Latest / In the Loop AI / Everyone Needs to Use OpenAI Codex... Until Claude Mythos Comes Out

Transcript

Everyone Needs to Use OpenAI Codex... Until Claude Mythos Comes Out · In the Loop AI

0:00
Josh: A few months ago, we told you to use Claude. Now we're telling you to switch
0:03
Josh: back because for those of you who aren't familiar, well, over Christmas break,
0:06
Josh: there was a major vibe shift where AI coding went from this like fun tool to
0:09
Josh: things that developers actually use when they're shipping code.
0:12
Josh: And even if you're not a developer, the amount of use cases and applications
0:15
Josh: that were created around that time were really strong.
0:18
Josh: And since then, Anthropic has gone on this generational run of shipping these
0:21
Josh: incredible products seemingly every single day that has turned Claude code into
0:24
Josh: this supercharged super app that is the place that EJS,
0:28
Josh: I know you've gone to, I've gone there too, in order to get all of our AI progress done.
0:33
Josh: Any work that we have, we've gone to Cloud Code.
0:35
Josh: Now, OpenAI has woken up. And over the last few weeks, Codex has shipped more
0:39
Josh: features than most companies ship in a year.
0:41
Josh: And I bet, I guarantee that you haven't heard of some of these features that
0:44
Josh: we're going to talk about in this episode.
0:46
Josh: The pendulum has fully swung back, or at least I believe, because I'm totally Codex-pilled.
0:49
Josh: And in this episode, we're going to kind of walk through the differences between
0:53
Josh: these two and why the model that you're using today probably won't be the model you're using tomorrow.
0:58
Josh: And I don't think we're going to convince you, but maybe we could show you why
1:00
Josh: you might want to consider using something else here.
1:03
Ejaaz: I just want to talk through some of the crazy stats here because the script
1:06
Ejaaz: has genuinely flipped. A few months ago, Claude Code was anything everyone could talk about.
1:11
Ejaaz: And every software engineer was using Claude Code. Every enterprise was installing it. It was crazy.
1:17
Ejaaz: But just over the last couple of weeks, specifically by the end of April,
1:21
Ejaaz: Chat GPT 5.5 was released, and that was plugged into the coding AI model.
1:26
Ejaaz: It's all one and the same.
1:27
Ejaaz: And OpenAI went on this code red run where they focus on nothing but building
1:31
Ejaaz: the best coding AI model and the best LLM.
1:34
Ejaaz: And the numbers show that it's worked.
1:36
Ejaaz: Over the last week, Codex has been downloaded over or installed over 46 million times.
1:42
Ejaaz: Cold code, under 500,000 times. Now, that is crazy to say, because if you look
1:46
Ejaaz: at the historical data, cold code downloads and installs has absolutely dwarfed
1:51
Ejaaz: Codex, but something changed over the last couple of weeks.
1:53
Ejaaz: That something was OpenAI putting out just a better model. You mentioned that
1:58
Ejaaz: you were Codex pill, Josh.
2:00
Ejaaz: I think so am I. I've spent the last couple of days playing around with Codex.
2:04
Ejaaz: This morning, we prepped a bunch of really cool demos, and it is just completely flipped script.
2:09
Ejaaz: But it's one thing saying it. It's another thing actually showing the direct
2:13
Ejaaz: comparison. So we created this visual artifact to kind of
2:16
Ejaaz: give you the scoreboard. And you can see it at the top here.
2:19
Ejaaz: It's OpenAI Codex at 11 and Anthropic Claude at two. But let me explain why.
2:23
Ejaaz: Okay. So number one, computer use.
2:26
Ejaaz: Codex and Claude code can use your computer.
2:30
Ejaaz: It can take over your desktop and it can like move your cursor around.
2:33
Ejaaz: Now, Claude pioneered it. There were the first ones there, but it was super slow.
2:38
Ejaaz: It kind of runs into a bunch of obstacles and you have to kind of like handhold
2:41
Ejaaz: it and improve it to do a bunch of different things.
2:43
Ejaaz: Codex is not only quicker than me, it's quicker than the average person.
2:47
Ejaaz: In fact, I can actually see the cursor move around so quickly.
2:50
Ejaaz: And it's like using a computer, but it's a superhuman and it can run pretty much 24-7 at this point.
2:56
Ejaaz: Long horizon autonomy. Codex can work for longer in a much more intelligent
3:01
Ejaaz: manner versus Claude code, which is, again, crazy to say because literally a
3:05
Ejaaz: month ago, it was the inverse of this.
3:07
Ejaaz: Claude right now can run for a decent number of times or amount of time,
3:11
Ejaaz: but not as long as Codex can.
3:14
Ejaaz: And then the last two that I want to talk about here is browser use.
3:17
Ejaaz: So Codex can take over your browser. It can do a lot more intentional things.
3:22
Ejaaz: It understands what it's looking at, very importantly. Previously, it could not do that.
3:25
Ejaaz: Claude can do the same, but not as intelligently. And then finally,
3:29
Ejaaz: ChatGPT Images 2.0 got released, what was it, like two weeks ago now? Oh, it's so good.
3:35
Ejaaz: Yeah, it's the image generation model from OpenAI, and it is absolutely astounding.
