57 Min Read

[The AI Show Episode 145]: OpenAI Releases o3 and o4-mini, AI Is Causing “Quiet Layoffs,” Executive Order on Youth AI Education & GPT-4o’s Controversial Update

By Claire Prudhomme on April 29, 2025

Podcasts
Featured Image

Wondering how to get started with AI? Take our on-demand Piloting AI for Marketers Series.

Learn More

After a quick spring break, Paul Roetzer and Mike Kaput are back, and the AI world definitely didn’t take a vacation. In this episode of The Artificial Intelligence Show, our hosts catch up on two weeks of major developments, including OpenAI’s surprising release of o3 and o4-mini, the accelerating wave of quiet AI-driven layoffs, and a new federal executive order on AI education.

Listen or watch below—and see below for show notes and the transcript.

Listen Now

Watch the Video

Timestamps

00:05:49 —o3 and o4-mini, and AGI

00:17:21 — AI-Caused “Quiet Layoffs” and Impact on Jobs

00:31:46 — White House Plan for AI Education

00:36:04 — Other OpenAI Updates

00:43:04 — Ethan Mollick Criticism of Microsoft Copilot

00:46:43 — Era of Experience Paper

00:54:23 — Chief AI Officers at Companies

00:58:54 — Anthropic Researcher Says Chance Claude Is Conscious

01:07:03 — xAI Funding and Updates

01:11:07 — Other AI Product Updates

01:13:40 — Listener Questions

  • I've been hearing about "AI Assistants” or “AI Agents.” Are these real things? Or just built-out versions of a Custom GPT?

Summary:

o3 and o4-mini, and AGI

OpenAI has launched two major new models: o3 and o4-mini, their smartest and most capable yet.

What sets them apart isn’t just better math or coding, it’s that they can now reason about when and how to use every tool inside ChatGPT. That means they can search the web, run Python code, analyze images, even generate visuals, all chained together, without needing a human to prompt every step.

The result? o3 sets new records across academic benchmarks and real-world tasks, while o4-mini offers lightning-fast, affordable reasoning, ideal for high-volume work.

Both models can think with images, not just about them, unlocking a new level of multimodal problem-solving. 

Right now, you can access up to 100 messages a week with o3 and 300 messages a day with o4-mini if you have a ChatGPT Plus, Team, or Enterprise account. (Pro users have “near unlimited access” according to OpenAI.)

AI-Caused “Quiet Layoffs” and It’s Impact on Jobs

Across industries, CEOs are quietly making a big bet, according to The Information, and that bet is: more AI, fewer jobs.

The publication writes: “Executives at more than half a dozen companies said AI has affected their hiring plans, though most were careful to avoid saying AI was effectively replacing existing employees.”

PayPal says its AI now handles 80% of customer service tickets, cutting support staff dramatically and saving hundreds of millions in the process. Cloud giants like Microsoft and Google are pitching AI as a full replacement for junior sales reps, IT staff, even software engineers. And firms are buying in, especially as fears of a Trump-era recession loom.

Executives admit that if a role can be automated, it's either being frozen or outright eliminated.

One EY leader put it bluntly: most clients now expect slower hiring or headcount cuts across their entire business. Meanwhile, firms like UWM and Fujitsu have built AI systems that double output without adding a single new employee.

It seems others are also starting to say the quiet part out loud. In an exclusive, Anthropic’s Chief Information Security Officer told Axios that the first fully AI employees are a year away. 

And a new startup called Mechanize, backed by heavy hitters like Nat Friedman and Daniel Gross, Patrick Collison, Dwarkesh Patel, Jeff Dean, and others has launched to develop “virtual work environments, benchmarks, and training data that will enable the full automation of the economy.”

White House Plan for AI Education

President Trump has signed an executive order to make AI education a national priority, starting from kindergarten.

The order creates a new White House task force that will coordinate AI programs across government, aiming to get foundational AI training into every K-12 school and expand opportunities for lifelong learning.

It also calls for a national AI Challenge to spotlight student innovation, and sets aggressive deadlines: within 90 to 120 days, federal agencies must launch partnerships with tech companies and universities, create online AI resources, and start funneling grant money toward AI-focused teacher training.

The plan goes beyond schools. It pushes for more AI apprenticeships, industry certifications, and even encourages high school students to earn college-level AI credentials.

This episode is brought to you by our AI for B2B Marketers Summit:

Join us and learn valuable insights and practical knowledge on how AI can revolutionize your marketing efforts, enhance customer experiences, and drive business growth.

The Summit takes place virtually from 12:00pm - 4:45pm ET on Thursday, June 5. There is a free registration option, as well as paid ticket options that also give you on-demand access after the event.

To register, go to b2bsummit.ai 

This week’s episode is also brought to you by MAICON, our 6th annual Marketing AI Conference, happening in Cleveland, Oct. 14-16. The code POD100 saves $100 on all pass types.

For more information on MAICON and to register for this year’s conference, visit www.MAICON.ai.

Read the Transcription

Disclaimer: This transcription was written by AI, thanks to Descript, and has not been edited for content. 

[00:00:00] Paul Roetzer: If you're a service provider, every time you put a proposal together, you need to be asking yourself, can o3 do this? Like I'm about to send a proposal to somebody for 10,000, 20,000, a hundred thousand, a million dollars, whatever it is. If this is a AI emergent business like ours would be, could they just use o3 to do this?

[00:00:20] Welcome to the Artificial Intelligence Show, the podcast that helps your business grow smarter by making AI approachable and actionable. My name is Paul roetzer. I'm the founder and CEO of SmarterX and Marketing AI Institute, and I'm your host. Each week I'm joined by my co-host and marketing AI Institute Chief Content Officer Mike Kaput, as we break down all the AI news that matters and give you insights and perspectives that you can use to advance your company and your career, join us as we accelerate AI literacy for all.

[00:00:57] Welcome to episode 145 [00:01:00] of the Artificial Intelligence Show. I'm your host, Paul Roetzer, along with my co-host Mike. Put, we are back after an extended break. We're on spring break. I was actually in Aruba, which was incredible. I had never been there. It was, if you have a chance to go, I would highly recommend it.

[00:01:15] My family and I went and, enjoyed it. Seven days, Mike. It was, it was crazy and It was, it was cool because like my kids are at the age where like. You know, downtime and kind of chills. They do Minecraft or whatever they're working on. And so I actually get like a couple hours on vacation each day to myself to just kind of think and work, be bigger picture things.

[00:01:36] Plus you got a seven hours in the planes. So I actually had like a, an incredibly productive and relaxing trip, which is like my favorite combination, my best guide. 

[00:01:46] Yeah. So I was like keeping up on the AI news and, you know, filling the, our sandbox for the podcast with stuff. Still did my newsletter, you know, a couple, a couple of Sundays, but there was no lack of things [00:02:00] happening while we were away.

[00:02:01] So, Mike and I are gonna do the all rapid fire approach again. so anybody who's new to the show, normally we do3 main topics where we, you know, kind of linger for seven to 10 minutes per topic. And then the rapid fire is usually like one to3 minutes. so to try and get through everything from the last two weeks, we're gonna try and go all rapid fire.

[00:02:21] Now. There are. A host of topics that each could definitely be a main topic. So I would say follow along here, but go to the show notes if you want to dive into anything that we cover. if you, if you aren't aware, the show notes always have the links to everything we cover. So if there's any topic that we don't get, you know, really far into today that you want to go explore more, you know, check out those show notes and go explore those topics as well.

