WEBVTT
00:00:02.240 --> 00:00:12.880
Any hour or day or week now, we'll get an official announcement from OpenAI that signals the biggest AI pivot we've seen since ChatGPT.
00:00:12.960 --> 00:00:17.440
And that's the official rollout of their AI super app.
00:00:17.760 --> 00:00:33.280
So the ChatGPT Parent Company, now with 1 billion weekly active users, made headlines recently when a senior executive off the record at least said three words to a reporter that will probably echo for months.
00:00:33.600 --> 00:00:35.759
Chat is dead.
00:00:36.320 --> 00:00:45.920
So any day we'll get the official announcement of OpenAI's new super app, their push to bring an autonomous coworker to millions of companies.
00:00:46.240 --> 00:00:49.359
But it's not just OpenAI making this push.
00:00:49.600 --> 00:01:02.399
Actually, Anthropic started the wave many months ago with their rollout of Claude Code and Claude Cowork, and every other big tech challenger is racing to get their version of an AI super app some momentum.
00:01:02.560 --> 00:01:05.439
Google released theirs in Anti-Gravity 2.0.
00:01:05.680 --> 00:01:09.680
Cursor has quietly become a serious contender in the space.
00:01:09.920 --> 00:01:14.239
And Microsoft also announced their upcoming super app as well.
00:01:14.480 --> 00:01:18.319
But that begs the obvious question: what the heck is a super app?
00:01:18.480 --> 00:01:20.799
And why should my company take note?
00:01:21.040 --> 00:01:26.799
Well, we're gonna tackle the latest work layer shift on today's show.
00:01:26.959 --> 00:01:31.120
So welcome to everyday AI and our start here series.
00:01:31.280 --> 00:01:35.120
But let's go over the big picture of what's happening now.
00:01:35.280 --> 00:01:43.680
So Anthropic started this desktop agencera with Claude Code and Claude Cowork in late 2025 and early 2026.
00:01:43.840 --> 00:01:54.159
But Anthropic hasn't really been able to build off of that momentum, at least when it comes to being the big name in the room, when it comes to an AI super app.
00:01:54.319 --> 00:01:59.120
And I think OpenAI's super app, which is essentially codecs, right?
00:01:59.200 --> 00:02:05.760
And it will uh be a little more tightly integrated into Chat GPT, will be coming any day.
00:02:05.840 --> 00:02:11.919
And I think that's really going to set the tone for this big pivot to the super app era, right?
00:02:12.080 --> 00:02:22.879
And Microsoft, I think, especially when you have OpenAI and Microsoft in the coming weeks or months shifting most of their agentic focus to a super app.
00:02:23.039 --> 00:02:26.560
I think the rest of the enterprise really needs to pay attention.
00:02:27.120 --> 00:02:39.439
So on today's show, we're gonna be diving into and stick with me for the next 25 minutes, and here's what you're gonna learn why Anthropic started the super app race, but they have kind of fallen out of immediate contention.
00:02:39.599 --> 00:02:45.520
You're gonna know why the models are no longer the moat and the super app harness matters more than ever.
00:02:45.680 --> 00:02:54.719
And you're gonna kind of know the state of the super app race, at least as it stands right now, right before we officially get it all kicked off.
00:02:54.960 --> 00:02:56.400
All right, let's get into it.
00:02:56.560 --> 00:02:58.560
Welcome to the Start Here series.
00:02:58.719 --> 00:03:05.120
This is everyday AI's essential podcast series to both learn the AI basics and to double down on your knowledge.
00:03:05.280 --> 00:03:10.479
So if that's what you're trying to do, we're at the uh the tail end of the start here series, right?
00:03:10.560 --> 00:03:13.280
Now we're in uh episode what are we on now?
00:03:13.439 --> 00:03:16.159
We're on uh volume 28 of this.
00:03:16.240 --> 00:03:18.319
So we'll probably do two more and wrap it at 30.
00:03:18.400 --> 00:03:20.800
But if you don't know the start here series, it's great.
00:03:20.879 --> 00:03:25.840
I would say start at volume one if you can, uh, and work your way up to where we are now.
00:03:25.919 --> 00:03:28.800
They're usually quicker episodes, about 30 minutes, give or take.
00:03:28.879 --> 00:03:31.199
But make sure you go to starthirseries.com.
00:03:31.360 --> 00:03:38.639
There, we put together a playlist where you can listen to every single episode of the Start Here Series so you don't have to go fishing around our website or anything like that.
00:03:38.719 --> 00:03:43.599
And that's also going to give you exclusive access to our inner circle community.
00:03:43.680 --> 00:03:46.800
Yeah, right now that's the only way to get open access to it.
00:03:46.879 --> 00:04:02.639
Uh, so you can not only sign up for our free inner circle community, which gives you access as well to our prime prompt polish uh version three course to learn context engineering 101 at your own pace, but also all of the start here series info.
00:04:02.800 --> 00:04:09.759
All right, and make sure if you didn't listen to our last episode, I think it's an important one in the start here series, which was volume 27.
00:04:10.000 --> 00:04:18.160
We went over the token maxing era and how it's over and talking about the new era of token efficiency and how your company should adapt.
00:04:18.240 --> 00:04:19.519
So make sure you go check that out.
00:04:19.600 --> 00:04:22.399
That's volume 27 or episode 789.
00:04:22.480 --> 00:04:27.439
But let's get into today's topic, which is the AI super app.
00:04:27.920 --> 00:04:33.920
So, what the heck is an AI super app and why does it matter?
00:04:34.160 --> 00:04:35.199
Well, I'll say this.
00:04:35.439 --> 00:04:46.480
Even though I don't think we've quote unquote arrived in the super app era per se, I will say it's safe to assume this is the next version.
00:04:46.720 --> 00:04:54.000
All right, and it is funny because we started the start here series in January, and we've actually seen kind of two shifts.
00:04:54.079 --> 00:05:12.639
Uh, so I think the first shift we saw in January was models becoming more and more agentic by default, and saying, hey, how the future of work is not uh reactively talking to AI chatbots, it's knowing how these models work, scheduling things out, right?
00:05:12.720 --> 00:05:18.879
And then we also had another episode talking about something I called agentic context carry, right?
00:05:19.040 --> 00:05:28.639
And that is the crucial role that uh not just models, but the harness that control the models play in taking context from your certain apps, right?
00:05:28.800 --> 00:05:32.800
So uh to boil this down to the the basics here, right?
00:05:32.879 --> 00:05:39.120
One of the biggest advantages of not just the upcoming super app era, but where we're at now, right?
00:05:39.199 --> 00:05:44.560
Before Microsoft and OpenAI start really highlighting uh and promoting their super apps, right?
00:05:44.720 --> 00:05:50.480
Where we're at now is we do have the system that can carry our context autonomously, right?
00:05:50.560 --> 00:06:02.240
So a lot of what the human duct work, uh duct tape work has been over the last two years is us humans who do use large language models is giving a chat our context, right?
00:06:02.399 --> 00:06:09.600
Connecting whether it was static files, dynamically uh shifting things, but then it's a lot of copying and pasting from different AI systems.
00:06:09.759 --> 00:06:26.720
But now, as these systems are living on our desktop and becoming more autonomous, that does leave uh kind of lead into the next wave uh going beyond this agentic context carry, and that's where the agents ultimately live, which is the desktop OS, right?
00:06:26.879 --> 00:06:30.079
That is where I think the super apps are going to play.
00:06:30.240 --> 00:06:37.600
And I don't know if this is going to be the quote unquote final frontier of the agentic era as we have officially entered into it.
00:06:37.759 --> 00:06:53.680
It may look a little bit different, but I do think at least uh the companies that are going to be most successful in late 2026 and early 2027 are those that are going to be investing in their companies using these super app systems on the desktop.
00:06:54.079 --> 00:06:57.680
So the core idea of this is simple, right?
00:06:58.959 --> 00:07:05.600
A super app is bringing all of your different work surfaces together under one agentic roof.
00:07:05.759 --> 00:07:11.040
So the big AI tech companies are all building their own super apps, they know the same truth.
00:07:11.279 --> 00:07:14.319
An AI chatbot is not a moat, right?
00:07:14.480 --> 00:07:17.199
ChatGPT.com is not a moat.
00:07:17.360 --> 00:07:19.120
The model is not the moat.
00:07:19.279 --> 00:07:23.120
They all want to own the execution layer, uh, right.
00:07:23.279 --> 00:07:29.680
And that is why every single big AI tech company is trying to build an AI super app.
00:07:29.839 --> 00:07:31.759
You could call it the harness, right?
00:07:31.920 --> 00:07:36.240
Kind of the uh the place where all of the tools come together, but it is much more than that.
00:07:36.319 --> 00:07:42.639
And I think we'll see uh as we really uh roll out and dig into what a super app actually is.
00:07:42.800 --> 00:07:46.639
But at its basis, I think it combines a couple of different things.
00:07:46.959 --> 00:08:03.759
It combines your company's context, your files, browser access, a file viewer, uh, automations, unified memory, uh, and approvals and automations all in one unified platform.
00:08:04.160 --> 00:08:09.360
And you, you know, we're gonna get into some of the nuances and the pros and the cons and all of those things.
00:08:09.600 --> 00:08:16.639
But the way I like to look at it now, and this is after I've been using uh especially codex heavily, right?
00:08:16.720 --> 00:08:23.199
I I also use the claw desktop programs heavily, but I do think codex is in the lead in the super app, right?
00:08:23.279 --> 00:08:25.519
And open AI has essentially said FYI.
00:08:25.600 --> 00:08:36.480
They're saying, yes, Codex is the early version of the super app, which is why I've been encouraging people to use it so far because I do think codex is very ahead of all of the other quote unquote super apps.
00:08:36.639 --> 00:08:45.679
We've seen that by how they're all the other big tech companies are kind of just copying and pasting exactly what OpenAI has done with Codecs because it is that good and it is that far ahead.
00:08:45.759 --> 00:08:51.440
But this is essentially the way I look at a super app right now, it is a full like intern, right?
00:08:51.519 --> 00:08:58.320
It is a full almost human coworker because anything that I can do, right, a super app can do.
00:08:58.559 --> 00:09:02.799
And I don't think this was the case in late 2025 with Claude Code.
00:09:02.960 --> 00:09:08.240
I don't think it was the case in early 2026 with Claude Cowork either, because now where we're at, right?
00:09:08.320 --> 00:09:26.480
So aside from what I said, dynamic access to your company's files, um, automations built-in, approvals built in, uh, and then having these kind of three different panes, and we're gonna talk about that a little bit more, but then also the ability to control a browser and the ability to control your entire computer.
00:09:26.559 --> 00:09:40.320
Because at that point, that's when you can kind of right say, okay, this system at least has the capability to be a true researcher, to be a true uh, you know, in terms of an output.
00:09:40.559 --> 00:09:46.480
This can output something like a smart human actually could because it has access to every single tool.
00:09:46.639 --> 00:09:46.960
All right.
00:09:47.039 --> 00:09:51.600
And there's obviously great downsides to this because with great power comes great responsibility.
00:09:51.759 --> 00:10:00.240
But essentially the super app gives read-write access to your entire computer, or it can if you give it those permissions and kind of go in YOLO mode, right?
00:10:00.320 --> 00:10:04.480
But it can download, it can upload, it can change different settings on your computer.
00:10:04.559 --> 00:10:05.120
Uh, right.
00:10:05.200 --> 00:10:08.240
You can even uh with codecs at least, right?
00:10:08.399 --> 00:10:14.799
I can, if my MacBook Pro is closed, I can still control my MacBook, you know, from my phone.
00:10:14.960 --> 00:10:22.480
So it also brings in that um, you know, that remote aspect as well, which is why I do think uh OpenAI's codex is a little more behind, right?