3:40
Ejaaz: In fact, it beat all the other predecessors, including Google's,
3:44
Ejaaz: what is it, Nano Banana 2.0 Pro, which previously held the lead.
3:50
Ejaaz: It beat it across every single benchmark.
3:52
Ejaaz: Anthropic, on the other hand, doesn't even have an image gen model.
3:55
Ejaaz: So, so far, it's crushing. Yeah.
3:57
Josh: Yeah, I think a lot of the best is now bundled into codex. The image gen for
4:01
Josh: anyone who uses any sort of visual work is unbelievable. And being able to use
4:04
Josh: that directly in your software is awesome.
4:07
Josh: One thing that you mentioned is the long horizon autonomy. I think that needs
4:10
Josh: a double clicking on because it's really impressive how well it works.
4:14
Josh: Traditionally, there's been this thing called a Ralph loop that we use.
4:17
Josh: It's actually named after the character from The Simpsons who is very persistent.
4:20
Josh: And it's basically a planning mode where you give the AI a goal and it will
4:25
Josh: continue to iterate towards that goal until it accomplishes it.
4:27
Josh: So like, let's say you want to build a Lego car or something and you give it the exact parameters.
4:32
Josh: It will go and go and go until it solves that problem and gives you exactly
4:35
Josh: what you want in a way that other AI models haven't.
4:37
Josh: Codex did that. And this is the only native implementation that you can get
4:41
Josh: of this long horizon thinking where it actually will go for days on end.
4:45
Josh: I've seen screenshots of some thinking for as long as 36 hours to accomplish the goal.
4:48
Josh: So if you have really difficult tasks, Codex is going to be really good at solving those.
4:52
Josh: Now, continuing to scroll down, there was another feature that was just released
4:55
Josh: this week called auto review.
4:57
Josh: And a huge pain in the ass for people who are creating code for working on complex
5:01
Josh: projects, whatever it may be, is you're constantly having to sit there and approve
5:05
Josh: things because the permission system is a little finicky, right?
5:07
Josh: You don't want to give it full access to your computer.
5:10
Josh: You also don't want to sit there and approving every time it wants to use Chrome
5:13
Josh: or every time I want to access your file.
5:15
Josh: So Codex created auto review and they rolled it out last week where the agent is kind of smart.
5:20
Josh: It knows which things are going to possibly be systemic existential threats
5:24
Josh: and which approvals aren't.
5:26
Josh: And it will just automatically approve all the things that aren't going to get you in a lot of trouble.
5:30
Josh: It creates a much easier user interface where you can just kind of walk away
5:33
Josh: from the computer for a little while and come back and things get done.
5:35
Josh: Memory and context is pretty strong i'd say
5:38
Josh: the one thing and we haven't mentioned many claude winners the
5:41
Josh: place where claude wins currently is on their open claw capability funny
5:44
Josh: enough because open ai bought open claw but dispatch is the mobile app feature
5:49
Josh: for claude in which you can actually engage with claude code remotely that doesn't
5:53
Josh: currently exist on codex and while the team has promised to ship that you don't
5:57
Josh: actually have that currently today claude has that also in terms of the personality
6:01
Josh: and ui claude is just so much better i think we're going to get into our.
6:04
Josh: Personal takes but whenever you're using an llm versus an actual
6:08
Josh: tool set or a harness claude is pretty great and the
6:10
Josh: ui is very warm so there's there's some kind of
6:13
Josh: instances in which claude is better but for the most part codex is
6:17
Josh: really just kind of crushing it and i've really enjoyed using it one of the
6:20
Josh: fun things is pets i mean just recently they released pets and claude also released
6:25
Josh: pets but these pets are a little bit different this is an example of angry dario
6:28
Josh: we're seeing on the screen and it's fun because you have this persistent character
6:32
Josh: that exists throughout your computer use.
6:34
Josh: And as you're engaging with Codex, it'll just kind of chat with you in the background
6:38
Josh: so you can see your progress, see where you're at.
6:40
Josh: It's fun, it's playful, and it just shows that they kind of care about the user experience.
6:43
Josh: Now, one feature I would guarantee most people don't know is Chronicle, EJS.
6:47
Josh: And you were just telling me about Chronicle and how cool it is,
6:49
Josh: how it kind of monitors your screen as you go. This seems like novel technology
6:52
Josh: that we haven't seen yet.
6:53
Ejaaz: Yeah, so one of the earliest episodes that we did here on Limitless was an interview
6:58
Ejaaz: with the folks at OpenAI that created...
7:01
Ejaaz: Something called, what was it called, Josh? Do you remember?
7:04
Ejaaz: It was like agent mode or personal mode, something like that.
7:08
Josh: Yes. It thought overnight for you, right? Yes.
7:10
Ejaaz: It basically took all the conversations that you'd had with ChatGPT the night
7:15
Ejaaz: before or the day before or the week before, and it created important context
7:20
Ejaaz: around you in the form of something called memories.
7:22
Ejaaz: This is where AI memory was birthed from OpenAI themselves, from the OpenAI
7:27
Ejaaz: team. And what it would do is it would feed you a report in the morning that
7:31
Ejaaz: would update you on information that it thought you would be interested to read about.
