Latest / In the Loop AI / The Coding Model Wars: Claude Opus 4.6 vs GPT-5.3 Codex

Transcript

The Coding Model Wars: Claude Opus 4.6 vs GPT-5.3 Codex · In the Loop AI

0:00
Ejaaz: 48 hours ago, Anthropic dropped Claude Opus 4.6, the world's most powerful AI model.
0:06
Ejaaz: And literally 20 minutes later, OpenAI dropped Codex 5.3, which is not only
0:10
Ejaaz: better, but also built itself.
0:13
Ejaaz: Now, to say both of these models are powerful would literally be the understatement of the century.
0:18
Ejaaz: By the time I'd eaten breakfast yesterday, one of the models had discovered
0:20
Ejaaz: 500 security flaws, which no one else had discovered before.
0:24
Ejaaz: And by lunchtime, a bunch of software stocks were down hundreds of billions
0:28
Ejaaz: of dollars out of fear that these models would replace entire teams.
0:31
Ejaaz: And it's actually already happened. These models can replace a team of 50 software
0:35
Ejaaz: engineers, rebuild Pokemon from scratch, and so much more.
0:38
Ejaaz: And in this episode, we're going to be doing a live demo side by side to show
0:42
Ejaaz: you which model is the best.
0:44
Josh: Yeah, this is pretty cool. I wanted to spend a lot of time this episode kind
0:47
Josh: of introducing people to these models, what they could do, how they work through
0:50
Josh: demos that we're going to perform ourselves.
0:52
Josh: These are definitely two frontier models but i think more importantly they're
0:55
Josh: frontier coding models and when people hear that i
0:58
Josh: think a lot of them get turned away because it seems like this complicated
1:01
Josh: thing like you need to be a developer in order to use them and we
1:04
Josh: are here to tell you that is not the case as from
1:08
Josh: one non-technical person to another i fed this
1:11
Josh: model a prompt i fed it some assets and
1:14
Josh: then i pressed play and what i got is a
1:16
Josh: side-scrolling game which was exactly what i asked for so on the screen now
1:19
Josh: you're seeing the one shot prompt that i fed this model to ask to create a side
1:23
Josh: scroller that was like mario that we can actually play so it has coins and i
1:28
Josh: don't think the gravity quite works what you're saying is that it understands
1:31
Josh: physics it is able to generate graphics and it plays like a pretty solid side
1:37
Josh: scroller and i created this in five minutes,
1:40
Josh: with one prompt and it actually works what.
1:43
Ejaaz: Was the prompt that you used josh
1:44
Josh: Yeah so i'll pause playing this game to
1:47
Josh: actually show you the the prompt it was very simple it was this
1:50
Josh: one paragraph i want you to make a game you can
1:52
Josh: use python or c++ whatever you find the most convenient a 2d
1:56
Josh: platformer that closely resembles super mario use the
1:58
Josh: attached background image and sprites found in the
2:01
Josh: asset folder take into account that the sprites don't come with transparent background
2:04
Josh: but pink ones so you need to filter the background and for those who are
2:07
Josh: watching you can actually see the sprites on my screen they were
2:10
Josh: just a series of assets that there was no context given as
2:13
Josh: to what each one of them was but the model reasoned through it it removed
2:16
Josh: the background and it actually generated a pretty good representation of
2:19
Josh: that now this was built one shot on codex which
2:22
Josh: is the new open ai mac application that just released this
2:25
Josh: week and i wanted to compare it to claude
2:28
Josh: so i have another instance here on the screen with claude this is using opus
2:31
Josh: 4.6 the newest frontier model that they just released this week and i want to
2:35
Josh: do an exact one-to-one comparison so i'm gonna launch the same exact prompt
2:39
Josh: we're gonna have that cook on codex or we're gonna have that cook in claude
2:42
Josh: code and in the meantime you just maybe we can kind of talk about more of what
2:46
Josh: these models do and how they work well.
2:48
Ejaaz: Before we do that actually um as you set this game up i ran it on claude opus
2:53
Ejaaz: 4.6 as well but with a slight twist okay
2:57
Josh: Let's see your output what do we have okay.
2:58
Ejaaz: Uh i don't know if you can see my screen
3:01
Ejaaz: but it is the exact game that you just created but i don't know if those characters
3:06
Ejaaz: look uh kind of familiar to you we have the uh hero protagonist character which
3:10
Ejaaz: is uh my beautiful face and my beautiful person ejaz um and we have uh who's
3:15
Ejaaz: this enemy over here that looks a lot like the bear guy
3:20
Ejaaz: and listen we can double jump here josh and i think yep i can crush you but every time i mean this
3:27
Ejaaz: Kind of jokes aside, this is insane. This took me like around three minutes to build end-to-end.
3:34
Ejaaz: I used the exact same prompt that you gave me.
3:36
Ejaaz: And we didn't have sprites ready-made of ourselves, right?
3:39
Ejaaz: We didn't have like cartoon images of ourselves. So I uploaded an image that
3:43
Ejaaz: we had taken, I don't know, like six months ago and said, hey,
3:46
Ejaaz: can you make game avatars out of this?
3:48
Ejaaz: It did it in 20 seconds. And then I said, could you add these to the game and
3:51
Ejaaz: replace the enemy with Josh and the protagonist with Ejaz? And it did it in a minute.
3:57
Josh: So here we go. That's pretty amazing. And these are really, these are just using
4:00
Josh: standard desktop applications. So what you're using right here,
4:03
Josh: this was done in Cloud Code, right?
4:05
Josh: You just went onto Cloud, the MacBook, the Mac app. You downloaded it. You put in the prompt.
4:10
Josh: You shared some assets. And now it built this amazing game in one single prompt.