00:10:22.559 --> 00:10:32.879
Because even if, you know, my my my laptop is closed or you know, if my coworker is not in front of the desk, I can still send a message and they can still get to work and report back to me.
00:10:33.440 --> 00:10:37.200
Let's clear up a little bit about the name super app first, right?
00:10:37.519 --> 00:10:52.639
Because I think that there's a little bit of a mix in terms of when you throw the word super app, because in the pre-generative AI era, or even in the early Chat GPT days, when people heard super app, they instantly thought of China's WeChat, right?
00:10:52.720 --> 00:10:57.600
So that is the super app, the uh original super app OG, right?
00:10:57.679 --> 00:11:01.039
So they have messaging, payment, shopping, banking, everything.
00:11:01.200 --> 00:11:08.240
And I think that that was kind of one of the big goals of X slash Twitter slash early groc, right?
00:11:08.399 --> 00:11:10.639
It was this everything app ambition, right?
00:11:10.720 --> 00:11:16.320
It was moving all of your day-to-day personal life into a single app like China's WeChat.
00:11:16.480 --> 00:11:26.000
But that's not what I'm talking about here when I talk about the new era of super apps, because I think for whatever reason, that dream has maybe been dashed.
00:11:26.159 --> 00:11:47.200
And maybe it's because most of the big companies have understood and realized wait, why would we go after this on the consumer side, all on a smartphone, when there's a much bigger opportunity to do this AI super app on the desktop side and to go after companies that are spending millions of dollars on this every single year.
00:11:47.279 --> 00:11:53.279
So a bigger way to capitalize in a faster way to roll this out across the enterprise as well.
00:11:54.240 --> 00:12:00.559
So now you know it's not the you know, quote unquote, traditional WeChat mobile super app.
00:12:00.639 --> 00:12:07.919
We are talking about an AI uh desktop companion, and it's better than a chatbot, right?
00:12:08.080 --> 00:12:16.559
I don't think I need to go into too much depth on why a desktop super app is better than a chatbot, but let me just go over the basics, right?
00:12:16.720 --> 00:12:18.799
A chatbot, I still have to go on.
00:12:18.960 --> 00:12:20.559
I have to make that decision.
00:12:20.720 --> 00:12:27.200
I have to go on to whether it's claw.ai or chat gpt.com or gemini.google.com or whatever.
00:12:27.440 --> 00:12:32.080
I have to proactively go in there and give information to a chatbot.
00:12:32.159 --> 00:12:38.000
I have to share the context, share my problem, share all these things with a super app as an example.
00:12:38.159 --> 00:12:39.919
You can have automations that run.
00:12:40.000 --> 00:12:41.600
And this is how I have it set up.
00:12:41.759 --> 00:12:53.120
So whether it's hourly, twice a day, twice a week, whatever you want it to do, you can have it go through with your custom instructions, every single important piece of your work life, right?
00:12:53.200 --> 00:12:57.759
And obviously, I need to put the giant, the giant grain of salt out there, right?
00:12:57.840 --> 00:13:08.000
Don't go and download uh, you know, codecs or you know, whatever the uh the Microsoft co-pilot version of their super app, I think it's gonna be powered by GitHub co-pilots.
00:13:08.159 --> 00:13:11.600
So they do have their version of that, but it's just not unified yet.
00:13:11.759 --> 00:13:12.399
Um, right.
00:13:12.559 --> 00:13:16.320
This isn't something where you're trying to use this without your company's permission.
00:13:16.399 --> 00:13:24.799
So if you are a business leader, if you're a decision maker at the enterprise level, you need to start saying, hey, is it time that we invest in a super app strategy?
00:13:24.960 --> 00:13:26.799
And the smart companies already are, right?
00:13:27.039 --> 00:13:34.080
Months ago, I was doing uh, you know, big trainings for companies that rolled this out to tens of thousands of employees.
00:13:34.159 --> 00:13:37.279
So if you think, oh, companies aren't doing it, yes, they are.
00:13:37.440 --> 00:13:38.639
Uh they definitely are.
00:13:38.799 --> 00:13:40.559
So if you are not, you are behind.
00:13:40.799 --> 00:13:44.000
But there's a lot of reasons why it beats your traditional chat bot.
00:13:44.080 --> 00:13:51.519
But I think the biggest couple are number one, um, it's always on and autonomous, right?
00:13:51.600 --> 00:13:54.720
So you can schedule uh these things to happen, right?
00:13:54.879 --> 00:14:07.120
I have ones that run uh about every four hours, I have runs that run every day, uh, I have different ones that run every single hour uh on codex, especially, because then it can share all of this context.
00:14:07.279 --> 00:14:19.120
So whether it's going through my email, uh you know, every hour, it's going through social media, every four hours, it's going through my my DMs, every 12 hours, and then it's just reporting back to me, and then I can make those decisions, right?
00:14:19.200 --> 00:14:26.159
So you can see right away how that's immensely more in uh helpful than a chat bot that is really just uh reactive.
00:14:26.240 --> 00:14:30.799
And I have to go in and the chat bot reacts to me, but I have to be the one being proactive.
00:14:30.960 --> 00:14:38.639
Where on the AI super app end, it is actually the AI super app that is the one being proactive in delivering me uh decisions, right?
00:14:38.720 --> 00:14:43.840
And saying, Hey, human, I followed my rules that you set out for me.
00:14:44.000 --> 00:14:46.799
Here's all the this the decisions that you need to make.
00:14:46.960 --> 00:14:53.360
I went ahead and built you this dashboard, I built you this deck, I built you uh, you know, all these things.
00:14:53.519 --> 00:14:56.639
You approve the one which is best, go through, give me your updates, etc.
00:14:56.879 --> 00:14:57.039
Right.
00:14:57.200 --> 00:15:05.519
So there's a lot of reasons why the AI uh app beats the uh you know the traditional chatbot.
00:15:05.679 --> 00:15:08.720
And the biggest thing, it's the interface, right?
00:15:08.879 --> 00:15:11.840
So let's talk about this kind of three-pane interface.
00:15:12.000 --> 00:15:16.879
So first, there's a lot of the you know, you have to tip your hats to the OGs, right?
00:15:17.039 --> 00:15:23.360
I think one of the companies that actually got the three-pane interface done really well early on was Google's notebook LM.
00:15:23.679 --> 00:15:40.000
Uh, obviously, Anthropic started the uh phase of the uh desktop agencorker uh with Claude Code, um, the desktop version in late 2025, and then with co-work in early 2026.
00:15:40.080 --> 00:15:44.559
But essentially, if you're not watching the video version of this, I want you to think of three panes, right?
00:15:44.799 --> 00:15:51.120
So on your left pane, that's your traditional chat and folder pane, but then you also have your automations.
00:15:51.360 --> 00:15:58.559
So that's where you have your different kinds of workspaces, that's where you organize, you know, all of your kind of agentic flows.
00:15:58.879 --> 00:16:02.000
In the middle, this is kind of your instruct panel.
00:16:02.159 --> 00:16:05.200
This is where you are actually chatting with the chat bot.
00:16:05.360 --> 00:16:10.320
So, right in traditional uh, you know, chat GPT web interfaces, this is the same.
00:16:10.399 --> 00:16:13.120
On the left side, you have all your chats, your folders, etc.
00:16:13.440 --> 00:16:18.399
And then on the panel to the right, that's your uh ability to chat, right?
00:16:18.559 --> 00:16:23.039
So the difference here in the big uh jump is the right panel.
00:16:23.200 --> 00:16:28.720
This new panel, uh, you know, you can call it the inspector panel, uh, whatever you want to say, the browser panel.
00:16:28.799 --> 00:16:37.360
But this is where you bring in that context um and you don't have to lose or degrade the agent's ability to bring it in.
00:16:37.440 --> 00:16:39.440
So, as an example, right?
00:16:39.840 --> 00:16:47.440
Um, the earlier days of Claude, um it had a very clunky browser use, uh, right.
00:16:47.600 --> 00:16:53.120
But now with today's version, with like Codex, it has a built-in browser there on the right side panel.
00:16:53.279 --> 00:17:06.799
So it can not only browse the web without having to, you know, connect a third-party tool or a second party tool technically by controlling a browser, but it has its own browser inside of the program.
00:17:06.960 --> 00:17:12.240
So it is much faster, it is much more accurate, it is much more token efficient, right?
00:17:12.319 --> 00:17:17.839
That's the big thing as we talk about token efficiency is going to be one of the biggest uh trends of the rest of 2026.
00:17:18.000 --> 00:17:25.359
To be able to bring in that inspector or browser is a huge piece of what makes a super app an actual super app, right?
00:17:25.440 --> 00:17:37.920
So not only bringing in the right side browser or the right pane or the inspect pane browser uh for the internet, but also for any files, uh, but also to preview anything that you're actually building.
00:17:38.079 --> 00:17:45.440
So at least for me, I've been trying to spend more and more of my time in that right side pane, right?
00:17:45.599 --> 00:17:53.440
So bringing in my browser sessions, so I can control that right side pane, or you know, codecs or the super app can control it as well.
00:17:53.599 --> 00:18:00.000
Um, I can bring in my files, my PDFs, my Word documents, uh, and then also preview different things I'm building there.
00:18:00.079 --> 00:18:01.920
But I can control that pain, right?
00:18:02.000 --> 00:18:05.680
Or Codecs can as well, or the super app of choice can as well.
00:18:05.839 --> 00:18:19.440
So you can see there by doing your work in an interface that the super app can see and work with you is obviously a huge advantage and one of the biggest steps uh toward the uh agentic work future.
00:18:21.359 --> 00:18:24.720
So let's quickly talk about the race right now.
00:18:25.440 --> 00:18:30.480
So I will say codecs and ChatGPT, they're ahead by a far margin.
00:18:30.640 --> 00:18:41.279
And I don't think anyone is going to dispute that because even uh Google literally copy and pasted codecs, right?
00:18:41.519 --> 00:18:50.480
Even in their launch video of anti-gravity 2.0 desktop version, they accidentally left something in the video, uh, the launch video that had a codex folder.
00:18:50.640 --> 00:18:53.039
It is a spitting image of codecs, right?
00:18:53.200 --> 00:18:58.160
Even with Microsoft's new GitHub Copilot, very much set up like Codex.
00:18:58.559 --> 00:18:58.960
All right.
00:18:59.039 --> 00:19:02.000
Uh cursor, very good uh in its own right.
00:19:02.160 --> 00:19:06.240
Claude Desktop, I I I would say it's a little far behind.
00:19:06.720 --> 00:19:10.720
And you know, you can even look at different benchmarks that talk about the harness.
00:19:10.880 --> 00:19:15.039
Uh, unfortunately, Anthropic has some of the best models in the world, right?
00:19:15.119 --> 00:19:16.720
Literally, the best models in the world.
00:19:16.880 --> 00:19:18.720
Their harness is bad.
00:19:19.279 --> 00:19:20.480
That's not my opinion.
00:19:20.559 --> 00:19:27.440
That's every single benchmark that looks at um that looks at uh harness, tool calling, etc.
00:19:27.680 --> 00:19:34.880
So yeah, the clawed models perform much better inside cursor as an example than they do inside of Clawed Desktop.
00:19:35.039 --> 00:19:35.839
Uh yeah, yeah.
00:19:35.920 --> 00:19:37.440
It's I don't know.
00:19:37.599 --> 00:19:40.559
I don't know if they design it to be token inefficient on purpose.
00:19:40.720 --> 00:19:41.359
I don't know.
00:19:41.440 --> 00:19:41.920
Uh right.
00:19:42.000 --> 00:19:44.160
And again, I'm not talking about the command line interface.