[00:02:50] So, ton to get to. first I wanna touch on, this episode is brought to us by the AI for B2B Marketer Summit, presented by, [00:03:00] intercept. So, if people, again, are kind of new to the show, don't know how this all works, marketing AI Institute is, Mike's Chief Content Officer of Marketing Institute.

[00:03:11] I'm the founder and CEO of Marketing, AI Institute and Smart Rec. So Marketing Institute I created in 2016 that, that business is very much focused on, education, but also events. And so we have four core events through Marketing Institute. One of them is our AI for B2B Marketers Summit, which is a virtual event we just announced the agenda, so you can go check that full agenda out.

[00:03:38] It's happening virtually on June 5th. So this is, we have three virtual summits that we do, AI for writers, AI for agencies, and AI for B2B marketers. And so this one's happening June 6th at noon Eastern. you'll learn real world strategies to use AI for growth, better content, stronger customer relationships.

[00:03:56] And thanks to Intercept and our sponsors, you, there's actually a free tick ticket [00:04:00] option, so you can register at b2bsummit.ai. You can go learn more about it, check out the full lineup and get registered. So, as I said, there is a free registration option. There's a, there's a paid option as well, but you can register for free.

[00:04:14] And again, thanks to intercept in our partners. For making that possible. And then the other, you know, our flagship event at Marketing Institute is MAICON. This is our marketing AI conference. We started this in 2019, was the first one. So MAICON 2025, we will be back in Cleveland, Ohio, which is our home base.

[00:04:34] October 14 to 16. we've already announced 23 speakers, dozens of breakouts, main stage sessions, and the four hands-on, immersive workshops that happen on the 14th. So this is our sixth year where we're bringing this, bringing thousands of marketers together. We're expecting probably 1500 plus this year.

[00:04:54] And, you can check that out ticket prices go up May 31st. They kind of go up every 30 [00:05:00] days roughly. So, you know, try and get in there in these early bird pricing specials. save yourself a few hundred dollars so you can go to MAICON.ai. That's MAICON.AI. To learn more about, about that event.

[00:05:13] We still have a lot more speaker announcements coming. We're working on some really exciting things. But again, you can check out the first couple dozen speakers and learn more about, the breakouts and main stage sessions that are gonna be coming up there. So, yeah, check out AI for B2B marketer Summit for a virtual event on June 5th, and then join us in Cleveland, October 14th to the 16th for Mayon 2025.

[00:05:35] Okay, Mike. We got new models, we got research reports, we got calls for interpretability from Rio Ade. We got a whole bunch going on education initiatives from the government. Like, let's kick it off. 

[00:05:49] o3 and o4-mini, and AGI

[00:05:49] Mike Kaput: All right, Paul, so, and as you a reminder for what Paul said previously, we are covering the last couple weeks of news.

[00:05:56] So some of this stuff may not have happened exactly just [00:06:00] last week, but it is stuff that we have not yet covered. And first up is a big one, which is OpenAI just launched two major new models. There's now o3 and o4 mini. So these are their smartest, most capable models yet. And what sets them apart isn't.

[00:06:18] They're just better at math and coding and writing. It's that they can also now reason about when and how to use tools inside ChatGPT. So these things come outta the box, ready to search the web run code, analyze images, generate visuals. All of this starting to be chained together without you needing to necessarily, you know, select a bunch of different tools or prompt every single feature or functionality.

[00:06:45] So o3 is setting new records across academic benchmarks and real world tasks. O four mini offers. Lightning fast. Affordable reasoning is ideal for high volume work, and both models can think with images, not just about them. [00:07:00] So there's a whole new level of multimodal problem solving. Right now you can access up to 100 messages a week with o3 and 300 a day with O four mini.

[00:07:12] If you have a chat GT plus team or enterprise account, according to OpenAI, pro users have quote near unlimited access to these models at the moment. Now, what's really interesting here, Paul, is that oh three in particular is making some serious waves due to just how intelligent this thing seems. There are some prominent voices out there, including the popular economists, Tyler Cohen, who have said straight up, they believe this model is essentially artificial general intelligence or AGI.

[00:07:47] So Paul, I know you and I have both been really impressed with o3. Maybe walk me through your initial impressions. Maybe give me a sense of what you think about all this commentary about it being actual AGI. [00:08:00] 

[00:08:00] Paul Roetzer: There's definitely been lots of the AGI, context. I think a lot of people starting to wonder if we're not, you know, really on this accelerated path to it.

[00:08:10] And if this isn't, I. Kind of an early preview. 'cause I think o3 PRO is gonna come out soon as well. Yeah. Like, I think, yeah. So there, there's a more powerful version coming. I've also seen quite a few reports that halluc hallucination rates are higher with o3. So just sort of a, a, you know, a, a, a user warning.

[00:08:29] It seems super impressive and it really is, but. Again, if you're depending on this thing for work that you're going to turn in for things that you're gonna put out into the public, you have to be very vigilant on the accuracy and reliability of the outputs. So just kind of a, a note there, a couple of people that surface for me when I was looking at reactions here, Alexander Wang, who we've talked about numerous times on the podcast, the CEO at scale ai, which is a company that works with all these big model companies to do the training, you [00:09:00] know, provide the data, things like that.

[00:09:02] for the training. So he said open AI oh three is a genuine, meaningful step forward for the industry. Emergent agent agentic tool use, working seamlessly via scaling reinforcement learning is a big breakthrough. It is genuinely incredible how consistently OpenAI delivers new miracles. Then Bob McGrew, who's the former chief research officer at OpenAI, tweeted that the defining question for AGI isn't quote, how smart is it, but quote, what fraction of economically valuable work can it do?

[00:09:34] The spotlight for oh three is on tool use because intelligence is no longer the primary constraint. The new frontier is reliable interaction with the external world. So just a reminder, like, you know, as we talk about AGI, and you know, again, people follow the show, know we have an entire new series dedicated to per like, kind of following this road to AGI and beyond.

[00:09:55] I think it's really important that people continue to remember [00:10:00] we don't need to reach it or agree on it, that we have reached it for it to transform everything. So just using o3 myself over the last week, you really start to increasingly see it doing the things that I would otherwise be paying advisors and consultants to do, or the things that we would traditionally be hiring someone to do.

[00:10:24] So, as an example, while I was in Aruba, we had to make a kind of a relatively quick decision on internet for the office. So we, you know, have internet in the office. we need to upgrade it. It is not my area of expertise. It's not something as the CEO of the company I've even had to think about for like five years.

[00:10:41] 'cause we did this before and it's been working fine. But as we're scaling up our company, we have to rethink how we're handling the internet, make it more reliable, more stable, things like that. So we get a quote from a vendor, Tracy, our COO sends it to me. She and I go back and forth. I've got questions, she's got questions.

[00:10:58] Neither of us are [00:11:00] experts in this field. So I was like, screw it. Like I'm just gonna go into o3 and like, let's just have this conversation. Hey, you're a senior IT advisor. We're trying to solve for this problem. And it analyze things in ways like I've been paying IT people for 25 years running my companies.

[00:11:15] It helped me understand more deeply how to solve this than any IT person I've ever talked to. And I was able to just like continue as I say, I don't understand this. Can you explain this for me better? Can you gimme examples of why I would care that this is the difference? And so rather than like me reaching out to my IT person and then waiting five hours for a response that I might not understand in the moment when I had 20 minutes, I just did it myself.

