:05incredible model.
5:06>> Yeah, GPT 5.6 passed my I don't know if
5:10you saw this, my I have a new repletable
5:13benchmark. So, can these
5:15>> can these models in one prompt create a
5:18lovable clone, which is a mobile app,
5:20like a Swift iOS app that builds web
5:24apps just like lovable and replet?
5:27>> When I asked Fable to do it, my prompt
5:29was like four sentences. I it's in my
5:31video. It's pinned on my Twitter if you
5:33guys want to see my prompt. But I
5:34basically said I want you to create a
5:36lovable clone. You use sand uh use um
5:40Daytona for sandboxes, Convex for
5:42database, and I think use anthropics
5:45models to generate the app. And in one
5:49shot, Fable didn't hesitate and it just
5:53did it
5:54>> in four prompts. I will say 5.6 six
5:56required me to create a longer prompt
6:00because I had to like give it more
6:01instructions.
6:02>> And that makes sense, right? Cuz again,
6:04people are comparing 5.6 to Fable 5. I
6:07don't think it's fair. I think the
6:09mythos class models are in a different
6:11class. If you notice every like four,
6:14five, six, like there's always been like
6:16something like it's been a monumental
6:18shift. The only like point shift that
6:21was monumental was Opus 45. I think Opus
6:2345 is when we all realize like wo like
6:25like this is not a joke right so I don't
6:28think it's fully fair to do like a
6:30onetoone comparison between 5.6 and
6:33Fable cuz Fable is just a monster and I
6:36genuinely think with the considering
6:38cost cuz it's important thing to
6:39consider cost we're not going to get
6:40subsidized forever considering cost I
6:43think it's very wise for people to think
6:45and start to aggregate okay what tasks
6:47can I ch can I tr trust a model like 5.6
6:51six with what task do I give Fable with?
6:53Right? If you're like, I don't know,
6:55changing the color of a landing page,
6:58maybe you don't use Fable 5 Max, you
7:00know what I mean? Maybe we throttle it
7:01down to low. Maybe we even give it to
7:02Opus 4.8. Right? So, Fable is an
7:06incredible model. Like some of the
7:07things, Riley, I built is I built a
7:09Daytona alternative, right? So, spinning
7:11up like your own sandboxes on a bare
7:14metal server. Um, I built uh game plane
7:17which is this dashboard where you can
7:19deploy your own like game servers like
7:21Vice City um Minecraft and like it's
7:25modded Vice City like I have like it's
7:27not fully functional but I'm trying to
7:29build like a multiplayer Vice City
7:31Fable's doing it like some of the stuff
7:32I wouldn't be able to do. Eve which is
7:35uh Verscell's agent uh framework I
7:38ported that so it can fully run on
7:40convex right this is something I could
7:43have done but it would have taken me 3
7:45or 4 months and this is like 3 or 4
7:47months of like this is what I'm focused
7:48on this is what I'm thinking about th
7:51this is all I'm grinding on it took it 8
7:53hours 8 hours one prompt it did the
7:56entire thing all test passed fully
7:58deployed on Verscell fully working chat
8:00app right so Fable is in a class of its
8:04own. And I think the comparable model
8:06that they're probably cooking up right
8:08now is GPT6.
8:105.6 is a class above Opus 48, a class
8:15above 55, much better than these models,
8:18but in terms of sheer intelligence and
8:21getting very hard tasks done, nothing
8:23touches Fable.
8:24>> Yeah. And it's it's also the same price
8:26as 5.5, too. So, it's
8:28>> which is incredible, by the way. I I
8:30don't think people like grasp that,
8:32right? like it it is difficult to build
8:34a very smart model. So kudos to
8:36Anthropic but it's very difficult to get
8:39price right to make it efficient and I
8:41think open AI is building that
8:43foundation of building efficiency and
8:46then they're just going to train the
8:47super duper new class GPT6 model. Right.
8:51So I I think cost is something people
8:53should really start thinking about. Um
8:55we're drinking the Kool-Aid too much
8:56because one day the subsidies will end.
8:59>> Yeah I I agree. I agree. What I do want
9:01to do is I I really want to selfishly I
9:04want to kind of get some of your
9:06workflows. I want to understand how you
9:08you how you've been using codecs, your
9:10workflow from a tactical perspective and
9:12some of your favorite skills that you've
9:14created. Yeah, I you know me I get it to
9:17scaladraw. So, there's a couple things
9:20that OpenAI, Chat GBT, Codeex, whatever
9:23these names are all over the place that
9:25do phenomenally well, right? The first
9:27one, and I beg people just use this, is
9:30computer use. This model is incredible,
9:35incredible at using your computer,
9:38right? So much so that I made the
9:40decision to get a mini a mini PC so that
9:43I can remote view into this and I could
9:46tell it to use this computer. It's that
9:48good. Now you might ask, what are you
9:50using computer use for? To fill out
9:52forms, to do this, that and the third.
9:53In terms of application development, my
9:56number one favorite thing to do right
9:57now is test my app. As simple as that.
10:01You don't even need a skill for this cuz
10:03the model and the agent are so good. So
10:06let's say I develop It does it
10:08automatically now. Like I notice I don't
10:10even ask the testing app. It just
10:12>> I I come back 30 minutes later I'm like
10:14why is it taking so long and I see that
10:16it's like testing it like over and over
10:18and over again finding edge cases and
10:20just doing it. Anyway, sorry to
10:21interrupt but it is the computer use and
10:24and the browser use it has is genuinely
10:27something that you cannot we could not
10:29explain it properly. You must experience
10:32it. It is incredible. Anyway, keep
10:34going.