00:19:44.240 --> 00:19:47.359
Um, so I'm not talking about the CLI uh side, right?
00:19:47.440 --> 00:19:52.640
I'm talking about the uh, you know, everyday knowledge worker who's downloading these desktop programs.
00:19:52.720 --> 00:19:58.400
But I'd say right now, if I had to call who's where in the race, I'd say number one is codex.
00:19:58.559 --> 00:20:01.599
Uh, I'd say 2A is probably cursor.
00:20:01.759 --> 00:20:03.920
I'd say 2B is clawed desktop.
00:20:04.000 --> 00:20:07.359
Uh, and then I'd say the rest is like to be determined.
00:20:07.440 --> 00:20:07.920
Uh, right.
00:20:08.079 --> 00:20:13.599
Anti-gravity, uh, not not a big fan uh of that one.
00:20:13.759 --> 00:20:16.079
Microsoft will see where they land.
00:20:16.240 --> 00:20:30.400
Uh, I do think their new GitHub co-pilot app will ultimately be the kind of the underlying harness that will uh uh kind of power their super app that CEO Satya Nadella uh previewed at the Microsoft Build Conference.
00:20:30.559 --> 00:20:32.400
Uh and then you have others, right?
00:20:32.720 --> 00:20:34.240
Grok has their new build.
00:20:34.319 --> 00:20:36.720
That's still too early, kind of in the beta phase.
00:20:36.880 --> 00:20:50.799
But I mean, just in the past month, you know, and with what Kodak has coming up, we've seen four of the six big players say this is where we're focusing a lot of our engineering efforts from here on out.
00:20:50.880 --> 00:20:59.519
And I'm not saying that the model can no longer become a part of the moat, but I think there is going to get a certain point, right?
00:20:59.599 --> 00:21:04.799
I'm trying to uh come up with a good way to verbalize this and maybe I'll do a show on it in the future.
00:21:04.960 --> 00:21:08.799
Uh, you know, the way I think of it is like 95-5.
00:21:09.200 --> 00:21:23.759
Pretty sure we're gonna get to pretty soon, uh, I'd say probably early 2027, we're gonna get to the point where only 5% of people can actually take full advantage of 95% of what a model has to offer.
00:21:23.920 --> 00:21:33.839
I think there's gonna be a certain point where your your average, even your technical person, is not going to be able to take advantage of 95% of what a model can offer.
00:21:33.920 --> 00:21:50.160
So that's why I think more and more, at least when it comes to short and medium-term enterprise adoption, I think it's actually more about the user experience of using the harness uh than it is about the model, which I think is actually one reason why cursor is actually nice, nicely positioned in this.
00:21:50.720 --> 00:21:55.759
So let's quickly go over why I think Codec leads right now and I'll give you quick takes on all of them.
00:21:55.920 --> 00:22:01.519
I think Codecs is by far, they they've set the the Bar very high.
00:22:01.680 --> 00:22:04.799
Um, mainly just because everything works.
00:22:04.880 --> 00:22:09.680
And they are the only one out of the bunch that has every single thing that I talked about.
00:22:09.759 --> 00:22:10.240
Uh, right.
00:22:10.319 --> 00:22:12.799
So the three panes uh all work, right?
00:22:12.960 --> 00:22:27.200
Specifically on the inspect or browser, they have the code preview pane, they have the file preview pane, and then they have the um the browser pane all connected, which I think only two others have.
00:22:27.359 --> 00:22:32.160
So not everyone has all of those different um elements, which I think all of them are very important.
00:22:32.319 --> 00:22:35.920
But the other reason why I think codecs is very far ahead, uh, two reasons.
00:22:36.160 --> 00:22:49.279
One for the laptop crowd, you can literally have your laptop closed and codecs can still run everything as if your laptop was open and logged in, which I still don't understand the wizardry that OpenAI did to make that work.
00:22:49.440 --> 00:22:49.759
All right.
00:22:49.839 --> 00:22:53.680
But then the other thing is being able to access this from your phone.
00:22:53.920 --> 00:22:58.799
Because as you are on the go, um I, you know, this is weird.
00:22:58.960 --> 00:23:02.160
I've never been able to get so much work done from my phone.
00:23:02.319 --> 00:23:15.039
So I think that's two reasons why, aside from the obvious design and UX of Codecs, which everyone is copying, but the fact that the computer use, the browser use are just a step ahead of everyone else.
00:23:15.200 --> 00:23:20.160
Um, the actual harness itself, but the ability as well, your computer can be closed.
00:23:20.240 --> 00:23:23.119
Uh, you it can be on the locked screen, logged out.
00:23:23.359 --> 00:23:28.160
Um, and to be able to access everything from your phone is huge, right?
00:23:28.240 --> 00:23:32.720
So, as an example, we're gonna talk about um Anthropic's Claw Desktop here.
00:23:32.799 --> 00:23:39.839
It's like they have dispatch where you can kind of access some things from your phone, but it's super buggy.
00:23:39.920 --> 00:23:43.680
And you actually can't, you know, as an example, you can't access everything.
00:23:43.839 --> 00:23:51.200
You get one thread that's just a dispatch thread, which to me wasn't very helpful, but I think codex really leads.
00:23:51.359 --> 00:23:59.440
But I think cursor, yeah, cursor might actually be one of the best challengers, at least right now, in the super app game.
00:23:59.680 --> 00:24:11.279
So not only have they, you know, kind of been in the super app space a little bit longer than everyone else, um, but they've had the developer feedback.
00:24:11.440 --> 00:24:19.279
Uh, obviously, now they have this partnership uh with SpaceX, uh, which just had the largest IPO in history.
00:24:19.440 --> 00:24:29.200
Uh, so you can't count cursor out because they're gonna have more and more resources um to build uh this this their kind of super app, but also their own model as well.
00:24:29.359 --> 00:24:34.880
Their new composer, uh, I think it's composer 2.5 model is really good.
00:24:35.039 --> 00:24:47.200
Um, it is token efficient, and you know, you can kind of get about 90% of what you can get out of like an Opus 4.8 for like 10% of the cost, which is why one of the reasons I think you can't count cursor out.
00:24:47.279 --> 00:24:48.960
They've been in this space the longest.
00:24:49.039 --> 00:24:52.000
Uh, their user experience and the harness is really good.
00:24:52.160 --> 00:25:02.000
It's been uh shaped and loved by uh many developers for multiple years, and they have the one big advantage where you can use literally any model.
00:25:02.240 --> 00:25:04.720
This is the only one where you can use any models.
00:25:05.039 --> 00:25:13.200
You can use, you know, the GPT models, you can use the claw models, you can use Microsoft's models, you can use Google's models and their own models.
00:25:13.279 --> 00:25:16.960
So, cursor, I think you have to pay close attention to.
00:25:17.119 --> 00:25:20.799
Uh, next, Microsoft's could be right.
00:25:20.960 --> 00:25:21.680
We don't know.
00:25:21.839 --> 00:25:24.880
Microsoft has said their super app is coming.
00:25:25.039 --> 00:25:31.519
I know people laughed at me uh like six months ago when I'm said I'm actually kind of bullish on Microsoft, right?
00:25:31.680 --> 00:25:47.359
Uh they we've seen some pivots that in early 2026 that was kind of against their strategy of 2025, uh, which was creating more and more products, where it seems like they are going to this super app strategy and trying to streamline and unify everything.
00:25:47.519 --> 00:25:49.440
So we don't know what it's gonna look like.
00:25:49.599 --> 00:25:54.319
I do think that their recent GitHub Copilot desktop app was well well received.
00:25:54.480 --> 00:26:04.000
I think was not well received, was their uh kind of coupling that with the uh unsubsidizing uh of their plan, right?
00:26:04.160 --> 00:26:07.759
So yeah, we talked about that in uh the last episode.
00:26:07.839 --> 00:26:17.920
But so I think a lot of people had a sour taste from GitHub Copilot's pricing moves, but their actual new desktop app that they rolled out for Mac and Windows, it's really good.
00:26:18.000 --> 00:26:24.640
And it does look like that's gonna serve like the underlying baseline uh of the feat of the future Windows Copilot super app.
00:26:24.799 --> 00:26:33.039
So they did talk about this at their build conference and bringing also kind of the new element that they're gonna be bringing into this is their autopilots, right?
00:26:33.119 --> 00:26:37.599
So this is their their version of an autonomous coworker that works on your desktop, right?
00:26:37.680 --> 00:26:42.319
It can read, write, uh, you know, you can schedule it and it can access all your Microsoft 365 data.
00:26:42.400 --> 00:26:48.319
So even though Microsoft's version isn't out yet, I do put this fairly high up.
00:26:48.480 --> 00:26:53.680
All right, let's talk about Claude in Anthropic because they deserve a ton of credit here.
00:26:53.839 --> 00:27:08.000
Because I think without uh Anthropic's move uh in late 2025 and early 2026, specifically with co-work, I don't know if we would be in this era of um agenc AI and AI super apps.
00:27:08.160 --> 00:27:17.359
So uh I do think even though they've they've fallen behind maybe a little bit, um, Anthropic Claude, you have to give them their shine on this, right?
00:27:17.599 --> 00:27:30.240
Because even I remember working with the early versions of co-work, I'm like, even though this is slow, even though it's clunky, even though it fails sometimes, when it did get something right, co-work was mind-blowing.
00:27:30.400 --> 00:27:39.599
And I do actually think, which maybe this is because I do this every single day, I do think the enterprise now is just realizing, right?
00:27:39.680 --> 00:27:51.839
Because I think the enterprise is always about three months behind kind of cutting edge AI, which is why I still think that there's this momentum uh around Claude Code and Claude Co-Work and you know, Claude's desktop offerings.
00:27:52.079 --> 00:27:56.960
But at least when you compare it to what's out there right now, specifically with codecs, I do think they're behind.
00:27:57.039 --> 00:28:02.000
One of the reasons why, and I've done this on the show before, I showed this, it's not unified, right?
00:28:02.240 --> 00:28:05.519
Claude Code has no clue what Claude Cowork is doing.
00:28:05.599 --> 00:28:07.759
And on the desktop, there's essentially three different modes.
00:28:07.920 --> 00:28:15.279
You have your standard chat, you have your Claude Cowork, you have your Claude Code, and none of the others know that any of the others exist.
00:28:15.359 --> 00:28:23.440
So there's no unified memory, which is one of the biggest downsides of something like an autonomous AI super app.
00:28:23.519 --> 00:28:24.799
Like it has to know everything.
00:28:24.880 --> 00:28:27.200
That's one of the things I think Codex is great at.
00:28:27.440 --> 00:28:32.000
Codex can actually go in and start new codex chats on its own, right?
00:28:32.160 --> 00:28:33.680
Which is great, right?
00:28:33.759 --> 00:28:39.599
When it comes from an organizational standpoint, you can't do any of that with uh Anthropic's desktop offerings now.
00:28:39.759 --> 00:28:51.039
Uh, the easy, I guess, the the the bright side for uh Anthropic in what they should be doing, because this would even change my per perception on the current AI race.
00:28:51.200 --> 00:28:53.359
All they have to do is do what everyone else has done.
00:28:53.519 --> 00:29:03.039
Go copy OpenAI, go copy paste codex, and that turns what I think uh the rest of 2026 and 2027 looks like.
00:29:03.200 --> 00:29:15.519
Um, if they don't, I do think open AI has the ability to really run away, uh, especially with the however many weeks or months or quarters we have until Anthropic and OpenAI both go public.
00:29:15.599 --> 00:29:29.279
I think OpenAI is gonna be way further ahead than most people are giving them credit for, unless Claude essentially uh or Anthropic copies uh co-work, which or sorry, copies codex, which I don't know why they haven't.
00:29:29.440 --> 00:29:34.640
Uh everyone else has done it because it's very popular, it's the best thing out, right?