[00:11:42] I just solved the thing. And so you start to realize like. I don't necessarily have to have deep expertise here. I know enough, having managed my internet as a CEO for 20 years, what I need and don't need. I just needed some guidance and like some, some frameworks to [00:12:00] help me make a decision. So in a matter of about 20 minutes talking to o3, I made a decision, replied to Tracy.

[00:12:06] I was like, okay, let's go and here's what we're gonna do. And then I shared that chat with Tracy so she could also see the context of why we were making that decision. . And then she could continue on and see if she had any other questions as well. So that is a prime example of something I would have absolutely paid an advisor for.

[00:12:23] Same time, I'm working on this massive organizational design strategy for our company because again, as we're kind of scaling up and new complexities around size of the staff, compensation models, all these challenges that we haven't had to really face, and even when I was running my agency, we peaked at like 20 employees or something.

[00:12:41] So I never had to design an organization that could scale to a hundred plus employees, which is what I'm now having to kind of like, envision is like, okay, we have to make decisions now that can get us to stable, growth from like 50 to a hundred employees if we choose to go that route. But now [00:13:00] I'm, I'm out of my league.

[00:13:00] Like this isn't what I've done. I haven't run a company with a hundred plus people. So again, I could pay and probably 50 to a hundred thousand dollars for the specific thing that I was looking to do, or I could do it myself with oh three, which is what I did. And what I realized in the process of doing this over a few days on vacation was.

[00:13:20] Rather than paying someone to give me a report and say, here's what you should do, that I would then have to sit there for hours reviewing, analyzing, trying to make sure I understood the recommendations so that I could then make an educated decision. I just did all the work myself with o3. Now, I, you know, I kind of knew the prompts to give it, like the questions to ask, but the main value of the project became my ability to critically question the outputs of the model.

[00:13:49] Be like, well, why are you saying that? Like, where are you getting this data from? And it would show me the citations. It's like, so it became this like immersive experience to where I am gonna have a much greater [00:14:00] confidence level in the final output because I was bought into the process and I was able to ask all my questions along the way in real time.

[00:14:09] And so it really just starts to change the way I think about. How we do knowledge work. And we talk about this mike a lot on the show, but like, these are very practical examples where I just saved myself probably a hundred plus hours of time and work. . And probably a hundred thousand dollars in expenses.

[00:14:27] And I actually feel better about the end product, which by the way, the other thing I'm gonna do is take that end product before I operationalize it, and I'm gonna use other models as critics to evaluate what I ended up at working with o3. So I'll take the final output, I'll put it into Gemini 2.5 and say, just basically start from scratch.

[00:14:46] Hey, here is the organizational structure I'm looking at. Here's the decisions I've made, please assess this for me. You know, please criticize or, or look critically at different areas and challenge these decisions. And now I'm not just dependent upon a single model that might be [00:15:00] hallucinating. I can actually vet it against one or two additional models that maybe take a different perspective.

[00:15:05] And again, I end up at the, you know, a place where I'm just more confident in the final product. So, I don't know, Mike like it just. Changes things, and I know you and I talk about this all the time, but like when you know this stuff, like when you can do what I just explained. When you start to look at problems in your business differently, because, you know, AI can help you do it, you can run a business in a, or a department or a team or a campaign in entirely different ways when you know how to work with these tools.

[00:15:34] it really is like hard to comprehend if you're not actively doing it, but it's so transformative. 

[00:15:42] Mike Kaput: Yeah, and I would also say too, I'm certainly biased here, but this alone justifies the cost of 200 bucks a month having easy, unlimited access to it. I mean, you just described, I just saved 5,000 in my IT bill.

[00:15:56] Yeah. One 

[00:15:58] Paul Roetzer: project I could write it. [00:16:00] 

[00:16:00] Mike Kaput: Yeah, exactly. And I also would just to kind of wrap up here, be really blunt, honestly, and say if you are a professional services provider, like a lawyer, an accountant, an IT consultant, what have you, you need to run and not walk to go spend $200, or you can get in a plus account in limited usage and put this thing through the paces of the hard questions clients ask you.

[00:16:25] Because I would be really starting to think about how do I become the guy that they then go hire after they've done this initial thing themselves? 

[00:16:34] Paul Roetzer: A hundred percent every, every time. If you're a service provider, every time you put a proposal together, you need to be asking yourself, can oh three do this?

[00:16:42] Like I'm about to send a proposal to somebody for 10,000, 20,000, a hundred thousand, a million dollars, whatever it is. . If this is a AI emergent business, like ours would be, could they just use o3 to do this or 80% of this? Because the answer is going to increasingly [00:17:00] be yes. As we get to a higher level of awareness and AI literacy for leaders at these companies right now, it's still early and we're still very much in kind of the early adopter innovators phase where very small percentages of companies and leaders are aware they can do this in place of hiring you, but it's going to change pretty quickly.

[00:17:21] AI-Caused “Quiet Layoffs” and Impact on Jobs

[00:17:21] Mike Kaput: And you know, the second topic we're covering is kind of related to this because. We're seeing some reports, especially according to the information that CEOs are quietly making a big bet. And the bet is the more AI we're using, the fewer jobs they're basically gonna need to hire for. So the information came out with a report where they wrote, quote, executives at more than half a dozen companies said AI has affected their hiring plans.

[00:17:48] Though most were careful to avoid saying AI was effectively replacing existing employees. So for instance, they cite all these case studies where PayPal says AI now handles 80% of customer [00:18:00] service tickets, cutting support staff dramatically. Cloud giants like Microsoft and Google are literally pitching AI as a full replacement for junior sales reps.

[00:18:09] IT staff, in some cases, even software engineers and executives are starting to admit that if a role can be automated in these ways, by ai, especially as they're fearful of kind of. Possible economic issues coming up through some of the policies we're seeing through some of the headwinds we're experiencing.

[00:18:27] If a role can be automated by ai, they're starting to say it's either going to be frozen or outright eliminated. So one, leader at ERNs and Young at EY put it bluntly and said most of their clients now expect slower hiring or headcount cuts across their entire business. And we've also seen a couple related reports where others.

[00:18:48] Are starting to kind of, let's call it, say the quiet part out loud. In an exclusive report with Axios Philanthropics, chief Information security officer said that the first fully [00:19:00] AI emplo AI employees are a year away. There's a new startup that's on our radar called Mechanize, which is backed by heavy hitters like Nat Friedman and Daniel Gross, Patrick Collison of Stripe, Esh Patel, Jeff Dean, and others.

[00:19:13] And their explicit mission is to develop, quote, virtual work environments, benchmarks, and training data that will enable the full automation of the economy. So Paul, there's kind of all these threads coming together, and I think you summed up the central point here really well. In a recent LinkedIn post where you said, quote, my belief is that quiet AI layoffs have been happening for the last six to 12 months.

[00:19:39] For instance, masked under a return to work policy, and they are accelerating. Companies have been replacing staff with ai or at minimum not hiring new staff due to ai. But they don't, just don't wanna admit it because it's bad pr. Can you maybe walk me through what you see going on right now? Like what do you expect to be happening in the near future [00:20:00] with AI's impact on jobs?

[00:20:02] Paul Roetzer: Disruption and displacement of jobs, which is what we've been saying for the last, you know, 18 plus months on this show is that this is coming. And I just don't think people were ready to hear it. Like, I don't think people wanted to admit it or I maybe they just didn't understand fully what these things were gonna be capable of doing, and they were just in denial that it was gonna be possible.

