What OpenAI Dev Day 2026 Means For AI Builders

OpenAI Dev Day 2026. What should AI builders look forward to from the OpenAI ecosystem? Today I have a guest and co-host of the pod, Hamel Husain, who has been in the AI industry as a pioneer for 20 years plus on ML before it was cool at GitHub, Airbnb, and early LM research doing work in that space. Hamel has worked on open source ML tools, debugging systems. He teaches the most famous AI evals course, thousands of students served and hundreds of companies, including OpenAI. And today I have him on the pod and we're going to go into technical deep dives for the entirety of the pod. And we're probably going to have a few episodes that spawn from this. I'm with Hamel at his home and we decided to just content max this whole session that we're together.

I'm really excited about the Codex, uh, ultra fast. So the Codex has fast mode, which bills you at 1.5x your token spend quota. Then they have now Ultrafast. I don't know how much Ultrafast. Okay.

Yeah, let's look at this together. So OpenAI developers on Twitter, Ultrafast is our fastest way to build with Astra yet in Codex. It runs up to 8x faster than Astra Standard and 4x faster than Astra Fast. Bring your ideas to life as fast as you could type them.

Now that is fast.

That's fast.

8x. So this is 4x faster than fast mode. You're gonna notice it. Does it say how it bills you for that? Like how much does that cost?

I don't know, but let's real quick look at the side-by-side.

Oh, sweet.

Standard versus UltraFast. Oh.

Simon Willison, they're benchmarking the hell out of it.

Look at that UltraFast, how many tokens per second you get. Wow. Okay. Let's see what some of the community says about UltraFast. So, okay, wait. EncodeX UltraFast generates code faster as you can move more quickly from idea to code to iteration. In the API, combine it with WebSocket mode in the responses API to build real-time experiences with Frontier Intelligence. I actually, real quick, from a product lens for any of the AI product managers out there, Think about the user experiences in your product experience where it's like voice mode or taking some text, and then you need that to be very responsive. I find that especially in cases of high degrees of emotionality or subjectivity, like health coaches, health agents, the person is anxious on the other line of the agent to get a response. There are ways you could think about this from a user experience perspective, but at the end of the day, they're looking for an answer. that is the value prop to them as quickly as possible. An answer that shows that the agent understood their case. I am, you know, dealing with bladder health and I'm a spinal cord injury patient. We're working on a use case with that with Clarity. And so how do I best know whether I have a UTI or not? And I want the agent to minimize the amount of time it takes to give me a highly efficacious answer. I could kind of see this working well, especially with some of the voice capabilities that OpenAI is coming out with. Like this WebSocket mode responses API, you get way, way more, quote, real time, which I know is sort of a thing in building software is you should always question the need for real time sometimes.

No, you know what, like with coding agents, even if you're just developing a web application, I don't know about you, but when I'm using my coding agent, if it's taking a while, I'll just go browse Twitter.

Yeah.

And then, you know, humans are really bad at multitasking. So I think actually it's good for everybody.

That this stuff goes faster?

Yeah, just speed is actually super helpful, even if you don't need like a real-time thing. The thing, I guess like one situation where I don't care is like batch processing of data or something like that, where they truly just—

It's like jobs.

It's like happening in the background.

Okay.

And I don't care if I'm like in the zone trying to make something, then yeah.

Okay. So for the coding agent use case where it's this back and forth, you're saying speed is part of the developer experience that is a really valuable pillar. Do you think we're gonna start to see this as a competitive edge from like the Frontier Labs and—

It already is a competitive edge. So like, this is one of the reasons I like Codex because their fast mode is included within your subscription. So like Claude Code fast mode, they have fast mode as well. That's not included in your Claude Code subscription. You have to pay extra for it.

Wait, wait, wait, for real? So it is super confusing as an AI builder to be like, what am I getting out of the $200 OpenAI versus $200 Claude Code? license. So you're saying the speed is not packaged with Claude Code, but it is packaged with Codex?

Yeah. If you open up like your settings in Claude and I can look at it too, just to confirm what it says. So if, yeah, if I open up Claude and I go to settings, go to usage. So you'll have like, you'll have a different quota for Fable than you do. Like, so their Frontier Intelligence will have a completely separate—

Run out of my tokens for the day, but yeah. So go ahead. What do you mean?

So there's so many different carves out— carve-outs in the Claude Code subscription. There's a, there's a separate carve-out for Fable, which is like, hey, they have a completely separate quota. So like once you hit the Fable limit, then all of a sudden you can hit a Fable limit without hitting the other limits. And it's super confusing. And then also I believe the same thing for real, for the fast mode, they have fast mode and that you can only pay for with credits. You can't pay for out of your subscription. Whereas Codex is a lot simpler to understand. It's just, Everything comes out of your quota. So like whether you want to use frontier models, you want to use Astra at the highest level of thinking, or you want to use fast mode, it just all comes out of this subscription. It'll just burn your subscription faster. And that is so much nicer because then you can decide what you want to do instead of like paying Anthropic like some money because you want, you clicked on something.

It's more opaque for sure. Like I'm looking at this speed up responses with fast mode, get faster Opus responses in Claude Code by toggling fast mode. Fast mode is a high-speed config for Cloud Opus, making the model up to 2.5% or 2.5x faster at a higher cost per token. So yeah, that's what they're saying. It will cost more with fast mode. And that is outside of the plans for Cloud Code. That's what Hamel's saying. So from an AI builder perspective, it's— I've always found this is true as well, being an AI builder myself, building on the front lines, is cloud cost and usage and billing is more opaque. That is definitely true. And I am seeing moves from OpenAI. Being a lot more developer-centric here, especially as we're seeing speed in using AI as a thinking partner, as when you're in the zone, as Hamill says, you're really in the zone and you want to go ship. And so not having to think about, oh, I'm paying $200 a month, my weekly session limits are this, my monthly session limits are this, and then I have to throw in another vector to assess like speed as part of my developer experience. It does get a little confusing. So Definitely gotta give the nod there to Codex, ChatGPT, and the OpenAI team for this. So let's read some of what the Twitter comments are saying. How can I— can we sign up? I tried signing up earlier. I tried signing up for the Codex 500, but it— I'm

gonna try signing up right now.

Let's try sign up right now.

Okay, so I'm gonna go to settings. I'm gonna go to usage and billing. This is in the desktop app, Codex. I'm gonna go to view plans. That didn't quite work. Let me try again. Settings.

Oh, it's available. I could upgrade.

No, just click the button though. Click the button. See if it works.