00:29:34.799 --> 00:29:36.240
Uh, so yeah, we'll see.
00:29:36.400 --> 00:29:38.480
We'll see what Anthropic decides to do.
00:29:38.720 --> 00:29:40.880
Next, Google Anti-Gravity.
00:29:41.039 --> 00:29:52.400
Uh, to me, again, aside from being a uh a I don't want to say anything uh bad about the team there, uh, because I do have good relationships with a lot of the people at Google.
00:29:52.799 --> 00:29:56.319
I don't think they should have released their anti-gravity 2.0.
00:29:56.480 --> 00:29:58.880
Uh it's it just wasn't ready.
00:29:58.960 --> 00:30:01.680
Uh it's it's there's hardly no features.
00:30:02.000 --> 00:30:04.880
Um I don't know.
00:30:05.039 --> 00:30:07.440
They rolled this out uh right after their I.O.
00:30:07.599 --> 00:30:08.160
conference.
00:30:08.240 --> 00:30:13.119
Uh you know, they rolled it out in conjunction with Gemini 3.5 Flash, which I thought was a good model.
00:30:13.200 --> 00:30:20.799
So they thought that, hey, the combination of this fast coding uh, you know, model, Gemini 3.5 flash, that can do this long autonomous work.
00:30:20.960 --> 00:30:25.839
Let's couple it with Google Anti-Gravity, and you know, it's gonna be this great splash.
00:30:25.920 --> 00:30:34.160
I I don't think it played out like that, mainly because I think Google kind of changed the flash model from what people expected versus what they were making it.
00:30:34.240 --> 00:30:49.039
Because I think the way people looked at Gemini 3.5 Flash was it's a cheap, fast model, and it's no longer cheap, it's still very fast, but it is expensive relatively compared to every other uh quote unquote flash model that Google has ever released.
00:30:49.200 --> 00:31:00.240
So, in that aspect, by tying it together with the anti-gravity 2.0 with the Gemini 3.5 flash, maybe hurt adoption of anti-gravity, but that and just the features aren't really there.
00:31:00.400 --> 00:31:02.160
Yes, it's an earlier version.
00:31:02.400 --> 00:31:05.119
Yes, they have basic automations.
00:31:05.519 --> 00:31:08.640
Uh, you know, you can use a couple other models, right?
00:31:08.720 --> 00:31:11.440
You can use the Claude models um in there.
00:31:12.640 --> 00:31:14.000
Not for me, right?
00:31:14.079 --> 00:31:15.519
Uh, just not enough features.
00:31:15.599 --> 00:31:20.559
It is like the most bare boned uh kind of super app you can think of, right?
00:31:20.799 --> 00:31:28.079
Where I think Codecs is robust, you know, even cursor and uh Claude Desktop, much more robust.
00:31:28.720 --> 00:31:32.880
Anti-gravity, uh, it's not super clunky, it's just there's like no features.
00:31:33.039 --> 00:31:36.880
So maybe it's great if you want to learn what a super app can do.
00:31:36.960 --> 00:31:45.599
And if your organization is heavily bulked into the Google ecosystem ecosystem, maybe it's a good place to start and get your kind of super app feet wet.
00:31:45.759 --> 00:31:54.400
But if you have the choice to use the others, I would not personally use Google Anti-Gravity 2.0, although I do hope that they continue to improve it in the future.
00:31:54.559 --> 00:32:00.319
And I think they will because, like I said, they did give it a lot of spotlight uh at their Google I.O.
00:32:00.480 --> 00:32:00.880
conference.
00:32:01.279 --> 00:32:06.000
All right, so let's begin to wrap this up and because we have to talk about the other side.
00:32:06.240 --> 00:32:10.079
I talked about all the reasons why, hey, super app is the next trend.
00:32:10.160 --> 00:32:14.000
Uh, it is the next wave of work, it is the execution layer, uh, right?
00:32:14.079 --> 00:32:30.160
It is the harness because I think most of the big companies know that the the the kind of the benefits or the advantages uh over new models is going to become less and less as a smaller and smaller uh section of the enterprise will even be able to squeeze out the juice, right?
00:32:30.319 --> 00:32:36.799
Let's just say, as an example, let's say Fable 5.1 or you know GPT 5.7.
00:32:37.200 --> 00:32:45.359
What percentage of people, what percentage of people honestly are going to be able to take true advantage of those point step releases?
00:32:45.440 --> 00:32:46.480
And I think not a lot.
00:32:46.640 --> 00:32:59.039
So I that's why I think the uh emphasis in the moat in the short and medium term is actually going to be on the user experience and giving people the harness to power all of this.
00:32:59.200 --> 00:33:02.799
But with that comes great responsibility, right?
00:33:02.960 --> 00:33:05.119
Essentially, let's talk about YOLO mode.
00:33:05.279 --> 00:33:11.680
Uh essentially, I can set up and you can set up most of these super apps to quite literally do anything.
00:33:11.759 --> 00:33:48.079
And when you talk about the ability to read and write any file, any folder, to use the terminal, you can accidentally, if you're not paying attention, to use to have browser use, to have computer use, if you don't have the right uh role based access controls, if you don't have the right user permission set up, if you don't have redundancies and backups, these AI super apps, if you're whether the person is trying to be a little bit too aggressive and automating too much of their work, whether someone's being a little lazy and not really paying attention, whoops, you can accidentally do a lot of harm.
00:33:48.240 --> 00:33:49.920
You can accidentally delete something.
00:33:50.000 --> 00:34:01.839
And if you don't have uh something properly permissioned, if you don't have the backups uh in place, yeah, you can the the the the the AI super apps, there's actually the risk is in the action, right?
00:34:02.079 --> 00:34:06.880
The power is actually the downs, the downside and the downfall as well, right?
00:34:06.960 --> 00:34:16.880
I did a training for a large uh organization uh a couple of months ago that was rolling out kind of a desktop super app to thousands of users.
00:34:17.039 --> 00:34:20.079
And they were asking me, hey, how should we be using computer use?
00:34:20.159 --> 00:34:21.920
And I said, Don't, right?
00:34:22.159 --> 00:34:22.639
Don't.
00:34:22.800 --> 00:34:28.239
You know, you shouldn't, you like some of the more powerful things you can go and change settings on the computer.
00:34:28.400 --> 00:34:46.159
If you're not um strict enough in your context engineering or prompting process, sometimes these systems, especially if you're using uh inferior model, they're gonna look at the end goal and they're gonna build a completely new path that might be a path that's detrimental to your group.
00:34:46.400 --> 00:35:07.039
So, you know, if you're not paying enough attention to learning, and if your company, your department, you personally, if you haven't put in the proper um time to even understand the agentic natures of these models by default, I wouldn't recommend, you know, letting these things loose and then having hands off, right?
00:35:07.280 --> 00:35:10.719
Human in the loop plus an AI super app is not a good combination.
00:35:10.800 --> 00:35:13.199
That's why I preach expert-driven loops.
00:35:13.280 --> 00:35:19.519
So if you are doing true human in the loop, that's where you give it a prompt, and oh, I'm gonna have it go run for 48 hours.
00:35:19.679 --> 00:35:27.280
Let's go token maxing, let's use computer use, browser use, everything, and just have it solve all my all my issues for the quarter, right?
00:35:27.360 --> 00:35:29.679
And just go truly hands-off YOLO mode.
00:35:29.840 --> 00:35:35.920
If you don't know what you're doing, if you don't know what that agent is capable of, things can go down a sour path.
00:35:36.000 --> 00:35:38.960
If you just go truly, oh, I'm the human in the loop, let me set this up.
00:35:39.039 --> 00:35:41.199
No, you have to have an expert-driven loop.
00:35:41.280 --> 00:35:52.880
You have to be very hands-on in learning and understanding these capabilities, reading the chain of thought, iterating, uh, sandboxing, you know, sandboxing uh your browser use at first until you fully understand it.
00:35:52.960 --> 00:36:04.559
Because it can, you can log into any SaaS application that you and your company use, even if there's not an API, even if there's not an MCP or a direct integration or a connector in these super apps, right?
00:36:04.719 --> 00:36:09.760
You can go log in and well, the super app can control whatever browser you're using.
00:36:09.920 --> 00:36:16.239
So the risk is immense if you are not properly learning uh and keeping up with all of the developments.
00:36:16.400 --> 00:36:18.000
So let's wrap here.
00:36:18.159 --> 00:36:19.519
Here's where I think you should start.
00:36:19.679 --> 00:36:20.880
Haha, start here.
00:36:21.039 --> 00:36:23.199
Uh, read only first, right?
00:36:23.440 --> 00:36:24.400
Add approvals.
00:36:24.480 --> 00:36:26.239
You need to be sandboxing these things.
00:36:26.320 --> 00:36:33.920
Uh instead of rolling these out to hundreds or thousands of employees, because there is going to be, I think, the the urge to do this, right?
00:36:34.239 --> 00:36:39.519
Whether it's in a couple of days or a couple of weeks when OpenAI unveils and opens up their super app.
00:36:39.679 --> 00:36:45.360
Same thing with Microsoft, as we undoubtedly see improvements from Anthropic and from Google as well.
00:36:45.519 --> 00:36:47.519
I think most people are going to want to rush in, right?
00:36:47.599 --> 00:36:49.920
They're like, oh my gosh, look at all these new capabilities.
00:36:50.079 --> 00:36:55.920
Don't read only, strict permissions, guardrails, training, learning and development.
00:36:56.079 --> 00:37:07.199
You have to be educating because I think this is where when I talked about agentic crash, uh, you know, in our 2026 um AI prediction and roadmap series, this is where agentic crash happens.
00:37:07.360 --> 00:37:17.199
When you have companies that give too much access, not enough training, uh, you know, too much power without proper responsibility and training.
00:37:17.360 --> 00:37:22.400
So you have to automate the low-hanging fruit with low risk.
00:37:22.559 --> 00:37:25.199
We've talked about how you can tear off the different risk.
00:37:25.280 --> 00:37:26.320
So that's where you start.
00:37:26.400 --> 00:37:36.559
Don't start with the highest capabilities, with the most handoff and autonomous um actions, because I know that's the that's the that's the sexy thing, right?
00:37:36.639 --> 00:37:42.880
You're like, oh my gosh, I can automate 90% of my work, have this thing work for 48 hours, and I can come back.
00:37:42.960 --> 00:37:44.719
And it did two weeks of deliverables for me.
00:37:44.880 --> 00:37:46.000
Oh my gosh, let's do it.
00:37:46.159 --> 00:37:47.280
Let's do this at scale.
00:37:47.360 --> 00:37:48.559
That's not where you start.
00:37:48.800 --> 00:37:49.280
All right.
00:37:49.440 --> 00:37:54.079
You have to be able to crawl before you walk, walk before you run.
00:37:54.159 --> 00:37:55.599
So don't start this with a sprint.
00:37:55.679 --> 00:37:56.960
You will get burned.
00:37:57.119 --> 00:37:57.519
All right.
00:37:57.679 --> 00:38:08.320
That is a wrap previewing the AI super apps and why every company is racing to create one, what they are, and hopefully how your company can properly take advantage of this.
00:38:08.480 --> 00:38:12.159
So if this was helpful, make sure you go to startheseries.com.
00:38:12.239 --> 00:38:23.599
That's going to give you access to our everyday AI inner circle community, a playlist with every single Start Here Series episode on there, and you can go connect and network with other business leaders who are trying to do the same.
00:38:23.679 --> 00:38:24.719
So thank you for tuning in.
00:38:24.880 --> 00:38:27.679
Hope to see you back tomorrow in every day for more everyday AI.
00:38:27.760 --> 00:38:28.800
Thanks, y'all.