7:35
Ejaaz: So say, for example, you were interested in the stock market,
7:37
Ejaaz: it'll give you an update on a bunch of advancements that had happened overnight
7:40
Ejaaz: or over the last week or whatever it might be.
7:43
Ejaaz: Right now, fast forward today, memory is embedded across every single AI model
7:47
Ejaaz: and tool. The reason why is context is so important.
7:50
Ejaaz: It's one thing a user asking for something explicitly and directly.
7:54
Ejaaz: It's a complete other thing for an AI to actually understand what you mean,
7:58
Ejaaz: the nuance in the sentence that you've created, and even better,
8:01
Ejaaz: to predict what you want.
8:03
Ejaaz: But there was still an obstacle, which was you needed to feed it the context
8:07
Ejaaz: and say, hey, Claude, hey, ChatGPT, can you remember this?
8:11
Ejaaz: OpenAI recently released a feature called Chronicle, where it observes what you scroll through,
8:18
Ejaaz: What you click on, what you type, and it builds its own context and memories
8:23
Ejaaz: around you without you needing to feed it, which actually led to a really cool
8:28
Ejaaz: prompt that you pointed out, Josh, or that you found, which was,
8:31
Ejaaz: what have I been doing very inefficiently on my computer, according to Chronicle,
8:35
Ejaaz: which is this new memory feature, make some recommendations,
8:38
Ejaaz: be direct, tell me what I need to hear. That's, that's pretty awesome.
8:42
Josh: Yeah. So this is alpha because I don't think a lot of people recognize that
8:45
Josh: this is a possibility because Codex and OpenAI didn't do a good job of explaining this.
8:49
Josh: When they released Chronicle, they said it's a way of the system to review your
8:52
Josh: code as you've gone because it's been taking sequential screenshots.
8:55
Josh: But it's the reality is, is that it's much bigger than this.
8:58
Josh: And I suspect they didn't market it this way because it could be a bit of a privacy issue,
9:01
Josh: but it's essentially constantly monitoring your screen and taking screenshots
9:05
Josh: of what's happening on your screen and interpreting it so it understands your
9:09
Josh: habits, the way that you work, the thing that you do.
9:11
Josh: And then you can ask it, what have I been doing very inefficiently on my computer?
9:14
Josh: According to Chronicle, make some recommendations, be direct,
9:17
Josh: tell me what I need to hear.
9:18
Josh: And it'll actually evaluate how you've been using your computer,
9:21
Josh: how long you've been scrolling on Twitter, perhaps, how long you haven't been
9:24
Josh: doing the things you're supposed to be working on, or just generally how to
9:27
Josh: improve your workflow and give you real feedback based on your actual actions that it's seen.
9:31
Josh: And I think this is a super powerful thing currently only available to pro members.
9:35
Josh: So if you pay for the $100, $200 a month subscription, you get access to this.
9:39
Josh: But I suspect this is the early signs of a very important feature they're going
9:42
Josh: to roll out, which is that entire computer monitoring system to improve your
9:48
Josh: system and also probably train the models to get better at engaging with your system.
9:51
Josh: But I found Chronicle to be one of those kind of secret features that not a
9:54
Josh: lot of people know about, but has a lot of upside if you use it to your advantage
9:57
Josh: and let it monitor what you're doing and improve your workflow on a day-to-day basis.
10:01
Ejaaz: So the point is, from both of these companies, Anthropic and OpenAI,
10:05
Ejaaz: we are getting feature releases every single week. In fact, every single day.
10:09
Ejaaz: And it's becoming, I'm being bombarded by this.
10:13
Ejaaz: And it's hard to keep track with all of this. So what is the number one litmus
10:16
Ejaaz: test for both of these models and products and companies?
10:19
Ejaaz: It's to actually use the thing. It's to build the thing.
10:23
Ejaaz: And we have two special demos that we have prepared for you that we're about
10:27
Ejaaz: to jump into. Now, Josh, can you guess what my first demo is about?
10:30
Josh: The theme. First one's a game. We're gamers, man. I want to play a game.
10:33
Josh: I want to see how well it does on a game.
10:34
Josh: I know we did this demo in the past months ago. It left a lot to be desired.
10:38
Josh: So I'm curious to see the current up-to-date status as it relates to Cloud Code
10:42
Josh: versus Codex. Who's winning on the one-shot game prompt?
10:45
Ejaaz: Indeed. Okay. So I am a nostalgic kind of guy. And so I was like,
10:49
Ejaaz: oh, back in the day, I loved Mario.
10:52
Ejaaz: So I want you, both of these models, to create the best Mario type or inspired
10:58
Ejaaz: game, a side scroller, but make it futuristic.
11:00
Ejaaz: Maybe add a little bit of neon, sprinkle a bit of neon in there,
11:04
Ejaaz: create levels. I want game design. I want there to be enemies.
11:07
Ejaaz: I want there to be pitfalls.
11:08
Ejaaz: And I also want there to be a scoreboard and also tell me how to do this thing.
11:11
Ejaaz: I want, give me the whole package.
11:13
Ejaaz: Basically, I fed this prompt or idea into ChatGPT and Claude.
11:18
Ejaaz: And I said, can you create a detailed prompt that I can then feed into your coding models?
11:21
Ejaaz: I then set each of the coding models to their highest settings.