4:13
Josh: And we're actually going to experiment further in this episode where we're going
4:17
Josh: to create a trading room that does actual real-time stock analysis.
4:21
Josh: So as I'm curating the prompts and as we're getting ready for that second demo,
4:24
Josh: maybe we could walk through what makes these models so exceptional.
4:27
Ejaaz: Yeah, well, you might actually notice the first difference on screen right now.
4:32
Ejaaz: If you notice, if you look closely, my avatar is kind of glitching out, right?
4:36
Ejaaz: And if you compare it to your Codex game that you just coded up,
4:39
Ejaaz: there's no glitches. It runs super smoothly.
4:41
Ejaaz: And the main takeaway here is Codex 5.3 is a superior coding model to Anthropic.
4:48
Ejaaz: And that's a sentence I never thought I would say, at least for the next couple
4:51
Ejaaz: of years, because Anthropic has held that prestige and title for so long.
4:55
Ejaaz: But since Code Red was initiated in open air around three months ago,
4:59
Ejaaz: Sam has devoted pretty much all his resources towards building the best coding model.
5:04
Ejaaz: And the benchmarks don't lie. It is a full 12 points on the software engineering
5:08
Ejaaz: benchmark ahead of Claude Opus 4.6.
5:11
Josh: That's a pretty significant difference.
5:12
Ejaaz: So I've actually pulled up a more general comparison between the two models here.
5:17
Ejaaz: And it summarizes it really well. So if we look at Claude's model,
5:20
Ejaaz: Opus 4.6, what's good about it?
5:22
Ejaaz: Well, they've 5x the context window.
5:26
Ejaaz: So it's gone up to a million tokens or rather characters that you can put in
5:30
Ejaaz: a single prompt, which if you want to understand how powerful this is,
5:33
Ejaaz: you can just put way more information into your initial prompt.
5:36
Ejaaz: It has much better context and memory. So you can end up cooking up much better
5:40
Ejaaz: products overall, which is very, very impressive and important to have.
5:44
Ejaaz: Number two, I would think about this as an orchestration model.
5:49
Ejaaz: So if you look at specific benchmarks, it is beaten OpenAI at GDP eval.
5:54
Ejaaz: GDP eval is a benchmark where they go out and they test a model's performance
5:58
Ejaaz: at a really complex task versus a professional human that would normally do that task.
6:04
Ejaaz: And the decision is, would you use the AI model or would you use the human?
6:08
Ejaaz: And in this case, you would choose Claude 4.6 over humans way more than you
6:12
Ejaaz: would choose OpenAI's latest model. So that's a really important thing.
6:16
Ejaaz: And the point around Claude's latest model is that it doesn't code as well as
6:21
Ejaaz: codecs, but it can orchestrate a bunch of agents and overall activity better than OpenAI.
6:27
Ejaaz: Now, if you look at Codex and OpenAI's new models specifically,
6:31
Ejaaz: It wins on the software engineering. It is simply a better software engineer
6:36
Ejaaz: than Claude is, which is a massive flip around and shows that it's a testament
6:40
Ejaaz: to how much resources and fine-tuning that OpenAI has been able to achieve.
6:44
Josh: And to the note on the quality of the models here, my prompt is done in Claude
6:49
Josh: code that I used, the same one that we used in Codex. And I'm going to run it
6:52
Josh: here for the first time now.
6:53
Josh: You can see on screen and we'll see what it looks like.
6:56
Josh: So underneath, we have our Codex version, which looks beautiful.
6:59
Josh: On top we have our brand new version that was just made by opus now i haven't
7:03
Josh: tried this yet so we're going to see what happens when i press space to start,
7:09
Josh: so it looks like opus has failed to create a
7:12
Josh: floor so i am just falling through the floor until the game ends um okay so
7:18
Josh: just based on this one demo alone this is a fairly significant difference where
7:21
Josh: gpt's codex has created a beautiful side scroller it doesn't have gravity but
7:25
Josh: i could just ask it to or it has gravity it's a little too much i could ask
7:28
Josh: it to lower it opus doesn't even work at all,
7:31
Josh: And again, the test was just a one-shot prompt. So I'm going to get back to
7:34
Josh: work prompting it again to build this new application, the trading application.
7:38
Josh: We'll follow up with that. But I think that's a funny kind of demo just to showcase
7:42
Josh: that one actually is kind of superior in the other in this one use case, at least.
7:46
Ejaaz: Yeah, I mean, you said it pretty clearly, which is Codex is the best coding AI model.
7:53
Ejaaz: And I have to like, I can't emphasize that enough because OpenAI for a long
7:57
Ejaaz: time was behind Anthropic and by a massive margin. and in some way,
8:01
Ejaaz: shape, or form, they've been able to catch up.
8:03
Ejaaz: Now, what's interesting here is both companies have focused on each other's goals.
8:10
Ejaaz: So when Anthropic was typically meant to be the leading frontier model in coding,
8:15
Ejaaz: it now has decided to focus on what OpenAI was really good at,
8:18
Ejaaz: which is overall orchestration and being a better generalized model, right?
8:23
Josh: They're taking each other's lunch. Yeah, exactly.
8:25
Ejaaz: OpenAI has decided to eat Anthropic's
8:27
Ejaaz: lunch and say, okay, we've got the generalized stuff sorted out.
8:30
Ejaaz: Let's try and figure out the coding specific niche, highly defined,
8:34
Ejaaz: professionalized functions. And it's produced the best coding model.
8:37
Ejaaz: So it's kind of a weird win-win for both labs.
8:41
Ejaaz: And what's awesome about this is they both now have really well-rounded,
8:46
Ejaaz: but also very specialized models.
8:48
Ejaaz: And the reason why this is important is, and this is like kind of maybe my hot take,
8:54
Ejaaz: I don't think the coding models matter, Josh. I actually don't think the generalized models matter either.