00:00:02.240 --> 00:00:12.880
Any hour or day or week now, we'll get an official announcement from OpenAI that signals the biggest AI pivot we've seen since ChatGPT.
00:00:12.960 --> 00:00:17.440
And that's the official rollout of their AI super app.
00:00:17.760 --> 00:00:33.280
So the ChatGPT Parent Company, now with 1 billion weekly active users, made headlines recently when a senior executive off the record at least said three words to a reporter that will probably echo for months.
00:00:33.600 --> 00:00:35.759
Chat is dead.
00:00:36.320 --> 00:00:45.920
So any day we'll get the official announcement of OpenAI's new super app, their push to bring an autonomous coworker to millions of companies.
00:00:46.240 --> 00:00:49.359
But it's not just OpenAI making this push.
00:00:49.600 --> 00:01:02.399
Actually, Anthropic started the wave many months ago with their rollout of Claude Code and Claude Cowork, and every other big tech challenger is racing to get their version of an AI super app some momentum.
00:01:02.560 --> 00:01:05.439
Google released theirs in Anti-Gravity 2.0.
00:01:05.680 --> 00:01:09.680
Cursor has quietly become a serious contender in the space.
00:01:09.920 --> 00:01:14.239
And Microsoft also announced their upcoming super app as well.
00:01:14.480 --> 00:01:18.319
But that begs the obvious question: what the heck is a super app?
00:01:18.480 --> 00:01:20.799
And why should my company take note?
00:01:21.040 --> 00:01:26.799
Well, we're gonna tackle the latest work layer shift on today's show.
00:01:26.959 --> 00:01:31.120
So welcome to everyday AI and our start here series.
00:01:31.280 --> 00:01:35.120
But let's go over the big picture of what's happening now.
00:01:35.280 --> 00:01:43.680
So Anthropic started this desktop agencera with Claude Code and Claude Cowork in late 2025 and early 2026.
00:01:43.840 --> 00:01:54.159
But Anthropic hasn't really been able to build off of that momentum, at least when it comes to being the big name in the room, when it comes to an AI super app.
00:01:54.319 --> 00:01:59.120
And I think OpenAI's super app, which is essentially codecs, right?
00:01:59.200 --> 00:02:05.760
And it will uh be a little more tightly integrated into Chat GPT, will be coming any day.
00:02:05.840 --> 00:02:11.919
And I think that's really going to set the tone for this big pivot to the super app era, right?
00:02:12.080 --> 00:02:22.879
And Microsoft, I think, especially when you have OpenAI and Microsoft in the coming weeks or months shifting most of their agentic focus to a super app.
00:02:23.039 --> 00:02:26.560
I think the rest of the enterprise really needs to pay attention.
00:02:27.120 --> 00:02:39.439
So on today's show, we're gonna be diving into and stick with me for the next 25 minutes, and here's what you're gonna learn why Anthropic started the super app race, but they have kind of fallen out of immediate contention.
00:02:39.599 --> 00:02:45.520
You're gonna know why the models are no longer the moat and the super app harness matters more than ever.
00:02:45.680 --> 00:02:54.719
And you're gonna kind of know the state of the super app race, at least as it stands right now, right before we officially get it all kicked off.
00:02:54.960 --> 00:02:56.400
All right, let's get into it.
00:02:56.560 --> 00:02:58.560
Welcome to the Start Here series.
00:02:58.719 --> 00:03:05.120
This is everyday AI's essential podcast series to both learn the AI basics and to double down on your knowledge.
00:03:05.280 --> 00:03:10.479
So if that's what you're trying to do, we're at the uh the tail end of the start here series, right?
00:03:10.560 --> 00:03:13.280
Now we're in uh episode what are we on now?
00:03:13.439 --> 00:03:16.159
We're on uh volume 28 of this.
00:03:16.240 --> 00:03:18.319
So we'll probably do two more and wrap it at 30.
00:03:18.400 --> 00:03:20.800
But if you don't know the start here series, it's great.
00:03:20.879 --> 00:03:25.840
I would say start at volume one if you can, uh, and work your way up to where we are now.
00:03:25.919 --> 00:03:28.800
They're usually quicker episodes, about 30 minutes, give or take.
00:03:28.879 --> 00:03:31.199
But make sure you go to starthirseries.com.
00:03:31.360 --> 00:03:38.639
There, we put together a playlist where you can listen to every single episode of the Start Here Series so you don't have to go fishing around our website or anything like that.
00:03:38.719 --> 00:03:43.599
And that's also going to give you exclusive access to our inner circle community.
00:03:43.680 --> 00:03:46.800
Yeah, right now that's the only way to get open access to it.
00:03:46.879 --> 00:04:02.639
Uh, so you can not only sign up for our free inner circle community, which gives you access as well to our prime prompt polish uh version three course to learn context engineering 101 at your own pace, but also all of the start here series info.
00:04:02.800 --> 00:04:09.759
All right, and make sure if you didn't listen to our last episode, I think it's an important one in the start here series, which was volume 27.
00:04:10.000 --> 00:04:18.160
We went over the token maxing era and how it's over and talking about the new era of token efficiency and how your company should adapt.
00:04:18.240 --> 00:04:19.519
So make sure you go check that out.
00:04:19.600 --> 00:04:22.399
That's volume 27 or episode 789.
00:04:22.480 --> 00:04:27.439
But let's get into today's topic, which is the AI super app.
00:04:27.920 --> 00:04:33.920
So, what the heck is an AI super app and why does it matter?
00:04:34.160 --> 00:04:35.199
Well, I'll say this.
00:04:35.439 --> 00:04:46.480
Even though I don't think we've quote unquote arrived in the super app era per se, I will say it's safe to assume this is the next version.
00:04:46.720 --> 00:04:54.000
All right, and it is funny because we started the start here series in January, and we've actually seen kind of two shifts.
00:04:54.079 --> 00:05:12.639
Uh, so I think the first shift we saw in January was models becoming more and more agentic by default, and saying, hey, how the future of work is not uh reactively talking to AI chatbots, it's knowing how these models work, scheduling things out, right?
00:05:12.720 --> 00:05:18.879
And then we also had another episode talking about something I called agentic context carry, right?
00:05:19.040 --> 00:05:28.639
And that is the crucial role that uh not just models, but the harness that control the models play in taking context from your certain apps, right?
00:05:28.800 --> 00:05:32.800
So uh to boil this down to the the basics here, right?
00:05:32.879 --> 00:05:39.120
One of the biggest advantages of not just the upcoming super app era, but where we're at now, right?
00:05:39.199 --> 00:05:44.560
Before Microsoft and OpenAI start really highlighting uh and promoting their super apps, right?
00:05:44.720 --> 00:05:50.480
Where we're at now is we do have the system that can carry our context autonomously, right?
00:05:50.560 --> 00:06:02.240
So a lot of what the human duct work, uh duct tape work has been over the last two years is us humans who do use large language models is giving a chat our context, right?
00:06:02.399 --> 00:06:09.600
Connecting whether it was static files, dynamically uh shifting things, but then it's a lot of copying and pasting from different AI systems.
00:06:09.759 --> 00:06:26.720
But now, as these systems are living on our desktop and becoming more autonomous, that does leave uh kind of lead into the next wave uh going beyond this agentic context carry, and that's where the agents ultimately live, which is the desktop OS, right?
00:06:26.879 --> 00:06:30.079
That is where I think the super apps are going to play.
00:06:30.240 --> 00:06:37.600
And I don't know if this is going to be the quote unquote final frontier of the agentic era as we have officially entered into it.
00:06:37.759 --> 00:06:53.680
It may look a little bit different, but I do think at least uh the companies that are going to be most successful in late 2026 and early 2027 are those that are going to be investing in their companies using these super app systems on the desktop.
00:06:54.079 --> 00:06:57.680
So the core idea of this is simple, right?
00:06:58.959 --> 00:07:05.600
A super app is bringing all of your different work surfaces together under one agentic roof.
00:07:05.759 --> 00:07:11.040
So the big AI tech companies are all building their own super apps, they know the same truth.
00:07:11.279 --> 00:07:14.319
An AI chatbot is not a moat, right?
00:07:14.480 --> 00:07:17.199
ChatGPT.com is not a moat.
00:07:17.360 --> 00:07:19.120
The model is not the moat.
00:07:19.279 --> 00:07:23.120
They all want to own the execution layer, uh, right.
00:07:23.279 --> 00:07:29.680
And that is why every single big AI tech company is trying to build an AI super app.
00:07:29.839 --> 00:07:31.759
You could call it the harness, right?
00:07:31.920 --> 00:07:36.240
Kind of the uh the place where all of the tools come together, but it is much more than that.
00:07:36.319 --> 00:07:42.639
And I think we'll see uh as we really uh roll out and dig into what a super app actually is.
00:07:42.800 --> 00:07:46.639
But at its basis, I think it combines a couple of different things.
00:07:46.959 --> 00:08:03.759
It combines your company's context, your files, browser access, a file viewer, uh, automations, unified memory, uh, and approvals and automations all in one unified platform.
00:08:04.160 --> 00:08:09.360
And you, you know, we're gonna get into some of the nuances and the pros and the cons and all of those things.
00:08:09.600 --> 00:08:16.639
But the way I like to look at it now, and this is after I've been using uh especially codex heavily, right?
00:08:16.720 --> 00:08:23.199
I I also use the claw desktop programs heavily, but I do think codex is in the lead in the super app, right?
00:08:23.279 --> 00:08:25.519
And open AI has essentially said FYI.
00:08:25.600 --> 00:08:36.480
They're saying, yes, Codex is the early version of the super app, which is why I've been encouraging people to use it so far because I do think codex is very ahead of all of the other quote unquote super apps.
00:08:36.639 --> 00:08:45.679
We've seen that by how they're all the other big tech companies are kind of just copying and pasting exactly what OpenAI has done with Codecs because it is that good and it is that far ahead.
00:08:45.759 --> 00:08:51.440
But this is essentially the way I look at a super app right now, it is a full like intern, right?
00:08:51.519 --> 00:08:58.320
It is a full almost human coworker because anything that I can do, right, a super app can do.
00:08:58.559 --> 00:09:02.799
And I don't think this was the case in late 2025 with Claude Code.
00:09:02.960 --> 00:09:08.240
I don't think it was the case in early 2026 with Claude Cowork either, because now where we're at, right?
00:09:08.320 --> 00:09:26.480
So aside from what I said, dynamic access to your company's files, um, automations built-in, approvals built in, uh, and then having these kind of three different panes, and we're gonna talk about that a little bit more, but then also the ability to control a browser and the ability to control your entire computer.
00:09:26.559 --> 00:09:40.320
Because at that point, that's when you can kind of right say, okay, this system at least has the capability to be a true researcher, to be a true uh, you know, in terms of an output.
00:09:40.559 --> 00:09:46.480
This can output something like a smart human actually could because it has access to every single tool.
00:09:46.639 --> 00:09:46.960
All right.
00:09:47.039 --> 00:09:51.600
And there's obviously great downsides to this because with great power comes great responsibility.
00:09:51.759 --> 00:10:00.240
But essentially the super app gives read-write access to your entire computer, or it can if you give it those permissions and kind of go in YOLO mode, right?
00:10:00.320 --> 00:10:04.480
But it can download, it can upload, it can change different settings on your computer.
00:10:04.559 --> 00:10:05.120
Uh, right.
00:10:05.200 --> 00:10:08.240
You can even uh with codecs at least, right?
00:10:08.399 --> 00:10:14.799
I can, if my MacBook Pro is closed, I can still control my MacBook, you know, from my phone.
00:10:14.960 --> 00:10:22.480
So it also brings in that um, you know, that remote aspect as well, which is why I do think uh OpenAI's codex is a little more behind, right?
