A (6:45)

Yes, Paul. So OpenAI launched GPT 5.5 this past week. They call it a quote, new class of intelligence for real work, empowering agents built to understand complex goals, use tools, check its work, and carry more tasks through to completion. It is OpenAI's first fully retrained base model since GPT 4.5 and the first API model from the company to ship with a 1 million token context windows. So pricing comes in at $5 per 1 million input tokens, 30 bucks per 1 million output tokens that roughly double GPT 5.4. There's a GPT 5.5 Pro variant at $30 per 1 million input and $180 per 1 million output. On a bunch of benchmarks, GPT5.5 took the top spot on the Artificial Analysis Intelligence Index. It had a score of 60, which is three points ahead of Claudopus 4.7 and Gemini 3.1 Pro Preview. It leads the Browse Comp benchmark at 90.1%, Frontier Math tier 1 through 3 at 52.4% and it also posts an 84.9% on their GDP VAL benchmark, which is measuring how AI is good, how good it is at doing real work. Sam Altman framed this release as saying, hey, we believe in iterative deployment. Although GPT 5.5 is already a smart model, we expect rapid improvements. There were a couple people also reported after having early access some of the results they were getting. So Aaron Levy, we talked about a bunch. CEO of Box said the model saw a 10 percentage point jump in accuracy on their most complex knowledge work evals, the lovable team. The vibe coding tool lovable. They reported a 23% reduction in tool calls per request. I called it the most capable model for people taking on complex builds with technical depth. So Paul, a lot of stuff we can kind of unpack here. Just kind of curious about your broader thoughts here. I mean just again another new model, but there is a big emphasis. OpenAI stated just outright about agentic coding, computer use, knowledge work and early scientific research. They said those were areas where these gains of the model were especially strong. And I don't know if you could more succinctly put like a series of trends of like exactly where AI seems to be going.

B (9:08)

We've talked a lot recently about OpenAI refocusing. You know, they, you know, cutting the Sora app. They're thinking about robotics but not heavily invested in it quite yet. They dropped the idea of having like a social network. So they're doing their best to try and refocus. I think in large part due to the success of Claude. You know, if we go back to the start of the year, not only did Claude all of A sudden start getting a lot of headlines and a lot of attention for the quality of its work. Not only in coding though, but in, in knowledge work like, and we talked about it so much on this show, Mike, of the ways we've been using Claude and it just seems to have been post trained really well to do knowledge work, to do strategy documents and research papers and, and, and so OpenAI has been watching anthropic making gains and seeing their revenue skyrocketing and then a lot of it's coming from their work with enterprises. And I'll share a little bit more about, you know, my last couple weeks but you know, I was at the Google Next event this week and every person I talked to was using Claude. I mean they have co pilot licenses, they have Gemini licenses, but I didn't talk to anybody that wasn't at least experimenting with Claude as well. And in some of the cases I was talking to massive like Fortune 50 enterprise leaders in some cases who are in charge of AI within their organizations and they're giving people cloud access on top of everything else. So like OpenAI seeing this, they're hearing this, it's why they're, they're having to not only like do all these deals with the consulting firms, but they have to focus on the real work. And so when you read the post that they put out about this release, it's very obvious, as you said, like where they're going. So it said we're releasing GPT 5.5, our smartest and most intuitive to use model yet, and the next step toward a new way of getting work done on a computer. GPT 5.5 understands what you're trying to do faster and can carry more of the work itself. It excels at writing and debugging code, researching online, analyzing data, creating documents and spreadsheets, operating software and moving across tools until a task is finished. Instead of carefully managing every step, you give GPT 5.5amessy multi part task and trust it to plan, use tools, check its work, navigate through ambiguity and keep going. Now we're going to talk a lot about agents on this episode, but this is the kind of stuff people have been using codecs and Claude code for and things like that in Gemini. But what they're saying is the average knowledge worker wasn't seeing those same capabilities. You had to be a developer, you had to be a technical person to get those capabilities. Which is what we've been stressing on the show is that these like Claude cowork, open call, these things they're great for developers. Like, you have to be technically minded. We're trying to talk to the people who are outside of that world who are trying to just go in and build an agent and then they get into like an anti gravity, like, what the hell do I do with this? Like, it's not intuitive. So where OpenAI is obviously going here is moving in that direction of bringing those coding capabilities in a more reliable, secure way right into the platform that the average knowledge worker would use. So they continued, they said the gains are especially strong in agent decoding, computer use, knowledge work, and early scientific research. Because the model is better at understanding intent, it can move more naturally through the full loop of knowledge work, finding information, understanding what matters, using tools, checking the output, and turning raw material into something useful. And then just some quick context here. Mike. I listened to this core Memory podcast with Ashley Vance, which I think it's a new podcast, and if I'm not mistaken, it was a gated podcast, like you couldn't get it. And then someone had proposed, like, well, why don't you raise money or something and make it open? And someone paid $100,000 to unlock this podcast. And so just this episode, so it was with Sam Altman and Greg Brockman. So Ashley Vance sat down with the two of them, and I think it was the first time they've ever actually done an interview together. So on my flight back from Vegas on Wednesday, I listened to this, and I'll just highlight a couple things because this came out as a prelude to 5.5, but Sam and Greg were obviously talking about some of the things they were doing. So Sam talked a lot about the tech, but said they haven't connected the dots enough on what the abundant future will look like. I thought this was fascinating because an episode or two ago I was saying how there was a PR problem in the industry and how they were all talking about this abundance and yet no one understood what that meant. So I was fascinated to hear Sam basically echo exactly what I was saying. And he was like, we're not doing a good enough job as an industry making it tangible for people, what this amazing future is that we're envisioning. He also said, they're not far away from a model that knows the complete, complex complexity and context of your life. This is the memory component. And I think this is a really important thing for people to understand. And so when you're using 5.5, they're obviously starting to rely more on memory, but they're also relying more on the fact that the Memory is just going to get better. And so when you have models like 5.5 and eventually 6 that have full context through memory, and they also, like, are able to continually learn, which I'll talk a little bit more about in a minute, the need for prompting in the ways we've become adept at prompting goes out the window. Like, you don't need to do context and interview me, and all these things that have become standard ways of prompting because it knows everything already. And so prompting literally just becomes, hey, do that report for me that, you know, I have to do on Sunday nights. And it's like, okay. And it goes and does the thing. And then Greg, along those lines, talked about personal AGI, which is the first time I think I've heard him talk about it in this terms. So what they're saying is, rather than like a universal AGI as this model, and then, you know, the next generation models come out, it starts to know you so well that it feels like general intelligence to you because it does have this full context and memory and ability to learn from what you're doing. And so in that vein, they talked a lot about this idea of still this jagged intelligence that we still are on this age where sometimes these things feel superhuman. And then like, it gets hung up on a stupid thing and you're like, oh, it's no smarter than a preschooler when it comes to this thing, but it's superhuman at this other stuff. And then they just really talked a ton about agents. So Greg said at the moment, they're at the transition agents. Agents are going to do all the work. They specifically highlighted context, computer use, and memory as the core components. They want to bring codex, the coding capabilities of codex to everyone. And that's what I think we're going to start to see. We'll talk about the agents specifically for this new workspace. Agents in a moment. They want personal AI that is not only feels like AGI, but it's proactive. It actually anticipates what your needs are going to be. And it does things in the background for you and surfaces things like, hey, you asked for this last week. I went ahead and ran this for you, like that kind of stuff. So the interview's worth listening to. It's nothing groundbreaking like I was expecting with the two of them together. They were going to talk about a whole bunch of things they'd never talked about. But they did get into sort of the evolution of the relationship, the evolution of Greg's role and what he's doing moving forward. And then they did talk a little bit about the Elon Musk lawsuit and how, how painful it was for both of them personally. And one, because Greg's personal journal has got like, yeah, you know, put in as evidence. So like real personal stuff was out there. But Sam did say at the end that his biggest fear right now is that Elon's gonna drop the lawsuit like the day before it starts. Because Sam's like, we went through the hell basically for this. Like, I want it all out there now. Like all of our lives have been put out for everybody. Let's have this trial and let's hear, let everybody hear what really happened. So if I could totally see Elon dropping the lawsuit, just messed with them enough to like make their lives miserable and then like ask her it. But if this goes to trial, man, it's going to get, it's going to get messy and make for some pretty interesting conversations.