10:35>> Yeah. So like uh you know there have
10:37been talks about loops and stuff. One of
10:39my favorite loops that I've been using
10:41and you can use this with really any
10:43agent that has computer use but I find
10:45that open codeex is the best at this
10:47because their computer use is just great
10:49is as follows. I'll connect something
10:51like uh like linear like I use linear
10:54for like issue tracking. You can use
10:55whatever notion whatever whatever right
10:58and I will basically tell my agent
11:01you'll build this feature copy paste or
11:03you can even connect the MCP tool and
11:05attach it as an agent. And this is the
11:08loop I run. I literally tell the agent
11:10hey build this feature the success
11:13criteria for this feature is you use
11:15computer use if it does not succeed fix
11:19repeat and once it's done come back to
11:22me right. Um, this is me by the way. I
11:25just asked, now you might wonder why
11:27that popped up. I just asked Chad GBT to
11:29run this app locally and then I'm going
11:32to say, hey, can you test this app using
11:35computer use end to end? And I'm going
11:39to hit enter. So I'll get it to use
11:42computer use as an evaluation step,
11:44right? Cuz if I'm building a feature and
11:46that feature is somewhat UI centered,
11:48right? It's a form submission or
11:50anything like that, um, code isn't just
11:52enough, right? I need my agent to QA
11:54test this. And even in like a a a a
11:56normal computer engineering team, there
11:59are people who are going to, you know,
12:00there are QA people who test this,
12:02right? In my very old corporate job,
12:04Riley, I used to build a feature and
12:06there was one person's job. Their whole
12:08job was to break what I built, right? I
12:11can now do that with Codeex. And now
12:13you're going to see codeex itself test
12:16this app out.
12:19>> It's just testing it out. It can spin up
12:20the browser and it's just testing it.
12:22>> So it Yeah. So it opened it up. It says
12:24the browser pass succeeded. It basically
12:26made sure that it worked on browser. So
12:28it's just going right. And the cool
12:30thing about Codeex's computer use is I
12:33can do my work
12:35>> while it does it. Right. With anthropics
12:37computer use, it actually takes over
12:39your mouse and your screen and you kind
12:41of have to let it work. Right. So
12:42>> this is very important. This is very
12:45important when you use computer use.
12:47Sorry to echo what you're saying because
12:49I think it's really important. When you
12:50use a computer use when you use computer
12:52use on codecs, they can do it in the
12:54background. When you use it on clawed
12:57code, you'll notice there's this like
12:59kind of like brown gradient that takes
13:02over your computer and you're just like,
13:03ah, I have to wait for this agent to be
13:05done using my computer. But Codeex,
13:08you'll I'll switch to a new app tab and
13:11I'll see that it's working in the
13:13background, which is which is a
13:15mindblowing experience, which makes me
13:17want to get the most powerful computer
13:18ever. And and and I want to just I want
13:21to have 100 computers with agents
13:23working on them. Like it's crazy.
13:24>> And this is the thing, right? Like
13:26computer use and and I might pivot on a
13:28point here. Like you really need to
13:30start thinking bigger, right? If you
13:33have an agent that can use your
13:34computer, well, my question for you is
13:37what are the things that you can now do,
13:39right? What are like, let's say there's
13:41like a spreadsheet with thousands of
13:44rows and there's certain things that
13:45even an MCP connector couldn't do. You
13:48can now do that with computer use. Let's
13:51say you want your app fully tested end
13:53to end. You can now do that. There's so
13:56much you can do with computer use. Right
13:58now, I have two browser instances. I
14:00don't know if you could see my screen
14:01and maybe this isn't the best product to
14:03pick to pick but it spun up two
14:05different browser instances and it's
14:07basically running commands on this
14:09virtual terminal I have right and look
14:12it says launched and left a visible
14:14machine backend so this is stuff that
14:18again these things were people's jobs
14:21right I remember in my old corporate job
14:23Riley I had someone whose sole job again
14:26I build a feature they run the feature
14:28on their on their computer and they
14:30would spam and break it. Now, one thing
14:32that I do is I'll ask the agent to
14:34create a test plan for my app cuz it has
14:37context of how my app is and test it
14:39using computer use. So, it's going to
14:41use the app. It's going to identify, you
14:44know, what features work and what don't.
14:46And guess what I could tell it to do? I
14:48can tell it to fix my features. So, I
14:50would say homework for everyone
14:52watching. At least at least, please for
14:54me at least 30 minutes just just spam
14:57computer use. Tell it to do the most
14:59random things. If you have an app being
15:00built, tell it to test out your app
15:02using computer use. This is I would say
15:04the biggest up for me in terms of
15:07workflow.
15:08>> Can I add one thing here? The best way
15:10to use computer use if you are not a
15:13developer and if you have no interest in
15:14vibe coding, let's say you're a marketer
15:16and and I honestly think I use this way
15:18more for marketing tasks than I do
15:21coding tasks now. Um, but OpenAI
15:23released a new feature. It's called
15:24record and replay. What this does is it
15:28allows you to tell the agent to record
15:31your screen so that you can just use
15:33your computer for up to 30 minutes and
15:35then it will take that recording and
15:38convert it into a skill so that the
15:40agent can do it later. I 100% think that
15:43this is the future of of computer use
15:46for knowledge work tasks. It's not going
15:48to be writing this perfect prompt and
15:50like writing out an SOP for an AI agent.