00:10:22.559 --> 00:10:32.879
Because even if, you know, my my my laptop is closed or you know, if my coworker is not in front of the desk, I can still send a message and they can still get to work and report back to me.
00:10:33.440 --> 00:10:37.200
Let's clear up a little bit about the name super app first, right?
00:10:37.519 --> 00:10:52.639
Because I think that there's a little bit of a mix in terms of when you throw the word super app, because in the pre-generative AI era, or even in the early Chat GPT days, when people heard super app, they instantly thought of China's WeChat, right?
00:10:52.720 --> 00:10:57.600
So that is the super app, the uh original super app OG, right?
00:10:57.679 --> 00:11:01.039
So they have messaging, payment, shopping, banking, everything.
00:11:01.200 --> 00:11:08.240
And I think that that was kind of one of the big goals of X slash Twitter slash early groc, right?
00:11:08.399 --> 00:11:10.639
It was this everything app ambition, right?
00:11:10.720 --> 00:11:16.320
It was moving all of your day-to-day personal life into a single app like China's WeChat.
00:11:16.480 --> 00:11:26.000
But that's not what I'm talking about here when I talk about the new era of super apps, because I think for whatever reason, that dream has maybe been dashed.
00:11:26.159 --> 00:11:47.200
And maybe it's because most of the big companies have understood and realized wait, why would we go after this on the consumer side, all on a smartphone, when there's a much bigger opportunity to do this AI super app on the desktop side and to go after companies that are spending millions of dollars on this every single year.
00:11:47.279 --> 00:11:53.279
So a bigger way to capitalize in a faster way to roll this out across the enterprise as well.
00:11:54.240 --> 00:12:00.559
So now you know it's not the you know, quote unquote, traditional WeChat mobile super app.
00:12:00.639 --> 00:12:07.919
We are talking about an AI uh desktop companion, and it's better than a chatbot, right?
00:12:08.080 --> 00:12:16.559
I don't think I need to go into too much depth on why a desktop super app is better than a chatbot, but let me just go over the basics, right?
00:12:16.720 --> 00:12:18.799
A chatbot, I still have to go on.
00:12:18.960 --> 00:12:20.559
I have to make that decision.
00:12:20.720 --> 00:12:27.200
I have to go on to whether it's claw.ai or chat gpt.com or gemini.google.com or whatever.
00:12:27.440 --> 00:12:32.080
I have to proactively go in there and give information to a chatbot.
00:12:32.159 --> 00:12:38.000
I have to share the context, share my problem, share all these things with a super app as an example.
00:12:38.159 --> 00:12:39.919
You can have automations that run.
00:12:40.000 --> 00:12:41.600
And this is how I have it set up.
00:12:41.759 --> 00:12:53.120
So whether it's hourly, twice a day, twice a week, whatever you want it to do, you can have it go through with your custom instructions, every single important piece of your work life, right?
00:12:53.200 --> 00:12:57.759
And obviously, I need to put the giant, the giant grain of salt out there, right?
00:12:57.840 --> 00:13:08.000
Don't go and download uh, you know, codecs or you know, whatever the uh the Microsoft co-pilot version of their super app, I think it's gonna be powered by GitHub co-pilots.
00:13:08.159 --> 00:13:11.600
So they do have their version of that, but it's just not unified yet.
00:13:11.759 --> 00:13:12.399
Um, right.
00:13:12.559 --> 00:13:16.320
This isn't something where you're trying to use this without your company's permission.
00:13:16.399 --> 00:13:24.799
So if you are a business leader, if you're a decision maker at the enterprise level, you need to start saying, hey, is it time that we invest in a super app strategy?
00:13:24.960 --> 00:13:26.799
And the smart companies already are, right?
00:13:27.039 --> 00:13:34.080
Months ago, I was doing uh, you know, big trainings for companies that rolled this out to tens of thousands of employees.
00:13:34.159 --> 00:13:37.279
So if you think, oh, companies aren't doing it, yes, they are.
00:13:37.440 --> 00:13:38.639
Uh they definitely are.
00:13:38.799 --> 00:13:40.559
So if you are not, you are behind.
00:13:40.799 --> 00:13:44.000
But there's a lot of reasons why it beats your traditional chat bot.
00:13:44.080 --> 00:13:51.519
But I think the biggest couple are number one, um, it's always on and autonomous, right?
00:13:51.600 --> 00:13:54.720
So you can schedule uh these things to happen, right?
00:13:54.879 --> 00:14:07.120
I have ones that run uh about every four hours, I have runs that run every day, uh, I have different ones that run every single hour uh on codex, especially, because then it can share all of this context.
00:14:07.279 --> 00:14:19.120
So whether it's going through my email, uh you know, every hour, it's going through social media, every four hours, it's going through my my DMs, every 12 hours, and then it's just reporting back to me, and then I can make those decisions, right?
00:14:19.200 --> 00:14:26.159
So you can see right away how that's immensely more in uh helpful than a chat bot that is really just uh reactive.
00:14:26.240 --> 00:14:30.799
And I have to go in and the chat bot reacts to me, but I have to be the one being proactive.
00:14:30.960 --> 00:14:38.639
Where on the AI super app end, it is actually the AI super app that is the one being proactive in delivering me uh decisions, right?
00:14:38.720 --> 00:14:43.840
And saying, Hey, human, I followed my rules that you set out for me.
00:14:44.000 --> 00:14:46.799
Here's all the this the decisions that you need to make.
00:14:46.960 --> 00:14:53.360
I went ahead and built you this dashboard, I built you this deck, I built you uh, you know, all these things.
00:14:53.519 --> 00:14:56.639
You approve the one which is best, go through, give me your updates, etc.
00:14:56.879 --> 00:14:57.039
Right.
00:14:57.200 --> 00:15:05.519
So there's a lot of reasons why the AI uh app beats the uh you know the traditional chatbot.
00:15:05.679 --> 00:15:08.720
And the biggest thing, it's the interface, right?
00:15:08.879 --> 00:15:11.840
So let's talk about this kind of three-pane interface.
00:15:12.000 --> 00:15:16.879
So first, there's a lot of the you know, you have to tip your hats to the OGs, right?
00:15:17.039 --> 00:15:23.360
I think one of the companies that actually got the three-pane interface done really well early on was Google's notebook LM.
00:15:23.679 --> 00:15:40.000
Uh, obviously, Anthropic started the uh phase of the uh desktop agencorker uh with Claude Code, um, the desktop version in late 2025, and then with co-work in early 2026.
00:15:40.080 --> 00:15:44.559
But essentially, if you're not watching the video version of this, I want you to think of three panes, right?
00:15:44.799 --> 00:15:51.120
So on your left pane, that's your traditional chat and folder pane, but then you also have your automations.
00:15:51.360 --> 00:15:58.559
So that's where you have your different kinds of workspaces, that's where you organize, you know, all of your kind of agentic flows.
00:15:58.879 --> 00:16:02.000
In the middle, this is kind of your instruct panel.
00:16:02.159 --> 00:16:05.200
This is where you are actually chatting with the chat bot.
00:16:05.360 --> 00:16:10.320
So, right in traditional uh, you know, chat GPT web interfaces, this is the same.
00:16:10.399 --> 00:16:13.120
On the left side, you have all your chats, your folders, etc.
00:16:13.440 --> 00:16:18.399
And then on the panel to the right, that's your uh ability to chat, right?
00:16:18.559 --> 00:16:23.039
So the difference here in the big uh jump is the right panel.
00:16:23.200 --> 00:16:28.720
This new panel, uh, you know, you can call it the inspector panel, uh, whatever you want to say, the browser panel.
00:16:28.799 --> 00:16:37.360
But this is where you bring in that context um and you don't have to lose or degrade the agent's ability to bring it in.
00:16:37.440 --> 00:16:39.440
So, as an example, right?
00:16:39.840 --> 00:16:47.440
Um, the earlier days of Claude, um it had a very clunky browser use, uh, right.
00:16:47.600 --> 00:16:53.120
But now with today's version, with like Codex, it has a built-in browser there on the right side panel.
00:16:53.279 --> 00:17:06.799
So it can not only browse the web without having to, you know, connect a third-party tool or a second party tool technically by controlling a browser, but it has its own browser inside of the program.
00:17:06.960 --> 00:17:12.240
So it is much faster, it is much more accurate, it is much more token efficient, right?
00:17:12.319 --> 00:17:17.839
That's the big thing as we talk about token efficiency is going to be one of the biggest uh trends of the rest of 2026.
00:17:18.000 --> 00:17:25.359
To be able to bring in that inspector or browser is a huge piece of what makes a super app an actual super app, right?
00:17:25.440 --> 00:17:37.920
So not only bringing in the right side browser or the right pane or the inspect pane browser uh for the internet, but also for any files, uh, but also to preview anything that you're actually building.
00:17:38.079 --> 00:17:45.440
So at least for me, I've been trying to spend more and more of my time in that right side pane, right?
00:17:45.599 --> 00:17:53.440
So bringing in my browser sessions, so I can control that right side pane, or you know, codecs or the super app can control it as well.
00:17:53.599 --> 00:18:00.000
Um, I can bring in my files, my PDFs, my Word documents, uh, and then also preview different things I'm building there.
00:18:00.079 --> 00:18:01.920
But I can control that pain, right?
00:18:02.000 --> 00:18:05.680
Or Codecs can as well, or the super app of choice can as well.
00:18:05.839 --> 00:18:19.440
So you can see there by doing your work in an interface that the super app can see and work with you is obviously a huge advantage and one of the biggest steps uh toward the uh agentic work future.
00:18:21.359 --> 00:18:24.720
So let's quickly talk about the race right now.
00:18:25.440 --> 00:18:30.480
So I will say codecs and ChatGPT, they're ahead by a far margin.
00:18:30.640 --> 00:18:41.279
And I don't think anyone is going to dispute that because even uh Google literally copy and pasted codecs, right?
00:18:41.519 --> 00:18:50.480
Even in their launch video of anti-gravity 2.0 desktop version, they accidentally left something in the video, uh, the launch video that had a codex folder.
00:18:50.640 --> 00:18:53.039
It is a spitting image of codecs, right?
00:18:53.200 --> 00:18:58.160
Even with Microsoft's new GitHub Copilot, very much set up like Codex.
00:18:58.559 --> 00:18:58.960
All right.
00:18:59.039 --> 00:19:02.000
Uh cursor, very good uh in its own right.
00:19:02.160 --> 00:19:06.240
Claude Desktop, I I I would say it's a little far behind.
00:19:06.720 --> 00:19:10.720
And you know, you can even look at different benchmarks that talk about the harness.
00:19:10.880 --> 00:19:15.039
Uh, unfortunately, Anthropic has some of the best models in the world, right?
00:19:15.119 --> 00:19:16.720
Literally, the best models in the world.
00:19:16.880 --> 00:19:18.720
Their harness is bad.
00:19:19.279 --> 00:19:20.480
That's not my opinion.
00:19:20.559 --> 00:19:27.440
That's every single benchmark that looks at um that looks at uh harness, tool calling, etc.
00:19:27.680 --> 00:19:34.880
So yeah, the clawed models perform much better inside cursor as an example than they do inside of Clawed Desktop.
00:19:35.039 --> 00:19:35.839
Uh yeah, yeah.
00:19:35.920 --> 00:19:37.440
It's I don't know.
00:19:37.599 --> 00:19:40.559
I don't know if they design it to be token inefficient on purpose.
00:19:40.720 --> 00:19:41.359
I don't know.
00:19:41.440 --> 00:19:41.920
Uh right.
00:19:42.000 --> 00:19:44.160
And again, I'm not talking about the command line interface.