Okay.

'Cause it was erroring out for me before.

For all the AI builders out there, I'm testing out upgrading to maximum power, $500 a month for Codex. Let's try it out. Upgrade. What happened?

Oh, yo.

Okay. It's

gonna do it.

All right, let's go. Wait, I might get the upgrade. Oh my gosh. Please upgrade. I need more tokens. Actually, I gotta do work, so. Uh, as, as the audience could see, I've been like, oh dude, I'm in it.

I can't even get an upgrade.

GPT Pro 500. Let's go, dude. We're here. We're out here. We got ChatGPT Pro 500 on this date, only a few hours after the announcement of ChatGPT Pro 500. So breaking news for AI builders, OpenAI at Demo Day announced ChatGPT Pro 500 plan with 25x usage capacity and access to ultra fast mode. No 5-hour limit as well. I just successfully upgraded to ChatGPT Pro 500 a few hours after the announcement. They seem to be slow rolling the deployment of this so that you could buy and opt in. They also sort of cheekily decreased the value of their current $200 offer by half. So you get half the amount of usage. Before it was 20x the Plus plan. And now it's only 10x the Plus plan. I think that they're really leaning into, hey, push you to use the Pro 500 plan, sell your enterprise into the business plan for Teams. While we're here on the topic, actually, somebody made with Codex a little OpenAI Dev Day 2026, what's shipping? Should we go over that real quick, Hamel?

Yeah, let's do it.

Number 1, DOTS, Always-On AI Agents. This is a competitor product to Muse and GrokBot finding product-market fit with personalized agents. where you ship off a task and it just gets work done across your apps. You could run it on your computer or your cloud, keep working over time with continuous threads. They also released GPT-6.1 Sol. We just talked about that a little bit, but that's a new flagship model for agentic coding, computer use, and professional work. Ship higher quality stuff, better coding benchmarks at cheaper rates. Ultrafast, we just spoke about that too. This is a high throughput model for real-time and high volume workloads. So We just showed that previously on the pod, but go take a look. Up to 8x faster token generation. And as Hamel said, 20+ years in the business of building AI and ML on the front lines, he was saying when he's in the zone, really looking for that fast token generation output so that AI is a thinking partner with you. You go back and forth, you try something, you get feedback, you think of other things. Having that experience be part of a core part of the developer experience is really key. Private intelligence, number 4, business-grade privacy stack for sensitive data and regulated use cases. This from a product lens and strategic lens to me means that they're moving further and further market into large enterprise, and that has been their core play all year against Claude and Anthropic. Number 5, Codex, your coding companion now in the cloud. So this is moving workloads into the cloud. We saw a lot of success in the Meta landscape for AI. With Groq and Cursor moving workloads to the cloud, I really like that meta. Instead of having to like always keep my device on or whatever, like I've been on a trip and then my Mac mini at home dies and then I have to cycle it and I need to push some buttons and be there to do it. I can't get it live. So Codex moving a lot of their workloads to the cloud and optionality in the cloud is such a great thing for developers. Number 6, API and developer primitives. They have a decisions API, which is in preview, reasoning layer for classification, routing, and choosing an agent's next action. I can see that being very useful and probably a competitive product to the System 1 Jev that came out for classification and routing downstream for agent usage and application layer usage. Agents API with computer use. So I've been really impressed by Codex's computer use. I've had it just like type slacks. I know you can use the API for that, but just anything like I'll have it use a virtual machine. If I don't want to go into a virtual machine for one of my clients and I have it like write updates to Teams or whatever and DM people. So good. Bedrock managed agents so that they're furthering their investment there. A few other things to rattle off and then I'll get Hamel's take. ChatGPT has a shared workspace. So ChatGPT Space, this is like multiplayer shared home with files, agents, and team context. Looks like they're going into Company Brain a little bit in that territory. Pages, turn ideas into structured documents with AI, looking like they're going into Notion's territory there and making it more of a corporate collaborative environment. Collaborative slides, really interesting that this has now become a commodity. I remember back in the day, slides were how we created narrative and story and got influence and buy-in and shared strategy and metrics and stuff. Now it's just a commodity that they just release. Teams and shared tasks, self-explanatory. ChatGPT everywhere. So now it's in Slack. Bring it to Teams, have a meetings plugin. Oh, that's actually low-key kind of crazy. So there's all these tools out there, Otter AI, Fireflies AI, Granola AI, Quill. I'm using a bunch right now recording the pod, and they just decided we're gonna go in the meetings game and keep all that first-party data in our ecosystem, which I think is a smart move. And it does seem like that's very buildable now, especially with their investments into voice tech. Shareable profiles. It's another one. Sign in with ChatGPT. Hamel will have much to say on this. I'll ask him in a second. But this is use your ChatGPT identity, similar to sign in with Google, sign in with Microsoft, sign in with Facebook, sign in with Apple. But in this case, it's bring your own tokens. So the idea being an application developer or an AI builder could build an AI or LLM-based app and allow ChatGPT sign in for their identity provision. And also allow the user to bring their own tokens to bill. That sort of abstracts away the need for like pricing policy and usage rates and stuff like that and pricing model innovation or updates on the application layer. Pro 500 plan, I just upgraded. Y'all saw that. That's sick. I'm going to have to run some workloads soon. OpenAI Marketplace. So they're going into the demand and supply game of their ecosystem, trying to create more value into their ecosystem by opening up. Their ecosystem for more building. Hamel, what are the top 3 things that draw your attention when it comes to OpenAI's Dev Day here? I'm curious.

Yeah, I mean, I think the signing in with ChatGPT seems like the most exciting thing. I did check into it and it says it's only open to limited partners. And if you try to go to the page, it says like, you need to contact your OpenAI representative. I don't even know what that— I don't have one of those.

I need one.

Yeah, I don't have one. I probably will never have one. But in theory, that's the most exciting feature because if you're building like an application, that's one of the problems you have to continuously sort of contend with is how do you figure out your token pricing? You know, you need to meter it, you need to make sure users don't abuse it, blah, blah, blah. And then you have to factor that into the pricing of your software is like the Pricing of intelligence. And so it's just so nice sometimes to be able to let the user bring their own intelligence to your software. So, um, it would be cool if that was rolled out more generally. I'm really excited to try, like a lot of this stuff is not available, like Dot. We could not get access to it yet, so it's really hard to know, but I'm excited to try it.