15:53It's simply going to be recording your
15:56screen. And a lot of these tasks are
15:58repetitive and annoying. Like there's so
16:00many tasks that take me 10 minutes
16:01manually on a computer that if I could
16:03delegate to an agent, I would be so much
16:05happier. But so like the full process is
16:07literally like uh I want you to record
16:10my screen and turn what I do into a
16:14skill. Like literally that's all you
16:16have to do. I digress. Let's get back to
16:17the list here.
16:18>> Number two is loops. Now, I know this is
16:22a lot of controversy behind loops, but
16:24hear me out. I think they're loops make
16:28sense in certain card categories when
16:30you're building your application. And
16:32because
16:34codeex mixed with GBD 5.6 soul, super
16:38powerful model, one thing that you can
16:40do is you could have an automated uh
16:42review cycle, right? So every time um
16:46like you push changes or you make an
16:48update, you can tell the uh the GPT 5.6
16:52sole model to review the code and to
16:55assess it and give you a score from like
16:571 to 5 or 1 to 10. Now I use an external
17:00code review agent, but you don't even
17:02have to do that. You could tell Soul
17:03itself to give a review of the code it
17:06changed. I would do this in a new thread
17:08and then give it a review, one out of
17:10five. Most of the time, you'd be
17:12surprised. The agent that wrote the code
17:14itself will review itself pretty harshly
17:16and it will tell you, "Oh, this is a
17:17four out of five or this is a three out
17:18of five." Sometimes it's a five out of
17:20five. And then what I would do is I
17:22would tell the agent, keep fixing until
17:26you give yourself a five out of five.
17:31Another type of loop, I want I know I'm
17:33jumping ahead, but these two tie
17:34together is using automations. So, if
17:37you go on the cursor tab right here, um
17:40I think they cursor tab, but uh codec
17:43tab. Sorry.
17:45>> I I I I did wrong. I did wrong. Um if
17:48you go to the codeex tab, I think they
17:50changed it to schedules now. I can um I
17:55can basically say at 8 am every day, I
17:59want you to review all my open PRs and I
18:02want you to give me feedback on which
18:04ones I should close, which one I should
18:06keep open, which ones uh need further
18:09review. Um I want you to do a thorough
18:11security review of every single PR and
18:13give me a report. And if you find
18:16there's something necessary to fix on
18:17the spot, spin up a thread and fix it.
18:21So just something like this right and
18:25mind you this with computer use you know
18:29people talk about software factories
18:32ladies and gentlemen your software
18:33factory is codeex the only thing that
18:36was missing is a super powerful model
18:39like GPT 5.6 6 soul. So you can write an
18:42automation like this works every single
18:45morning specific time. It will review
18:47all your open PRs. It will review all
18:50security issues that you have and you
18:52just have a report. So you know we talk
18:54about being managers and managing
18:56software factories. That really wasn't
18:58possible until models like 5.6 and Fable
19:01came about. What really is the issue now
19:04is one's imagination. And I think what
19:08we what I would like for people to watch
19:10from this is like how far can I push
19:12this? Like what's the weirdest thing
19:14that I can build? What does my heart
19:16desire? You know what my heart desired?
19:18I wanted to build a GTA and let me show
19:20you some screenshots. I wanted to build
19:22a multiplayer GTA V city game cuz they
19:25don't exist. And I like GTA Vice City,
19:27the classic one. And I'm going back and
19:30forth with this. It forked a Vice City
19:34fork that runs on browser and then I
19:36asked it, "Hey, I want multiplayer." It
19:39went about four hours and guess what it
19:41did? I have multiplayer. The one screen
19:44is the gentleman standing. The other
19:46screen is the person in the car. Now,
19:49there were some bugs. We had to go back
19:51and forth. But can you imagine, Riley,
19:53as someone who runs a team, can you
19:54imagine how many engineers that would
19:56have taken? How much time that would
19:57have taken? Right? So all these tools
20:00that you have access to the computer use
20:02and all this stuff what this has now
20:04allowed you and me is to build bigger
20:06and wilder things. So going back cuz I
20:09know I go off on tang can we go back to
20:11this? Can I add something here? I I want
20:13to make this tangible for non-coding
20:15tasks. Can you go to the excaladraw real
20:17quick?
20:18>> Um just just talking about loops here.
20:21So basically what he did is he basically
20:24told the agent to review its work. give
20:28it a score and it basically decide like
20:30if it doesn't meet its own certain uh
20:34doesn't meet some certain criteria it
20:36loops back and it starts again until it
20:38reaches a certain score that's basically
20:40the essence of a loop and this can be
20:43applied to anything in and so like
20:46obviously I'm a marketer right I'm a
20:48founder
20:48>> actually I would love to know yeah how
20:49do you use loops in marketing Riley
20:51>> so this is this is the way we do it so
20:53there's there's a let's let's take ads
20:56for example so There's an there's a an
20:59API called foreplay. Very weird name for
21:02a company, but it's called foreplay, and
21:05it allows you to scrape ads from any
21:07company, and you can basically pull the
21:10videos, pull the transcripts from all of
21:12the top performing ads, and you can sort
21:15them by longest running. And if a
21:17company's running their ads for the
21:19longest, you know that that one's
21:20probably converting the most, right? you
21:22can you can't tell how well they're
21:23converting or what the ROI is, but if
21:25they're running an ad for a really long
21:27time, you know that it's a it's a good
21:30ad. And so what you can do is is you can
21:32basically um in marketing is you can
21:35pull all of your competitors really good
21:37ads or really well performing ads and if
21:40you're asking AI to create um 20 ad
21:44scripts, right? 20 ad scripts for your
21:47company, right? And you you can give
21:48codeex all of the context of your
21:51company and it writes these 20 scripts.