A (17:14)

Yeah, I bet Greg Brockman is regretting keeping a journal at this point.

B (17:18)

Yeah, he kind of glossed over it. It's like it is what it is, but I mean, no one wants their personal thoughts out there in the world like this.

A (17:28)

All right, so our next big topic this week. OpenAI has launched workspace agents in ChatGPT this past week. So the company kind of calls these an evolution of custom GPTs, positions them as shared agents that can handle complex tasks and long running workflows across tools and teams. So teams build an agent once essentially and use it together inside ChatGPT or Slack the moment with the agent improving over time. Now agents are being powered here by codecs running in the cloud so they keep working even when the user is offline. They can run on a schedule or they can be deployed directly into Slack channels to pick up requests as they come in. OpenAI is shipping pre built templates here for finance, sales and marketing agents with out of the box connections to things like Slack, Google Drive, Microsoft Apps, salesforce and more. The availability of these is a research preview right now for ChatGPT Business, Enterprise, Edu and Teacher plans with a gradual rollout across business and enterprise over the next several weeks. The feature is off by default for enterprise workspaces pending admin enablement and the pricing appears to be free for the next couple weeks after which they shift to kind of a credit based model but they're still not, have not yet disclosed kind of the rates and things here. On the governance side, OpenAI is shipping with these role based admin controls over who can create and share the agents. There's required human approval for sensitive actions like sending communications or modifying records, and a compliant API that exposes every agent's configuration and runs and safeguards as well against prompt injection attacks. So Paul, I know this is something you and I have been talking about quite a bit this week. You've done a little initial experimentation with this. Any thoughts? Like, how big a deal is this?

B (19:19)

This is one of those things where you initially look at like this might be a really big deal and I'll, I'll give some brief context. So I was as I mentioned I was at Google Next this week and that it was all about agents. Like literally everything. Every talk from the leaders of Google about agents. Yeah and one of the things they previewed was this agent designer. And then I actually sat in a masterclass where you could build agents with this agent designer and I was like, this is slick. Like this is really cool direction. Unfortunately it's not available. Like I don't know when it's coming, but sometime later I think it's in some sort of a research preview mode. So almost everything that Google showed was for developers. So it's like Vertex, AI, Anti gravity, things like that. And you need some elements of technical capabilities and you probably need it involved. So I was like, oh, like just that this is cool. Oh wait, I'm disappointed again. And then that same day OpenAI announces these agents as does Microsoft announced their agents. So we'll get to that in a minute. So I see the ChatGPT one and I'm like oh my gosh, that's, that's amazing. Is that actually available? Like can we get this? So I go into my ChatGPT account and sure enough, there, there it is. And I was like awesome. And so you just click on like so again I'm in our team account for Chat GPT and I just click an agents. It's in the left column and then I can click Browse agents and I can do browse templates and you immediately get a sense of what's, what's possible. Now it shows there's also recent uses. So you can look and see that. You can see it built by me agents and the SmartR x directory in our case. So if you've ever gone into the custom GPTs area, it's kind of like that. But for agents I would say like it's the easiest way to kind of envision how this works. But the beauty here is they have these pre built templates and I'll just read three of them quickly to you because it gives you A sense of what, what's going to be possible. So they have a template and you can start with a template or you can create your own by just using words like hey, I want a keynote. Keynote Abstract writer. So they have a chief of staff and this is how they describe it. Prepare a high signal operating brief from schedule inbox and team chat context. Great for users who want sharper priorities. Meeting prep to do, capture, source link, follow up guidance and requested email or chat Follow through in one concise daily artifact. And then you can connect it to Google and Microsoft calendars, Microsoft email and teams and Slack. They have a data analysis, one that's again a custom or a template agent. A data analysis plugin arranged around the life of an analyst rather than a tool checklist. Use it to sharpen the question, write and improve SQLs, inspect the shape of a data set, build clear visuals, prototype dashboards, and run a final quality pass. So basically just teach it skills that are specific to what that person would do, in this case the agent and then one other one sales assistant agent. Use generalized sales workflows for account intelligence, competitive research, value engineering, meeting prep, follow up, pipeline planning, seller coaching. Great for teams who want stronger prep, clearer strategy and better execution across the deal cycle. And then it shows you a bunch of capabilities. And actually I'll do one more customer support agent. So this is a generalized customer support workflow for ticket triage, case investigation, response drafting, escalations, customer research and knowledge creation. So now with each of these you can connect it to things. So I just picked these because every one of those, if we connect it to HubSpot, completely changes our workflows and potentially our staffing plans.

A (22:46)

Yeah.

B (22:47)

So if these things actually work in a reliable environment that I as a CEO am okay with us experimenting with it. It completely evolves the way I think about how we're going to do our hiring this year and how we're going to analyze it. And the thing I keep coming back to is this need for, to somewhat centralize. We'll talk a little bit more about this in the next topic too with this agent usage. But this idea of like centralizing the building of these things. And so what I did is on the flight back, I messaged Mike and Jeremy on our team and I put a calendar invite for next week and I was like, we're just going to run a lab on this, like kind of like a hackathon lab. And like let's just take an hour together and figure out what these things can do. Yeah. And so Jeremy and Our team's looking into the connectors and trying to make sure we're you know, good from, from a perspective, like a safety perspective to do these things. And then we'll, we'll actually do this. Like we'll spend an hour next Friday like hacking together and like let's, let's pick a couple of these agents, let's build something and see what happens. And it again like I, I don't want to overstate this but if these things actually work like this goes back to when we first got like some form of workspace studio agents in Google and then it was, it's like they're fine, but they really are just, this is a few months back. They're for like automating email stuff and like some calendar things. But it's okay. They're just like rules based things though. It's nothing too crazy. This is a different level. This is truly doing the work and you know, the ability to build agents for each role in the company, it really just starts to change how I think about this because it's so easy to do. Like you could literally train anybody to do this. Even, even somebody who's like, has been hesitant to do anything with AI. Yeah, we could run an intro to AI 30 minute class. Here's what it is, here's how it works, here's what agents do and like let's build an agent for you in real time and you can just do these things in these lab environments. So I don't know, like I, you know, until we actually do this next Friday, Mike, and until we have time to like play around, I don't want to, you know, say this is transformative per se, but it has all the signs of being a very important thing. And then Microsoft did the same thing. Google with this agent designer is going to do the same thing. Like it's pretty clear that by fall of this year, if not sooner, depending on which platform you're on, they're all going to enable a knowledge worker, a non technical knowledge worker to build agents and run them.

A (25:10)

Yep. You know, it's really interesting to read through the announcement about these and start playing with them because what really occurred to me is it was a subtly important point to read that it's powered by Codex because if you're, if you're one of these more non technical users, which I am one, if you haven't used Codex or Claude code, this is why people are freaking out about those tools because it's a preview essentially and it's a different modality and not exactly the same as these agents, but they basically do the same types of things for non coding tasks. They do agentic work using files, code tools and memory to do skills to do way more than you can do with a prompt or just a chat. So I think people are about to wake up to what's possible here and just to kind of connect the dots like this is why we keep harping on about these tools. Because the game changes when you go beyond just chat, I think.

B (26:09)

Yeah. And it changes many things and organizational design, like I said. And yeah, it's again, I don't want to oversell it, but I said if you go back to episode 141 and even go back to episode 87 prior to that. Yeah, my projection was that AI agent explosion would happen 2025, 2026 would be the starting point of it and then that would continue on. And by 2027 we would completely transform work with agents. So this is something we've been known was coming for multiple years. We've been talking about this and I feel like we are, we are clearly in the very early stages of not just the agent capabilities for the technical people and for development work, but now bringing that to knowledge work to make it as simple as building a GPT, which leads me to like the usage and stuff because there's so many people who've never built a GPT. So like even that is, is advanced for most average users of this technology.