8:59
Ejaaz: I think they're both going off to something much bigger, which is creating the
9:03
Ejaaz: operating system for the future of work.
9:06
Ejaaz: They know that AI models and AI agents are gonna automate a ton of different
9:10
Ejaaz: industries and the industries are only gonna pick you if you can do both generalized
9:14
Ejaaz: work and hyper-specific work really well.
9:17
Ejaaz: That is coding and orchestration and managing your data.
9:20
Ejaaz: And now we have two amazing models dropped within 20 minutes of each other.
9:24
Ejaaz: That does exactly that to the highest performance metric that we've ever seen before.
9:29
Josh: They're pretty exceptional. So now for this next demo, I have it queued up here.
9:32
Josh: What we're going to do is, what I did is ask the model itself to build me a
9:36
Josh: prompt for this. So I wanted it to create me an AI stock portfolio war room.
9:40
Josh: And I asked, hey, I want to create this, create me a fully fleshed out prompt
9:44
Josh: that kind of should solve this problem with one shot.
9:48
Josh: So what I do is I loaded it up here in our Cloud Code app.
9:52
Josh: And then I also loaded it up into the codex app i created its own
9:55
Josh: project folder and now i'm going to hit send so both of
9:57
Josh: these things are thinking in real time we will check back
10:00
Josh: in once their outputs are done and we'll compare again the second version
10:03
Josh: which is more of a robust one i mean you'll see uh on
10:06
Josh: the cloud screen it has this whole list of to-dos that it wants to do it has
10:10
Josh: an entire plan there's nine different panels that it's going to build it's going
10:14
Josh: to do risk analysis matrix and portfolio action bars and all this stuff so we'll
10:18
Josh: let that cook and let's get back to what separates these what people have been
10:21
Josh: freaking out about on the internet more as these things get going could i.
10:24
Ejaaz: Take three minutes show you some wild demos yeah
10:27
Josh: Let's see what the internet's been demoing while we wait for hours to cook okay.
10:30
Ejaaz: Cool like listen our 2d mario inspired game was cool but imagine if i told you
10:35
Ejaaz: you could recreate the entire pokemon game including levels cities characters
10:41
Ejaaz: and creatures that you fight from scratch in about an hour and 30 minutes
10:46
Ejaaz: That's pretty impressive. That's what we're looking at right now.
10:47
Josh: Wow, it even has the fighting.
10:49
Ejaaz: Yeah, yeah, yeah. And buttons and the multimodal gameplay.
10:53
Ejaaz: And obviously this looks like it's been made by a child image wise,
10:57
Ejaaz: but it's probably going to take you, what, another couple of hours to make a
10:59
Ejaaz: really high fidelity game that you could probably run on your Nintendo Switch or whatever.
11:03
Ejaaz: It is just so impressive that we can do these things.
11:07
Ejaaz: Anyone can do these things with no previous background. Just upload a few images
11:10
Ejaaz: or generate a few images and you can create childhood nostalgic games that are
11:14
Ejaaz: worth billions of dollars, which is just super cool to see.
11:17
Josh: Yeah, one of the cool things that I think it's really important to note is how approachable this is.
11:21
Josh: Like for the recent example that we're having run right now on my screen,
11:25
Josh: all I did was tell it what I wanted and ask it to develop the prompt with me.
11:29
Josh: So even if it feels overwhelming, like you don't really know how to code,
11:32
Josh: you don't know how to prompt things, you can actually just ask the model to
11:35
Josh: help you generate the prompt, help explain to you how it works.
11:38
Josh: And it's a really easy way to build basically anything you can imagine.
11:41
Josh: It's not just games. It's productivity tools. It's CRM tracking.
11:45
Josh: It's whatever you want it to be so i think that's really interesting but it
11:48
Josh: also goes much more technical right i saw another crazy example with the compiler.
11:52
Ejaaz: Okay so for for the tech nerds
11:55
Ejaaz: out there that's been a lot of time coding you are going to
11:57
Ejaaz: be wowed by this um for one of their uh flagship demos for uh opus 4.6 the anthropic
12:04
Ejaaz: team decided to task the model with building a c compiler which is an incredibly
12:11
Ejaaz: complicated execution tool that is required to code up some of the most craziest types of apps.
12:17
Ejaaz: And they just walked away. And they just kind of like looked at it,
12:21
Ejaaz: monitored it, made sure that it wasn't going awry.
12:23
Ejaaz: And in two weeks, let me emphasize that,
12:26
Ejaaz: Two whole weeks, 14 days, it coded nonstop and built this compiler.
12:32
Ejaaz: Now, you might think two weeks is quite a long time. I want my thing done in an hour and a half.
12:36
Ejaaz: Well, let me hearken back to history where previously, if you wanted to create
12:40
Ejaaz: something like this, in today's world, it would take a team of around 50 or
12:44
Ejaaz: so humans, and it would take them a few months to build from scratch. That's today.
12:49
Ejaaz: But back in the day, it would technically have taken them around a decade to
12:54
Ejaaz: build and like thousands of people.
12:56
Ejaaz: So we have just kind of condensed the timeline to create really complicated
13:01
Ejaaz: tools in a matter of hours or weeks in this case.
13:05
Ejaaz: Now, the second thing I want to point out is the fact that these models can
13:09
Ejaaz: go untouched for two weeks is just insane.
13:13
Ejaaz: There was another stat that was released today by OpenAI with,
13:17
Ejaaz: sorry, yesterday with OpenAI is 5.2, I think, 5.2 high, I believe,
13:21
Ejaaz: where it can go pretty much 50% hit rate for 6.6 hours. a time horizon.
13:28
Ejaaz: So that means if you gave it any kind of complicated coding task,
13:31
Ejaaz: 50% of the time in 6.6 hours, it would get that done, completely done.