00:19:44.240 --> 00:19:47.359
Um, so I'm not talking about the CLI uh side, right?
00:19:47.440 --> 00:19:52.640
I'm talking about the uh, you know, everyday knowledge worker who's downloading these desktop programs.
00:19:52.720 --> 00:19:58.400
But I'd say right now, if I had to call who's where in the race, I'd say number one is codex.
00:19:58.559 --> 00:20:01.599
Uh, I'd say 2A is probably cursor.
00:20:01.759 --> 00:20:03.920
I'd say 2B is clawed desktop.
00:20:04.000 --> 00:20:07.359
Uh, and then I'd say the rest is like to be determined.
00:20:07.440 --> 00:20:07.920
Uh, right.
00:20:08.079 --> 00:20:13.599
Anti-gravity, uh, not not a big fan uh of that one.
00:20:13.759 --> 00:20:16.079
Microsoft will see where they land.
00:20:16.240 --> 00:20:30.400
Uh, I do think their new GitHub co-pilot app will ultimately be the kind of the underlying harness that will uh uh kind of power their super app that CEO Satya Nadella uh previewed at the Microsoft Build Conference.
00:20:30.559 --> 00:20:32.400
Uh and then you have others, right?
00:20:32.720 --> 00:20:34.240
Grok has their new build.
00:20:34.319 --> 00:20:36.720
That's still too early, kind of in the beta phase.
00:20:36.880 --> 00:20:50.799
But I mean, just in the past month, you know, and with what Kodak has coming up, we've seen four of the six big players say this is where we're focusing a lot of our engineering efforts from here on out.
00:20:50.880 --> 00:20:59.519
And I'm not saying that the model can no longer become a part of the moat, but I think there is going to get a certain point, right?
00:20:59.599 --> 00:21:04.799
I'm trying to uh come up with a good way to verbalize this and maybe I'll do a show on it in the future.
00:21:04.960 --> 00:21:08.799
Uh, you know, the way I think of it is like 95-5.
00:21:09.200 --> 00:21:23.759
Pretty sure we're gonna get to pretty soon, uh, I'd say probably early 2027, we're gonna get to the point where only 5% of people can actually take full advantage of 95% of what a model has to offer.
00:21:23.920 --> 00:21:33.839
I think there's gonna be a certain point where your your average, even your technical person, is not going to be able to take advantage of 95% of what a model can offer.
00:21:33.920 --> 00:21:50.160
So that's why I think more and more, at least when it comes to short and medium-term enterprise adoption, I think it's actually more about the user experience of using the harness uh than it is about the model, which I think is actually one reason why cursor is actually nice, nicely positioned in this.
00:21:50.720 --> 00:21:55.759
So let's quickly go over why I think Codec leads right now and I'll give you quick takes on all of them.
00:21:55.920 --> 00:22:01.519
I think Codecs is by far, they they've set the the Bar very high.
00:22:01.680 --> 00:22:04.799
Um, mainly just because everything works.
00:22:04.880 --> 00:22:09.680
And they are the only one out of the bunch that has every single thing that I talked about.
00:22:09.759 --> 00:22:10.240
Uh, right.
00:22:10.319 --> 00:22:12.799
So the three panes uh all work, right?
00:22:12.960 --> 00:22:27.200
Specifically on the inspect or browser, they have the code preview pane, they have the file preview pane, and then they have the um the browser pane all connected, which I think only two others have.
00:22:27.359 --> 00:22:32.160
So not everyone has all of those different um elements, which I think all of them are very important.
00:22:32.319 --> 00:22:35.920
But the other reason why I think codecs is very far ahead, uh, two reasons.
00:22:36.160 --> 00:22:49.279
One for the laptop crowd, you can literally have your laptop closed and codecs can still run everything as if your laptop was open and logged in, which I still don't understand the wizardry that OpenAI did to make that work.
00:22:49.440 --> 00:22:49.759
All right.
00:22:49.839 --> 00:22:53.680
But then the other thing is being able to access this from your phone.
00:22:53.920 --> 00:22:58.799
Because as you are on the go, um I, you know, this is weird.
00:22:58.960 --> 00:23:02.160
I've never been able to get so much work done from my phone.
00:23:02.319 --> 00:23:15.039
So I think that's two reasons why, aside from the obvious design and UX of Codecs, which everyone is copying, but the fact that the computer use, the browser use are just a step ahead of everyone else.
00:23:15.200 --> 00:23:20.160
Um, the actual harness itself, but the ability as well, your computer can be closed.
00:23:20.240 --> 00:23:23.119
Uh, you it can be on the locked screen, logged out.
00:23:23.359 --> 00:23:28.160
Um, and to be able to access everything from your phone is huge, right?
00:23:28.240 --> 00:23:32.720
So, as an example, we're gonna talk about um Anthropic's Claw Desktop here.
00:23:32.799 --> 00:23:39.839
It's like they have dispatch where you can kind of access some things from your phone, but it's super buggy.
00:23:39.920 --> 00:23:43.680
And you actually can't, you know, as an example, you can't access everything.
00:23:43.839 --> 00:23:51.200
You get one thread that's just a dispatch thread, which to me wasn't very helpful, but I think codex really leads.
00:23:51.359 --> 00:23:59.440
But I think cursor, yeah, cursor might actually be one of the best challengers, at least right now, in the super app game.
00:23:59.680 --> 00:24:11.279
So not only have they, you know, kind of been in the super app space a little bit longer than everyone else, um, but they've had the developer feedback.
00:24:11.440 --> 00:24:19.279
Uh, obviously, now they have this partnership uh with SpaceX, uh, which just had the largest IPO in history.
00:24:19.440 --> 00:24:29.200
Uh, so you can't count cursor out because they're gonna have more and more resources um to build uh this this their kind of super app, but also their own model as well.
00:24:29.359 --> 00:24:34.880
Their new composer, uh, I think it's composer 2.5 model is really good.
00:24:35.039 --> 00:24:47.200
Um, it is token efficient, and you know, you can kind of get about 90% of what you can get out of like an Opus 4.8 for like 10% of the cost, which is why one of the reasons I think you can't count cursor out.
00:24:47.279 --> 00:24:48.960
They've been in this space the longest.
00:24:49.039 --> 00:24:52.000
Uh, their user experience and the harness is really good.
00:24:52.160 --> 00:25:02.000
It's been uh shaped and loved by uh many developers for multiple years, and they have the one big advantage where you can use literally any model.
00:25:02.240 --> 00:25:04.720
This is the only one where you can use any models.
00:25:05.039 --> 00:25:13.200
You can use, you know, the GPT models, you can use the claw models, you can use Microsoft's models, you can use Google's models and their own models.
00:25:13.279 --> 00:25:16.960
So, cursor, I think you have to pay close attention to.
00:25:17.119 --> 00:25:20.799
Uh, next, Microsoft's could be right.
00:25:20.960 --> 00:25:21.680
We don't know.
00:25:21.839 --> 00:25:24.880
Microsoft has said their super app is coming.
00:25:25.039 --> 00:25:31.519
I know people laughed at me uh like six months ago when I'm said I'm actually kind of bullish on Microsoft, right?
00:25:31.680 --> 00:25:47.359
Uh they we've seen some pivots that in early 2026 that was kind of against their strategy of 2025, uh, which was creating more and more products, where it seems like they are going to this super app strategy and trying to streamline and unify everything.
00:25:47.519 --> 00:25:49.440
So we don't know what it's gonna look like.
00:25:49.599 --> 00:25:54.319
I do think that their recent GitHub Copilot desktop app was well well received.
00:25:54.480 --> 00:26:04.000
I think was not well received, was their uh kind of coupling that with the uh unsubsidizing uh of their plan, right?
00:26:04.160 --> 00:26:07.759
So yeah, we talked about that in uh the last episode.
00:26:07.839 --> 00:26:17.920
But so I think a lot of people had a sour taste from GitHub Copilot's pricing moves, but their actual new desktop app that they rolled out for Mac and Windows, it's really good.
00:26:18.000 --> 00:26:24.640
And it does look like that's gonna serve like the underlying baseline uh of the feat of the future Windows Copilot super app.
00:26:24.799 --> 00:26:33.039
So they did talk about this at their build conference and bringing also kind of the new element that they're gonna be bringing into this is their autopilots, right?
00:26:33.119 --> 00:26:37.599
So this is their their version of an autonomous coworker that works on your desktop, right?
00:26:37.680 --> 00:26:42.319
It can read, write, uh, you know, you can schedule it and it can access all your Microsoft 365 data.
00:26:42.400 --> 00:26:48.319
So even though Microsoft's version isn't out yet, I do put this fairly high up.
00:26:48.480 --> 00:26:53.680
All right, let's talk about Claude in Anthropic because they deserve a ton of credit here.
00:26:53.839 --> 00:27:08.000
Because I think without uh Anthropic's move uh in late 2025 and early 2026, specifically with co-work, I don't know if we would be in this era of um agenc AI and AI super apps.
00:27:08.160 --> 00:27:17.359
So uh I do think even though they've they've fallen behind maybe a little bit, um, Anthropic Claude, you have to give them their shine on this, right?
00:27:17.599 --> 00:27:30.240
Because even I remember working with the early versions of co-work, I'm like, even though this is slow, even though it's clunky, even though it fails sometimes, when it did get something right, co-work was mind-blowing.
00:27:30.400 --> 00:27:39.599
And I do actually think, which maybe this is because I do this every single day, I do think the enterprise now is just realizing, right?
00:27:39.680 --> 00:27:51.839
Because I think the enterprise is always about three months behind kind of cutting edge AI, which is why I still think that there's this momentum uh around Claude Code and Claude Co-Work and you know, Claude's desktop offerings.
00:27:52.079 --> 00:27:56.960
But at least when you compare it to what's out there right now, specifically with codecs, I do think they're behind.
00:27:57.039 --> 00:28:02.000
One of the reasons why, and I've done this on the show before, I showed this, it's not unified, right?
00:28:02.240 --> 00:28:05.519
Claude Code has no clue what Claude Cowork is doing.
00:28:05.599 --> 00:28:07.759
And on the desktop, there's essentially three different modes.
00:28:07.920 --> 00:28:15.279
You have your standard chat, you have your Claude Cowork, you have your Claude Code, and none of the others know that any of the others exist.
00:28:15.359 --> 00:28:23.440
So there's no unified memory, which is one of the biggest downsides of something like an autonomous AI super app.
00:28:23.519 --> 00:28:24.799
Like it has to know everything.
00:28:24.880 --> 00:28:27.200
That's one of the things I think Codex is great at.
00:28:27.440 --> 00:28:32.000
Codex can actually go in and start new codex chats on its own, right?
00:28:32.160 --> 00:28:33.680
Which is great, right?
00:28:33.759 --> 00:28:39.599
When it comes from an organizational standpoint, you can't do any of that with uh Anthropic's desktop offerings now.
00:28:39.759 --> 00:28:51.039
Uh, the easy, I guess, the the the bright side for uh Anthropic in what they should be doing, because this would even change my per perception on the current AI race.
00:28:51.200 --> 00:28:53.359
All they have to do is do what everyone else has done.
00:28:53.519 --> 00:29:03.039
Go copy OpenAI, go copy paste codex, and that turns what I think uh the rest of 2026 and 2027 looks like.
00:29:03.200 --> 00:29:15.519
Um, if they don't, I do think open AI has the ability to really run away, uh, especially with the however many weeks or months or quarters we have until Anthropic and OpenAI both go public.
00:29:15.599 --> 00:29:29.279
I think OpenAI is gonna be way further ahead than most people are giving them credit for, unless Claude essentially uh or Anthropic copies uh co-work, which or sorry, copies codex, which I don't know why they haven't.
00:29:29.440 --> 00:29:34.640
Uh everyone else has done it because it's very popular, it's the best thing out, right?