A (27:13)

I want everyone to keep this discussion in mind as we get into this next topic because, you know, Paul, we've been, you and I have been talking quite a bit informally this week about agents at large and how you actually deploy them inside a real business today. So a couple updates that came out and we're going to kind of get into what this discussion about agents has looked like for us personally over the last couple weeks. But first up, some things that kind of spurred this discussion. So first we saw Jason Lemkin of Saster, who owns and runs Saster, posted a pretty widely shared take this past week about their use of agents in how they run that event. And some really interesting stuff on podcasts and on posts online where he's basically talking about using all these specialized AI agents to essentially run different parts of the company. They use Artisan for outbound, qualified for inbound, agent, Force for reactivation. They use agents for new customer acquisition. At the same time, Microsoft also, like you had mentioned, made Copilot's agentic capabilities available generally across apps and Copilot, and also we talked about just how OpenAI is rolling out workspace agents. Google is hyping up agents at Google Next, which we'll talk about. But these land in the middle of this bigger conversation you and I have been having about kind of where we're at on all this and the open questions around AI agents, because there's like no shortage of voices we hear from them out there asking these like some version of the question, why aren't you all in on agents? Like, why aren't you doing every possible thing you can with agents right now? And Paul, I don't know, correct me if I'm wrong, we're not anti agent. I feel like they're 100% the future. And we're actively experimenting with general purpose things. Agents like Claude Code and Codex. We have not gone all in yet on things like openclaw, but there's always like really important open questions and nuance that I feel like people are just shoving under the rug here about, like, what does actual production usage look like? What about security? What are the specific use cases that actually matter for a business? And the usage question you just alluded to, like, how do we price the usage of these things? So, Paul, let's just like get into this. Like, where do you want to start here we've talked about.

B (29:31)

Yeah, So I mean, really what happened is I got back late Wednesday night. You know, I'd already put this lab meeting on the calendar for Mike and Jeremy and I, and I hadn't had a chance to actually play with the agents yet. And so I got in the office Thursday morning and I was like, all right, let me just jump in ChatGPT real quick. So I jump in and I'm like browsing these templates and looking at its connections, like, oh my God, like this, this might be it. This might be what we've been waiting for. And then Mike came in the office and I was like, dude, look at this. I'm showing him these sample agents in these templates. And so again, coming fresh off of Google Next. I like, all of this is fresh in my mind because I met with some really interesting people and it's just that random. Like how, you know, sitting next to somebody at lunch randomly, or the person you're sitting next to at like the keynote and you just, you have these conversations and these are, you know, you might randomly run into like a person who's heading up gen AI adoption, managing token budgets at these major companies.

A (30:28)

Yeah.

B (30:28)

And she's like, well, what are you doing? Like, what? Like, what's happening at this company. What's going on with your developers? What's going on with like marketing, sales and customer success? Like how real is this stuff within enterprises? So these are the kinds of conversations I'll allude to often on the podcast. Like we're talking to the real people and there's this balance between developers who are hardcore, pushing the frontiers of everything that's possible, seeing into the future, a future that no enterprise is going to touch for a while. Like they aren't going to do those things. And so when we're talking about this stuff on the show, we're trying to talk to the practitioners and the business leaders who are the non technical people often who have to actually figure out what does this really mean. They're trying to solve for what are the token budgets we're giving our developers? And some people are like, oh, let's just do token maxing, burn all the tokens you want. And then I talk to somebody who's in charge of tokens and they're like burning through like our whole monthly budget in two days. Like, how are we supposed to budget for that? And they're going back to these vendors being like, we can't do this. This isn't a sustainable way to handle this. Then there's like the vendor selection. Do we go all in with Anthropic or do like, well, ChatGPT 5.5, like is that a good model? Should we be using that? Or is this new agent designer from Google going to be the thing? And should we just put all of our eggs in one basket with Google? So these are tough choices. The pricing models, getting back to the token budget. Like I've been transparent before about this. I just went into HubSpot today and we're like, we're already out of credits. I'm like, how the hell did you run out of credits already? It's three days old. Is the billing cycle like what, what did we do to run out of credits? And I actually like went in and I'm like trying to audit where did the credits go? Like, what are they being used for? And it makes no sense. And so I'm just like, God, this is so frustrating. And then you mentioned risks. The other thing we'll hear about, and the Saster episodes are amazing by the way. We'll put the link into it. They're like, here's what we're doing. We're using 20 agents for this. We have for that. And you start to realize that when you actually are on the frontier trying to Innovate with these agents within a real business. How the hell do you govern them? Like okay, now there's 20 agents running loose that have access to all these different connectors and these people have the freedom to just go get more whenever they want. And Mike can go get this subscription. Jeremy, that's. And so like now you actually have to manage these things and these agents, they function off of knowledge bases, they function off of skills. Those things get outdated. Like how are you managing those and updating them? Is that in a Google sheet? Like where are we doing all this stuff? And then at Google Next I watched a, a demo from the co founder of, of Wiz recent acquisition for, for Google Cloud and he was showing how they're actually managing the risk of these agents. And it was, it was beautiful. Like it was incredible to watch. But also makes you aware of how unprepared most people are for everything that goes into running and governing these agents as they get access to more and more data. So yeah, I don't know, I just keep coming back to like, I love these practical use cases like Saster is doing. It's inspiring stuff. Like it's really cool to hear these stories and in a real business that's like our business, I mean they run events. It's like it's close to home for me and I listen to what they're doing. It's like, oh, that's a pretty cool idea. But you also listen to them and they're being totally transparent about the fact that they're just figuring this all out because they'll build something in repl it and then like they launch it and then it breaks and they're like, what do we do now? Like how do we fix this? Like we have no idea what's even happening. And then they're going to talking to Claude and being like what, what broke? Like how do you. Because they're not the people who would usually take those things to production. And that's another element of this agent stuff. It's like we're being empowered to build these things but like I don't know how to take things to production and I don't know like how to deal with it if something breaks. So I don't know. Yeah, it's like we could literally go any direction with this. But those are just some of my thoughts for the week. Having spent the week seeing agents being debuted, hearing them talked about and then talking with real leaders at massive enterprises who are, they're nowhere near prepared to, to do this stuff with agents. Like outside of their developers. And even then it's like, it's like a free for all and they have no idea how to manage the tokens and which vendors to use. And so yeah, I don't know. It's it, it is the wild west right now. But the people are figuring it out, are getting a really fast event, a

A (35:03)

competitive advantage again, you know, I wrote down kind of as we were preparing, just a few big unanswered questions I have about agents, or let's call them at least not sufficiently answered. I'm going to share them really quick just in case they're helpful to people. But like, first is really, how can I more clearly think about different, let's call them types of agents because like, in a practical sense, the more I learn, the more there's not just one type of agent really. Like quad code runs agents to do things in real time with periodic guidance, partnership, handholding from humans. But that's like materially different in practice from something like Open Claw, which can do similar stuff, but does so persistently and autonomously. And I don't necessarily think one is better or worse. It's just that like, when I think about this, it's already nuanced that people aren't addressing where I'm like, no, it's not just like AI agents. It's like these are two at least very distinct paths to me and I'm sure there's others in here I'm missing. But like, I think there's more nuance to this. Like just because I'm not using OpenClaw yet, or like a 24,7 persistent agent I don't think necessarily means you're at a disadvantage. It just totally depends on the use case, right? So I think about that a lot and I'm still kind of trying to work that out on my own. I often also am thinking, like, what are the actual use cases for always on agents like OpenCL. That sounds really obvious to say. I could rat off 20 different ways you use these. And keep in mind, again reference the previous segment. I am not bearish on these. I think this is the future. But there's the real consideration, like if I have to worry about this thing all the time, if I have to manage it all the time or try to troubleshoot it if it breaks regularly, how is it remotely worth it for me to spend time on this versus something like, shouldn't I just be building out even better and more expansive skills for quad code or building the workspace agents in ChatGPT? I don't know the answer here, but, like, that's a real consideration for me. And then finally you just hit on this. Like, how in the hell do you pay for 24, seven persistent agents? I feel like there was this, like, honeymoon period because I think until really recently, you could just plug open Claw or something into your Quad Max account. Right. And use it that way so, like, you didn't have to just pay via API, I don't think. And you can't do that anymore. They, like, turned it off. So how on earth am I going to spin up like a $500 a month agent to, like, do my grocery list? I'm not thinking across that much, but I have no idea. That's the point. It could cost 5 cents. It could cost $5,000 a month. I genuinely have no idea how to gauge this. And that's just like a personal experimentation. Like, how in the heck do you figure this out as a business? Like, how would you. I mean, like, that's what you're getting at, right? It's like, there's no predictability here. You can't budget for this.