13:36
Ejaaz: And it would nail it 50% of the time, which is just such an impressive track
13:39
Ejaaz: record when you look back a year.
13:41
Ejaaz: And that time was, what was it like 30 minutes, maybe an hour.
13:45
Ejaaz: So every iteration, we see this thing double. It's just so insane.
13:49
Josh: Yeah, it's really, it's unbelievable and almost like intimidating how
13:52
Josh: capable and competent it is even for someone who
13:55
Josh: is a novel at writing code it's not about writing
13:58
Josh: code it's about being able to generate whatever you want it to so like if you
14:03
Josh: think of it you kind of in a way it abstracts the code away and allows you to
14:07
Josh: just speak the english language and get what you want from speaking english
14:11
Josh: and in a way that you understand and it will help walk you through the way one
14:14
Josh: of the things that i love about cloud code in particular is the plan mode.
14:17
Josh: If you leave a lot of things out of your prompt, it'll actually just continue
14:20
Josh: to prompt you with additional questions to understand where you want.
14:23
Josh: And one of the most fascinating things that I read about GPT's 5.3 codex in
14:29
Josh: particular is like you mentioned in the intro, it helps build itself.
14:33
Josh: And I don't think that can be overstated because this is the first model in
14:37
Josh: the history of OpenAI that has helped with the building and construction of itself.
14:42
Josh: And what happens as that starts to ramp up, right? If you think of each model
14:47
Josh: iteration as a flywheel, what is the constraint?
14:49
Josh: The two constraints are the speed at which a developer can actually build it
14:54
Josh: and then create the test for it and make sure that it's safe to ready to deploy.
14:57
Josh: And then it's the hardware that's required to actually train the model.
15:00
Josh: What we're seeing with Codex and Opus, which I really believe was kind of Sonnet,
15:04
Josh: is the incremental improvements.
15:06
Josh: Now, for the incremental improvements that don't require an entirely new training
15:09
Josh: run, the real constraint is the actual software and what you could squeeze out of it.
15:13
Josh: And when you have a model that's helping you build this
15:16
Josh: software that can think for 6 12 24 hours
15:19
Josh: at a time even longer and that is it kind
15:22
Josh: of creates this like self-fulfilling loop right where the models use the
15:25
Josh: new models to make the new models the future models
15:28
Josh: stronger and more powerful and better and i thought that was a really interesting
15:31
Josh: thing to note is that this is the first self propagating model where it ran
15:35
Josh: a lot of the test for itself it introduced new code that made itself better
15:38
Josh: and as we continue to see that you can start to imagine that vertical that like
15:43
Josh: exponential progress line going pretty close to vertical and things getting
15:46
Josh: really good like really really quick.
15:48
Ejaaz: I think what most people listening to this might think is that,
15:53
Ejaaz: well, what was different before?
15:56
Ejaaz: Well, previously, models would just kind of work in a very analog mode.
16:00
Ejaaz: You would just point it at a problem
16:02
Ejaaz: and it would just understand what the problem was and then solve it.
16:05
Ejaaz: But it lacked that awareness and wider context as to like what the wider vision
16:10
Ejaaz: and goal was to achieve and then figuring out stuff for itself.
16:13
Ejaaz: You always had to kind of handhold it. But now with its ability to kind of like
16:17
Ejaaz: understand what it's trying to do and look internally and say,
16:21
Ejaaz: huh, I made that mistake because of this error in my code.
16:24
Ejaaz: I'm going to now like rewrite my code and then I'll be better at it.
16:27
Ejaaz: It kind of functions similarly to a human. Now, I actually saw a great analogy.
16:32
Ejaaz: I forgot who wrote it, but it's
16:34
Ejaaz: fantastic. where if you imagine yourself standing on a sidewalk, right?
16:39
Ejaaz: And a Bugatti Veyron drives super fast by you at let's say 200 miles an hour,
16:45
Ejaaz: you'll be like, wow, that's kind of fast.
16:47
Ejaaz: And then two minutes later, another Bugatti drives by you at 300 miles an hour.
16:53
Ejaaz: You'll be like, wow, that's kind of fast. But you wouldn't really notice the
16:56
Ejaaz: difference between that 100 mile an hour difference, right?
17:00
Ejaaz: But if you were in the car strapped in, you would notice it is significantly
17:04
Ejaaz: improved. And that's how software engineers feel right now.
17:07
Ejaaz: Now, if you're someone that doesn't code all the time, you're not necessarily
17:10
Ejaaz: going to understand these impacts, but it's really important for those of you
17:13
Ejaaz: listening to this to figure out that this is massively impactful and will change
17:17
Ejaaz: the way that a lot of things are happening today.
17:19
Ejaaz: I mean, just take a look at this, right? This is a direct quote from someone
17:24
Ejaaz: who is building at a major tech company, Rakuten.
17:27
Ejaaz: And the quote here says, Claude Opus 4.6 autonomously closed 13 issues and assigned
17:32
Ejaaz: 12 issues to the right team members in a single day, managing a 50-person organization
17:38
Ejaaz: across six repositories.
17:40
Ejaaz: Josh, do you know who else is responsible for doing that?
17:44
Ejaaz: An entire team of product managers that each get paid a quarter of a million
17:48
Ejaaz: dollars in compensation automatically.
17:50
Josh: Minimum per year at least yeah their.
17:52
Ejaaz: Jobs are automated now
17:53
Josh: Well one of the earlier moments in
17:56
Josh: which i realized this was pretty profound is is when claude co-work they
17:59
Josh: said they built it with what just a hint like four people over the course of
18:03
Josh: 10 days and it was 100 built by the current model of claude which is opus 4.5
18:09
Josh: at the time like the the amount of leverage from these tools is so high but
18:14
Josh: it cuts both ways it's like if you can design and develop a product in 10 days,
18:19
Josh: then that means another company can probably do that in five.