[00:20:23] But it's absolutely what's happening. It's what I've been hearing sort of behind the scenes now for six to 12 months. It's what we're now seeing people saying publicly, it's not gonna be evenly distributed across industries. So I think that some people may hear this and be like, yeah, I'm just not seeing it.

[00:20:40] And it's like, that's fine. Like maybe in the legal industry they're just slow moving or banking or financial services, like whatever it may be. There. There's different reasons why different industries might not move as quickly. It's absolutely what's going to happen. And you know, I think we had this return to work policy was the natural cover initially.

[00:20:58] So it started in the [00:21:00] tech space. Like, it's like okay, you gotta be back in the office four days a week, three days a week. They know 20% of people are gonna refuse to do that and they're just gonna leave the job great. We just cut 20% of our staff without having to say we were replacing them with AI basically over the next two years.

[00:21:13] and now the latest cover is gonna be tariffs in the economy. Yeah. So, you know, things aren't great in the economy right now. There's increasing chatter that we're heading towards a recession. and that's going to give the impetus to say, well, we gotta cut costs anywhere we can. And if that means people, it's, it's people.

[00:21:30] Now again, they're not gonna say we're doing this because we don't think we're actually gonna need as many people 'cause we're gonna use AI to do a lot of this work. That won't be the lead talking point, but it's probably gonna be the underlying thing that's actually causing this is there's increasing confidence by C-suites and boards.

[00:21:46] They don't need as many people to do what they do, and I think they're probably right. This is, this is the thing I've been saying all along is like, we just don't need as many humans doing the current work. So like if you [00:22:00] take, you know, if all things being we say like, we do these 100 things, or we create these 100 widgets every month moving forward, we have 50 people creating those a hundred widgets, maybe we can do it with 20 instead.

[00:22:13] So if you're, if the creation of the output, the product or service remains flat, you just don't need as many people to do it. Now, if there's tremendous demand for what you do and you have a ton of growth and new markets and new products, then great. You, you may keep hiring people. You don't need as many, and your revenue per employee number should, in theory, be skyrocketing over the next two years because you should be able to create more output, generate more revenue per employee.

[00:22:38] If you don't, you're doing something wrong. If you don't start to generate a higher revenue per employee number, you are mismanaging your company and it's because you just don't need as many people. So this is why I've said like it's the best time ever to be a startup because you can just build more intelligently, you can build with fewer people, build with smarter processes.

[00:22:58] You don't even have to deal [00:23:00] with the quiet layoffs. . You just grow smarter. But if you're an existing company that has dozens, hundreds, thousands of employees, you got some really challenging times ahead to manage that headcount. now the mechanized thing, Mike, that you mentioned. Yeah, man. Like that's wild.

[00:23:17] Talk about just coming out and saying it like, yeah. So this is the kind of startup we have absolutely expected. I remember we talked a few episodes back, this is, I don't know, maybe like 10 episodes back, could be longer. We talked about, Y Combinator and how they were like investing in vertical agent companies.

[00:23:33] Yeah. This is exactly what we were talking about, that people are going to build companies that automate workforces by industry. Now Mechanize just comes out, strata says, we're just gonna automate the entire economy. Literally, they just, their mission. So even though we've known this was what people were going to build, it was what venture capitalists were going to invest in.

[00:23:52] It's jarring to actually see someone come out with the mission statement. So. Spend a a, a moment on this. And I know we're not doing main [00:24:00] topics, but man, we gotta talk about this one. So Mechanize was launched on April 17th via a Twitter post, by its founder who's a, a researcher, Tamay, Bess Belu. who is Epic AI was I think he created it, or he is lead researcher there, which is a research institute that investigates key trends and questions that shape the trajectory and governance of ai.

[00:24:23] So the startup goal, according to be Solu, is the full automation of all work and the full automation of the economy. Now in the tradition of recent ai, major startups like Safe Super Intelligence and Thinking Labs, their website is a single page of text with no images, nothing. It's just like text in some links.

[00:24:44] And in there they kind of go over the, some of the information that they tweeted, which is today we are announcing mechanized. And this is just direct quotes, A startup focused on developing virtual work environments, benchmarks, and training data that will enable the full automation of the economy. We will [00:25:00] achieve this by creating simulated environments and evaluations that capture the full scope of what people do at their jobs.

[00:25:06] This includes using a computer completing long horizon tasks that lack clear criteria for success, meaning there's no goal you can set. It's just like you gotta figure out what the milestones are along the way, coordinating with others and reprioritizing in the face of obstacles and interruptions. 

[00:25:21] Now, I'll pause for a second here before I continue. one of the things we've been talking about recently is the need for evaluations and benchmarks in AI to not be tied to IQ tests, but to actual jobs. This is exactly what they're doing. So they're using that to then inform the building of smarter models.

[00:25:38] Okay, continue. mechanize will produce the data and evals necessary for comprehensively automating work. Our digital environments will act as practical simulations of real world work scenarios, enabling agents to learn useful abilities through reinforcement learning. Now, here's where it gets kind of crazy.

[00:25:55] The market potential here is absurdly large workers in the US are paid [00:26:00] around 18 trillion per year in aggregate for the entire world. The numbers over three times greater, around 60 trillion per year. So this is their total addressable market is 60 trillion per year. The explosive economic growth likely result from completely automating labor could generate vast abundance, much higher standards of living and new goods and services that we can't even imagine today.

[00:26:22] Our vision is to realize this potential as soon as possible. So they're directly saying the thing no one has been willing to directly say, which is . They plan to take all knowledge work and you say how to do it as quick as possible. Now, the part that was somewhat shocking to me was the investors.

[00:26:41] Now you highlighted them, Mike, but like Nat Freeman is the GitHub, CEO. So he is the GitHub is Microsoft, right? Microsoft bought GitHub I bank, yeah, yeah, yeah. tech investor, Daniel Gross, Stripe co-founder and CEO Collison. You mentioned Duar Kesh who actually just did a podcast with these guys. .

[00:26:57] So I'll come back to that in a second. Jeff Dean, if you're not [00:27:00] aware of that name, is Google's chief scientist. Jeff Dean is like one of the godfathers here of like modern AI and then a couple of key investors. So. The fact that these people, these guys are all behind a company that is directly saying, we plan to intelligently automate all knowledge work in the economy.

[00:27:19] And I haven't seen a comment from any of these guys. Like I'm really curious to hear their positioning on this. But, so if you wanna go deeper on this, I know like I'm flying to Boston in two hours, so as soon as we get off this, I'm jumping on a plane to Boston. I know what I'll be listening to, which is Dwarkesh podcast with Tamay and Ege Erdil, who is, I think his partner in this.

[00:27:37] But he just dropped the podcast the day this got announced and then in a tech crunch article. Tamay referred to a research report that he and Egg a published in 2023 called Explosive Growth from AI Automation, A Review of the Arguments. . So now I think this paper is actually the prelude to [00:28:00] mechanize.

[00:28:00] And so in that paper, he said We examine whether substantial AI automation could accelerate global economic growth by about an order of magnitude 10 X akin to the economic growth effects of the industrial Revolu revolution. We identify three primary drivers for such growth, the scalability of AI of an AI labor force restoring a regime of increasing returns to scale.

[00:28:22] Two, the rapid expansion of an AI labor force. And three, a massive increase in output from rapid automation occurring over a brief period of time, we conclude that exclusive growth seems plausible with AI capable of broadly substituting for human labor, but high confidence in this claim seems currently unwarranted.