B (37:54)

Right. They've already shown in the last six months they're going to keep changing the pricing models. So then, you know, and I'm not saying they're going to do this in a deceitful way, but the way this traditionally works in business is you get somebody hooked and then you jack up the price.

A (38:06)

Yes.

B (38:06)

So, you know, let's say for us, we go next week, on Friday, we're like, oh, my God, these agents are incredible. And then we build a team internally that basically goes department by department and looks at workflows and problems, problems and goals and rocks and says, okay, we're going to centralize the building of agents because it's going to be too complicated. We have everybody doing their own thing. And let's get this small team together. We go through, we prioritize these things, we start tackling a couple workflows, a couple problems at a time. You build a bunch of agents, they're crushing it. They're part of our 20amonth per person plan. And then all of a sudden they're not. Now they're HubSpot's model where we're like burning credits. And I have no idea where the credits are going to.

A (38:45)

Yeah.

B (38:46)

And to your point, like, Maybe it's now 5,000amonth instead of 300amonth, but now I'm hooked. Like, now these things are built into our workflows, so. And maybe they don't change it in two months, maybe it's in a year when they figure this out and it goes back to that pricing. And I think I, you know, I'd said this to you yesterday morning, Mike. I'm like, I don't get how this isn't eventually a human replacement cost thing. Like it just seems like if there was a, a simple way for the labs to calculate the value of their own technology, which I don't think they're currently capable of doing, they would just charge more for it. So like for example, if I go into these agents next week and I figure out like, wow, we can actually build like a customer success assistant that's going to do these things each week, each month. And if I had to hire someone to do that, that would be like a hundred hours of work. That's a full time hire that this agent's basically going to do that work. And now let's go do the same thing for sales. Like we'll build an SDR agent and it's just going to basically do what an SDR would have done or like an event market or whatever. Like if we figure out a way to actually do it then I like, I would happily pay like if I knew as the CEO that that agent I just built or a collection of agents working together is doing the work of three people. And OpenAI came to me and said, hey, like built these agents, like, you know, the value of that would be 300,000 a year. We're going to charge you 3,000amonth instead of 20 bucks a month. I'd be like, yeah, all right, let's go. Like, okay. And so I feel like for finance to truly get involved and manage this process as these agents become more prevalent within organizations, I, I can't imagine how a token or credit based budget, whether you're constantly running into a limit, is, is in all possible or scalable for anybody. And I, I keep coming back to, it has to be simple, it has to be clear, it has to be understandable. I'm paying X a month. You're, I'm getting use of these things. And I don't know if it's just like, you know, These models get 10x cheaper each year. So maybe it's at some point over time maybe. Well yeah, maybe at some point you're just like 5.5 is good enough. Like these agents crush it. I don't need GPT7 and I know that it's going to cost you the lab 10x less to serve me this model in 12 months. So yeah, Just keep, let me stay on the old model, I don't know. Or maybe that's where the open source stuff comes in. It's like once we have an open source model, it's good enough. Like the deep seq. The numbers on deepseek are that it's basically like on par with some of these frontier models.

A (41:35)

Right.

B (41:36)

And so does it go back to the open source? Does it swing back where you're like, yeah, I'm happy with fifth generation models. Like, I don't, I don't need, I don't know. And I don't, I truly don't think the labs know because they've focused so much on building for developers that are cool with the token maxing model. And we're just going to pay for our tokens because they're used to that approach and I don't think they've yet solved for how to charge the way SAS traditionally would have. Like what is the evolution of a seat based license? Yeah, and then, yeah, then you're like HubSpot and you're like, okay, I'm just gonna build these agents and I'm just gonna connect them over to HubSpot. So I actually, I'm gonna get rid of a bunch of my seats because I don't, I don't need anymore and I can just access it through ChatGPT.

A (42:19)

Yeah, there's a lot more nuance to it than just go use agents.

B (42:24)

Yeah, yeah. And I think like sometimes you get pushback on, you know, the not trying to belittle the, the capabilities of these agents or not give them enough significance. I just think sometimes people don't have the nuance of what really happens in an enterprise and like how complex this is and that's we spend our time talking to these companies all the time who can't even get copilot rolled out or nobody's ever been trained how to even build a GPT or analyze a workflow and figure out where AI can fit into it. It's so messy. When you actually get into the real stories of adoption, it's easy to just see the technology and think, oh my God, everybody should be doing this. And it's like, no, they shouldn't. It's not ready for prime time yet. But if you can embed codecs and CLAUDE code right into the user interface that the average knowledge worker can use them, it changes everything.

A (43:19)

Oh yeah. And to your point, you mentioned to me in the office, it's like even on the GPT front, it's like very few enterprises have fully explored what is possible simply with GPTs or simply even with connecting standard chat to valid, useful data sources. Right. So it's like there's so much value to be accrued and created there. It's like, I'm not saying you don't need agents, and that's for sure where we're going. But why, why does it just have to be that this is also a path where I think it was overlooked because we're all, you know, in the Twitter or X. A high buffalo where everyone's like, oh my gosh, I'm running my entire company with agents. Which is amazing. Like, I'm sure some people are doing that, but the vast majority of people are not remotely close to that.

B (44:03)

If you're a native company and you can do that from the ground up, you can take those risks, go for it. Yeah, that's not the reality for the vast majority of companies. These, these ones that we want, AI mergent, they're trying to figure out how to work within legacy systems, legacy talent, legacy governance structures, highly regulated industries. Like, it's not the. It's not reality.

A (44:25)

Well, yeah, I mean, I won't harp on this, but just one more consideration. It's like, you know, it's so hard to predict what is worth investing time into anyway in AI, because a year ago someone would have been like, go build all your own agents. Then you might have done really well with that. But then OpenAI comes out with this and you're like, why did I waste any of this time? Also, the architecture behind some of, like, rag and things like that. I don't want to get over my skis on the technical stuff. Like, some of these methods are like, totally out of date now. So I should have spent six months figuring this out when I should have really just been probably building GPTs or skills or something. And then they flip the switch and I can just click a few buttons and make an agent in ChatGPT. It's a very hard. I'm not saying, like, that's the right path, but it's really hard to predict. Like, should you just actually wait until it becomes a little easier to do?

B (45:11)

Right, right.

A (45:13)

All right, Paul, so before we get into rapid fire, one more announcement for this week. This week's episode is also brought to us by AI Academy by SmartRx, which helps individuals and businesses accelerate their AI literacy and transformation through personalized learning journeys and our AI powered learning platform. We add new educational content weekly, so you will always stay up to date with the latest AI trends and Technologies and we wanted to spotlight this week our AI for Departments collection which right now features six core series and certificates designed to jumpstart AI understanding and adoption across departments. Right now we've got Marketing, Sales, Customer Success, hr, Finance, Operations. I just actually wrapped up Paul AI for Legal this past week. So I believe that would be coming out next week. Don't quote me on that, but very soon we'll have that done. These are the ideal launch pad for organizations that want to level up their teams and accelerate AI adoption and impact. I'm actually going to share a little later in the episode a few quick insights from the AI for HR series which I talked just as a note. Individual and business account plans are available now. You can also buy single courses and series for one time fees. So go to Academy SmartRx AI to learn more. Okay, Paul, first rapid fire. This past week you were at Google Cloud Next 26 in Las Vegas. That event wrapped up. Their headline announcement was Gemini Enterprise Agent Platform. This kind of full enterprise stack for building, scaling, governing and optimizing AI agents that basically effectively absorbs and replaces vertex AI. Going forward, this bundles a few things like a low code agent studio, an upgraded agent development kit, agent runtime, a persistent memory bank and some governance tools. It also has access to 200 plus models including Gemini 3.1, Gemma 4 and also outside models.

B (47:12)

Techcrunch in there I think.

A (47:13)

What's that?

B (47:14)

I think you can get Claude.