Oh, let's show, uh, on Dots, the intro to Dots by Sam Altman. And then Simon says, Simon says it does look very Muse-like. You get to pick its name, then it's represented by cute blob-like avatar. There's a demo video, a whole lot of voice interaction. So this is what DOTS sort of looks like. Very, you know, like GroqBot, like Muse. DOTS are powered by Astra, and they gave DOT an ambitious use case in this demo, migrating an API. Get our app off this inventory API before it shuts down. So this is presumably connected to your GitHub environment, and it goes and it traces what's needed. Mm-hmm. And then it just comes back with work done. The code is updated, tests are passing, I'm prepping the pull request. And here's the pull request on the right. Now, whether it worked or not, but it is interesting that they're saying, hey, this is going to be your personal assistant and it can also code. So it's sort of like Muse, but more developer focus is how I'm reading this. How are you reading this from your AI builder perspective, Hamel?

No, I like it. Like, you know, I've had fun playing with all of these things, like GroqBot, and some of the others. You know, the fact that's included in your subscription just makes it pretty compelling.

Yeah.

Especially for you since you just upgraded to $500 plan.

I just did that.

Yeah. You're going to need potentially some excuses. Burn tokens.

Yes, exactly. I'll need, I'll need to use all their stuff to really get value out of that $500 a month plan. Here's a live demo. So you not only ship PRs with Dot, you could also be like, hey, what time is my lunch party? And then Dot can act as an executive assistant or a chief of staff. I've been hearing this in other podcasts and in the AI native landscape, which is Muse really took the world by storm and gave people useful AI versus let me go get use out of it. Here is just use and value out of the box. You don't have to prompt, be a prompt god or prompt engineering god. You don't have to know what skills to use. You don't have to know what plugins. It just does work. And I think that shift in UX to this thing just does work for you is such a valuable one as well, speaking to broader sort of macro view, that it'll get the American public where we live bought into AI will help you and it'll make your life easier, which I think is so key in the landscape, especially as we see leaders like Dario and stuff. basically saying AI is going to kill us all. So it's nice that this stuff is coming out and it shows people, wait, it's valuable to you today. Like, you don't have to invest in stocks and AI companies and get value later. It's like, hey, this thing could be a chief of staff for you today. Wouldn't that be sick? Like, somebody that can't afford a chief of staff at work now has a chief of staff at home and for their entire life. What do you think of that use case? Like, what do you— maybe following on that, the chief of staff personal assistant use case, tying to improving public perception, 'cause it's something useful right now. Like, what time's my lunch party? What time's my kid's soccer camp? Like, those types of things are everyday people.

Yeah, I think it's everyday people. To be honest with you, it's not super exciting to me because I've been already doing these things with Codex. You can do, you can, you can do a lot of this with Codex now. Granted, it's really nice to have this nice clean interface. Um, like it's simple. Sometimes less is more. So like, go to this app, it's like super dead simple. You have all this stuff out of the way. The Codex app, you can already do a lot of these things. You can have scheduled tasks, you can have like children, child tasks, you know, all kinds of stuff. And I already do a lot of this stuff through that. So I'm like, am I going to switch?

Not so much for you.

Maybe not. Maybe I'll switch some things.

Maybe some—

it is, you know, especially if they can all talk to each other, like can Dot create Codex tasks? Okay. Then, then I'll be able to do it. One thing I just saw that I think might be interesting, this was also announced today, the Agents API with computer use.

All right. Computer use. Let an agent complete tasks in an OpenAI-hosted browser.

Yeah. And the thing that I think is really key about this is, okay, if you scroll down, so like if the task needs an account, it handles sign-in. So there's a way, you know how what's like really annoying about cloud tasks is like if it needs to log into something or it needs to do something, it's kind of a pain in the ass.

It's so annoying. Yeah.

And that's like one of the reasons I haven't used cloud that much, but now I'm starting to see this UX where You have the cloud thing running, the moment it needs you to log in, it will push something to you locally and have you do the login, like maybe on your phone, maybe on your desktop, and then it will figure out how to send those credentials to the cloud. And this API allows you to do that. Like it has a handle sign-in. You want to click the handle sign-in, you can see like how it does it. It's like, okay, tasks such as reading issues in a private GitHub repo require an authenticated browser. Your application handles sign-in. So users can choose a login method, enter credentials outside the chat. So sign-in can involve several requests. You know, it might ask the user for an email, enter an email address, then enter a verification code, et cetera. So that, you know, like, yeah, this is a way for you to like, I mean, I think they're already doing this in their cloud. I haven't tested it, but I've seen some videos of it. This allows you to do the same thing.

Yeah, that's really cool. I mean, I like the idea. So for AI builders out there that are trying to, for example, code on the go. Or for me recently, I went on a road trip and I had like a Starlink on and my laptop on in the back, and I just kept those on. And I had my phone, I was being very safe, had my phone and I had like Claude Code and Codex just remote into that. And I set up autonomous goal loops before I would start driving for hours on the road trip, and then I could just nudge them along through my mobile phone. But I've even heard of, you know, some of my founder friends, they'll like set up their laptop at home or something like that, and they, they go to the gym and they remote in. So the, the cool part about this is when it runs in the cloud, your MacBook doesn't have to be alive and your computer doesn't have to be alive. Your Mac mini doesn't have to be alive. And what I like about this feature from a developer experience or AI builder perspective is the OpenAI hosted environment. That just means that, hey, when I gotta log in, when I, my agent has to log into something of yours, the humans, like credentials or whatever. There's a very elegant way now to handle that UX on the go. That's how I view this. And then there's that continuity because it's in the cloud, exists way more value in continuity and waiting for the agent to finish things. Or now, you know, I would imagine it could talk to other agent runs in the cloud too and share information similar to the Codex UX right now. So that's a really cool thing. And another meta for AI builders where if you have, for example, heavy manual tasks in your workflows for your software, the software you build, like the product that you build, evolve, maintain. One of the metas that I'm seeing is, hey, can we make it easier on the user where they use an agent interface and on the backend we use computer use so that the user doesn't have to go and click through the app anymore, especially if it's an enterprise app or software which has a lot of legacy code where refactoring the UI is not so simple and the APIs are not quite there. And there's a lot of coupling between the front end and the back end. I could see Computer Use being a really good meta for those cases to reduce the time it takes to configure your software. For example, if you're building enterprise software, that's a bit legacy. Hamel, from your AI builder perspective and for other AI builders out there, any other implications on this Computer Use technology being more open through their APIs?

No, I think it's great. I use computer use. I mean, the reason why I'm so stuck on Codex is computer use.