[00:28:41] Key questions were made about the intensity of regulatory response, physical bottlenecks, and the production of economic value of superhuman abilities, and the rate at which AI automation could occur. So the key takeaway here is mechanize is not alone. There is the first ones to come out publicly and say, this is what they're doing.

[00:28:57] A Andreessen Horowitz is [00:29:00] probably funding 10 companies that are trying to do this exact same thing, like this is going to be pursued. I'm not saying it's achievable. I'm not saying like the to address of market is reliable. I'm just telling you, venture capital is going to pour hundreds of billions of dollars.

[00:29:16] If you look at something with a tens of trillions of dollars of market potential, that means they're willing to put in hundreds of billions in the next three years to pursue this idea. So I know this wasn't supposed to be main topic, but I won't talk about this without doing this. there's so much uncertainty about what this all means, and I get all the anxiety.

[00:29:37] The thing I keep coming back to is like, if this is blowing your mind, file it away, know it's happening, and go back to your work and just do the next best thing to increase your own literacy and capabilities here. Like it's not gonna happen overnight. It is, you know, I think I've said it's a, you know, it's kind of like climbing up a hill, [00:30:00] not falling off of a cliff at the moment.

[00:30:01] And so like you have a chance to sort of be out on the frontier here and like figure this stuff out as it's going. I get that it can cause anxiety, but like I wouldn't let that overwhelm you. I would just like go, go do the next things. 

[00:30:15] Mike Kaput: I would also say it, the silver lining to me here as well is if they think this is possible, you know your own job better than, very likely, better than these people do.

[00:30:24] So you can go. Start figuring this out for yourself. Not necessarily automating everything away, but whatever productivity and performance gains they believe are possible, you probably are the best suited person to go figure that out in your own chop. 

[00:30:38] Paul Roetzer: Yeah, and I do think that there's quite a window here.

[00:30:41] Like I don't see what they're trying to do is like a 2027 Yeah. Outcome. You know, I think it's gonna be by industry, but again, I think if you're in like AI research, financial analyst lawyer, like there's just gonna be some really obvious industries that this stuff's gonna hit sooner than others. And I would [00:31:00] not be ignorant to it.

[00:31:00] Like I think that's the key is like you have to educate yourself on not only what these these models are capable of, but what they're capable of in your industry. Because things like, we started this with, people are gonna stop accepting your proposals 'cause they know they can do the work cheaper.

[00:31:14] . People are gonna stop hiring you as a professional because they just don't need as many of you in their. In their company anymore. So you're gonna start to see earlier signs here, sort of like canary in the coal mine kind of stuff, like it's coming, but what they're envisioning isn't probably like a near term, like three to five year reality.

[00:31:33] Now, anything beyond that, as I've said on the road, AGI podcast, like I I can't help you, like, I don't know, beyond like three years. It's really hard to project the reality here. 

[00:31:46] White House Plan for AI Education

[00:31:46] Mike Kaput: In our next topic, president Trump has signed an executive order to make AI education a national priority starting from kindergarten.

[00:31:55] This new order creates a new White House task force that will [00:32:00] coordinate AI programs across government, aiming to get foundational AI training into every K to 12 school and expand opportunities for lifelong learning. It also calls for a national AI challenge to spotlight student innovation and set some aggressive deadlines within 90 to 120 days.

[00:32:18] Federal agencies must launch partnerships with tech companies and universities, create online AI resources, and start funneling grant money towards AI focused teacher training. This plan also aims to go beyond schools. It pushes for more AI apprenticeships, industry certifications, and even encourages high school students to learn to earn college level AI credits now.

[00:32:44] Paul, I found it really interesting. The federal government is actually starting to at least talk about this in a serious way. how much substance do you think there is to this initiative? 

[00:32:54] Paul Roetzer: I don't know. I mean, it's the first time I've heard this administration say anything on this topic, so it [00:33:00] kind of came out of nowhere in my opinion.

[00:33:02] I don't know who's actually the driver of this. I think it's, it's a very smart initiative. Like I say, I don't know who is the sponsor of this. Like I don't, I'm not sure where that's coming from. but it's the kind of initiative that we've been calling for on the show for a couple years that the government had to get involved in the eye literacy.

[00:33:20] This is like absolutely essential, and I like the idea at least, like, again, that much is known about this. It's like, Hey, in 90 to 120 days, come back with a plan, is basically what this executive order says. But we have to teach the teachers, we, we have to make the technology accessible to students, which we've seen.

[00:33:37] I think open philanthropic. Google have all kind of made their, models free for college students. I think in the last like, you know, month, we've seen announcements around that. we have to teach the responsible use of it. This can't be handled the way enterprise adoption has been, which is, hey, here's a thousand licenses to co-pilot.

[00:33:54] Go figure it out. Like, if we're gonnAGIve the technology starting at kindergarten all the way up, [00:34:00] we have to actually teach the students and the teachers how to use the technology. So, as I said, like I'm, I'm a bit skeptical because It's the first I'm hearing of this. Yeah. And like, I don't, AI was not mentioned on the campaign trail once by, by either, you know, potential administration.

[00:34:18] So this idea of like this massive investment in AI literacy, while I love it, I don't necessarily know that they're really committed to it or that they even really understand the importance of it. Whoever wrote this seems to, but I don't know that the administration at large actually. Believes this is like critical.

[00:34:36] I really would love to be wrong on that though. Like I'm totally open-minded and I'm very optimistic about the approach. in the fact sheet, they said the executive orders to create new educational workforce development opportunities for America's youth fostering interest and expertise in AI from an early age.

[00:34:52] Love that. That's great. . early training in AI will de demystify this technology and prepare America's students to be confident participants in the AI [00:35:00] assisted workforce propelling our nation to new heights. Absolutely. Again, whoever's writing this, it might be o3, I don't know, but like whoever's writing this, that gets really good.

[00:35:08] preparing our students to be leaders in AI technology requires investing in our educators a hundred percent. providing them with tools and knowledge to both train students about AI and utilize the tech in the classroom. And then they said lifelong learners also need new resources to develop technical skills for rapidly evolving work environment that increasingly incorporates digital technology.

[00:35:27] So again, on the surface, this sounds like a really, really positive direction. anything that involves more government interest, action, and funding around AI literacy, I am absolutely for. So 

[00:35:39] Mike Kaput: yeah, 

[00:35:39] Paul Roetzer: I would love to see this come to light and be real and to actually have like full government support. 

[00:35:46] Mike Kaput: And just to reiterate our previous topics, again, if you haven't used o3, it a hundred percent could have written that with the right input.

[00:35:52] So no doubt. Go test out. 

[00:35:54] Paul Roetzer: Hey, hi. Hi. It's really important to America's youth. Go. Yeah. Create a fact sheet and draft an executive order. I say [00:36:00] that because 

[00:36:00] Mike Kaput: of how good it is. Like it is that good. It's really 

[00:36:02] Other OpenAI Updates

[00:36:02] Mike Kaput: good. All right, so for our next topic, I'm going to run through a bunch of. Open AI related updates because they had a ton going on since our last episode.

[00:36:13] And then Paul kind of just let you weigh in on whichever of these you find the most noteworthy. 

[00:36:17] Paul Roetzer: Sounds good. So 

[00:36:18] Mike Kaput: first up on Friday, April 25th, Sam Altman posted that GPT-4o got an update that quote, improve both intelligence and personality. So according to the company's model release notes, this included making what they call subtle changes to the way it responds.

[00:36:36] But this may not be that subtle because a bunch of people online claim that right after the update, 4o's personality became kind of annoying and very focused on being basically like a yes man. Like, like telling you only what you want to hear in this like really annoying, enthusiastic way. And Altman actually responded to these claims and said the company is working on a [00:37:00] fix there.