A (47:15)

Yes you can. Claude as well is in there. And then TechCrunch actually framed the platform as Google's response to, to things like Amazon's bedrock Agent Core, Microsoft Foundry. They have a bunch of launch customers using this. They paired the platform with a refresh Gemini Enterprise app. They made a $750 million commitment to their 120,000 partner network to accelerate agentic AI deployments. There's also a new agent marketplace. So Paul, you were at the event. What was your read on what Google announced? I mean it seems, I think agent is safe to say probably the word of the year here at this point.

B (47:55)

That's for sure. Yeah. So I'm part of the Google Cloud leaders circle. So you get, it's like an invite only thing and, and so I get a day with like Google's leaders. And so that was Tuesday. I got to sit through some pretty amazing talks including the opening talk from Thomas Kiran, the, the CEO of, of Google Cloud. And it was very apparent from the jump that they're, they're all like everything's agents. So he said the goal is to Make Gemini Enterprise the best place to run and manage agents. And then in his opening keynote at Google Next, he said bringing AI to every employee and every workflow was like the goal they were focused on. Now the thing you always have to differentiate with Google Cloud is like again, when they're talking to developers and when they're talking to the non technical users and a lot of the things that they traditionally would announce is like focus more on that developer audience. A lot of things they've built like Vertex and Anti Gravity, they are not for your average user. You really need technical capabilities to get into them. One of my favorite sessions at the Leaders Circle event was Google AI at Google. So they were basically had some of their key people who are working on AI transformation, AI strategy at Google talking about what they're doing. So I'll highlight a couple of those real quick. So Ryan Vaugh, who's the VP of AI Transformation, talked about these lighthouse workflows. And so he was saying they're basically trying to focus on moving from just tasks into the actual workflows and they want each business unit focused on two workflows. So they're all about like prioritizing where the impact could be. And I really like this concept. It's something we've talked a little bit about ourselves internally. You want to get past the cost savings, focus on growth and innovation, which, you know, I obviously love that thinking. He talked about this analogy of going from fishing where you're throwing lots of lines in the water, trying lots of use cases, to farming where you're getting very strategic and deliberate. And then one of my favorite things, the echo that we say all the time is this idea of reimagining work. So seeing significant changes in how teams work together, they're starting to field experiments within AI native work labs, which that might have stuck with me when I was thinking about doing this lab, I don't know. But he also talked about how it's so difficult right now to predict change and that the lines between roles are starting to blur. We've talked a lot about that on the podcast how like as the CEO, I all of a sudden have the ability to do people's jobs because I can just go in and like, yeah, use Claude. It's like I'm getting annoyed someone's not ready, like, oh, I'll just do it tonight, I'll do it myself. And so they're seeing smaller teams emerging where these blurred roles are sort of allowed to blossom. I guess it's like cool that everybody can kind of do each thing. There was also Joshua Spanier, who's the VP of AI and marketing strategy at Google. He said they're even within Google. They were struggling to get everyone internally to use the technology, which again, is kind of counterintuitive to a lot of people, but it's not. If you've spent time with these labs themselves, they're just like us. They have marketers and salespeople and CS people. Who doesn't mean just because you work at Google that you're like AI forward necessarily. So you're there to do job. So you talked a little bit about that and how they started a dedicated AI team that was actually in charge of like, the contracts, the data sets, the tools, the systems. And so that team builds out a suite of tools that then is shared with teams to use. So it goes back to that idea of ChatGPT agents. Like maybe we just build agents and we, we say, hey, sales team, here's your three agents and here's CS team. And that's a big question for me moving forward and I think for all of our listeners that we often think about is are we centralizing the building of AI capabilities and then like distributing them to teams, or are we allowing everybody just sort of do their own thing? He just said they relied less on individuals to figure things out. They made a big investment in ad, creative development and testing, and they're seeing a massive impact on cost and performance. And then he said something I thought was really interesting. No one joins Google. And I wrote any or any company myself to just be efficient. Like, no one's goal in their job is to be as efficient as possible.

A (51:51)

Right?

B (51:51)

So that's why they focus on trying to bring the creativity and innovation. And then the one other note I'll share, I was really excited about this one. So Jeff Dean was the closing talk at the Leaders Circle the first day. And if you don't know Jeff Dean, we've talked about him on the show many times. He was the 30th employee at Google. He's actually the one who named Gemini. And the name came from the merging of Google Brain Team and DeepMind. So it's like the sisters, like the Gemini. And he said that even then. So again, going back to this, how mature are agents? He said his words starting to see glimpses of the agent economy, meaning we are still early. He highlighted specifically the lack of reliability and trust agents that we should all have in agents right now for giving them access to credit card information, filing systems, all of these things. So again, we Say this on the show, but this is Jeff Dean, an authority on the topic, saying agents are early. You have to be very cautious with them. You have to be conscious of what you're giving them access to. But they're getting really good and we're seeing glimpses of them making an impact. And then on breakthroughs for AGI like, how close are we again? There's very few people in the world more qualified to actually talk intelligently about this topic. He said he thinks we're still one to two major away, which echoes what Demis Hasaba says. And when talking about what does he think like, that key is he alluded to that he thought continual learning was likely one of them. Now, having listened to Jeff and others for the last 15 years, I've been studying AI. Usually if they pick something, it means that it's something they've been working on and they've made advancements on and continue learning. To me, I've said this many times in the last 12 months. I think that that is the unlock that most of these labs think if they can solve for continual learning, that the model doesn't stop once it comes out of its training, that it actually learns like humans do from experience and inputs and outputs. It constantly changes its own weights and gets smarter and more capable. That is, that might be the final unlock. And my guess is DeepMind has made progress on this, and I would imagine the other labs have as well. It's also a very complex thing to put into the world because it can lead to the fast takeoff concept that we're probably not prepared for. So really cool stuff. They do an amazing job at those events. I mean, Google's just incredible. And Google Cloud puts on the leader circle is great. And then the event itself, I was only able to stay for the first half of the first day of the actual next conference, but even that was awesome. And then a final note, I did so Sarah Kennedy, who's a good friend of mine, she led a panel with Shaun White and Bryson DeChambeau, Shaun White the Olympian and Bryson the golfer. It was awesome. Like hearing those two guys talk about what they're doing with AI in their sports. But just seeing them and their personalities was really cool. Like, Bryson's kind of one of those people, like, a lot of people like to not like Bryson. When you sit there and listen, like, I don't know how you couldn't like the guy. I mean, it was. It was a really cool story. And I was, I was like, excited to kind of get to experience that.

A (55:05)

That's awesome. All right, next up in less positive news, a leaked internal memo this past week revealed that Meta is trying to basically inst install tracking software on their US Employees computers to capture mouse movements, clicks, keystrokes, occasional screenshots across a designated designated list of work apps and sites. And the memo frames the rollout as a way to teach AI models to use computers by giving them real examples of how people actually use them. And CTO Andrew Bosworth described the end state as one where our agents primarily do the work and our role is to direct, review and help them improve. The memo assures staff the tool will not read or read files or attachments, will not be used for performance evaluation, and will not learn incidental personal information picked up from the screen. There are reports that Meta staff are protesting the rollout internally. I wonder why. Separately, Meta also it leaked and then Meta had to announce it. I believe that it is going to cut roughly about 10% of its workforce with layoffs beginning May 20. There are additional cuts expected in the second half of 2026. A big part of this is the cuts are part of their effort to run the company more efficiently and the chief people officer Janelle Gale told People that it was to allow us to offset the other investments we're making. I would just like to, without over speculating, point to what other investments Meta is making. There's exactly one that is quite large and that is its CapEx guidance of $115 billion to $135 billion that is basically on AI infrastructure that is nearly double what it spent in 2025. So AI somewhat adjacently is probably responsible for some of this. So Paul, the kind of reason we're talking about this be curious about your thoughts. First on are they just basically trying to train agents to do what the humans are doing and then get rid of the humans? Also like, what do you think of the cuts and the layoffs due to the capex and investments they have to

B (57:16)

make to stay current cuts and layoffs expected? I would expect more. Not just them, that's obvious. And that's going to continue the, the monitoring of employees. I'll. I mean it's not confident. So this isn't new. I'll say that.

A (57:36)

Yeah.