18:23
Josh: And it starts to lower the competitive threshold for these companies to catch up.
18:28
Josh: And it starts to raise the bar of what is possible.
18:32
Josh: Like if you could build something that profound in 10 days, what can you build
18:35
Josh: over the course of six months?
18:37
Josh: Like, can you really build something fantastic that has a moat that like actually
18:42
Josh: delivers on the total power that you have by leveraging this AI?
18:46
Josh: It's going to be interesting to see because i mean what we're finding even with
18:49
Josh: the the codex and opus dual launch is that these companies are right next to
18:53
Josh: each other and if one publishes something,
18:55
Josh: profound or something that attracts a lot of users they're just a few days and
18:59
Josh: a few prompts away from copying it and that's like a pretty difficult thing
19:03
Josh: to compete against on on the software front well.
19:06
Ejaaz: That's why if we look at the stock market over the last couple of days like
19:10
Ejaaz: it's down trillions of dollars and i'm not exaggerating if you look at microsoft
19:13
Ejaaz: over the last two weeks, the stock is down 20%. It's trading like a meme stock, which is just insane.
19:20
Ejaaz: And the reason why that is, is a lot of investors are anticipating that these models,
19:27
Ejaaz: Specifically Opus 4.6 and Codex 5.3, will just create the tools that these billions
19:34
Ejaaz: of dollars worth of SaaS companies have spent or valued their entire lives on
19:38
Ejaaz: in a couple of seconds, just as you described.
19:40
Ejaaz: Now, the counter argument to this, Josh, is, and Jets of Wine actually kind
19:45
Ejaaz: of went live at a conference and spoke about this and made this point,
19:49
Ejaaz: If you're an AI agent or AI model that is capable of building these tools, right?
19:54
Ejaaz: Why would you rebuild the tool every single time you do a function?
20:00
Ejaaz: Surely you would just access the best tool and use it.
20:03
Ejaaz: So there's a bit more nuance where AI models aren't just gonna recreate your
20:07
Ejaaz: entire software stack if you are at a Fortune 500 company.
20:10
Ejaaz: That kind of doesn't make any sense. There are a bunch of tools that are hyper-optimized to do that.
20:14
Ejaaz: But what it will do is it will connect all of these tools and silos in a much more effective way.
20:20
Ejaaz: And maybe that requires rebuilding parts of it.
20:23
Ejaaz: Maybe it requires kind of connecting different ways, but not rebuilding the entire tools.
20:26
Ejaaz: And whatever operating system that ends up becoming will be the most sticky
20:31
Ejaaz: and valuable company ever.
20:33
Ejaaz: Now, that could be Salesforce, or it could be someone completely different,
20:36
Ejaaz: a startup that we haven't even heard of. And I think that's really important
20:38
Ejaaz: to understand, but people are experimenting.
20:41
Ejaaz: And if you look at this graph right here, which is may not look insane to some,
20:44
Ejaaz: but is insane to me at least, 4% of daily GitHub commits are now clawed code.
20:50
Ejaaz: That was, I think, 5% of what it is today two months ago.
20:57
Ejaaz: So the ascent has just been insane. These companies are adopting it and they are using it.
21:02
Josh: Yeah, the number is just going to keep going up and there's no reason why it
21:05
Josh: wouldn't. It's such a testament. One, the speed.
21:08
Josh: It feels like we're strapped in that car and now we're flying.
21:11
Josh: Two, an outsider might not look like it. It certainly feels like that
21:14
Josh: on the inside and i think a lot of people are starting to notice this and get
21:17
Josh: a little nervous about it too like look at this example on the screen right
21:20
Josh: now this is a prompt from gpt 5.3 codex which basically created an entire minecraft
21:27
Josh: clone in a single prompt and it looks awesome and it works really fast and it
21:32
Josh: was super lightweight and
21:34
Josh: And it says, I also tried on Opus 4.6, but for some reason it got stuck.
21:38
Josh: But you can build anything that you want very, very quickly,
21:42
Josh: like very cheaply as well.
21:44
Josh: What Opus 5.3, or Opus 5.3, I'm getting them all mixed up.
21:49
Josh: What GPT 5.3 Codex offered is double the rates, the double the token rates for
21:55
Josh: the next couple of months.
21:56
Josh: So you actually have the freedom for their $20 a month plan to go and build whatever you want.
22:00
Ejaaz: Can I maybe deliver a hot take, Josh?
22:02
Josh: Yeah, what do you got?
22:03
Ejaaz: I think the most exciting part about these model releases aren't the models themselves.
22:08
Ejaaz: Largely, I think the models are kind of similar in capabilities.
22:12
Ejaaz: They are around the same coding benchmarks, and they can roughly do the same
22:16
Ejaaz: things. They can spin up a bunch of agents and orchestrate themselves.
22:19
Ejaaz: The bigger picture, which I think a lot of people missed, was both companies,
22:24
Ejaaz: Anthropic and OpenAI, are at war with each other.
22:27
Ejaaz: And they're trying to basically build and own the operating system for work,
22:31
Ejaaz: which isn't just a model. it's a software suite.
22:34
Ejaaz: So this week alone, OpenAI didn't just release this new model.
22:37
Ejaaz: They released the Codex app, which is a desktop Mac app, which is kind of like
22:42
Ejaaz: a command line interface, which makes the coding experience way better.
22:45
Ejaaz: And they also launched an enterprise platform called Frontier,
22:48
Ejaaz: which allows Fortune 500 companies to basically take this magical model and
22:54
Ejaaz: give it to non-coders and let them do magical things. Now,
22:58
Ejaaz: All of these products together creates a very sticky experience where it starts
23:02
Ejaaz: to make sense for software engineers and non-software engineers to use these products.