11:25
Ejaaz: So what you're about to see is the best of the best for the most
11:28
Ejaaz: detailed prompt that they came up with and let's see what they
11:30
Ejaaz: did so step number one or example number
11:33
Ejaaz: one is called opus 4.7 so
11:37
Ejaaz: this is called code at the highest setting with their latest model
11:40
Ejaaz: um okay it took the prompt pretty literally it's titled this neon plumber moon
11:46
Ejaaz: base run which is obviously mario inspired and it said hey this is a demo edition
11:49
Ejaaz: by the way this is not production ready what i like about this is it's giving
11:53
Ejaaz: me the instructions but how does the game actually play out let's see it looks
11:57
Ejaaz: good can you see me here josh i
11:59
Josh: Can yes i can and it looks like.
12:01
Ejaaz: The animations are pretty good i'm jumping around i think
12:04
Ejaaz: i'm like a little robot i can see my feet pitter-pattering now
12:07
Ejaaz: i'm guessing this thing is about to kill me so let's see if i can jump oh i
12:10
Ejaaz: can jump there we go that's awesome um one bit can i kill this guy oh yes i
12:16
Ejaaz: can now one bit of feedback i've noticed is uh i can't double jump and it told
12:22
Ejaaz: me in the menu that i could double
12:23
Ejaaz: jump so that's weird so the physics hasn't really paid off can i die
12:28
Josh: Oh, it certainly looks like you could die.
12:30
Ejaaz: I can die. Great. Okay. So that is Claude's attempt at it. What's your feedback
12:35
Ejaaz: on this, Josh? I think the graphics are pretty good.
12:37
Josh: The graphics are great. For one shot, I mean, granted, this is only one single
12:41
Josh: prompt. So for one prompt, it created great graphics.
12:43
Josh: It had sound design that actually sounds pretty accurate to what you would expect in the game.
12:46
Josh: It has similar principles. It's following gaming principles.
12:50
Josh: You kind of understand what looks dangerous, what doesn't.
12:52
Josh: You knew that those spikes were going to hurt you and they hurt you.
12:55
Josh: The logic seems to be a little bit flawed i think it's having problems with gravity or at least that
13:00
Josh: double jump functionality because it looks like those coins that you probably
13:03
Josh: want to collect you can't actually reach because you can't do the double jump
13:06
Josh: so in terms of logic not so hot in terms of visuals aesthetics in terms of i
13:11
Josh: mean how good this game is from one shot very impressive yeah.
13:14
Ejaaz: I think it's important to understand that i started from zero it literally asked
13:20
Ejaaz: me to give it a folder to build in and the folder was completely empty.
13:23
Ejaaz: So all the visual renderings, all the graphics, the animation style,
13:27
Ejaaz: the scoring system, the way that the avatar moves and looks was created from
13:32
Ejaaz: scratch from a bunch of characters from this AI model.
13:35
Ejaaz: So this is Claude Code's current best attempt and it is way better than what
13:39
Ejaaz: we tested out and honestly demoed on this show about a month ago.
13:42
Ejaaz: But now let's see what OpenAI's ChatGPT 5.5 codex at the highest possible setting cooked up
13:49
Josh: Okay and this is using the same prompt correct so you just fed the model the
13:52
Josh: same prompt identical but identical right oh god i'm excited i hope codex did
13:56
Josh: well because now now that i'm a fan i'm gassing it up it better perform here okay.
13:59
Ejaaz: So this is gpt 5.5's attempt now you
14:02
Ejaaz: might notice that this isn't the entire browser that's because
14:05
Ejaaz: codex has a very unique feature which is not only
14:08
Ejaaz: can it do all the coding in a single app for you but it
14:11
Ejaaz: has an in-app browser so it can
14:14
Ejaaz: live test the thing in the app without you needing to go to google chrome or
14:17
Ejaaz: whatever but anyway we have the starting screen here it has also called it neo
14:22
Ejaaz: neon plumber moon base run it looks a little more rudimentary from the start
14:27
Ejaaz: but i do like the background animation josh we didn't get this in the previous
14:31
Ejaaz: one or at least not this side scrolling thing well let's
14:33
Ejaaz: Oh.
14:34
Josh: Oh, this is nice.
14:35
Ejaaz: This is nice. I think this has good logic.
14:39
Ejaaz: Wait, but this is no music. There's no music. I can't double jump.
14:44
Ejaaz: Might be a skill issue. Might be a prompt issue.
14:46
Josh: Let's have a look. Did it say you can double jump?
14:49
Ejaaz: That's a good question, actually.
14:50
Josh: This is a fully playable game.
14:53
Ejaaz: Yes. And I like that it's like zoomed in. There's like... Oh,
14:58
Ejaaz: we got the boost. I can jump on the platforms. Let's see if I can kill this guy.
15:02
Josh: Yes. nice okay.
15:03
Ejaaz: And can i jump the gap there's a scoring system
15:05
Josh: You could see your hearts oh dude this is way better power.
15:09
Ejaaz: Up wait oh my god i want the power up i'm still gonna go back double jump
15:13
Josh: You can you could go back go back to the last platform oh god.