21:53It can basically compare the scripts
21:55that it wrote to the high-erforming ads
21:58and give each one a score and you can
22:00have codecs go in a loop until all of
22:02them are like a certain criteria or
22:05comparable to the other ads. That's just
22:08one example. Another example is YouTube
22:11thumbnails. And the only way to do this
22:13in marketing, right? Because if I don't
22:16know if you would agree with this, but
22:17coding is is a lot more um it's a lot
22:20easier to say good job or bad job, you
22:22know, and that's why design that's why
22:24that's why AI models are bad at design.
22:26It's because it's ambiguous as to
22:28whether it is a good job or a bad job.
22:30Whereas um you know, design um scripting
22:34that's why that's why AI is not good at
22:36script writing is because these
22:37companies that are training it, it is
22:39ambiguous whether it's a good job or a
22:40bad job. How do you reinforce an AI to
22:42do a good job if the diff the people
22:44doing reinforcement on it all have
22:47differing opinions? No one knows what's
22:48good where coding it's very easy to
22:50tell. And so the only way to do that is
22:52to gather really highquality examples uh
22:55and give it to an AI and then tell it to
22:57evaluate its work in comparison to all
23:00of those examples. Right? If you give
23:01codeex 20 different thumbnails that you
23:04like, it has a really good sense of what
23:07you like and it can grade the thumbnails
23:10that it creates according to the
23:12thumbnails that you like and you could
23:13just say like keep going until the
23:16thumbnails that you generate are similar
23:18to those 20. Does that make sense? And
23:20so that's kind of like how we're
23:22thinking about it. And with one tool
23:24that I have Codeex is it can use the
23:27Gemini
23:29video ingestion feature right Gemini you
23:31can upload a video to Gemini and it can
23:34actually like watch the video. It can
23:36pull subtitles from it. It can see all
23:38the frames and so CEX has a tool which
23:40uses Gemini and it can analyze my video.
23:43So, we have UGC creators and and other
23:46creators for our company will submit
23:48videos and I'll just have Codeex analyze
23:50them and it will compare those videos
23:52that they submit to the top performing
23:57um videos that are run by other
23:59companies and it'll judge them. And so
24:01these are kind of the loops that we're
24:03starting to create is comparing stuff
24:06that we do as humans in our company and
24:09then also stuff that the AI generates
24:11and compare it to really high quality
24:12examples and it basically loops and
24:14edits them until it reaches some quality
24:17bar. Does does that make sense?
24:19>> Yeah. And like to further like just
24:22solidify what you just said, a loop is
24:24as simple as follows. You have a well-
24:26definfined prompt, right? In this case
24:29for Riley, it was, you know, I want a
24:31YouTube thumbnail. You have a feedback
24:33engine, right? In his case, it was 20
24:36other thumbnails that were really good
24:38and he wanted these thumbnails to be
24:40similar. In like a developer's case or
24:43you were building an application case,
24:44it's telling the agent to use computer
24:47use. And honestly, nowadays, you don't
24:48even have to tell it. It will use
24:50computer use. use the app and if it
24:52functions just as well then it will
24:55that's that's basically your feedback
24:57engine and then you have a well-
24:59definfined success criteria and you're
25:01basically telling the agent until we
25:03reach here using the feedback engine
25:06keep on going right so loops apply to
25:09anything any industry any space you just
25:13need to know what your what it is that
25:15you want most people don't know what
25:17they want most people don't even know
25:18what they want for lunch today you need
25:20to know what success uccess looks like
25:21cuz the agent's not going to tell you
25:23what success looks like. The agent
25:24doesn't know what success is and then
25:26you're going to give it a mechanism for
25:28it to give feedback. Right? In Riley's
25:29case, it was he had top performing UGC
25:32videos that were used as comparison. In
25:35uh building case, it would be using
25:37computer use or your you specify that
25:39you want the feature to do X, Y, and Z.
25:42And this basically the agent itself will
25:45keep on looping. You don't even have to
25:47tell it to loop. It will keep on going
25:49until it reaches what is defined as a
25:52success criteria. Right? And this is how
25:54you use something like Codex in any
25:57single field in any industry with the
26:00tools that it has, with the model that
26:02it has and this capability. There's
26:04honestly nothing we can't achieve
26:06nowadays.
26:07>> 100%. And and I think that's why plugins
26:10are so important, you know, like and I
26:11talked about the ads plugin. There's
26:14also another API. It's called scrape
26:16creators. And scrape creators allows you
26:19to literally download any video from any
26:22platform. So I could say uh I want you
26:24to look up know that. What's it called?
26:27>> Scrape create. Scrape creators. You can
26:29look up um you could say to any agent,
26:32you could be like, "Hey, I want you to
26:33go to Greg Eisenberg's uh Instagram and
26:36download his top 10 videos, right?" And
26:38it will just scrape them. And because of
26:40my Gemini skill, right, which which
26:42allows Codeex to gather as much
26:45information from the video as possible
26:46because to my knowledge, Gemini is the
26:49only one where you can actually video,
26:51you can upload a video to the model.
26:52>> Yeah. I don't think any other model can
26:54do that.