What are the top use cases you use computer use for with Codex?

There's a lot of things out there in the world that don't have APIs. You know, you have to click on something and click on some buttons and do something, whatever. And computer use is like basically how you can have AI act like a human. You know, maybe 10 years from now you won't need computer use because We'll just collapse all the user interface to be all agent-friendly.

Yeah.

And the idea of computer use doesn't have to be a thing, but for now, yeah, it's just the world is very jagged in that you need to like click a button and scroll and do all this nonsense.

I love how you say it. We're such in a bubble where we're like, all this is, this clicking stuff is for the past humans and that I'm not doing that anymore. If you talk to somebody That's not in SF or like New York. I mean, I know you're in Portland, but you're very connected to SF. These people are like, how do I use AI? And that's where Muse and like Dots, I think, are meeting those users, or Grokbot.

Potentially. I mean, I would say if you look at the statistics, right, like most people are not even paying for AI, which blows my mind because if you're not paying for AI, your experience of AI is completely different.

Yes, totally.

You might think it's not that good. You're like, uh, what is it? Like, I tried this thing, it didn't work. Well, yeah, you're not using like anything modern per the standards of AI. So yeah, I think you're right. Like people's perception of AI is totally different than what—

Than ours.

Than our experience.

We have a very different experience. Yeah. We're like, use computer use because I can't be fucked to go into that anymore. They don't have an API for this. Go reverse engineer the API. And it's like, Your typical mom, like taking their kids to like soccer practice is not thinking about this. They're like, can you just, can somebody just help me save time on literally anything? 'Cause my day's so chaotic. And I think that's where like apps like Muse and Grokbot and Dots are trying to compete for that attention and that behavior. Yeah. Where it's like struggling moment from a product perspective, jobs to be done, like perspective, struggling moment. Oh my gosh, I got this, all this chaos to organize around my family life and the home life and my work life and trying to progress my career. How do I make it easier? And so now we have these like very easy chief of staff agents.

I just gave you a link. You might want to open that. That seems like a really good comprehensive link. Everything that was released.

Oh, cool.

That's where I found out about some of these things that we've been talking about, like the, like the API for computer use.

Oh, very cool.

You know, of course they have, yeah, they have some of these other things. You know, they have like a new Slack thing. Now the new Slack thing, you have to have a business account.

Oh, Slack. Oh, ChatGPT.

I don't have like business.

Oh, like the Claude tag competitor.

It's a Claude tag competitor.

Yeah, I see. So @ChatGPT, we're making it easier to work with ChatGPT where team conversations happen. Mention @ChatGPT in a channel thread or DM in Slack or Teams. It could use tools connected by your admin and your own tools for permissions. So you gotta be business and enterprise plans. This is really cool. Backing up a little bit, we just nerded out a lot about computer use. I actually think it would be super useful if I just shared One example of computer use. So this is Codex. This is a loop that I have Codex run on. It goes to my waitlist for our open source Clarity Harness that we are working on. What it does is anytime the waitlist has an update, well, this runs on a loop, so every few minutes it runs this loop. It will go to my waitlist and it'll look up the latest person that added themselves to the waitlist. And check on the leads and then look for their personal email, look for their company role, their LinkedIn, and then add that to my CRM, my customer relationship management tool of choice. So right now I'm using HubSpot. Then it creates unsent Gmail drafts, and then I could go ahead and send those Gmail drafts to these leads that are really warm right now because they just signed up to my waitlist. It also goes in and actually sends a message. To the LinkedIn lead that I— here's one. Saw you joined the Clarity waitlist. Stoked to connect. What problems are you currently dealing with that resonate with Clarity? So right when they hit my waitlist, I have this run every few minutes and it checks the latest waitlist subscriber or signup, and it goes and finds their LinkedIn. If it's successful with that, it sends a connection request and it also uses—

What model are you using for this recurring thing that happens every minute?

Oh, let's take a look.

Don't tell me you're using like Astra High.

I'm using Soul Medium.

Okay.

Yes. Is that okay?

I don't know. I mean, it's good enough. If you're using Soul Medium, that probably suggests like, because it just came out, their new models.

I should probably switch to 6.1 Lite, I would imagine.

Mm-hmm. I don't know. I don't know what like the legacy Soul pricing is. You might wanna check that cuz you might, it might actually even be cheaper than—

oh, I see.

The new thing sometimes is cheaper than the old thing.

True, true. So you could see here, I have a bunch of these. Uh, this is just always going. It uses computer use because I'm not using LinkedIn APIs or anything fancy like that. And then it—

I don't even know if there is LinkedIn APIs.

I have no idea. I imagine they're pretty tight, tight knit.

I don't think there is. It's like, yeah, it's like kind of a lockdown service. These are, this is an example. This is a classic example where you have to have computer use.

Because there's no API or easy to get to API. I think you have to be like a social media aggregator app.

You can reverse engineer like. The web requests. However, you don't want to necessarily use that approach cuz you want your AI when it's like staging a message on LinkedIn to make sure it's formatted correctly, et cetera, you know, with the right spacing.

Mm-hmm.

And I found that when I try to like overdo it, sometimes it can't do what it needs. Whereas like, cuz some, some applications are just not made for agents. Yeah. Like LinkedIn is not made for agents.

No, it was made in the Human all the time.

I mean, it seems like it's made for agents because every reply you get on the site is an agent.

True. Yeah, yeah, definitely. Actually, this is a really interesting use case. So you mentioned when the API is not readily available, you'll like reverse engineer it, ask Codex to do that. Let's give an example to AI builders on the podcast. So if I were to reverse engineer it, what would you say to Codex? Is this right? So

Yeah, I will use my skill. So let me send you a link to my skill.

Oh, send, yeah, send me the skill. And for folks out there, uh, Hamel, if you're an AI builder and you want to reverse engineer an app or website's APIs so that your coding agent can use it, listen up. The reverse engineering site skill, a reusable agent skill for discovering and validating website internal APIs through instrumented browser traffic and direct HTTP replay. So I'm going to live just go to my Codex and say, use this skill to help me reverse engineer LinkedIn APIs so that we—

Okay. So I wouldn't try to like just say LinkedIn APIs, because that's the kind of unbounded problem. Like say like what you want to do, like what actions do you want to take on LinkedIn? And you wanted to say like reverse engineer the Requests required to perform those actions or something.