23:07
Ejaaz: And it becomes incredibly sticky, which results in billion-dollar contracts, right?
23:11
Ejaaz: Anthropic has done the same thing over the last two weeks.
23:14
Ejaaz: They released Claude Cowork, they released agent teams this week,
23:18
Ejaaz: and then they released this new model.
23:20
Ejaaz: They're going after the same thing, which it kind of makes sense why they're
23:23
Ejaaz: releasing Super Bowl ads that are kind of shitting on each other now.
23:26
Ejaaz: It makes a lot of sense. And so the point is, if they can own this operating
23:31
Ejaaz: system, this future of work, they will basically be the most valuable company.
23:35
Ejaaz: And I think it's going to be when it takes most.
23:37
Josh: I have to interrupt you here. We have some developments on our prompts that
23:39
Josh: we've been working on, our AI stock war room. Let's go. That I'm going to have
23:42
Josh: to share on the screen right now.
23:44
Josh: So currently what it's doing is it's asking to do some quality assurance testing.
23:48
Josh: So you'll see it actually used a it's taking over control of my browser and
23:52
Josh: it's asking to make prompts on the screen. So you can see all of this that you're
23:56
Josh: seeing right here is generated live, and it's doing an actual real-time debug
24:00
Josh: of the product that it made.
24:02
Josh: It's clicking around, it's resizing things, it's going through the links,
24:05
Josh: and it's running real quality assurance testing on the actual product.
24:09
Josh: It's really amazing to see.
24:12
Josh: This was all just built all these visual charts and they're all accurate so
24:15
Josh: right now we're looking at nvidia we have a chart and i'm not going to mess
24:17
Josh: with it because it's doing the real-time manipulation to do quality assurance
24:20
Josh: checks but it's actually clicking through it's making sure the
24:23
Josh: stats are accurate it's making sure all the widgets work and look it has this
24:28
Josh: amazing graphs already it has sentiment analysis 85 percent of people are bullish
24:32
Josh: on nvidia it has recent signals from the news it has the assessment a risk assessment
24:38
Josh: matrix where it shows the like export controls and chip controls.
24:41
Josh: It has revenue and earnings every single quarter, charted, competitive moats.
24:46
Josh: It has sector comparisons. It's like, this is unbelievable.
24:50
Josh: And it just generated this in a single prompt. And I just find it really funny
24:53
Josh: that we can actually watch this do it in real time.
24:55
Josh: So you'll see in this prompt, it's clicking through, it's taking screenshots of what it's seeing.
25:00
Josh: And then it's digesting, analyzing, and understanding what it made,
25:04
Josh: what it messed up and what it actually still has left to finish.
25:07
Josh: And it generated everything, all of this in real time as we're recording this episode.
25:12
Josh: So fascinating.
25:14
Ejaaz: Wow, it reminds me of some of the research platforms at the former companies
25:19
Ejaaz: that I used to work at and they would pay, I'm not joking, millions of dollars
25:22
Ejaaz: a year to get access to these types of platforms that would give them analysis
25:26
Ejaaz: like what you're showing on the screen right now.
25:28
Josh: And you just built it from scratch. From scratch, and look, it's doing this.
25:32
Josh: I'm not even touching my keyboard. I just searched for Apple and now I'm sure
25:35
Josh: if I go over to the prompt,
25:36
Josh: it's taking screenshots of apple it says apple dashboard
25:39
Josh: looking great let me scroll to see the new three column button row layout and
25:42
Josh: it's checking the button rows and it's really unbelievable like we have the
25:47
Josh: investment thesis the bull case for it the bear case for it catalyst and timelines
25:51
Josh: it has wwdc built in it has the iphone 18 launch props um set up for september,
25:57
Josh: It's like so cool. It's absolutely unbelievable. And now this is a real tool
26:01
Josh: that I'll be able to use to type
26:03
Josh: in whatever stock I want to look at and actually get some analysis on it.
26:07
Josh: Now, I'll go over to Codex over here and it looks like Codex is taking its sweet time.
26:12
Josh: It's still zero out of six tasks completed. So it might take a little while
26:16
Josh: for us to get a visual on that, but it's just amazing to watch this happen in
26:19
Josh: real time as at least Cloud Code and Opus 4.6,
26:24
Josh: does some quality assurance testing live by taking over my browser and running
26:27
Josh: it for itself. I just think this is like, this is amazing.
26:31
Ejaaz: It's magic. Something I just noticed in your Opus chatbot screen when it's going
26:37
Ejaaz: through its thinking, it seems to have like spun up a few different agents or
26:41
Ejaaz: instances of its own self to pull this off.
26:44
Ejaaz: Like I think if you scroll up, like I saw a few kind of like prompts that like
26:47
Ejaaz: suggested that that's what it was doing,
26:49
Ejaaz: which I think is, underscore is a very important point that both of these models
26:53
Ejaaz: can do, which is they can spin up multiple versions of the same model and task
26:59
Ejaaz: it with different things to run in parallel.
27:02
Ejaaz: What this means is you can get a really complicated product like what you're
27:05
Ejaaz: seeing on the screen right now in a matter of minutes because it's running in parallel.
27:11
Ejaaz: So imagine having a bunch of computer science geniuses that you can just duplicate
27:15
Ejaaz: immediately and run at a fraction of the cost of electricity, the cost of inference.
27:20
Ejaaz: And now you start to see why all these NVIDIA chips and stuff are worth so much.
27:24
Ejaaz: Because you want to do cool stuff like this. This is insane.
27:26
Josh: It's actually incredible. Okay, so now I want to test it on Tesla.
27:29
Josh: So I'm going to choose Tesla and see if it actually can do it in.