26:54>> Yeah. Which is so the most underrated
26:56that's like the only thing I would use
26:57Gemini for. They're kind of behind on
26:59everything else. And I don't mean to
27:00trash them, dude. I Faith Logan
27:02Kilpatrick, if you're for some reason
27:03watching this, which you're not. I I
27:05love you. Keep going. Um but um yeah
27:09that like if you give the agent the
27:11ability to go find high-quality examples
27:13and then you just tell it like hey I
27:15want I want you to create video like
27:18make sure this is as good as Greg
27:20Eisenberg's video make sure that this is
27:21good as RPN's videos on Instagram. It'll
27:25just go to Instagram, G, download his
27:28videos, watch them, understand what good
27:30looks like, analyze the script, and then
27:33it will make sure my videos are edited
27:34in a same way. Um, and that's kind of
27:37how I and then even this is really
27:39useful for manual processes as well. So,
27:41uh, yeah, that's just kind of how I've
27:43been using it for marketing.
27:45>> Yeah. So, loops again, well-defined
27:47prompt, feedback, uh, engine, and then a
27:50well-defined success criteria. This
27:53again, I didn't know he was doing that
27:54for marketing. I just learned something
27:55new. I've been doing all this stuff
27:56manually, right? I silly me, I should
27:58have had a loop going on, right? So,
28:00this literally applies to anything. It's
28:02just you just have to have a mindset
28:04shift and start really to think how far
28:06can I push these models and codecs. And
28:10I think that helps a lot.
28:11>> Yeah. Um, so this list, so we had
28:13computer use, loops, anything else.
28:16Number three, um, and you kind of said
28:19this and I want to be careful on what I
28:21mention here, but number three is tools,
28:24right? Like, and when I mean tools, like
28:26from a development standpoint is what
28:28stack are you using? I think, um, you
28:30know, back in the day, back in the good
28:32old engineering days, um, we would very
28:35much take pride in building a lot of the
28:38stuff and owning the stuff. Now, it's
28:41gone to the point where speed really
28:43matters and we don't really care for
28:45that, right? So, I want to make sure I'm
28:47using tools and things of that nature
28:50that really really help uh me develop
28:53whatever software I want to develop,
28:55right? Using the right tool for the
28:56right job, right? If you notice the
28:59agents for the most part default to
29:00React when you tell it build me a web
29:02app, right? With login this, this that
29:04and the third. I think having sort of
29:06like that sort of knowledge and knowing
29:07what it is that you need to build the
29:09thing that you're trying to build is
29:11necessary. If I want to build 3D stuff,
29:14the first thing that comes to mind is
29:15Blender. Blender has an MCP. I go to
29:17Blender set, copy the MCP instructions,
29:19and then boom, I paste it. Riley wanted
29:22to do video injection. I didn't even
29:23know like a model could do that, right?
29:26So, he goes to Gemini, right? Using
29:29Gemini. Um, I'm only going to mention
29:31this cuz you said this and I I I I will
29:34preface from the beginning. I do work
29:36there, but seriously, seriously,
29:38seriously, convex when developing
29:40applications. And I know people are
29:42going to call me a shill, Riley, but
29:44sincerely, there isn't an app I don't
29:45build without confix. And this one
29:47reason is the only thing. It's entirely
29:49code.
29:51And I'm getting to the point where I
29:54want the agent to do more of the work,
29:57more of the like like the the the the
29:59tasks, and I really just want to be the
30:02manager who kicks their feet up and
30:04makes decisions, right? So what tools
30:07can I give my agent that gives it full
30:09access that gives it a large surface
30:11area of things to do like that gives it
30:13full power but that allows me to just
30:16make decisions so that it I just say yes
30:18or no right so the tool stack that you
30:21use is important to this the stuff that
30:23you want to work so when you combine
30:25computer use with loops with the right
30:27tools for the job I don't think there's
30:29an app that people can't build I think
30:32the one thing that blocks a lot of
30:33people is not using the right tool for
30:35the right job. And what's interesting is
30:37if you ask codeex what tools are best
30:40for you, Codex itself will tell you,
30:42okay, this works for me, this works for
30:43me, and these are the reasons why.
30:45>> And I think this gets at kind of what um
30:47I think GMO Roush from Verscell talks
30:50about a lot. He talked about this on the
30:51Naval podcast that there's this whole
30:54building block economy building around
30:56agents where agents don't have to build
30:58any like a lot of things from scratch.
31:01Companies are building these building
31:02blocks that AI agents can just use. And
31:05if you know what you want to use, like
31:07you brought up Eve earlier, which is
31:08Versel's um like uh the agent framework
31:11and you can just spin up an agent,
31:13right? in the way that I uh and I I was
31:15using their AI SDK I think is what it's
31:19called for my Replet clone. That's what
31:21um and so when I asked um codeex to
31:25build replet it used it used convex for
31:30the database which so it just used this
31:32external tool and then it just used the
31:34AI SDK from Verscell um to have an AI
31:38that's the agent that generates the code
31:41inside the app and it just used these
31:44and then it used Daytona for the sandbox
31:46because you need a sandbox to generate
31:48and run the the code for every single
31:51app. And that's basically
31:53like all the agent did was kind of
31:56connect these pre-built building blocks
31:58like Legos. And that's how it creates
32:00all these different things.
32:02>> And and one thing I I will say that most
32:04of these tools like you mentioned,
32:06Convex, Eve, what's interesting is and
32:09especially with Eve, I really like Eve.
32:11And the reason being is the decision
32:13behind how an agent is constructed. Now,
32:16this might seem very high level, but
32:17it's so simple. If you notice, Riley,
32:21you the the file names, the folder names
32:24kind of give you a description as to
32:25what it does. Agent.ts. So maybe this is
32:27where the agents defined skills folder.
32:30Oh, this is where the skills are going
32:31to live. Tools, this is where the tools
32:33are going to live. The sandbox, this is
32:34where the sandbox is going to live.