You've been helping me with and reverse engineer how to do this programmatically instead of with computer use like we have been doing. All right. So I'll just fire this off and I'm using SOL Medium. Use this skill to help me find the LinkedIn connection from my warm lead, resend signup. So that's speaking to my Loop and then send them one of my DMs and reverse engineer how to do this programmatically. Okay, cool.

This is a, see, this is a situation where you want fast.

This is, yeah.

Maybe you should turn on, maybe you got like all kinds of, uh, now I have

all sorts of stuff, so let's, let's

try it out. Don't do, don't do ultra. No, no, don't do that. Uh, maybe SOL. You got what kind of fast you got? Let's see. Click that. Yeah. Okay. Do this. Maybe like tell it to stop and do continue.

Continue now that I'm on 6.1 and faster mode.

I would just say continue, but hey, you're being nice.

I'm being nice to the agent. Okay, so let's continue with the reverse eng site skill. By the way, drop a comment in whatever you're viewing this on. Just say HAML if you're interested in this reverse engineering site skill. We'll send you a link in DMs to that.

Nice.

So we're seeing. It's working through. I have read its capture and replay steps. I'll identify the connected waitlist lead, send one DM, and verify the API request by reading the message back.

There you go. That's, that's what it's supposed to do. It's supposed to take the action as a human would.

Yeah.

And then

sniff,

observe the, you know, request response pattern and say like, oh, okay. Like this request response pattern is like this thing. Okay. I'll just programmatically like create a script that can do that. Of course, it like might need your cookies or whatever. And so the skill, like, you know, has instructions around all that. It looks like it's trying to do something that's getting blocked. Just maybe it's something else though. It's a Chrome extension that you have. I think that's causing a problem, but it shouldn't be a big deal.

Yeah. So it's actually just going through, like Hamel said, as a human would, and it's trying to take these actions and then it's looking at the console and it's trying to reverse it. That's really cool. Clearly I have some rogue extension that is messing up all my console logs.

I got some kind of extension that It's causing a lot of problems, man. Let me give you a tip.

Yeah.

Press—

Oh, Command, Command.

Command, Command. There you go.

Command, Command.

App shot. Do an app shot.

There we go.

There we go.

So real quick for the AI builders out there, if you find yourself living like a cave person doing these screenshots and then dragging it over, here's what you could do instead is an app shot. So click Command and Command on your keyboard, and what you'll get is one of these guys where it actually takes the whole app. window, not just the snippet that you screenshotted. So in this case, it would've scrolled through, it'll have the context, this particular app shot, but the context of everything in this window, which is really cool.

This is one of my favorite things to do is to drive Codex with Codex. So like now you're gonna have it create a task, separate task.

Yeah, I'll steer it as well. That's another thing that I really like about Codex versus Cloud Codez actually is a queued prompt can be injected right into the work. in parallel to then steer the work agentically. So in that case, if I just type something, I could click steer. I'm gonna throw that away, but that's exactly—

I use the side chats as well quite often. Oh yeah. I find it to be like—

What do you use side chats for?

Oh, it's like, oh, what's going on here? Why are you doing this?

Like a by the way.

I don't want to interrupt it, but I wanna be like, why did you do this? Or what does that mean? Or whatever. Side chat.

Got it. Quick, quick little side chat.

If the side chat becomes something worthy of a main chat, I'll say, hey, like Create a new task that does this thing that we discussed.

Yeah, that's really cool. So while we're waiting for this, uh, Hamel, I'm sort of curious, and I bet other listeners of the podcast, people that are trying to become more AI-pilled and use agents for making their life better, their work life or their life life, what are you doing with Codex where you said, hey, DOTS sounds cool, but I'm already doing a lot of this in Codex? You mentioned like managing your life. Agentically. Can you talk through some use cases?

Yeah, there's a lot of things like students in my course, they have to like apply into Discord and like supply their email. And that email has to match like their Maven email. And sometimes there's confusion around that. So Codex will like take a look at the application. It will see if there's a miss— if there's a match, it'll just accept the application. If there's a mismatch, it will look at it. It will try to make a guess as to what is going on. It will stage an email. with all the information to the student and it'll say, hey, like I saw that you tried to do XYZ, you might be doing it wrong. Please like try this, you know, try to fix it.

I see. Okay.

And, or sometimes like students will have administrative issues, like, you know, they'll be trying to upgrade something and they'll be really confused. And so Codex will just like go research their enrollment history, how much they paid, It'll go look at my Stripe, it'll look at Maven, it'll look at my email history with them. It'll compile it because I, you know, it'll compile everything that happened with that student and then help me draft an email like, hey, this is your status in the class. Like, this is what's going on, blah, blah. And like, give the student kind of all the information and give me confidence as well. Got it.

And how much time, how much time does that save you right there? That one workflow? Like you couldn't really do that in math, right?

I didn't want to do that. Like, yeah, actually like my EA was out. I'm like, let me just do it with AI.

Yeah.

It's like, uh, so I just like had my AI and I have it just run hourly to like do that stuff.

Okay.

Um,

okay. What about like, okay, so that's managing your business, which is your evals course, right, on Maven. By the way, shout out to Hamel. I learned from him and then I took those skills and I was able to make money for that. So if you want to make money being an AI builder, I would take Hamel's course. And it is one of the best courses I've taken by far. Hamel's not paying me to say this. I just always sing his praises because he has such a great course. Anyway, outside of the course and your business in AI education, Hamel, what are some personal life cases? Because you have a family, you have young kids, you got to stay coordinated with your partner. Like anything around like the personal—

Oh yeah. Anytime I need to teach my kids something, like I need to teach my little one how to read A clock, analog clock.

Okay.

And so I made a little app for that.

You just vibe coded an analog clock?

Yeah, just vibe coded it with like challenges.

Oh, cool.

You know, like, okay, this is the time. What does it mean? Okay. Like, you know, just make it good for learning. She does like, my older one does Montessori and they have interesting ways of learning math, which is very tactile. It can be, but it can be very confusing for her because she's learned math the traditional way first. And so like, again, like more educational apps. I created an app that, like a web application that shows her why seasons, like how seasons work, like the tilt of the earth and the sun and how the sun is like hitting the earth and how it interacts also with rotation. Um, that was like really hard for her to understand, but like you can visualize it. with like motion graphics, do it in high fidelity. And then now I have, I can teach, you know, her things. So like, yeah, my, one of my favorite use cases is education. My wife uses it for education too. She, she makes songs for the children.

Okay.

To explain topics. So like, yeah, one of her favorite things to do with AI is like help with enrichment.

Oh, okay.

Of education.

Yeah. Yeah. Yeah. Okay. So it sounds like personalized learning that is interactive is a core use case in your household.