B (57:37)

Yeah. So there's another social media company. I did a talk two years ago and after I explained computer vision and the ability for things to be recorded and then analyzed and using training data, I actually had an employee from that, it was a different social network company, come up and be like is that why they've been recording everything. I've been on my computer, so. And then she explained to me how they were using the data, she thought, but she wasn't aware that this was even possible. So this isn't new. You know, it's not surprising at all. I, I think that at some point, you know, you have to think about the kind of organizations you want to run and the kind of talent you want to recruit and, and retain. And at some point the best talent is going to have choices to make about where they go to work. And if you, you know, if you're cool, going to work for a company that, you know, is tracking literally everything you do and likely using it to train your replacement, is that, is that motivating? Like, it goes back to that thing I just said about, you know, Google and saying, like, hey, listen, we're not. Nobody comes to Google to work, to be efficient. Like, it's not like nobody wants to go work to like, like watch an agent do their job. Like, I'm picturing like an assembly line and I'm just sitting here, like just eight hours a day. I'm just watching it click around and do things. And that doesn't sound like a fun career. So I don't know, like, I get what they're doing. I understand this is, I mean, it's meta. Like, they're gonna be on the edge of this and they're gonna do things that a lot of people are gonna hate and they're gonna get bad PR about it and gonna have pissed off employees. And that's the, the story of their history. Like, they've always done things that were, people felt were beyond the line of acceptability and they seem comfortable with that and it's just part of who they are. But I think every other company is going to have to make these same choices because what they're doing is possible. Like, if you want to do a consulting firm or an agency, or you want to pick operations or HR or finance in your own company, this tech exists. Like, you can train them up and you can build agents based on what people do there. There's a startup last year, I forget the name of it. This is what they did. Like, this is. They sold this technology to enable you to do this. Yeah. So, yeah, if this is new to you, sorry, like, this has been going on for a couple years and it's going to get tons of funding from VCs to do this. It's going to get ton of tons of payments to consulting companies to implement this. And they will absolutely Use it to reduce their workforces. Like there's no other reason to do it. So. Yeah, I mean, and I'm not even trying to be hard on Meta here like this. It's just the reality, like, and that's so much of the time when we're doing things like this or having these like more hard conversations about the reality. We're just trying to share with people like what the reality is. And if you're working for a company that's doing this, there is no other reason. Either it's either performance or to. To train on what you do for your job.

A (60:52)

Right.

B (60:54)

I can't think of a third thing that you would do for that. Why, why else you would do it? Yeah, so I think it's just, it's just a, I guess an awareness thing and you got to know the kind of company you're working for and what their intentions are with AI and ideally you want to like understand the responsibly I principles and whether or not they're a human centered company. That's why I think it's important just for people to have levels of awareness and then educate other people about these things. Because our, you know, listeners to our show are more likely to get this stuff like that. They already knew some of this.

A (61:25)

Yeah.

B (61:27)

But all your peers, your family, your friends, they don't know this stuff. And so sometimes it's just us trying to do our part to share it so that, you know, other people can go and educate people about it.

A (61:36)

Yeah. And to be clear, the intention behind this segment is not to pile on Meta specifically, because it's not anything new that companies monitor what their employees do on their work machines. Often that has happened well before AI I think what is just really fascinating to me is like, oh, this isn't just for security purposes anymore. They're just coming out and saying there's another. To your point, there's one of two reasons either employees following guidelines are performing, I. E. Are you doing work on your computer? Are you doing anything wrong on your computer? Which has existed for a decade now. Major enterprises. But there's this new lane where it's like, oh, okay, this is training data for. Exactly. For computer use agents. That gets really murky really quick.

B (62:22)

And if I'm not mistaken there, I don't think we covered this, but I'm pretty sure two or three weeks ago Elon Musk like changed the terms of employees at XAI and they had to agree to have everything.

A (62:32)

Oh, really?

B (62:33)

Yeah. For this purpose, like, it's all about training Data for Grok. And, and that's the thing is like they're not even necessarily using this just for their own purposes. They're using this to train their models. Like so the work they do. So imagine if you can collect every interaction that your marketing team, your sales team, your CS team, whatever. And you also happen to be a company that trains AI models. You don't have to go license that data because what's happening in other instances is the training labs, like a scaling AI are paying lawyers and consultants to sit there and have their stuff done. Like so not for a company they work for, but saying, hey, we'll pay you 500 an hour to like track everything you do on your computer for a few days. And then they're taking that to then train the models to do that, the job of those people. So yes, that is the new thing. To your point, Mike, it's like performance tracking and monitoring usage. That's not new. Using it as training data and data to then replace those people is new.

A (63:33)

Yeah, and I didn't even connect the dots until you just said it. Like, this has to have Alexander Wang fingerprints all over it. Exactly what.

B (63:42)

This is what they were doing at scale.

A (63:43)

Yeah, you're right. You're right.

B (63:44)

Yeah.

A (63:46)

Okay, well, our next topic this week we actually. Well, I guess it was technically this week because we recorded on Monday we covered Apple CEO transition at the end of the last episode because that had broken right before we started recording that John Ternus was going to succeed Tim Cook on September 1st of this year. In the days since, a bit more has come out, especially on that kind of AI angle of what Apple's doing with AI. So Cook and Ternus had in all hands at the Steve Jobs theater. Cook interestingly addressed some health rumors head on. He told employees, hey, I'm healthy. Energy's high. Plan to be in the role for a long time. Ternus teased an incredible roadmap ahead. He said AI is going to create almost unlimited potential for the company. According to Bloomberg, Ternus has already overhauled the hardware engineering organization around what he calls a new AI platform designed to speed up product development and improve device quality. On the same day as the CEO announcement, Apple promoted Johnny Shrugie to a newly created Chief Hardware officer role combining hardware engineering and hardware technologies into one organization. CNBC read this reshuffle as kind of a sprint to build in house chips for devices, with Apple doubling down on silicon for on device AI. Obviously we've talked about a bunch Apple's new and improved AI powered Siri, which has been delayed a couple times, is now expected to debut at WWDC in June of this year. They have a multi year deal now with Google reportedly worth around a billion a year to power the new Siri on Gemini. So CNBC is kind of framing this transition as you know Turnus facing this defining challenge which is obviously Apple does more than just AI, but his job is kind of fix the company's AI strategy it sounds like. And Paul, obviously it's so early here but given the new details, like what is the kind of your initial read on do you think he is the right guy for the job to fix Apple's AI problems? Where do you see this go?

B (65:48)

Yeah, time will tell but everything I've heard about him from following online is just extremely positive. Sounds like everybody's known he was going to be the guy he, everybody's saying he's the right guy. I watched a crazy clip where he was doing an interview about like the cinema, the cinema display, like the thing, you know, his first major project there. And he was talking about when he was at the, I think it was at the manufacturer or whatever, they were piecing it together and they had designed the screws in the back of the display that no one ever is going to see to have like 21 grooves in them. It was like a very specific number and he actually like took the screw out, took a magnifying glass and found that they had 30 grooves instead of 21 and made him redo it. Like it's just like they were trying to stress how like what a perfectionist, like a Steve Jobs type product guy he is. So it seems like that's what they're getting. And like I said probably in the last episode, I, I think like if they weren't comfortable with the roadmap they have to execute, it wouldn't be the time. So they're obviously very comfortable here. Interestingly, at Google Next, Thomas Kirian, when he was doing his opening keynote on the actual first day of the conference, he did mention Apple. They just put the Apple logo up and everybody's like cheering and then he just said about them being a preferred provider for their models and that was it. Like there's no big thing. He didn't go into a ton of detail. He talked a little bit about Siri, but it was basically like that partnership that we've talked about previously on the show. So I don't know, like again as a long time Apple user fan, I'm excited. It seems like Wall Street's liked it so far. I Don't. I mean, I think their stock's been doing pretty well since the transition, which isn't always a given when you have a CEO change.

A (67:25)

Right.

B (67:26)

So, yeah, I don't know. Everything seems positive and I hope I've said many, many times, like, I just, I love a working Siri. I'd love Apple intelligence to really be intelligent. Like, I think it, you know, it's billions of users that would get to experience AI in an entirely new way. And I think a very positive and exciting way if Apple solves how to do it the right way on the iPhones and all their devices, you know, AirPods and watches and glasses and everything else they've got.