15:16
Ejaaz: I died i'm going i'm going to the last platform here
15:19
Josh: We go it looks like they're sequentially gaining height which is interesting,
15:23
Josh: oh but okay so if i'm comparing these two i'm actually i'm not feeling very
15:25
Josh: let down this is good aside from the music not existing which we may not have
15:29
Josh: explicitly asked um this it looks like the logic plays better the actual gameplay
15:33
Josh: is usable this is a full i don't know if it's glitching or if this is you glitching no no.
15:38
Ejaaz: That is it's glitching it's glitching a bit
15:40
Josh: Okay so it's still there are some edge case errors yeah but
15:43
Josh: this is different in the sense that you have your hearts clearly projected you
15:46
Josh: have a score system that's clearly in place you're able to get these power-ups
15:49
Josh: they work they function i mean this is a very clean and functional game so i
15:53
Josh: would give this to codex i think the experience perhaps the design of claude
15:58
Josh: was better and And perhaps the music,
16:01
Josh: I mean, music was definitely better versus none, but Claude,
16:04
Josh: in terms of just, or Codex, in terms of just coding logic and making a better
16:07
Josh: game, I give, I give this Codex. Do you have a take?
16:10
Ejaaz: Yeah. So on the build side of things, I had a much more pleasant experience.
16:16
Ejaaz: Using codex as well so i think codex wins
16:19
Ejaaz: on this um i one-shotted it in the true sense
16:22
Ejaaz: where i just gave it a single prompt and codex didn't ask for
16:24
Ejaaz: any permissions it just kind of went on and did the thing i saw
16:28
Ejaaz: it it's thinking and at points where it was unsure it thought amongst itself
16:34
Ejaaz: and then made the decision to progress forwards whereas with cloud code it would
16:39
Ejaaz: come to me now that might just be a developer engineer's preference right like
16:43
Ejaaz: if you're building a production ready app for like, I don't know,
16:46
Ejaaz: a big company that you work for, you probably want to have more hands-on involvement.
16:50
Ejaaz: Whereas if you're just building a game like we did today, where I don't really
16:52
Ejaaz: care what it ends up looking like or what it does, then the hands-off preference
16:56
Ejaaz: is probably something that you would use Codex for. But I think Codex wins this.
17:01
Josh: So for our second demo, we have this handwritten piece of paper that I actually
17:04
Josh: wrote and took a picture of.
17:06
Josh: I didn't. It's GPT Image Gen 2.0, but it looks like it's handwritten.
17:10
Josh: The handwriting was too nice.
17:11
Ejaaz: Josh. That was the giveaway.
17:12
Josh: Yeah, my handwriting is far sloppier than this. But the idea is that you can
17:15
Josh: even write things on the back of a napkin and you could turn that into an application.
17:19
Josh: So what we did here is we just asked for it to create a generic limitless dashboard
17:23
Josh: application on the back of a piece of paper, fed it into the model, and this is what we got.
17:27
Josh: So it looks like it did a pretty good job.
17:30
Josh: I could tell this is Claude before you even tell me which model it is because
17:33
Josh: it has the standard design principles.
17:36
Josh: Claude design is so basic and
17:39
Josh: it's so predictable where like okay i've seen this
17:42
Josh: dashboard before it looks like it was a mission success there's a
17:45
Josh: lot of text on this page a lot of stuff going on a lot
17:48
Josh: of graphics i give a lot of credit for kind of inferring what
17:51
Josh: we would want to be seeing from something like this where we have a
17:54
Josh: proper trip budget i don't think we asked for a trip budget um but okay i think
17:59
Josh: it looks like it made it did a lot of inferring right like it kind of made a
18:02
Josh: lot of assumptions but in the end of the day it did take what we had on the
18:05
Josh: napkin and it turned it into a pretty generic dashboard of sorts based on very
18:10
Josh: limited information that we gave it.
18:12
Ejaaz: I think the issue with this is we asked for something
18:15
Ejaaz: completely different it created a dashboard um but
18:19
Ejaaz: we asked it for it to be based around the limitless podcast and
18:22
Ejaaz: it created a travel planning board so i don't know
18:24
Ejaaz: whether that was a a prompt issue or whether we just fed
18:27
Ejaaz: it the wrong image but but here we go here is where we're
18:30
Ejaaz: at um now let's take a look at what openai did okay so here we have the same
18:37
Ejaaz: prompt fed into gpt 5.5 and it's funny i can instantly tell this is GPT-515
18:43
Ejaaz: because it's cleaner and it's not neon and it's not trying to go for some futuristic spin.
18:48
Ejaaz: It looks very simplistic. This is actually a website or app that I would probably
18:53
Ejaaz: be more inclined to engage with.
18:55
Ejaaz: It's also more visually perceptive to me, right?
18:58
Ejaaz: Like, what do I have at the front here? It's this five-day trip that I want to go on.
19:03
Ejaaz: It's giving me the basic information that I need to know at the start.
19:06
Ejaaz: It has a bunch of different tabs as well.
19:08
Ejaaz: But again, it isn't what I specified on the napkin. So I think this might be
19:12
Ejaaz: a skill to show on our side, Josh. But otherwise, like, look at these graphics.
19:15
Ejaaz: They're like really good. One thing I've noticed is stylistically,
19:19
Ejaaz: although both models create very different looking things, the animation style looks the same.