Yeah, because I have 2 kids.

Yeah. How old are your kids?

5 and 8.

Okay. Anybody else that has a 5-year-old or an 8-year-old or anywhere around there, and if you're trying to teach them things and integrate AI learning to bolster their own schoolwork and learning, that's a great use case that Hamill just named, which is, hey, for things like teaching them how to read an analog clock or mathematics from an interactive perspective that's more practical.

Yeah, check out my wife's site. My wife does a lot of this. It's called Wonderwise. Science. And, uh, yeah, this is her site. So she's been making these songs. I think I need to—

Let's listen to one. Okay. Wait, there's a pod. Okay. Yeah,

it's a pod.

Okay. So let's, uh, let's try one.

She narrates the beginning and then she creates a song.

When the pressure builds up, the magma bursts out of the top as lava, along with ash and hot gases. As the lava cools, it hardens into new rock. Over many eruptions, these layers of rock build up and form the shape of a volcano. Some volcanoes erupt often, while others stay quiet for hundreds of years. Most volcanoes are found where the edges of Earth's huge plates meet, places where the ground is moving and changing. So volcanoes are a part of how our planet builds and reshapes itself. Next time you see a picture of one, remember, it's the Earth's way of letting its magma rise and lava flow. Next is a song that I wrote about this.

Enjoy! Under the ground where it's hot and deep, the Earth is rumbling in its sleep. Bubbles of magma start to grow, up they come. And up they go. Deep inside where rocks can melt, the earth makes magma all by itself. Pressure builds, it needs to go through cracks.

Yo, this is a banger. This is so good. It's really good. So good.

She puts a lot of time into these songs. It's not just like slop created. It's like she iterates on it like from like a lot for like sometimes for days until she likes it.

I could tell.

And the reason why this was created is we were just tired of listening to bullshit in the car, like Cocomelon or whatever that kids listen to.

Yeah.

Complete garbage.

Yeah. Wait, I love this use case, which is get people off of garbage childhood education content and just make your own because you're the parent and now you can use English to make a little app or make a podcast or make music that's educational in nature. That's— and our kids love

it. They're like, oh, can you play this song? Can you do this thing? And it's so good.

That's so good. That's so good. I love that. Okay, folks, check out Wonderwise if you're interested. Kids educational music. That's by Elifaya. That's Himmel's partner. And that is such a cool use case, which is, hey, I am a parent and I would like to help enrich my kids. And maybe they go to school for a certain kind of education. And I want to augment their education and learning in the household, and I want them to not doom scroll. So let's, let's do this instead. So use AI, Codex, Cloud Code, Cursor, whatever, to create this kind of solution, like help personalize their education, help personalize their enrichment. In Hamill's case, they're making science songs for curious kids. That's from Alafaya. Educational podcast that answers big questions through narration and catchy original songs. You just heard that banger. Which is How Our Volcanoes Formed, episode 2. So go check that out. Really cool stuff. And, and yeah, actually I see like she has like a bunch of 5-star ratings on here on the pod, so she's doing great.

Yeah. Let me show you another example. I'll share my screen real quick. Okay. So I mean, this one, it was just like, hey, how do you like understand the seasons? So I just had to iterate on it based upon my daughter's specific confusion about The relationship between the sun and the earth. And it's just like a little app that you, where you can like have the sun go around and like kind of shows you, like for example, added this view from the sun here.

Yeah. That's—

So it could really like explain to her like, okay, this is why the Southern Hemisphere is like getting warmer. 'Cause it's like tilted towards the sun. But you know, it's like, and then, you know, what, what it looks like when it's equal sunlight on both halves and the fact that it's, but it's still tilted, but it's not tilted like away from the sun. It's just tilted like on his axis, um, whatever. And then I had this like 3D version, which is kind of fun. Um, and let me just hide that thing. Let me see if I can. Yeah. So it was just like having it rotate so you could get like a, you know, like, okay, like what does it look like?

Yeah.

You know, that's all it is.

That's really cool.

And it's just like, it's not like I'm, these are like meant to replace anything. It's just a, this is a visual aid for me to have a discussion.

Yeah. And I like that, that what you said, which was it helps to speak to the particular confusion point that your daughter had.

Yes.

Like you don't get that level of personalized learning or enrichment unless you probably pay out the ass for a teacher to like one-on-one personalized education.

There's so many little apps like that, little things like this where it's like, oh, let's explain this concept. You know what, we can do a much better job with a very good visual aid.

For sure.

It's the same idea with annotate. You know, I always tell people annotate, uh, your data with your own custom annotation app. This is like, why not? You can have customized stuff that helps you do your work.