32:35Channels, this is where, you know, the
32:37agents communications channel are going
32:39to be configured. Connection some. So
32:41they constructed the framework with file
32:44names. And the reason why they did that
32:46is the agent, guess what? has access to
32:49all the files. So when it comes to
32:51making changes and reproducing things,
32:53it just makes it so simple, right? And
32:55that's why like build like using tools
32:58that really care about how your agent
33:00uses it is very important. And to
33:03Gearmo's point, there is going to be a
33:04whole economy where some people's
33:07companies are literally plugins,
33:09markdowns, and MCPs, right? And the
33:12focus is how good can they agent use it?
33:15Can I give it one prompt and can it
33:17build everything from scratch? And the
33:18reason why the Riley benchmark passes
33:21with Replet, the new models pass with
33:22Replet is because the models know of
33:24these tools and instead of building the
33:26thing from scratch or using AWS or using
33:29whatever low-level primitive, they're
33:32just going to build using these tools
33:33and it just makes it so much easier. And
33:36this is why we still need developers. I
33:38still want developers, really smart
33:39people to maintain those things, right?
33:40I don't think we've got to the point
33:42where AI could fully maintain those
33:43things. But that's an important point to
33:46realize that we I want to give my agent
33:48a tool that it can fully use and
33:50maximize on. And sometimes this comes
33:52through experience. So I would be
33:54someone who uses a ton of tools and see
33:56what you like
33:57>> about that. Yes, 100%. Um was there uh
34:01another was there number four or was
34:03that the the finality of the list?
34:05>> Well, number four is this. Number four
34:07is and and this is going to be
34:09controversial. Use multiple threats. I
34:13think there is this um like in
34:16instinctively you have this idea where
34:18like you're in a flow state you're going
34:20backto back with the agent you know
34:22you're there's this notion of sticking
34:25to one thread may I present that using
34:29multiple threads does two things first
34:31and foremost you get used to this idea
34:35of compartmentalizing
34:37features that don't touch each other or
34:39work that doesn't affect each other so
34:42that you can work on multiple things at
34:43the same time. But number two, what it's
34:47important to realize is different
34:49threads is a fresh pair is is fresh
34:52context. So now I'm basically using the
34:55agent at its most capable state. If I
34:57got a fresh agent versus an agent with
35:01120k tokens context already taken, the
35:04fresh agent will outperform any single
35:06day of the week. And what's cool with
35:07Codex, right? I don't know if most
35:09people know this. You can spin up
35:11multiple threads from a single thread.
35:13Now, um I I I did this on a on a work
35:16codeex account, so fortunately I can't
35:17show everybody, but one thing that I did
35:20was I built uh a full thing at work and
35:23then I told in the thread, "Spin up as
35:25many threads as they are features. Use
35:28computer use to test each one. Let me
35:30know if everything passes." 10 threads
35:33were generated, Riley. And then I went
35:36to hang out with my wife and let Codex
35:37do the work. Right? So management is
35:40important.
35:40>> It's the only tool to my knowledge. When
35:42you're in Codeex, you can ask Codeex
35:45directly. You can say, "Hey, I want
35:47you," you could literally just list all
35:49of the different things. I want you to
35:51do these 10 tasks in 10 different
35:53threads and it will spin up a new
35:55session. You'll see them being spun up
35:57in the side and it's incredibly
36:00interesting. One other thing that you
36:01can do with the multiple threads is you
36:03can say to codeex, hey, I want you to
36:06please summarize all of the threads that
36:08we have today and create a little
36:09document of all the things that we've
36:11created. It'll just read them. So it so
36:14by default, yeah, you have these
36:15isolated threads, which is nice. But if
36:17you ever want context from other
36:19threads, you can just say, "Hey, earlier
36:21today we were working on these different
36:23things." And since I do a lot of
36:24knowledge work and I'm constantly doing
36:26my email, you know, you know, you use it
36:28as a builder and that's great, right?
36:30you're a builder at heart. You know, I'm
36:31a marketer. I'm kind of operate a
36:33company like from the marketing
36:34perspective. So, I'm just in all these
36:36different softwares. I'm constantly
36:37emailing different people, um, writing
36:40different scripts. It's really useful to
36:42just get an AI to just summarize all the
36:44different things I've done. And it'll
36:45just create a concise one pager. And you
36:48can even ask it to include little links
36:51to the thread in your little document
36:53that it creates because each session has
36:55its own thread. Be like, you created
36:57this here. And if you click on it, it'll
36:59take you to that thread directly, which
37:00is really a really cool um a really cool
37:03feature in my opinion. Can you open up
37:06codeex again? I want to discuss three
37:09things that chat GBT talked about and
37:12this is their new sites. There it is.
37:14There it is. You see sites on the sites
37:16right there. Okay.
37:17>> Okay. Okay. There you go. There you
37:19learn. So, so okay. So this is a new
37:21feature that um OpenAI announced which
37:26uh they announced this a few weeks ago
37:27but it was only available for teams
37:29right so notice how when you clicked on
37:30that now it says create a website that
37:32and just say that is hello world let's
37:34just do one that's really quick um and
37:36then make it like the fastest model that
37:38we can do so we can show it.
37:40>> Yeah. Oh I love this by the way although
37:43>> it's lovely. Very lovely. Yeah. So you
37:45can fire this off and what this will do
37:47is this will create a hosted site that
37:50you can send to people, right? This is
37:52going to be hosted on the internet and
37:54you can share this with people and I
37:55believe that as part of their
37:57announcement yesterday, this is
37:58available to everyone. This was only a
38:00teams feature. Now I think you can host
38:02sites and share them with friends um
38:05like you can do on something like replet
38:06or lovable, right?