[00:37:01] OpenAI also launched GPT-4o.1 in the API, which. Is only available in the API won't, won't be available in ChatGPT, the tool. And it's focused specifically on real world developer needs. So it's got made huge leaps they claim in coding following instructions and it has long context understanding with a context window of up to a million tokens.

[00:37:26] OpenAI is also apparently building a social feed inside ChatGPT according to some either rumors or facts being reported on by the information. this new feature would let users post and share how they're using the chat bot, basically kind of like a mini social network. Internally, what they are calling, making a post in the feed, posting to the feed is called a yeet.

[00:37:51] Yes, that is really the word you are thinking of. And this idea is to help chatGPT's massive user base, which is now over 500 million [00:38:00] people a week. Better understand what this chat bot can actually do now. Also according to internal projections seen by the information, OpenAI expects to hit 125 billion in annual revenue by 2029 and 174 billion by 2030.

[00:38:20] That is roughly the size of something like Nvidia or Meta today, and they're betting a lot of that growth will come from agents and new products like shopping assistance and free user monetization. They even project that by 2029 agents alone could bring in 29 billion a year. They could be selling high end AI workers ranging from $2,000 a month, knowledge agents to $20,000 a month research agents. Now also, openAI is getting taken to court again by another major publisher, Ziff Davis. The company behind sites like PC Mag Mashable [00:39:00] Lifehacker has filed a lawsuit accusing OpenAI of copyright infringement and trademark dilution. They are seeking hundreds of millions of dollars in damages according to insiders, and OpenAI says it is using the material in a way that is grounded in fair use.

[00:39:18] Now, last but not least, a number of AI leaders have added their name to an open letter calling on US State Attorney's General to investigate open AI's plans to transition from a nonprofit to a for-profit company. This letter is titled Not for Private Game, and it's signed by literally dozens of top AI researchers, legal experts, and even Nobel Laureates.

[00:39:41] Some of the notable signatories include Jeff Hinton, one of the godfathers of ai, ai ethicist, and researcher Margaret Mitchell. And at least 10 former employees as far as I could count from open ai. And the letter argues that the company's plan to restructure a hand [00:40:00] control to a for-profit entity violates its original nonprofit mission, which we've talked about before, which was to ensure that AGI benefits all of humanity.

[00:40:09] They think allowing this structure to go forward could essentially allow private investors to capture and monopolize the value of AGI. So Paul, it's been a busy couple of weeks as always for open ai. Did you, did any of these like jump out to you as particularly noteworthy? 

[00:40:26] Paul Roetzer: Yeah, I mean we could obviously talk about any one of these at length, so I'll just, I'll stick with the 4o thing.

[00:40:31] you know, I think one, it's interesting to note, just they have this iterative deployment. Plan, that's their strategy at opening eyes. Just like, just keep putting things out into the market, see what happens, see how people respond to it. So Sam basically just tweets like, Hey, we've made some updates to 4o.

[00:40:46] No context at all as to like what those updates are. Yeah, like how it's different. But if you go to, they have a model release notes page, which they honestly don't really keep that updated as regularly as well. But what I saw his tweet, I'm like, there's gotta be something more to this. [00:41:00] So I went there and they had in fact put an update a little bit of what it was.

[00:41:04] So what the update was on April 25th is when this, they published this, it said, we're making additional improvements to GPT-4o optimizing when it saves memories and enhancing problem solving capabilities for stem. We've also made subtle changes to the way it responds. Making it more proactive and better at guiding conversations toward productive outcomes.

[00:41:25] We think these updates help GPT-4o feel more intuitive and effective across a variety of tasks. We hope you agree. And then as you alluded to on April 27th, after lots of feedback of this thing is really annoying, Sam actually tweeted this would've been Sunday night, the last couple of GPT-4o updates have made the personality sycophanty and annoying, even though there are some good parts to it and we are working on fixes asap.

[00:41:50] And then the part I thought was most interesting, he said at some point we'll share our learnings from this. It's been interesting. 

[00:41:57] This just alludes back to the thing we [00:42:00] talk about a lot on the show is like, they don't know how these things work. Like they don't know why it became annoying and like there's something that happened, some changes that they made where this thing all of a sudden just started.

[00:42:13] Becoming a yes man. Like you said, like it's just like, oh, you're great. Like I love you. Like, oh, that's so brilliant, and rather than being like a critic and helping you, and so, but like how that happened and what they have to do to like, try and fix it from being annoying. They don't know. And like, they gotta kind of go in and try and like, figure this out.

[00:42:31] and that's just weird. and you know, so we'll talk a little bit more about this in an anthropic topic come coming up. But like these models, they aren't programmed the way traditional software was programmed to just follow instructions, right? Like they, they have, this sounds weird, but like, they have a mind of their own and sometimes it's on the research trying to figure out why they do what they do.

[00:42:56] And sometimes it's just not very obvious why. And like, what needs [00:43:00] to be done to fix it back. 

[00:43:04] Ethan Mollick Criticism of Microsoft Copilot

[00:43:04] Mike Kaput: So in some Microsoft related news, in response to a post from Microsoft, CEO, Satya Nadella. Prominent AI expert, Ethan Mollick has criticized the company's co-pilot AI tool or product. Now, Nadella recently posted about a bunch of new features within co-pilot that he was excited about.

[00:43:26] This included its researcher and analyst agents, which we covered on a past episode, and in reply, Molik said quote, Microsoft keeps launching copilot tools that seem interesting, but which I can't ever seem to locate. Can't find them in my institution's enterprise account, nor my personal account, nor the many co-pilot apps or co-pilots to apps or agents for co-pilots.

[00:43:48] Each has their user interfaces. So we wanted to just quickly highlight this because it does come from a prominent, credible voice in ai. I mean, Ethan Mollick is one of the top people out there to follow. I [00:44:00] mean, Paul, this just certainly doesn't seem like a good look for Microsoft. It starts to kind of explain a disconnect We've heard from people about the value that's actually being created by copilot.

[00:44:11] What do you think? 

[00:44:12] Paul Roetzer: Again, we don't use copilot internally, so I can never speak directly to, you know, copilot experience. What I can tell you is if you have kids and you've ever tried to manage their Minecraft account through Microsoft, you know exactly how this goes. Like, it is the craziest thing ever, how complex it is to manage Microsoft accounts, especially if it's across multiple products.

[00:44:35] So if the copilot experience is anything like being a parent of a child who uses Minecraft, like good luck. what I will say is contextually we have Google Workspace, so I can speak to Gemini's experience. It sounds like Gemini's probably a little bit better, but I suffer from the exact same thing with Google.

[00:44:55] . So they announced what vo I think like the video model [00:45:00] supposedly was available in Gemini. I don't have it. And like I, and then I saw a tweet like five days later. It's like, oh, like we're starting to roll it out. It's gonna take a little while. It was like, okay, well that would've been done nice to lead with like when you announce that VO is now available.

[00:45:14] And then I always laugh because I have my personal Google account with Gemini that generally gets this stuff before our workspace account gets it. And so I'll go in, like, I just have both tabs open. It's like each day it's like, oh, nope, not there yet. You never know when you're going to get the thing or like which version you have open.

[00:45:35] AI seems to probably do this best in terms of like, they roll the models out fastest to their, their, their customers. Like if they say something's coming, it usually happens pretty quickly. It's still confusing as hell. Like I still don't, and same thing, I have a personal ChatGPT account. I pay the 200 a month for, and I have my business account.