Yeah, totally. Okay, awesome. We, we kind of went on a tangent a little bit. We did get it going where direct requests now work and Codex was able to reverse engineer The connection and the reply. I still need to do a little bit more to harden this because it was, we were sharing screen and I was taking over the Brave browser to share screen. So I got to look at this more. There's a lot here. Let me share again the Dev Day recap. This is end on this note because we're already deep into OpenAI's new releases for their Dev Day 2026. So just announcing they have DOTS. Remarkable, capable, always-on agents like we talked about. They have GPT-6.1 SOL, major upgrade to GPT-6 SOL, exceptionally strong performance, delivers near-astro intelligence to everybody at a fifth of the cost. AI builders want to look at that. Ultra fast. We talked about this key part of the development experience, and you saw that in the Codex demo I just gave. We switched it to faster because this is one of those things where it's like, I want something live on the pod to show. So obviously go fast. They have a rate of 300 tokens per second. Up to 6x faster in the API, 8x faster in Codex. Ultrafast Astra is available today. Private intelligence. We talked about this a little bit, but this is for enterprise offerings and businesses using Frontier AI. Zero data retention with private safety processing. Very cool. New tools for developers in Codex and the API. So Codex moves to the cloud. That's really cool. Run Codex wherever you need it, computer, remotely from phone, or in the cloud. From any device. Reusable dev environments help tasks start quickly with your creds and your secrets. Shared setups are available. Refreshed Codex CLI. So the Codex CLI now lets you start and steer tasks with your voice. This is cool. We didn't cover this before. The new /agents view makes it easier to delegate work and track multiple tasks at once. Also improved everyday flows to help you edit prompts, resume sessions using work trees. So here's the inspect, edit, and run code from your terminal by Codex CLI. Here's a little demo of this. You could work against your local repo, stay in control, compose with scripts and CI, and you could talk to it with your voice. And we all know the GPT voice models are really, really good. So I'm really excited to check this out. Just rattling off the rest of this, code review has gotten better code review experience. You could read summaries, explore diffs, ask Codex about issues. Codex has a security cloud now, giving defenders a better set of tools to harden their infra. Decisions API. So I kind of see this as a response to the Jeff System 1 model coming out and shaking the AI world, focusing Luna's intelligence on a specific set of user-defined questions with finite predefined answers. Agents API now has computer use. Bedrock managed agents powered by OpenAI. So now they have this on AWS. Plugins. So more customization in ChatGPT with plugins. And here's one example with Canva. You can access this directly in Codex now, and this moves them into more marketplace ecosystem territory. Improved plugin creation, submission, and discovery. So allowing any builder to create plugins and distribute onto the Codex ecosystem. Sites can now host plugins. So you can add supported ChatGPT plugins to the sites you build. That's really cool. And I could see that compounding in a big way and start to compete with people like Base44, Where, hey, I want to build an app and have all the plugins, all the integrations already figured out. I don't got to custom bake those from scratch every time. MCP events for plugin automations. So adding support for proposed MCP events spec so plugins could start automations when something happens in a connected app. Actually, yes, if it connects to LinkedIn and I get a connection or connects to Resend and I get a waitlist signup, then it no longer has to be on a loop. And that makes it a little bit more programmatic and deterministic. That's really cool. Improving human and people work together, collaboration together. ChatGPT Space is your new home for teams to collaborate with AI to get work done. Create a dedicated space for teammates, ChatGPT, and your dot to build on shared knowledge. Pages moving into Notion territory, collaborative slides, like we said earlier, moving into Google Slides, Gamma, PowerPoint territory to share knowledge and synthesize artifacts with each other. Create teams, share tasks. Bring people together to do all the above, and then a Claude tag competitor with tag ChatGPT or @ChatGPT in your Microsoft Teams or your Slack instance. Meetings plugin is another big one that I think is a little slept on, and it's like a footnote of this release, but I'm kind of curious about how the valuations of some of these AI note-takers or recorders are going to go after seeing the Frontier Labs move into that space. And lastly, shareable profiles. Letting people showcase sites and plugins on your profile, kind of like GitHub profile is how I'm reading this. A few more. Sorry, I said lastly, sign in with ChatGPT. We talked about this earlier, but here's an example. Connect ChatGPT and Devin, and that allows you to use your ChatGPT plan, allowing Devin or the consuming application to consume usage directly from your ChatGPT plan. I think that's such a boon for the developer experience of building apps to connect to AI models, uh, in a particular—

there's quite a few things here. There's like, okay, there's Devin, Notion, Vercel, T3, OpenClaw, Dactyl. I don't know what Dactyl is. So that's like a lot of things like that makes me think that they might make it more widely available.

I think it will be. I think this just follows what I see as like an early access play with design partners for a customer advisory board to get value first, make it repeatable, and then move downmarket and allow, you know, people with smaller consumer plans to do this as well, or your indie builder sort of persona. So it's saying identity is available globally, plan usage is available to Plus and Pro users and participating tools. Let's look at that.

It's really interesting, right? Like, okay, you can use it with Cognition's Devin. So what does that mean? Does that mean you should go from the $200 Cognition plan down to $20 and then ramp up your ChatGPT subscription just like you did? And then just does it eat these other things?

Yeah. Does it eat these other things? I don't know that it immediately eats these other things, but it creates a pathway for OpenAI to eat these other things.

Okay. So like in the Devin case, why wouldn't it eat it? Like, cause you don't need the, you can just maybe use a cheaper Devin tier if you can all of a sudden now

use your tokens. Yeah. I think it's going to cause consumers of this ChatGPT sign-in to rethink the way that their pricing is. And I just suspect that this allows somebody that's building harnesses for coding agents, like that's what Devin is, it's just like a harness on top of these models. They don't build their own model to what I'm aware of. So they have to, by definition, compete at the user experience layer because they're not competing at the efficacy.

I think it might actually force people to not charge by token.

Yeah. Like rethink pricing, right? Because you can't,

You're just gonna buy your electricity somewhere else.

Yeah, exactly.

And you're gonna come, you have to, you're gonna, okay, pay for the appliance, whatever.

Yeah.

But no, not, so that's interesting.

I think OpenAI is positioning to be an electrical bill provider. I think that's what they're doing here. I think this is one little step and I bet 6 months from now we redo this podcast and I suspect that we're going to see more and more apps using continue with ChatGPT or sign in with ChatGPT. And that creates an open space for them. To encroach on these application kind of territories. I believe it's going to force harnesses like Devin or AMP or whatever. In this case, what is it like?

I'm curious. Let's see if we can find— let's see what the Devin folks are saying about this. Maybe there's some commentary.

This is cool. You can even manage limits. So it is like electricity or it's like bandwidth on the internet. You could centrally manage your limits.

Okay. So pull up this tweet, the cognition, and let's see what they have to say about this.

All right. So we're just recapping this. And the very last thing is the sign in with ChatGPT. And we're kind of curious, how does this fit in with players like Devin? So Devin Cognition says their Twitter handle, starting today, you can now use your ChatGPT subscription directly in Devin. You sign into Devin with your ChatGPT Plus or Pro plan. All OpenAI model usage in Devin will draw down from your quota. So AI builders using Devin, listen up. This is the blog post to sign in with ChatGPT. Starting today, you can sign in if you have an eligible ChatGPT Plus or Pro plan. You sign in, it draws down your tokens from ChatGPT, and it looks like this. You'd use your ChatGPT plan and then put up PRs. Devin Cloud. OpenAI models are part of the model mix Devin Cloud uses in Lite, Normal, Ultra, and Fusion modes. So once you connect your ChatGPT account, your plan pays for the OpenAI model share of each session while other models draw down from your Devin also takes your connected planet into account when choosing models, leaning more on models in the mix so that more of each session runs on the usage you already have. They're particularly strong at computer use testing agents. I definitely may use them in browser. Interesting. And it says here, as a result, Astra spends fewer tokens using Fusion, which you can think of Astra as the lead and V2 as the sidekick. The lead plan review and most of the implementation work to the free sidekick. So in this case, your task and your check-in plan will last longer. The artificial analysis coding agent index shows that using Fusion with Astra and SWE2 is 39% less expensive than running a session with Astra alone. What do you think of this? Do you think OpenAI is creating openings to eat up players like DevOps? Like, I don't see how Cloud Code doesn't. Slap back with, hey, you can also do this now.