A (67:52)

Yeah, I was going to say a very longer term, but people, you know, we, we included have talked about Apple's, like, fall from grace and AI, but, like, they're also half a chance away from cracking the code on AI wearables. They're like the best people that do it. And if they do that, it's like, game over. Like, it's a whole different cable game.

B (68:10)

Dude, the data they have is insane. Like, I don't. There's so many things Apple does where you don't because they don't feature it. They have to, like, find these things. And I was, I was analyzing steps the other day. Like, I love the Health app in Apple. It's incredible. And I, I track everything. I've shared my personal story about my heart and how it, like, you know, kind of found something with that. But they track, like, things like distance between steps. Like, it just like. And it's like, how, like, it's all. It's either in my watch or my phone that they're getting the data from. And the fact that it has this kind of data and, like, you just realize the depth of data they can capture from these wearables or from the, you know, the phone in your pocket, whatever it may be. And then you start to imagine, like, my goodness, like, what could they do with that data? Yeah, if they have the intelligence baked in. So if you're. I'm serious, like, if you've never done it before, go into the Health app and just click, like, show all data and just look at the metrics they have on you. It's wild.

A (69:12)

And then an experiment I did which worked somewhat well, is then have Claude code go build some things to connect to that data and then tell you some stuff about it, which is interesting. It's a lot of trial and error involved. Not perfect. You probably just get the same thing Have Apple watch. But it was a fascinating experiment, Apple watches.

B (69:31)

And if you've ever had, like I was a watch guy before, like, I collected watches, like, nice watches. I stopped because, like, I the utility of wearing an Apple watch every day. I hate when I would not have the data for a couple hours. Like, you know, you put a nice watch on or whatever to go do a keynote or something and it's like, ah, damn, I don't have my heart rate while I was talking. And like, sometimes you want to see that. It's like, does my heart rate go up when I'm on stage? Like, I'd be curious. So, yeah, I just. I love that. So good.

A (69:59)

All right, so next up, we have our, you know, now regular segment we're doing on our AI use case spotlights here at SmartRx, where every week we're trying to give you a quick look under the hood at some real uses for AI that we're exploring, building or deploying in our own work and sometimes in our own personal lives. So, Paul, I just have a really quick use case to share this week. If you have anything to share, we can kind of talk through that too.

B (70:24)

Yeah, go for it.

A (70:25)

So for me, actually, I stole this one. The use case is not mine, but I don't think the person will mind me stealing it because it's actually from Taylor Rady, our director of research, who's taking the lead this year on SmartRx's State of AI for Business reports. So typically, we have done for five years in a row a State of Marketing AI report through Marketing AI Institute and SmartRx, where we've surveyed hundreds and then thousands of marketers and business leaders on AI adoption and usage. This year we decided to really expand that out to all functions of a business. So we've got. We just closed the survey. We have basically almost, I think, over 2100 responses, the most we've ever had, spanning every function, industry and company size. So we are knee deep in creating the actual report. And it's really interesting because Taylor is taking the lead on this. This year I'm kind of overseeing some stuff and reviewing it. But, you know, I think I had shared last year, or maybe early. Yeah, last year that, you know, the report alone used to take hundreds of hours to do all the manual data analysis, writing, understanding, synthesis, you know, years and years ago in 2024 and 2025, I cut that down to probably a few dozen hours, which felt like an incredible win. Taylor did the report in like a day this year, and I'VE looked at it a cursory fashion and it's really good. And we're obviously going through with a fine tooth comb with human oversight for this. And there's human complexity and tone and writing and rewriting and reworking. But, like, she was able to cut this down another order of magnitude and how long it took. And the cool thing is it wasn't just about time this year because, you know, in past years I've been like, oh my God. Thankfully it didn't take me this long. I gotta run to the next thing. This year, like, me and Taylor have spent a huge amount of awesome time spent going really deep on two things. One, how can we ask even smarter questions of the data and go further and deeper on this stuff to create an even better report? So we're not, say we're reallocating the time, we're not actually netting out with less time here, but it's going to be 10 times better. But also as part of kind of building out our research function at SmartRx, like, how can we blow the doors off activating this report both internally for sales, customer success, Everyone else academy, and externally across a ton of different channels, which is an area we've historically struggled with because it takes so long to do all this stuff. So really it is night and day. Even I was blown away last year by what the models could do, especially Gemini and Claude, with both data analysis and writing. This year, it doesn't even come close. They just smoked what we were able to do last year and it's just jaw dropping. I mean, I'm just continually reminded, like, I know this, I see this every day. But then, and something like this is just so cool to see how good this stuff has gotten. And it's really cool because like, recurring use cases, compound, we've done this every year now that we've been able to for several years, like using AI for parts of this, it just gets more and more every year and the results just compound and compound and compound. And it's, it's incredible to see. So super excited about that. We're releasing this in a few weeks here, so we'll have more on that and more announcements around that. But we're really excited.

B (73:45)

I can't wait to see it. For one. And two, as someone who has personally spent hundreds of hours in pivot tables building that report, I love to hear the stories of how we are solving for making it more efficient.

A (73:56)

Oh, and I will just note too, at our AI for Writers summit that is coming up in a Couple weeks

B (74:04)

if you market May 7.

A (74:06)

So if you go to marketingai institute.com, go to events, you can see there's a free registration option. Taylor is actually giving a talk about exactly how she did this. Super tactical. You can learn, you know, step by step how you can do this for yourself too. So go check that out.

B (74:21)

That's awesome. Yeah, let's do a quick one. This is. I actually, I forgot I ran this. It's funny. So sometimes I'll just go into like chat GPT and see what are the recent prompts I gave. So apparently like I said, I forgot I did it. I think this was last night or this morning. I had seen a Jason Calcanis maybe tweeted about like how we were going to have all these like new companies created and that was going to create all these jobs and. But not everybody's really made out to be an entrepreneur. And so just that like spur of the moment, like, I'm like, I'm not tired of this argument. Like I'm actually an advocate of this idea that entrepreneurship is, is maybe the thing that balances out the job loss. But I found myself wondering, are we seeing any signs of that yet? Like, are we seeing an increase in startups? So I just literally went into deep research in ChatGPT and I gave it the prompt. I said, one of the theories about how the economy will account for job losses driven by AI is that we will see a rapid increase in entrepreneurship and the number of startups created. Is there any data showing an increase in startup creation of the last 12 to 18 months? So I, I actually haven't gone through and read this whole thing yet, but it went through 33 citations and 341 search. Took 23 minutes to write me a report on startup creation, AI displacement and entrepreneurship since late 2024. And it has a bunch of charts and methodology and sources. And so I guess I'll just use it as a reminder of like, hey, sometimes that's a great use for AI is like curiosity. It's like I wonder and it can be at the most random moment and you can just like set. I mean deep research is an agent. Like it's going and doing its own thing. It's taking actions to like it builds a plan and then it goes and take actions. This is a form of an agent and it just goes to work and it does it for 23 minutes and then I forget I did it until I come back into here. But yeah, I mean sometimes those are the best use cases is just that spur the moment. Hey, I wonder if I could do this thing or if I could come up with this idea or if I could, like, create this visualization and then just throw it in there and see what happens. So be a fun one for me for the week. Now I gotta go read this, right?

A (76:26)

That's the key. All right, so one other recurring segment we've started doing is each week we spotlight one of the courses in AI Academy to give people kind of real, actionable takeaways from the course, whether or not you ever end up becoming an Academy member, just to give you some of the value for free that we're creating in AI Academy. So, Paul, I'm going to go through this week our AI for HR course series very briefly and kind of share some takeaways there.

B (76:51)

Sounds good.