[00:45:57] I never know which thing is in which account and which [00:46:00] models underlying, you know, custom GPTs. And so again, like Mike and I live this stuff 24 hours a day, and I'm lost half the time. So like if you're a listener and you're like, oh, I don't have no idea, like, which model is Gemini using? Or what am I supposed to do?

[00:46:14] And where's this vo? Like, welcome to the club. Like, it is, it is rough. And it sounds like if you're a Microsoft user, it may be worse than all of them. I don't know. 

[00:46:22] Mike Kaput: And sorry to poke more fun at Microsoft, but I saw a great post the other day that someone said, you can go find if a startup has actual customers by seeing if they have a Microsoft login option.

[00:46:33] Because no one would build this on their own. It would only be a customer request. So I think that maybe there are some issues there. All right. 

[00:46:43] Era of Experience Paper

[00:46:43] Mike Kaput: Next up a new paper from two AI researchers, one of whom works at Google DeepMind. Paints a very interesting vision of AI's future. So according to this paper, AI is about to enter what the researchers are calling the era of experience.

[00:46:58] So here's the [00:47:00] idea. They say that until now, most AI models have been trained on human generated data, you know, writing code, papers, whatever. But that data is running out. And crucially, it only gets AI to human level of performance, not beyond it. The next leap, according to David Silver and Richard Sutton, the researchers behind the paper will come from AI learning the way we do through its own experiences.

[00:47:26] They say that in this new era, AI agents won't just answer questions, they'll interact with environments, set long-term goals, adapt strategies, and even form memories across months or years. I. Instead of being judged by human preferences, they'll optimize based on real world outcomes, like how much they've improved a health metric, for instance, the scientific discoveries they've made, or energy efficiency they've achieved.

[00:47:52] So the researchers said this is a huge shift we need to prepare for. Because there are major risks. Agents could act autonomously [00:48:00] for long periods, making it harder for humans to intervene, but it also offers a safety benefit. Experiential agents can adjust if their goals or environments change, rather than getting stuck in perhaps destructive loops of behavior.

[00:48:15] So the bottom line, according to them, is quote. Ultimately, experiential data will eclipse the scale and quality of human generated data. This paradigm shift accompanied by algorithmic advancements in reinforcement learning will unlock in many domains, new capabilities that surpass those possessed by any human Now.

[00:48:35] Paul, this can, you know, maybe get a little denser forward thinking. But there's, or seems like there's a really important point here that despite the breathtaking rate of AI progress so far, these researchers seem to be saying we have barely scratched the surface of what's possible. 

[00:48:54] Paul Roetzer: Yeah. I mean, generally language models have gotten us to this point, [00:49:00] but, all the AI research labs seem to agree that they are not the end game.

[00:49:04] Like they, they are, you know, some like Jan Koon think they're a distraction. Like he literally has been on record saying like, don't, if you're coming out of college now, don't work on language models. They're not the future. But, there's different beliefs as to like what the unlock is. It, silver is like a legendary AI researcher.

[00:49:24] Yeah. So he led the Deep re DeepMind, AlphaGo effort. So, again, go watch the AlphaGo movie if you haven't seen it, and you'll understand, you know, the context here. But he was the lead researcher on AlphaGo and I believe on Alpha Zero, which came after. So the difference was AlphaGo was trained to play the game of go through examples from like top players.

[00:49:45] Alpha Zero was not given examples, like it learned to play a number of different games and solve problems without human data. What they realized was the human data may actually bias the system, that the systems might be able to learn better without the [00:50:00] prior human data. So we wrote about AlphaGo in our book Marketing Artificial Intelligence in 2022.

[00:50:06] We actually quoted Silver in the book. So I'll, I'll read this excerpt, real quick because I think it gives a glimpse into what Silver is referring to, about what they're working on and what DeepMind is focused on moving forward. So, again, I'm just reading an excerpt here from Marketing Artificial Intelligence.

[00:50:25] Cade Metz, an author and technology correspondent with the New York Times, was in Seoul, South Korea, covering the match for Wired Magazine. This is referring to the AlphaGo match. In 2016, he spoke with deepminds David Silver, the lead researcher on the AlphaGo project about Move 37 Met. Summarize what happened in this way.

[00:50:42] So AlphaGo learns from human moves and then it learns from moves made. When it plays itself, it understands how humans play, but it can also look beyond how humans play to entirely different levels of the game. This is what happened with Move 37. AlphaGo had calculated that there was a [00:51:00] one in 10,000 chance that a human would make that move.

[00:51:03] But when it drew on all the knowledge it had accumulated by playing itself so many times and looked ahead to the future of the game, it decided to make the move anyway, and the move was genius. In AlphaGo, the movie Silver Said of Move 37. That AlphaGo quote went beyond its human guide, and it came up with something new and creative and different.

[00:51:25] But in the documentary, silver also made the point that this is not human versus machine, but rather human plus machine quote. AlphaGo is human created, and I think that's the ultimate sign of human ingenuity and cleverness. Everything that AlphaGo does, it does because a human has either created the data that it learns from, created the learning algorithm that learns from that data, or created the search algorithm.

[00:51:49] All of these things have come from human. So really this is a human endeavor. So the reason I share that is because this kind of goes back to like the topics we've been building on throughout this [00:52:00] episode. This is the future. Like they think that they can build systems that can go into any industry, any job, and learn potentially to do it better by running simulations of it, by basically learning from itself, by identifying reward mechanisms.

[00:52:17] Because if you give something where there's like a finite outcome, like, you know, the end game, you know, the goal is this and we want you to achieve that. So if you give the AI a goal, it can work towards that goal and then it can know if it achieved it. But if you're playing into a, a, a game like business, which has this like infinite ending, like we were just, I was just listening to a BG two podcast where they talk about like finite versus infinite outcomes and like business is infinite.

[00:52:41] There's no like end goal, like there might be a near term revenue goal or something like that, but like winning isn't like an end point. So the idea of being able to put these AI systems into these environments where there's just, they have to figure out what the reward mechanism is. It's not always just we, we achieve this outcome.

[00:52:59] There's like these [00:53:00] in like difficult things to define along the way. And what they're saying is we can build systems that can figure that stuff out. They can find the reward mechanisms for themselves. They can create their own data, they can run simulations, and they can learn better than if humans were to provide the data for them or just learn from the best humans.

[00:53:17] And so the challenge today of AI systems is they can't invent something new. There is, there's nothing like they can, they can connect dots just the same way a human would of like all these things and create a new product idea. But they can't invent new physics. They can't like invent, a a, you know, a new proof in, in math.

[00:53:34] Like they don't come up with something that isn't somewhere in the training data. The belief is they can, like, there's no reason that they wouldn't be able to do that. And so the approach DeepMind is taking as well as other labs are probably gonna try and pursue this. But Google has a distinct advantage in this, in that they invented this.

[00:53:50] Like they have been doing this for 15 years. so if you wanna understand what DeepMind is working on, where they're going, go study Alpha Zero. Like, [00:54:00] and there's a podcast that just came out, the DeepMind podcast that actually as David Silver, I listened to this on the flight home from Aruba, where it's called is Human Data Enough, and he actually like, tells the story of what they're working on.

[00:54:11] David does a great job of like talking in non dense scientific ways. Like it's, it's really good. listen, so we'll drop the link to the YouTube video in the podcast, in the show notes. 