Well, I mean, okay, so Cognition did this willingly and they're like, they're promoting it. They're like, go do this.

Before we end the pod, let's kind of speculate on this because it's an interesting thread. What happens now that OpenAI and Claude, Anthropic will probably position as the electricity? What do you think is going to happen?

I think it's really good because this like idea of like paying for tokens from someone else just kind of sucks. Like, I don't want someone to do Like to have a token markup is kind of like a whole mess that developers have to deal with. Like if you're creating software and you want to offer intelligence as part of your product, it's just, it's just a pain in the butt to say like, okay, I'm gonna do token pricing plus plus or something like that. Or do like a flat, flat fee subscription and just give me enough headroom so I don't have to worry about like the long tail. Of like some guy like going crazy with the intelligence or whatever. You just have to worry about that, right? 'Cause that's like a variable cost. And you know, the variable costs, like they may not, may not make big people paranoid, but they definitely make indie hackers and small people and entrepreneurs like super paranoid when they're creating software. They're like, oh, okay. They have to, it just adds like tremendous complexity to meter all of these things. And to keep track of all of these things, prevent abuse, blah, blah, blah, whatever. And so like the fact that you can bring your own tokens, I think that really simplifies things. That's a big accelerant to people building tools, 'cause now you can just build a tool and you don't have to worry about that. Yeah. You know, like that's, that was preventing a lot of people from building tools, honestly. 'Cause you had to like build all this infrastructure to handle the, to dole out the intelligence in a way that wouldn't, that couldn't be abused and that wouldn't like, impact you financially. So I think it's like amazing.

Good for the— so in this world, the analogy is OpenAI is your electric— electrical provider. You buy electricity and you could use that in different appliances like dishwashers or, you know, your microwave or whatever. And Devin becomes an appliance in sort of your developer environment household, so to speak, in this analogy. So you're saying it's better for OpenAI, better for the consumer like me, and it's better for Devins? Like appliance makers, is that right?

I think it's good. Now, there's some people that would fight back and say, well, you don't want to get locked in. And I don't really see why it needs to be locked in if more and more people offer this kind of thing where you can sign in with OpenAI or sign in with Anthropic and sign in with whoever, sign in with Fireworks, whatever it is, right? Like, um, but I think this is a really cool way of leading the way. So yeah, I think it's really good. I think it's really good for developers, like really, really good. More people can now develop applications that have intelligence in them.

Yeah. And I think that brings more competition not to figuring out markup pricing models versus let's use all that capacity of experimentation cycles and our brain as a team to make the best appliance, like for this particular use case. I could see that argument. I could also see the argument on the indie dev side though. Of like that, Devin's not an indie dev by any means. They're fucking multi-billion dollar unicorn, but like for any indie dev, it could also be like, well, I probably know that they're not just going to come eat my lunch money. And I think the answer there is like, you don't really know.

You never really know. But like, you know, so this, why won't Google build it? Why won't Amazon build it? That's, that's been, that's been an objection for like years. And if you are afraid that Amazon's going to go build it, Google's going to build it, just stop building. Like, because what are you going to do? Like, you should just keep building, build good stuff, you know? Don't get caught up in like, is Amazon going to build this? OpenAI going to build it? Just build it.

Totally. And a piece of advice out there for all builders, if you haven't worked in corporate, then you don't know how slow they can move, especially in times like this where executives up the chain that are 12 layers removed from the actual engineering shipping with the coding agents. They're like thinking faster, better, cheaper, stronger, you know, there's a lot of churn that a company has to go through. Like, look at Workday. They were recently bought out for $2 billion private equity to take off the public markets so that they could actually slow down to speed up later. Like, there's a lot of—

Yeah, Lenny had a quote the other day, somebody interviewed, I think he said like, if you want To compete with a big player, all you have to do is find the product intersection of 2 VPs. Because 2 VPs in a lot of these companies are not going to collaborate with each other. They're competing with each other. So

all you

got to do is, okay, where's the, where's that product intersection?

And that's your wedge or beachhead.

Yeah. Because that, they're not going to do that. They're not going to collaborate.

Yeah. So for the AI builders out there, the smaller ones, don't look at this as Hey, OpenAI is now going to eat my lunch money. Look at it as, hey, I just, I don't have to solve for price markup or innovate on pricing models or markup tokens to other people. I'm not a token reseller. I could focus on the niche problem that is economically valuable for my consumer, no matter what. So, okay. With that said, we covered a lot of ground today on the podcast. We covered a ton of built-in meta around agent swarms. Using your brain and hands for planner worker sort of metas. Hamel gave us lots of good examples of using the different reasoning modes, keeping your life simple. He also gave us really cool meta for how to reverse engineer a website when it's not giving you the API access that you need and you're too lazy like I am, and everybody will be, and you just want to use a coding agent and you don't want to go to another UX/UI interface and click a bunch of buttons. Hamal has a really cool use case that I thought around his own family personal life, which is, I want to enrich my kids' education. They're 5 and 8, and how do I do that and augment their existing education and meet them where they need to be met? Kind of like thinking about it like a super highly personalized AI tutor or the best human educator you could possibly have dedicated to your child. So if you learn to use some of these tools, you could use Codex to Create your own podcast or create your own whatever sort of medium that gets your child where you need them for their own personalized education. I think that's such a cool use case. And of course, we went over the litany of things that OpenAI has dropped at their recent Dev Day and all the demos there. And the last thing we spoke about was the implication that OpenAI is becoming more of the electrical provider and providers like Devin are becoming More of the appliance provider. So something that you want to marinate on and think about as an AI builder. Catch us for more episodes. Okay. So the next episode that we're going to record is going to be a lot of fun. We're going to go into more technical deep dives. I think we're going to do interviews on AI, AI evals, how to objectively and metrically improve your agents. I'm very excited for that. And hopefully also do some— Hamel, for joining the pod. And for everybody listening, hit subscribe wherever you're at, Spotify, Apple Podcasts, YouTube, and follow Hamel. I'll have links to all of his stuff and his content in the show notes. And please take his course if you are an AI builder trying to increase your economic value with your skills in this AI landscape, which is changing all the time. But the thing that isn't going to change is, can you objectively and measurably improve your LLM-based products? That's not going to change. So these skills are A bunch of hard-earned secrets. You could take his class and learn just like I did. So anyway, thanks, Emil, for being an awesome guest.

Thanks for having me. Thank you.

Have a great day. Take care, AI builders. Bye-bye.