19:26
Ejaaz: Have you noticed that even with the game previously that we just demoed,
19:29
Ejaaz: the avatar looked the same.
19:31
Ejaaz: It was given the same sort of title and the objects interacted in the same way
19:35
Ejaaz: we're seeing this here so maybe it's just a change in quality i actually prefer gpt 5.5 on this one
19:41
Josh: Yeah this is crazy i'm just going to suspect
19:44
Josh: there was a prompt issue there where yes like we clearly we asked for something
19:48
Josh: that we didn't actually want but here it is i think if you're just comparing
19:51
Josh: them apples to apples uh chat gpt and codex is like no-brainer 10 times better
19:56
Josh: i far prefer this if you look at the original napkin photo this is much more
20:01
Josh: accurate to what the design looked like on that original piece of paper.
20:04
Josh: And then if you also just compare the general design, this is far easier to understand.
20:07
Josh: It's just a lot less dense. It's designed better. I wouldn't even say this is really...
20:12
Josh: A fair comparison it seems like codex just like completely crushed this and
20:15
Josh: it has all the functionality built in it looks good i am giving another win
20:19
Josh: to codex here that's two for two.
20:20
Ejaaz: Wow look i've got like a re-optimization uh toggle at the top and it actually
20:25
Ejaaz: updated i wonder where it's pulling that data from it's already hooked
20:29
Josh: Into data look at that yeah impressive stuff.
20:31
Ejaaz: Very very cool now one major reason why both of these models have advanced so
20:36
Ejaaz: rapidly over the last couple of months is something known as the ai model harness
20:41
Ejaaz: now you have the AI model, which is something that you and I have interacted with quite a lot.
20:45
Ejaaz: It's via ChatGPT or Claude itself.
20:48
Ejaaz: But there's an added layer that you can put on top of this model,
20:51
Ejaaz: which comes in the form of prescripted prompts that are engineered to make the
20:56
Ejaaz: model act in a particular way.
20:58
Ejaaz: But it's also the environment that the model works in.
21:01
Ejaaz: It's also the policies that you set to make sure that the model acts and behaves
21:05
Ejaaz: and sounds in a particular way.
21:08
Ejaaz: That's why we talked about Claude's
21:09
Ejaaz: personality earlier being better than ChatGPT. It all plays into the
21:15
Ejaaz: We figured out was it's an entirely new product category on its own.
21:19
Ejaaz: In fact, Cursor had some news over the last couple of days where they made their
21:25
Ejaaz: harness, Cursor SDK, available via API.
21:28
Ejaaz: And the reason why this is such a big deal is critics criticized Cursor for
21:33
Ejaaz: being an AI wrapper, which meant that Cursor doesn't have a model of its own.
21:37
Ejaaz: It would just create this harness, a set of prompts environments around, say, Claude or ChatGPT.
21:45
Ejaaz: And so people would say, cursor isn't actually special. Turns out the wrapper
21:49
Ejaaz: or the harness actually made these models way more intelligent.
21:52
Ejaaz: In fact, if you added cursor's harness on top of GPT 5.5 and Claude Opus 4.7
21:58
Ejaaz: right now, you end up with a smarter, more intelligent, more efficient model
22:02
Ejaaz: than the actual base models themselves.
22:04
Ejaaz: Now, remember, AI Labs spent hundreds of millions of dollars to train these
22:07
Ejaaz: models and to create the best thing and put their best foot forward.
22:10
Ejaaz: And still you have a startup which is worth, what is it now,
22:13
Ejaaz: $10 billion right now, potentially being acquired by XAI for $60 billion,
22:17
Ejaaz: creating a better model on top.
22:19
Ejaaz: So the harness and the AI model are arguably one and the same at this point.
22:23
Ejaaz: And it's just a valuable moat to point out that these models aren't just better
22:26
Ejaaz: at coding because of the base model itself. It's because of this thing known as a harness.
22:31
Josh: Yeah. And the harness is the difference maker when it comes to building this super app.
22:35
Josh: It's like every single company is trying to build the super
22:38
Josh: app the all-in-one application that kind of serves
22:41
Josh: as your operating system anytime you need to engage with ai
22:43
Josh: this is the place that you could do it and it's all encompassing it's
22:47
Josh: all in one now one of the best applications we've seen for this in the early
22:50
Josh: days has been something like open claw where it's this extension of what an
22:53
Josh: operating system could look like starting with ai at the foundation and open
22:58
Josh: claw did a really amazing job of that now in some news this week you can now
23:02
Josh: use your chat gpt account to generate tokens with OpenClaw.
23:06
Josh: So previously you had to use the API, whether you were using Anthropic or OpenAI
23:09
Josh: or any of the other models, and it was pretty expensive. It costs a lot of money.
23:13
Josh: Now, thanks to Sam Altman this week announcing, you can actually use your account
23:16
Josh: connected with it. And I think this is the beginning of a multi-step plan.
23:20
Josh: To really integrate OpenClaw directly into Codex in a way that Anthropik can't.
23:25
Josh: Because if you'll remember, OpenAI owns OpenClaw.
23:28
Josh: They bought Peter and granted OpenClaw will stay open source forever,
23:31
Josh: but they have the ability to actually integrate directly into their products.