38:08>> But notice how you had to you you go to
38:10sites. There's one thing I noticed and I
38:12click create create new site and it's a
38:15prompt with like I'm I'm assuming this
38:17is like some sort of like combination of
38:18skills and MCP. Yeah. So design picture
38:21app. So the reason why I bring this up
38:22is this is how capable like this is how
38:26like a capable model is important. A lot
38:28of this stuff a lot of the cool updates
38:30they made were product facing. It's
38:32still the same underlying model and it
38:34just shows you that soul in particular
38:37covers a large surface area. And I go
38:39back to the first point I made in our
38:40previous part. The best coding model is
38:43the best everything model. Sorry.
38:45>> No, 100%. You don't even need to go to
38:47the sites tab, too. Like you can press
38:49command N on on codec.
38:51>> Hold on. Let me
38:54>> You can press command N and you can just
38:56Wait, just Yeah. Get rid of that, I
38:57guess.
38:58>> No, that that's what happened when I
38:59pressed command N.
39:00>> Yeah. Yeah. Yeah. It took you to It was
39:02already written. So like you just go to
39:04a new chat.
39:05>> Command. Yeah. And then you just type at
39:07sites
39:09>> and there you go. You type at sites and
39:11then you could just build a site. You
39:12can vibe code it and you can and like
39:15you can host it on their platform and
39:16then once you like it, you could host it
39:19on Verscell if you're used to using
39:20Verscell, but this is an easier way for
39:22people who aren't technical to vibe
39:24code. Like they've basically turned
39:26Codex into a full vibe coding platform
39:28with this and like it allows you to
39:29iterate. It gives the model an ability
39:31to use this plugin. And I think if you
39:33were to ask for a website, it would just
39:35automatically host it in this and then
39:37the brow
39:38>> and then you can open this whatever the
39:40link that it creates like you can open
39:42it in the browser which is really cool.
39:44Another
39:44>> another thing I want to bring up while
39:46it's loading actually. Can you open up
39:47the side panel? Hit the thing on the top
39:49right on the right side.
39:51>> On the right side. Yes. Here.
39:52>> Yeah. So you can open up a browser. Um
39:56just open up the browser.
39:57>> By the way, really good. Really really
40:00good. So as of this latest update now it
40:03can open up multiple tabs. So like I I
40:06was using it
40:07>> know that.
40:08>> Yeah. So you can open up multiple
40:10browser tabs.
40:11>> So you can open up new uh multiple
40:13browser tabs now. Um and yeah it's a
40:17full browser. So one of the
40:19announcements they made yesterday is
40:20they're sunsetting atlas. So OpenAI is
40:22sunsetting atlas which they used to have
40:24a team working on atlas which is the
40:26browser and they've cut or at least
40:28merged that team. They basically sunset
40:30the private app and they're putting all
40:32of their effort in making the inapp
40:34browser really good uh inside codeex.
40:36And so that is what makes me so bullish
40:38on AI agent native sites is like so many
40:42people who use chat GPT are now going to
40:45have this tool and like when they ask
40:46for something instead of it opening up
40:48Chrome in some side browser it's just
40:50going to open it up in the browser right
40:52there.
40:53>> And you used a term for this. You were
40:54using a term for this. What was it like
40:56mega app or super app? Super app. Right.
40:58So, so s a super app is is this is
41:01codecs, right? Which you can think of as
41:03like an AI operating system. Um, an
41:06agent native app or a mini app is an app
41:10that is meant to be used by an agent
41:12first. Like I think it's like agent
41:14first. Dan Shipper was talking about
41:16this on this last he's been building
41:17agent native apps. These are apps that
41:19he never goes to
41:21>> unless he's using an agent. an agent can
41:23basically pull up this little email mini
41:26app that shows him the app that the AI
41:28drafted and he can just send them off or
41:30like provide feedback to codeex. And so
41:33these are these apps that I think are
41:34going to emerge which are just apps that
41:37are made to be used by agents with you
41:40like your agent can show you things but
41:43like there's buttons in it but it's
41:44meant to be agent native and so I think
41:46that might be what you were asking
41:47about. The way that I see it is a it is
41:49an operating system. We're seeing the
41:51early early glimpses of what an AI
41:54operating system if you were to buy a
41:55computer and it just had this loaded up
41:58it could you could basically do
42:00everything and you just chat with an AI
42:02a thing happens or an interface loads up
42:05you use the interface and then you move
42:07to a different chat uh interface and so
42:09yeah that's what makes this is so
42:10exciting for me is like we're literally
42:12watching a kind of it's just a new way
42:15of using a computer we're just in the
42:17early days of it so it feels kind of
42:18like chat GPT but
42:20If you get good enough at using tools
42:22like this, I'm not even just saying use
42:24codecs, even though I do think it's like
42:26the best in this paradigm, you you'll
42:28really start to become more agent native
42:30and you'll be like, "Wow, this can
42:32literally do anything." Anyway, we
42:34covered a lot today. Uh Ross, Mike, um I
42:37guess my closing question to you is
42:39this. What are you excited about in
42:41terms of codecs? like what are the next
42:43few things you want to work on and and
42:45like in order of like in terms of
42:48acquiring skills around codecs, what do
42:50you what do you want to do next?