00:29:34.799 --> 00:29:36.240
Uh, so yeah, we'll see.
00:29:36.400 --> 00:29:38.480
We'll see what Anthropic decides to do.
00:29:38.720 --> 00:29:40.880
Next, Google Anti-Gravity.
00:29:41.039 --> 00:29:52.400
Uh, to me, again, aside from being a uh a I don't want to say anything uh bad about the team there, uh, because I do have good relationships with a lot of the people at Google.
00:29:52.799 --> 00:29:56.319
I don't think they should have released their anti-gravity 2.0.
00:29:56.480 --> 00:29:58.880
Uh it's it just wasn't ready.
00:29:58.960 --> 00:30:01.680
Uh it's it's there's hardly no features.
00:30:02.000 --> 00:30:04.880
Um I don't know.
00:30:05.039 --> 00:30:07.440
They rolled this out uh right after their I.O.
00:30:07.599 --> 00:30:08.160
conference.
00:30:08.240 --> 00:30:13.119
Uh you know, they rolled it out in conjunction with Gemini 3.5 Flash, which I thought was a good model.
00:30:13.200 --> 00:30:20.799
So they thought that, hey, the combination of this fast coding uh, you know, model, Gemini 3.5 flash, that can do this long autonomous work.
00:30:20.960 --> 00:30:25.839
Let's couple it with Google Anti-Gravity, and you know, it's gonna be this great splash.
00:30:25.920 --> 00:30:34.160
I I don't think it played out like that, mainly because I think Google kind of changed the flash model from what people expected versus what they were making it.
00:30:34.240 --> 00:30:49.039
Because I think the way people looked at Gemini 3.5 Flash was it's a cheap, fast model, and it's no longer cheap, it's still very fast, but it is expensive relatively compared to every other uh quote unquote flash model that Google has ever released.
00:30:49.200 --> 00:31:00.240
So, in that aspect, by tying it together with the anti-gravity 2.0 with the Gemini 3.5 flash, maybe hurt adoption of anti-gravity, but that and just the features aren't really there.
00:31:00.400 --> 00:31:02.160
Yes, it's an earlier version.
00:31:02.400 --> 00:31:05.119
Yes, they have basic automations.
00:31:05.519 --> 00:31:08.640
Uh, you know, you can use a couple other models, right?
00:31:08.720 --> 00:31:11.440
You can use the Claude models um in there.
00:31:12.640 --> 00:31:14.000
Not for me, right?
00:31:14.079 --> 00:31:15.519
Uh, just not enough features.
00:31:15.599 --> 00:31:20.559
It is like the most bare boned uh kind of super app you can think of, right?
00:31:20.799 --> 00:31:28.079
Where I think Codecs is robust, you know, even cursor and uh Claude Desktop, much more robust.
00:31:28.720 --> 00:31:32.880
Anti-gravity, uh, it's not super clunky, it's just there's like no features.
00:31:33.039 --> 00:31:36.880
So maybe it's great if you want to learn what a super app can do.
00:31:36.960 --> 00:31:45.599
And if your organization is heavily bulked into the Google ecosystem ecosystem, maybe it's a good place to start and get your kind of super app feet wet.
00:31:45.759 --> 00:31:54.400
But if you have the choice to use the others, I would not personally use Google Anti-Gravity 2.0, although I do hope that they continue to improve it in the future.
00:31:54.559 --> 00:32:00.319
And I think they will because, like I said, they did give it a lot of spotlight uh at their Google I.O.
00:32:00.480 --> 00:32:00.880
conference.
00:32:01.279 --> 00:32:06.000
All right, so let's begin to wrap this up and because we have to talk about the other side.
00:32:06.240 --> 00:32:10.079
I talked about all the reasons why, hey, super app is the next trend.
00:32:10.160 --> 00:32:14.000
Uh, it is the next wave of work, it is the execution layer, uh, right?
00:32:14.079 --> 00:32:30.160
It is the harness because I think most of the big companies know that the the the kind of the benefits or the advantages uh over new models is going to become less and less as a smaller and smaller uh section of the enterprise will even be able to squeeze out the juice, right?
00:32:30.319 --> 00:32:36.799
Let's just say, as an example, let's say Fable 5.1 or you know GPT 5.7.
00:32:37.200 --> 00:32:45.359
What percentage of people, what percentage of people honestly are going to be able to take true advantage of those point step releases?
00:32:45.440 --> 00:32:46.480
And I think not a lot.
00:32:46.640 --> 00:32:59.039
So I that's why I think the uh emphasis in the moat in the short and medium term is actually going to be on the user experience and giving people the harness to power all of this.
00:32:59.200 --> 00:33:02.799
But with that comes great responsibility, right?
00:33:02.960 --> 00:33:05.119
Essentially, let's talk about YOLO mode.
00:33:05.279 --> 00:33:11.680
Uh essentially, I can set up and you can set up most of these super apps to quite literally do anything.
00:33:11.759 --> 00:33:48.079
And when you talk about the ability to read and write any file, any folder, to use the terminal, you can accidentally, if you're not paying attention, to use to have browser use, to have computer use, if you don't have the right uh role based access controls, if you don't have the right user permission set up, if you don't have redundancies and backups, these AI super apps, if you're whether the person is trying to be a little bit too aggressive and automating too much of their work, whether someone's being a little lazy and not really paying attention, whoops, you can accidentally do a lot of harm.
00:33:48.240 --> 00:33:49.920
You can accidentally delete something.
00:33:50.000 --> 00:34:01.839
And if you don't have uh something properly permissioned, if you don't have the backups uh in place, yeah, you can the the the the the AI super apps, there's actually the risk is in the action, right?
00:34:02.079 --> 00:34:06.880
The power is actually the downs, the downside and the downfall as well, right?
00:34:06.960 --> 00:34:16.880
I did a training for a large uh organization uh a couple of months ago that was rolling out kind of a desktop super app to thousands of users.
00:34:17.039 --> 00:34:20.079
And they were asking me, hey, how should we be using computer use?
00:34:20.159 --> 00:34:21.920
And I said, Don't, right?
00:34:22.159 --> 00:34:22.639
Don't.
00:34:22.800 --> 00:34:28.239
You know, you shouldn't, you like some of the more powerful things you can go and change settings on the computer.
00:34:28.400 --> 00:34:46.159
If you're not um strict enough in your context engineering or prompting process, sometimes these systems, especially if you're using uh inferior model, they're gonna look at the end goal and they're gonna build a completely new path that might be a path that's detrimental to your group.
00:34:46.400 --> 00:35:07.039
So, you know, if you're not paying enough attention to learning, and if your company, your department, you personally, if you haven't put in the proper um time to even understand the agentic natures of these models by default, I wouldn't recommend, you know, letting these things loose and then having hands off, right?
00:35:07.280 --> 00:35:10.719
Human in the loop plus an AI super app is not a good combination.
00:35:10.800 --> 00:35:13.199
That's why I preach expert-driven loops.
00:35:13.280 --> 00:35:19.519
So if you are doing true human in the loop, that's where you give it a prompt, and oh, I'm gonna have it go run for 48 hours.
00:35:19.679 --> 00:35:27.280
Let's go token maxing, let's use computer use, browser use, everything, and just have it solve all my all my issues for the quarter, right?
00:35:27.360 --> 00:35:29.679
And just go truly hands-off YOLO mode.
00:35:29.840 --> 00:35:35.920
If you don't know what you're doing, if you don't know what that agent is capable of, things can go down a sour path.
00:35:36.000 --> 00:35:38.960
If you just go truly, oh, I'm the human in the loop, let me set this up.
00:35:39.039 --> 00:35:41.199
No, you have to have an expert-driven loop.
00:35:41.280 --> 00:35:52.880
You have to be very hands-on in learning and understanding these capabilities, reading the chain of thought, iterating, uh, sandboxing, you know, sandboxing uh your browser use at first until you fully understand it.
00:35:52.960 --> 00:36:04.559
Because it can, you can log into any SaaS application that you and your company use, even if there's not an API, even if there's not an MCP or a direct integration or a connector in these super apps, right?
00:36:04.719 --> 00:36:09.760
You can go log in and well, the super app can control whatever browser you're using.
00:36:09.920 --> 00:36:16.239
So the risk is immense if you are not properly learning uh and keeping up with all of the developments.
00:36:16.400 --> 00:36:18.000
So let's wrap here.
00:36:18.159 --> 00:36:19.519
Here's where I think you should start.
00:36:19.679 --> 00:36:20.880
Haha, start here.
00:36:21.039 --> 00:36:23.199
Uh, read only first, right?
00:36:23.440 --> 00:36:24.400
Add approvals.
00:36:24.480 --> 00:36:26.239
You need to be sandboxing these things.
00:36:26.320 --> 00:36:33.920
Uh instead of rolling these out to hundreds or thousands of employees, because there is going to be, I think, the the urge to do this, right?
00:36:34.239 --> 00:36:39.519
Whether it's in a couple of days or a couple of weeks when OpenAI unveils and opens up their super app.
00:36:39.679 --> 00:36:45.360
Same thing with Microsoft, as we undoubtedly see improvements from Anthropic and from Google as well.
00:36:45.519 --> 00:36:47.519
I think most people are going to want to rush in, right?
00:36:47.599 --> 00:36:49.920
They're like, oh my gosh, look at all these new capabilities.
00:36:50.079 --> 00:36:55.920
Don't read only, strict permissions, guardrails, training, learning and development.
00:36:56.079 --> 00:37:07.199
You have to be educating because I think this is where when I talked about agentic crash, uh, you know, in our 2026 um AI prediction and roadmap series, this is where agentic crash happens.
00:37:07.360 --> 00:37:17.199
When you have companies that give too much access, not enough training, uh, you know, too much power without proper responsibility and training.
00:37:17.360 --> 00:37:22.400
So you have to automate the low-hanging fruit with low risk.
00:37:22.559 --> 00:37:25.199
We've talked about how you can tear off the different risk.
00:37:25.280 --> 00:37:26.320
So that's where you start.
00:37:26.400 --> 00:37:36.559
Don't start with the highest capabilities, with the most handoff and autonomous um actions, because I know that's the that's the that's the sexy thing, right?
00:37:36.639 --> 00:37:42.880
You're like, oh my gosh, I can automate 90% of my work, have this thing work for 48 hours, and I can come back.
00:37:42.960 --> 00:37:44.719
And it did two weeks of deliverables for me.
00:37:44.880 --> 00:37:46.000
Oh my gosh, let's do it.
00:37:46.159 --> 00:37:47.280
Let's do this at scale.
00:37:47.360 --> 00:37:48.559
That's not where you start.
00:37:48.800 --> 00:37:49.280
All right.
00:37:49.440 --> 00:37:54.079
You have to be able to crawl before you walk, walk before you run.
00:37:54.159 --> 00:37:55.599
So don't start this with a sprint.
00:37:55.679 --> 00:37:56.960
You will get burned.
00:37:57.119 --> 00:37:57.519
All right.
00:37:57.679 --> 00:38:08.320
That is a wrap previewing the AI super apps and why every company is racing to create one, what they are, and hopefully how your company can properly take advantage of this.
00:38:08.480 --> 00:38:12.159
So if this was helpful, make sure you go to startheseries.com.
00:38:12.239 --> 00:38:23.599
That's going to give you access to our everyday AI inner circle community, a playlist with every single Start Here Series episode on there, and you can go connect and network with other business leaders who are trying to do the same.
00:38:23.679 --> 00:38:24.719
So thank you for tuning in.
00:38:24.880 --> 00:38:27.679
Hope to see you back tomorrow in every day for more everyday AI.
00:38:27.760 --> 00:38:28.800
Thanks, y'all.