23:34
Josh: And I suspect that's what we're going to see.
23:36
Josh: In fact, we even got some confirmation from another post from one of the Codex
23:39
Josh: developers who replied to a post that was saying, Codex only needs a native
23:44
Josh: editor, an iOS app, a full browser, and OpenClaw.
23:47
Josh: And the developer, Tebow, said all of this and more
23:50
Josh: is coming to which sam altman retweeted it so we are
23:52
Josh: indeed getting open claw inside of codex we're getting a
23:55
Josh: mobile ios apps that you can access it remotely and soon
23:58
Josh: there's going to be no reason to really use a different app
24:01
Josh: because it's going to be all-encompassing now are there still downfalls yes
24:04
Josh: computer use 20 faster on codex but yesterday i was playing around with it i
24:08
Josh: told it to increase the volume of my music and it took 10 minutes to do it because
24:13
Josh: it tried to increase the slider on spotify even though it was max without actually
24:17
Josh: increasing my system audio so it's still a little dumb, but it is getting better.
24:21
Josh: And I think this leads me to this post that I really love, the vanilla maxing
24:25
Josh: post we have to talk about.
24:27
Josh: Which starts by saying, you should 100% be vanilla maxing. Just use the tools
24:32
Josh: as they're handed to you. That's it.
24:34
Josh: Because a lot of people, and I've found this personally, and in fact,
24:37
Josh: I've been caught by this personally, is that you try to get caught up in using
24:40
Josh: all these different repos and these skills and these plugins,
24:42
Josh: when the reality is, is if you just wait, the AI labs are shipping fast enough,
24:46
Josh: they'll just integrate it into your own native application.
24:48
Josh: So I'm vanilla maxing you, Jess.
24:50
Ejaaz: I'm totally vanilla maxing as well, dude. Like, listen, OpenClaw,
24:55
Ejaaz: when it was hyped up, was incredibly impressive and still is incredibly impressive.
25:00
Ejaaz: It opened up an entirely new product market and segment. That's why OpenAI acquired them.
25:05
Ejaaz: But something's majorly changed over the last couple of months,
25:08
Ejaaz: which is OpenClaw has kind of fallen off. No one talks about it anymore.
25:12
Ejaaz: People who are complaining about the errors and bugs that we're facing have
25:16
Ejaaz: kind of gone silent because they've just grown bored and they don't want to
25:18
Ejaaz: put their energy and effort into it.
25:20
Ejaaz: And the reason why is because although these tools are very frontier level,
25:24
Ejaaz: they can't actually be scaled to a practical use.
25:27
Ejaaz: You don't feel safe integrating OpenClaw into your desktop where you have personal
25:31
Ejaaz: files. I've seen horror stories where they access credit card data and expose
25:35
Ejaaz: that or where they deleted old wedding photos and the wife was super angry,
25:39
Ejaaz: a bunch of the stuff like that.
25:40
Ejaaz: If you are able to get given or access to a tool that comes under a branded
25:45
Ejaaz: reputation, such as ChatGPT, Codex, or Claude Cowork, where it kind of like
25:52
Ejaaz: takes over your computer, but in a sandboxed environment.
25:54
Ejaaz: I know that NVIDIA also released NemoClaw, which is like the enterprise-grade
25:58
Ejaaz: secure version of OpenClaw.
26:00
Ejaaz: You're vanilla maxing. That is the way to do it. And there's no need to rush
26:03
Ejaaz: ahead and lose all your data as a consequence. So that's basically it for the episode.
26:09
Ejaaz: We wanted to give you a comprehensive guide and insight into Codex GPT 5.5 versus Claude Opus 4.7.
26:16
Ejaaz: There's a lot of numbers in there, but basically the best coding models from
26:18
Ejaaz: both sides to see which is better.
26:21
Ejaaz: And the truth is, there isn't a clear winner right now. I would say it's probably
26:25
Ejaaz: Codex GPT 5.5, but the narrative switched so recently that maybe,
26:31
Ejaaz: maybe Claude can still catch up. And the only reason why I say that,
26:34
Ejaaz: Josh, is that's the only reason
26:36
Ejaaz: there's a model that we haven't discussed or demonstrated yet because we can't.
26:39
Ejaaz: It's called Claude Mythos.
26:41
Ejaaz: It was kind of pseudo-released about a few weeks ago.
26:45
Ejaaz: And on all benchmarks, it is technically better than 5.5.
26:50
Ejaaz: But the reason why we can't demo it is we can't get access to it.
26:53
Ejaaz: And the reason cited by Anthropic was because it's too dangerous.
26:57
Ejaaz: It's a cybersecurity risk. In fact, it wasn't just Anthropic saying it.
27:00
Ejaaz: It was Peter Heskett of the US Department of War also saying this,
27:03
Ejaaz: right? So there's concerns around that.
27:05
Ejaaz: OpenAI has created a mythos level type model here, but has made it available
27:09
Ejaaz: to everyone. And so the argument could be made that it's just because Anthropic
27:12
Ejaaz: doesn't have enough compute.
27:13
Ejaaz: So there's a lot of rumors around this, but I'm excited to get my hands on the
27:16
Ejaaz: best models from each of these and compare them directly.