42:52>> Yeah, I think like it it goes back to
42:54like and I tweeted this. It goes back to
42:55what I said earlier like I'm really
42:56starting to think bigger, right? I'm
42:58really like starting to like those like
43:01dream ideas, apps or businesses that I
43:04had in my mind that are like, man, maybe
43:06I don't have the intelligence level or I
43:07don't have the time or I maybe I just
43:09don't have the wit to figure out all the
43:11like, you know, nitty-gritty stuff. I
43:13really can with like, you know, with the
43:16latest and greatest model and with a
43:17harness like Codex, right? So, the big
43:19thing for me right now, Riley, is I'm
43:20like writing this running list of like
43:22things I've wanted to build, whether it
43:24was a kid or whether I was now. And I'm
43:27just building stuff. Like, I have like
43:28eight to nine different threads that are
43:31running that are building things. So, to
43:33me, like my eyes have widened um mostly
43:36because of Fable TBH, but then using GPT
43:405.6 Soul, because we're going to lose
43:41Fable in a bit, it made me realize that,
43:43oh, I have a model that's really, really
43:46good for a fraction of the cost. that's
43:48able, needs a little bit of guidance,
43:50but that's still able. So, me, my mind
43:52has expanded. Now, in terms of like
43:54things that I need to learn and skills
43:57I'm trying to acquire. I think it's the
43:59things the models are not good at,
44:00right? So, design, I'm really starting
44:02like I have a a running catalog of like
44:05really good design, right? So design
44:07inspo I'm trying to train myself on like
44:09okay colors typography spacing because
44:12believe it or not even as good as the
44:14cloud models are with design
44:16at at a certain point we'll start to
44:18realize what are the AI sites and what
44:20are not right so design is a big one and
44:23the other thing is actually writing like
44:25and being able to explain stuff like as
44:27the model gets better like sure it'll
44:30have intuition on what you mean but
44:31imagine you and I had the ability to
44:33communicate effectively explain exactly
44:36what it is we want, what it is what we
44:38expect, and what the feedback loop is. I
44:40find myself that the better I'm getting
44:42at explaining these things, the better
44:44outputs I'm getting. So, design and
44:46writing is a big thing that I'm trying
44:47to improve myself with. But in terms of
44:48what I'm doing, I'm really just trying
44:50to push even further. What can I build?
44:53What can I do? And one final thing I'll
44:55say is nowadays I've realized most
44:57products are markdown files. Most
45:00products are markdown files. So, don't
45:03spend time building a markdown file.
45:05spend time building something that is
45:07actually useful um and important and if
45:11those things are difficult well now you
45:13have access to the latest greatest
45:14model. So that's that's pretty much my
45:17thoughts around what's going on right
45:18now.
45:19>> Yeah. And yeah and for me I'm just going
45:22to be honing my marketing workflows. I'm
45:23going to spend a lot of time on this. I
45:25think this is the most under
45:26underexplored things like using it for
45:28operations and marketing and basically
45:30using it as like my operating system.
45:33And um
45:34>> I think you're one of the few people who
45:36who got jumped in with the engineers and
45:39learned how to use the tools with them,
45:41but now you have this unexplored field
45:43that a lot of people in the space are
45:44not really privy to these tools. And
45:46even if they are, they're not really
45:47that good. Cuz I'll see the content on
45:49Instagram and whatever. And sometimes I
45:51cringe at the advice that people like,
45:54you know, give. Like I remember I saw
45:56this one video where the person was
45:57like, "Oh, like you can use uh GLM 4
46:02point whatever and you could save a lot
46:03of money. It's such a great model." And
46:05I was like, "No, it's not a great
46:06model."
46:07>> The education the education gap is
46:10massive. And I I think that's one thing
46:11I wanted to bring up earlier uh which is
46:14you can build all these different
46:16products. You can do so many things and
46:18you uh I honestly think the biggest
46:20opportunity is education. I think
46:23there's a huge education gap in the
46:25market right now. I don't think there's
46:26that big of a tool gap. In fact, I think
46:28the tools are progressing too fast for
46:30people to actually adopt them. And so
46:32that's why I'm focusing more on content
46:34now is because I've I think ed people
46:37are not properly educated and people
46:40want to be educated. There's just a
46:41massive amount of terrible content out
46:44there especially on short form
46:45platforms. And so if you can figure out
46:47a way to be a voice of reason, teach
46:49people how to use these tools and then
46:51even go into businesses and just help
46:52them out by like by implementing these
46:55tools and teaching them the workflows,
46:56you can make so much money, especially
46:59as these tools get better at general
47:01knowledge work tasks. Actually, it's
47:02already good enough. It's just about how
47:04do you actually like conceptualize the
47:07operating system and configure loops but
47:10for marketing tasks or configure
47:12automations but for marketing tasks. um
47:14or configure the tools that we brought
47:16up, plugins, what plugins and tools do
47:18you need to use um and and and like what
47:21is the workflow? And so that's what I'm
47:22really focused on right now. Agentic
47:24workflows for non-coding tasks and how
47:27to conceptualize it and teach people at
47:29scale. Um that's just kind of like my
47:31main focus.
47:33>> So make sure you're subscribed to the
47:34Riley Brown channel.
47:36>> Yeah, let's go. Anyway, it's always a
47:38pleasure to have you on the pod. Um I'll
47:40hope to have you on soon. Um we should
47:42do this like every few months, dude. I
47:43love it.
47:44>> I'm I'm game again. Thank you, Riley,
47:46for having me. I really appreciate the
47:47opportunity. And thank you everybody for
47:49watching.

About this transcript

This page contains the full transcript of OpenAI Just Merged ChatGPT and Codex. This Changes Everything. by Riley Brown, generated from the public captions YouTube serves with the video. The transcript has 9,485 words across 1,302 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.