A (76:52)

So what's really cool and interesting and also a little scary in AI for HR is that it is really at the front lines of how AI is changing traditional systems. So in our research and in creating this course, I'm the one who taught this, we found that, you know, it's AI is creating chaos across the core work of hr. I mean, not only is it creating huge opportunities for HR as a function, but they're running into real issues where candidates and employees are using AI too. And it's not necessarily bad to use AI in your job search, but it leads to all sorts of, like, really messy questions because we're seeing hiring signals get really compromised because candidates are using AI to not only game the system, but also just really, really kind of hack their way through the process. And it's like you can't use these traditional signals anymore to see if someone actually knows what they're talking about. So, you know, employees themselves, even after hired, are using AI to do their work in ways managers can't see. This is affecting everything from resumes to performance reviews to just overall productivity. And HR professionals have a really tough job right now, and that's kind of the big macro trend. And one of the practical takeaways that we teach in this course is for your average HR person, this can feel deeply overwhelming. There's so much going on in AI. There's so much to learn. They're already dealing with the fallout in a negative way sometimes with how the hiring process has changed. One way that we kind of teach and walk you through in this course is just a really simple framework to get started thinking about. Okay, I know ChatGPT does this. I've heard about Claude over here. Like, how do I wrap my head around the opportunities for me in my job. And we use this framework called just pretty simply the three A's. And the three A's are this like sequential order to think about AI, automation, augmentation and acceleration. So first you want to start looking at things like where can AI handle low level, low hanging fruit, repeatable work that you can literally have it do for you in order to save time. Because that's where 99% knowledge workers and HR professionals especially are really stuck as they are drowning in like reactive admin work that is not the best and highest use of their time. So automation is a key initial step. And you know, like back to that discussion, productivity is not everything, but freeing up some time so we can be more innovative can be really helpful. And then second is augmentation. So looking at, we walk you through a series of questions on how to actually surface these opportunities. Augmentation is using AI as a co pilot. So let's say you freed up time by automating some things with AI, you then can start doing more of the work you're meant to do. More strategic, more high value stuff. Well, AI can actually augment you there to supercharge and just accelerate the value you create there, which benefits you and helps you do better work, not just faster work. And then finally, over time, after you are effectively automating and augmenting your function, as the case may be, acceleration is kind of the bigger picture stuff, right? The AI agents, the more transformative projects. That's where we then walk you through thinking about not just what AI can do for you or how AI can make you better, but what AI can enable. That just was not possible before. So we're talking, you know, we go through a bunch of use cases and examples of that in the course. One really interesting one is, is I believe Shopify uses an internal talent marketplace completely driven by AI to match people internally to different roles. That's kind of a really structural, long term, almost sci fi use of this technology that completely upends how the company actually works. So that's kind of the practical starting point, is kind of running your work and asking yourself a series of questions through those three lenses to actually sequentially, step by step, without biting off more than you can chew, actually see some real value from AI right out of the gate.

B (81:04)

A lot of stuff we need to be applying to our hr.

A (81:07)

Exactly. Right, right, Yeah. I have to admit, I mean some of the stories even in this course and the case studies and even just some of the research, even stuff that didn't make it in. You're just like, I would not want to have to figure this out. Yeah. Right.

B (81:22)

Recently, where, you know, there's a lot of that overwhelm feeling.

A (81:28)

Yeah.

B (81:28)

Of trying to not only figure it out for yourself internally, how are we going to use it, but how do we manage and like recruit and hire people who are obviously using it in the process or. Yes, it's a very dynamic space right now.

A (81:41)

All right, Paul, last but not least, we've got a bunch of AI product and funding updates. So I've got these, a bunch of these teed up like last week. There are a lot of things going on. I'm going to run through these real quick and if there's anything that jumps out, you let me know.

B (81:55)

Go for it.

A (81:56)

All right, so first up, OpenAI launched ChatGPT Images 2.0, its first image model with native thinking capabilities. It can search the web, generate up to eight consistent images from one prompt, and produce, this is important, produce readable text that is accurate at 2K resolution. It is widely seen right now as like the number one image model out there and is making a lot of waves. OpenAI also launched ChatGPT for Excel and Google Sheets. This is a sidebar App that lets +Pro business and enterprise users build, edit and analyze spreadsheets in natural language and pull in connected ChatGPT apps alongside their data. OpenAI also announced Codex Labs plus partnerships with with Accenture, PwC, Infosys and other global system integrators to deploy codecs across large engineering organizations. Anthropic and Amazon expanded their partnership with up to 5 gigawatts of new AWS compute for Claude, a fresh $5 billion investment from Amazon. There may be up to 20 billion more following on that and a $100 billion 10 year commitment from Anthropic to AWS, plus direct availability of the Quad platform inside AWS. Anthropic also added a memory feature to Claude managed agents, which is now in public beta that lets agents retain and build on learnings across sessions via file based storage. And Anthropic is apparently running a live pricing test on a roughly 2% of new signups, with existing Pro and Mac subscribers unaffected as the company experiments with how CLAUDE code access is packaged across tiers. So also figuring out that pricing problem we were talking about about Google rolled out an upgraded version of its Deep Research agent built on Gemini 3.1 Pro, adding a new max tier for extended asynchronous reasoning, MCP connections to proprietary data sources and native in report charts and infographics Google also signed a new multi billion dollar cloud deal with Miramorati's Thinking Machines Lab, giving the startup access to Google Cloud infrastructure. And Microsoft, as we talked about, has made Copilot's agentic capabilities generally available in word, Excel and PowerPoint. So this just a little more detail here. And on the product side, this allows Copilot to take multi step app native actions directly inside documents, spreadsheets and decks for Microsoft365 copilot premium personal and family subscribers. At Adobe Summit 2026, Adobe rebranded Experience Cloud as CX Enterprise and introduced CX Enterprise coworker and yet another trend of agents, an agentic AI layer that orchestrates customer experience workflows across Adobe's stack. SpaceX struck a deal giving it the right to acquire AI coding startup Cursor for $60 billion later this year or pay 10 billion if it walks away from the acquisition while the two companies collaborate on model training using Xai's Colossus supercomputer.

B (85:01)

That's a wild one.

A (85:01)

That is a wild one, right?

B (85:03)

I'm not going to get into it because we're running on time here, but that one might be worth unpacking. There's. There's a lot to that story.

A (85:10)

Yeah.

B (85:12)

Another time,

A (85:14)

Tencent and Alibaba are in talks to invest in Chinese AI Lab deep seats first ever funding round at a valuation now of more than $20 billion, with Tencent reportedly pushing to 10 take as much as a 20% stake. Moonshot AI, which we've talked about in the past, released Kimik 2.6, a new open source coding model that claims state of the art scores on certain benchmarks and can run 4,000 plus tool calls across 12.12 plus hours of continuous execution. And finally, Zapier launched Zapier Benchmarks, a new AI evaluation suite anchored by automation bench that tests agents on end to end business workloads across sales, marketing, operations, support, finance and HR using deterministic scoring grounded in 2 billion plus monthly tasks from 3.7 million Zapier customers.

B (86:09)

Maybe we'll have Dan Slag and talk about that at Macon.

A (86:11)

Maybe. There you go. That would be awesome. I would love to pick his brain about that.

B (86:16)

I know that's Dan's domain, but Zapier's got a lot going on right now. We were talking about their, with their internal literacy stuff. Not like a week or two ago, right?

A (86:23)

Yep. All right, so Paul, that is it for this week. One quick final announcement here. Like we said at the top of the episode, this week's Pulse survey will be live when you listen to this at SmarterX AI forward slash pulse. We're going full on agent this week just like the topic. So we're going to ask about things about where's your organization at when it comes to deploying AI agents today and also what is holding your organization back from deploying AI agent more than you are already today. So I'll be very interested to see that, Paul, based on the answers from this week as well. But thank you for breaking everything down for us. Another busy week. I know we've done two episodes this week, so I feel like I've, I've got a pretty good pulse on what's going on.

B (87:10)

Yeah. And I was actually home for like two days in a row for the first time.

A (87:13)

Right.

B (87:14)

Yeah. And so yeah, next time we're together we'll, we will actually be back in town. So enjoy your travels. Good luck at Experience Inbound. And I'll be off to, I think at the time this drops, I'll be doing the Aquea Engage event and then we'll be back. We'll be back in Cleveland. I'll see you for our ChatGPT agents lab next week. That will report on the next episode. Yeah, I'm super anxious. I hope it's everything I think it could be.

A (87:41)

I, I'm very excited.

B (87:43)

All right, everyone, have a great week. Thanks for listening to the Artificial intelligence show. Visit SmarterX AI to continue on your AI learning journey and join more than 100,000 professionals and business leaders who have subscribed to our weekly newsletters, downloaded AI blueprints, attended virtual and in person events, taken online AI courses and earned professional certificates from our AI Academy and engaged in the SmartRx Slack community. Until next time, stay curious and explore AI.