27:31
Ejaaz: A non-controlled environment. This UI is so cool.
27:33
Josh: It's very pretty. What the hell? This looks great. Okay, so here we have Tesla.
27:37
Josh: It has the charts. We're going to click through the charts. It has the one-week
27:39
Josh: chart, the one-month chart, the three-month chart. That looks fairly accurate.
27:44
Josh: It has the price-to-earnings ratio, the 52-week high, 52-week low.
27:48
Josh: So it looks like at one point it was trading at $4.88, now it's trading at $3.89.
27:52
Josh: The bull case for Tesla, RoboTaxi and FSD driving licenses could unlock $500
27:57
Josh: billion in revenue by 2030.
27:59
Josh: It has the RoboTaxi service launch in Austin that it's preparing for.
28:03
Josh: And let's see the sector comparison. So it's comparing it to Rivian, Baidu, Toyota, Ford.
28:09
Josh: It has the competitive moat where it says it's most strong in brand power,
28:13
Josh: IP patents, and cost advantages.
28:15
Josh: You can see the revenue, the estimate per share earnings.
28:19
Josh: Sentiment is much worse on Tesla than it was on Apple. It's at 52% right now.
28:24
Josh: And it looks like, as it relates to the risk assessment, devaluation and competition
28:29
Josh: and execution are all very high risk.
28:33
Josh: And that's probably an accurate assessment, although I'm not sure the competition
28:36
Josh: is really a problem. The execution is certainly going to be an issue.
28:39
Josh: But it's just amazing to see how well it does. And it even gives it a verdict.
28:44
Josh: So the AI verdict on Tesla is,
28:47
Josh: It's a hold. Tesla's optionality is enormous, but current valuations already
28:51
Josh: prices in multiple moonshots.
28:53
Josh: Execution on RoboTaxi will be the key catalyst. That sounds about right.
28:57
Josh: And it's amazing that we just built this with a single prompt without any oversight from me.
29:03
Josh: And it works. It actually works. It's really just unbelievable how capable these things are.
29:08
Josh: And now I have a dashboard that anytime I want to make a decision,
29:10
Josh: I can type in the ticker and get all this um optionality it even has menus that
29:16
Josh: work look at this profit margins pe ratios market cap wow pretty unbelievable it's.
29:21
Ejaaz: It's a reactive in real time bloomberg terminal oh wait for the modern age
29:26
Josh: There's um there's another feature here that looks like you could compare stocks
29:30
Josh: let's see if this actually works here so if i type in let's say apple's ticker
29:35
Josh: and i hit go will that compare the two now it looks like that doesn't work very
29:40
Josh: well oh my god but it has moving average lines and everything. This is pretty robust.
29:43
Ejaaz: I know it's like the traded and investors dream. Just crazy.
29:48
Ejaaz: Kind of like a side note on this, but like,
29:50
Ejaaz: The fact that Tesla's down and everyone's kind of like bearish on this company,
29:54
Ejaaz: even though they're like rumored to be merging and stuff like this.
29:57
Ejaaz: Like the point being is there's an asymmetry between what the market is seeing
30:02
Ejaaz: and what these inventors and builders are seeing.
30:06
Ejaaz: These AI labs have created what they define as pretty much a low form of AGI.
30:12
Ejaaz: You literally have an AI model that is building the next version of itself.
30:16
Ejaaz: That by description is like a super genius and it's only limited by the function
30:21
Ejaaz: of energy and compute, right?
30:24
Ejaaz: And then investors are looking at this and saying, huh, Amazon and Google are
30:28
Ejaaz: about to spend a combined $500 billion worth of CapEx this year.
30:32
Ejaaz: Kind of bearish, that's a lot of money. So there is a real investment opportunity
30:36
Ejaaz: here to really understand the difference of what these things can actually do.
30:40
Ejaaz: And that might lead to a lot of like opportunities to invest.
30:43
Ejaaz: I don't know, but I know that I'm buying Tesla today and a bunch of google stock
30:46
Josh: Yeah i mean look at this google valuation one this chart looks absolutely gorgeous
30:50
Josh: but two um the ai verdict is a buy even the ai thinks google is a buy because
30:54
Josh: they just have um alphabet offers the best value in mega cap tech dominant ai
30:59
Josh: capabilities diversified growth and a cheap valuation if search mode holds and.
31:03
Ejaaz: Yeah give me the week give me the week
31:05
Josh: Let's see the weekly chart here do you want some moving average lines as well
31:08
Josh: because we could drop those in please let's.
31:10
Ejaaz: See let's see i'm actually super yeah look see it's had a slight dip Markets are so reactive. Crazy.
31:15
Josh: Yeah, and I think to the point of the CapEx, markets are viewing that as a scary, high-risk statement.
31:22
Josh: But while that's true, I also think it's a testament to the fact that scaling
31:27
Josh: laws are going to work, and the largest companies in the world are betting on
31:31
Josh: the continuation of them working.
31:33
Josh: And the shared consensus between all of these large-cap companies deciding to
31:37
Josh: spend record CapEx this year,
31:40
Josh: is a testament to the fact that things are only going to go faster.
31:44
Josh: And they believe that the more money they put in, the more outputs they will get.
31:47
Josh: And they're going to continue to put their foot on the gas. So I think any question
31:51
Josh: that anyone had, if these scaling laws could continue to hold up and we could
31:55
Josh: continue to be on the path to whatever AGI looks like and beyond,
31:58
Josh: I think that was answered this week through these earnings reports.
32:00
Josh: And the overwhelming answer is yes, it's true.
32:02
Josh: It is likely that this is going to happen and everyone is betting their entire company on it?
32:06
Ejaaz: I think we have done a great job, if I pat ourselves on the back virtually,
32:11
Ejaaz: Josh, of showing what these models are capable of.