27:18
Josh: Yeah. And the compute's actually been degrading. So I think I want to wrap this
27:21
Josh: up on like, what do you actually currently use?
27:23
Josh: What is the limitless production stack? How are we using these AI models?
27:26
Josh: And for me, at least, it's not even close. I'm codex-pilled.
27:29
Josh: I'm fully switched over. I am codex superior domination. It's going to be the month of codex.
27:35
Josh: Maybe Anthropic will have a comeback, but that's not happening until at least
27:39
Josh: June, July, because this month is codex month.
27:42
Josh: So I've been using codex for basically everything, all of the difficult tasks that I need.
27:46
Josh: What I have found is that GPT 5.5 as an LLM, as a language model,
27:51
Josh: as a chatbot is a little bit inferior to Opus 4.7, which I believe to be the
27:55
Josh: better model if you're just chatting with an AI.
27:57
Josh: I like its personality it's warmer it's more precise it normally
28:00
Josh: gets the idea of what i want so if i am building a complex
28:03
Josh: project opus 4.7 is the orchestrator and
28:06
Josh: codex is the actual implementer the executor of this code of this plan i've
28:11
Josh: also noticed that opus 4.7 is a bit inferior to 4.6 at a few things and i think
28:16
Josh: this is another piece of alpha here i actually use opus 4.6 whenever i'm doing
28:20
Josh: anything relating to writing or word ingestion so one of the projects i've been doing recently.
28:25
Josh: Is andre carpathy he created this like wiki for
28:29
Josh: your own person where it ingests files and it kind of writes
28:32
Josh: these summaries for you and it creates a personal knowledge wiki i use opus
28:35
Josh: 4.6 exclusively for that because opus 4.7 i think is far inferior at summarizing
28:40
Josh: and kind of rewriting these topics that i use in my obsidian so that's kind
28:44
Josh: of my stack i use opus for llms codex for everything else it just what are you
28:48
Josh: currently optimizing for what.
28:49
Ejaaz: Are you planning so it's two things when i have
28:52
Ejaaz: a uh my stack is actually way more diverse when
28:56
Ejaaz: it comes to just like the research side of things only because i'm
28:59
Ejaaz: using the ai that's like available readily wherever i am
29:02
Ejaaz: right so if i'm on x a lot and i see breaking news i'm
29:05
Ejaaz: just tapping grok because honestly it's a recent model i think it's like what
29:09
Ejaaz: was it 4.3 at this point uh is actually pretty good and they have multiple agents
29:13
Ejaaz: that are kind of like running at this right but for the core bulk of the work
29:17
Ejaaz: i've started shifting towards gpt 5.5 for the research because 5.5 research
29:23
Ejaaz: things for so much longer. And it has a much more in-depth discussion.
29:26
Ejaaz: In fact, I tested it out today because I was curious about the AI power stack
29:32
Ejaaz: and what stocks I should be investing in to get exposure to the power grid lines
29:36
Ejaaz: that are currently constraining AI data centers, right?
29:38
Ejaaz: And I was like, all right, I gave a detailed prompt to both Claude Opus 4.7
29:41
Ejaaz: and 5.5 and 5.5 completely cooked 4.7.
29:46
Ejaaz: And it gave good reasoning why, whereas 4.7 did not.
29:51
Ejaaz: I had to ask it more question. So all in all, I think 5.5 is my preference right
29:56
Ejaaz: now. I still use 4.7 because of the personality.
29:59
Ejaaz: It's like less of an AI type of voice versus GPT 5.5.
30:02
Ejaaz: But again, I feel like OpenAI is on a generational run right now,
30:06
Ejaaz: and they might just kind of fix this in the next couple of hours at this point.
30:10
Josh: Yeah, it's coming. It's coming quick. And I think now is a good time to kind
30:13
Josh: of get familiar with Codex to understand the way it works.
30:15
Josh: And as they implement these features, you'll be able to adopt them within the hour, within the day.
30:19
Josh: It's pretty amazing. And it's been fun to just experiment. It's been fun to try something new.
30:23
Josh: And it's, again, competition is just better for everyone. So the end winner
30:27
Josh: of this is the user, because for as low as $20 a month, you get access to all
30:31
Josh: this frontier intelligence, all these capabilities.
30:33
Josh: And it's just, it's really been unbelievable to watch. So that is the comparison, Codex versus Opus.
30:38
Josh: If you have not tried both of them, I encourage you to give it a try.
30:42
Josh: Test the prompts against one another. If you have any type of work that you
30:45
Josh: need, if you're working on a computer at all, chances are you can use AI to
30:49
Josh: help you do your job even better.
30:50
Josh: Or you could just use it to help you do hobbies and side projects that you've
30:54
Josh: always wanted to do. So give it a try.
30:56
Josh: Let us know your preference, codex, cloud code. Which one is it going to be?
30:59
Josh: I think that's probably it for the episode. Thank you guys so much for watching.
31:02
Josh: If you enjoyed it, please don't forget to share with your friends.
31:05
Josh: Let them know which model they picked. And also don't forget to rate it five
31:08
Josh: stars on your favorite podcast listening platform. Any final thoughts, EJS, before we go?
31:13
Ejaaz: No, that's it. Thank you guys so much for listening and we'll see you on the next one.