32:14
Ejaaz: And remember, it's been less than 48 hours that these models have been alive.
32:18
Ejaaz: In fact, I think it's been like 36 hours. So if any of you are interested in
32:23
Ejaaz: trying these out, I cannot urge you enough to go out and try these things.
32:28
Ejaaz: Try to solve a problem that you're finding at work or try to solve a problem
32:32
Ejaaz: that you're finding just in your casual leisure time to code up a hobby or a
32:36
Ejaaz: project in a matter of seconds. It's so, so easy.
32:39
Ejaaz: And it'll put you at an advantage to understand how these tools work and why
32:43
Ejaaz: they're really changing the world as we see it around us, why stocks are dumping,
32:46
Ejaaz: why some stocks are pumping.
32:48
Ejaaz: But yes, go demo it. Let us know what you actually end up building.
32:52
Ejaaz: Josh and I are trying to give you more live demos in a lot of the episodes that we put out.
32:56
Ejaaz: And with every other model release and feature that drops, we are going to be
32:59
Ejaaz: trying and testing these things so we can bring to you exactly what these things
33:03
Ejaaz: can do and show you kind of like the benefits and disadvantages,
33:06
Ejaaz: what's real and what's really not.
33:08
Josh: Yeah. And I can't stress this enough. The best way to stay on top of things,
33:12
Josh: the best way to feel like you're not being left behind is just to use the tools
33:15
Josh: as they come out and to understand them and what makes them different.
33:18
Josh: And for a single subscription to ChatGPT or to Claude, you can access tools
33:24
Josh: just like this and build stuff just like this.
33:26
Josh: I'm not, this wasn't like an incredibly difficult technical challenge.
33:29
Josh: You just ask it what you want and you ask it to help you.
33:32
Josh: And it will actually walk through and help you through the process and build whatever you want.
33:35
Josh: So the most important thing for anyone listening is just to train that muscle and to get familiar with,
33:41
Josh: these tools and these skills that you're able to leverage them to your advantage,
33:45
Josh: however it may best fit in your life.
33:48
Josh: And that's what kind of we wanted to share with us.
33:50
Josh: Like, it's simple. You download the app, you log into your account,
33:52
Josh: and you're on your way. It's really
33:54
Josh: not as difficult as I think a lot of people make it seem like it is.
33:58
Josh: And I mean, this beautiful dashboard is a testament to that.
34:02
Josh: Okay, so Ejaz, it also looks like our codex output
34:05
Josh: has finished itself so we have here on the
34:08
Josh: screen we have opus which we saw which is
34:11
Josh: really a lovely dashboard but it seems like codex
34:14
Josh: now has its own version that we could quickly compare so maybe we'll try we'll
34:17
Josh: go to our favorite google we'll type google in and we'll click analyze and kind
34:20
Josh: of see how this compares i find it funny how they've they've merged on the same
34:24
Josh: type of design style but yeah oh okay this whoa this is interesting this is
34:29
Josh: different so it has the moving averages select oh is that,
34:34
Josh: Okay, yeah, so it has the charts.
34:35
Ejaaz: Is that accurate?
34:36
Josh: It has the PE ratio. Yeah, that's what I was looking at. Let's go to that one-week chart and see.
34:40
Josh: I have some questions about these. It looks pretty right.
34:46
Ejaaz: Okay. That looks very wrong.
34:48
Josh: Yeah, the one you're a little confused about. Let's compare it to Claude here.
34:53
Josh: Let's go to Google and we'll analyze that. Well, it thinks we can look at the
34:56
Josh: rest. So it looks like it emulated pretty well.
34:59
Josh: It has the verdict. It has the same stats.
35:02
Josh: The risk assessment matrix is... good but you could see like some of the text
35:06
Josh: you can't really read because it's black on black um but nonetheless pretty
35:11
Josh: interesting they both succeeded.
35:12
Ejaaz: Yeah i mean as we said before like these models are very equally capable and
35:18
Ejaaz: you know maybe it's just the way that you prompt something or uh the way that
35:22
Ejaaz: some of these things work but largely they kind of achieve the same goal and
35:25
Ejaaz: same quality um and like listen like we're talking about like minor discrepancies here
35:31
Ejaaz: I can't wait to see what we will build with this. Like, this is insane.
35:34
Josh: It's amazing. Both of these one-shot prompts didn't touch anything.
35:38
Josh: And here we are. I do think that Google, when your chart is wrong,
35:40
Josh: I think Claude got that one right.
35:41
Josh: But we overall both succeeded in the mission. Both look great.
35:44
Josh: And both are just excellent models.
35:46
Ejaaz: Amazing. Okay, well, that's it. Wherever you're listening to this,
35:49
Ejaaz: if it is on YouTube and you're watching our lovely faces, or if you're listening
35:53
Ejaaz: to us on Spotify, Apple Music, or wherever you listen to us,
35:56
Ejaaz: please subscribe, give us a rating, leave us some comments.
35:59
Ejaaz: We love your feedback and we respond to pretty much every single comment because
36:04
Ejaaz: we're trying to figure out how to make this show better and bring you the content
36:07
Ejaaz: that you guys deserve and want.
36:09
Ejaaz: Turn on notifications because we are releasing more and more videos every week
36:13
Ejaaz: on the hottest topics as they come out.
36:15
Ejaaz: We also have the sickest newsletter ever where one of us will either write a
36:20
Ejaaz: essay or give you the five top highlights of the week.
36:22
Ejaaz: So if you don't want to watch any of these videos, you can just read and digest
36:24
Ejaaz: that and you'll know everything that you need to know in AI and frontier tech.
36:29
Ejaaz: Thank you for listening, and we will see you on the next one.
36:31
Josh: See you in the next one. Peace.