[00:54:23] Chief AI Officers at Companies

[00:54:23] Mike Kaput: Our next topic is a new report from Digiday, which shows that major brands and agencies are racing to a point Chief AI officers or c Aios.

[00:54:36] Which is a role that's quickly moving from a novelty into actual table stakes at these companies. So in this report, they show in the past year companies like General Motors, MasterCard, PWC. Zocdoc, Accenture, they've all hired AI chiefs and so have some of the ad giants like WPP and indie agencies.

[00:54:58] And the reason [00:55:00] is these companies wanna move beyond experimentation and actually realize efficiencies. And to do that, they need dedicated leadership. So goes their argument to get AI initiatives off the ground. So in some cases, some people are working on this internally. Like at WPP for example, the AI Chief Daniel Holm says his role is about placing the right AI bets internally.

[00:55:23] Others are focused on customers real world impact, like PWC, where their AI chief is working directly with people in their ecosystem to scale ai. And some are just focusing only on helping their companies rebuild processes with an AI first mindset. So, paul, we first actually talked about the trend of the rise of the Chief AI officer I looked like way back in episode 82 in February, 2024, because the New York Times did a story on this.

[00:55:54] So back then we kind of talked about how these titles were probably going to become more [00:56:00] common, but also that it would be interesting to see how long they last because AI is or should be the responsibility of every single executive, not just one leader. Have your thoughts evolved here in the last year or so?

[00:56:15] Paul Roetzer: Yeah, I mean, I, we're hiring for one right now, so I'll, I'll, I'll put the job description in. Like, because I thought deeply about this role as I was building out like our, our organizational structure as I was thinking about the things that are needed. So the way I think about SmarterX, which is, you know, you know, we talk about it as like a sister company, a marketing institute.

[00:56:35] It's, it's really more of like the parent company marketing institutes, like a marketing, focused. Area within SmarterX. But like we think of SmarterX as an the AI transformation company. So we want to educate and empower leaders to reimagine business models, reinvent industries, and rethink what's possible.

[00:56:50] We wanna do it with ourselves as well. So like everything we do, we look at it and say, okay, is there a smarter way to do that? Is there a smarter way to build that department? There's a smarter way to run that campaign. Is there a [00:57:00] smarter way to build that process? What SmarterX means? It's like smarter version of any everything.

[00:57:05] And so to do that, you, you need to have, and again, I think every company should be doing this. You should be saying, what is the smarter way to build our business, to build this team, build this department. So I think that a chief, a officer is the logical role that should lead that. Now, does that person need to be highly technical?

[00:57:24] I don't know, like I could see arguments for it actually being like a marketing person or it being, you know, I don't know if not finance. like it could, it could be technical, but it doesn't necessarily have to be. So in our world, I actually saw it as a combination of like a CIO role and IT, and IT, and elements of a CTO role.

[00:57:42] So I was seeing it being more technical. I'll, I'll read real quick the description. So role overview, it says the chief AI officer, again, this is straight from the SmarterX website, is responsible for spearheading AI driven innovation and automation throughout the organization. The role will focus on developing and deploying AI agents intelligent [00:58:00] automation of processes and workflows, and optimizing our technology stack to enhance operational efficiency, productivity, revenue, growth, and innovation.

[00:58:09] The chief ai, several, work closely with leadership and cross-functional teams to ensure AI is leveraged effectively to scale our offerings and create sustainable competitive advantages. And then it goes into responsibilities of strategy and innovation, agent and intelligent automation, tech stack optimization infrastructure, and it cybersecurity and compliance, which becomes increasingly important with the AI agent side, AI education and adoption data strategy.

[00:58:33] and then AI tool app and product development, which is more of an exploratory phase. So, yeah, like if you want to go look at it, we'll put the link in there, SmarterX ai, and you can go look at 'em. But, yeah, so I do, I I think this is gonna become like a standard C-suite role. I think it's gonna become a prominent part of it.

[00:58:54] Anthropic Researcher Says Claude is Conscious

[00:58:54] Mike Kaput: All right, so we alluded to this next topic a little earlier. as AI [00:59:00] systems grow smarter, some researchers are asking a surprising question. If AI become conscious, should they have rights at Anthropic? The company behind Claude. This is not science fiction. Last year, according to the New York Times, they hired their first AI welfare researcher, Kyle Fish, to explore whether their models might one day deserve moral consideration.

[00:59:24] Today's ai, according to Phish, probably isn't conscious though he does say that there's a 15% chance in his estimation that Claude or another current AI system is conscious. But he thinks that in the next few years as models develop more human-like abilities, companies may need to take this idea much more seriously.

[00:59:45] So to do that, to actually figure out how do you tell if an AI is conscious? Phish suggests combining mechanistic interpretability, which is studying how, essentially how the AI's thought process and brain works with behavioral probing [01:00:00] and watching what models prefer or avoid over time. Now, it has never been more important, regardless if you think this is necessary, or science fiction or crazy.

[01:00:12] It's actually still never been more important to understand how the models actually work, especially as they become more powerful. And this comes from Anthropic, CEO, Dario Amide. In a new essay he wrote called The Urgency of Interpretability. In it, he argues that being able to increasingly understand what's actually going on in AI systems regardless if they become conscious or can become conscious or not, will help us more easily manage the dangers of AI like it possibly acting deceptively or exhibiting other dangerous capabilities.

[01:00:47] Now Paul, this topic, this conversation can really quickly get into like technical and philosophical weeds, but I think Aade put the bigger point here really well in that [01:01:00] essay, he said quote, people outside the field are often surprised and alarmed to learn that we do not understand how our own AI creations work.

[01:01:09] They are right to be concerned. This lack of understanding is essentially unprecedented in the history of technology. 

[01:01:17] Paul Roetzer: Yeah. So this isn't fringe research. This can sound really crazy. Yeah. I think it's just still a bit of a taboo topic, kind of like the quiet AI layoff thing. Like it's happening, but like no one wants to talk about it because people kind of think you're nuts if you talk about this.

[01:01:32] And, it just kind of veers into the sci-fi stuff. It plays into the fears people have of like, what they've seen in Hollywood the last 30 years about AI is like. What if it is conscious and aware and then it needs, its, you know, needs rights and it needs all, it is just weird. Like it gets very bizarre, very fast.

[01:01:49] Now, 15% being like maybe Claude is, my guess is that was a number he was comfortable saying publicly. I'm going to [01:02:00] guess that he probably feels it's higher than that. That that was like a, I can't say 50%, so let's say 15, because it just seems like a sort of arbitrary number. so I, all I would say here is like, again, we're, Mike and I aren't living in the labs, like pushing these systems to try and determine things like this.

[01:02:19] So this is like, you know, us observing the space and listening to lots of interviews and things like that. But based on everything I know this is a legitimate concern or at least a legitimate possibility in the minds of leading AI researchers. To define consciousness, just so we're all on the same page here.

[01:02:37] Few different definitions. state of being awake and aware of one's surroundings. The awareness or perception of something by a person. The fact of awareness by the mind of itself and the world. So the way I think about it is like knowing it exists. Like, yeah, the simplest way I think about it's like the model, knowing that it's an AI model in a world talking to humans, right?

[01:02:57] Like it's aware of that. So [01:03:00] when I say this is legitimate, just two weeks ago, DEMA, Saba was on 60 Minutes, and this question was asked of Demi Sabas, who I consider the most reputable, authentic AI researcher in the world. Like Demis is like, you know, mind Mount Rushmore. Demis is like the first one. So like, if, if he's talking about it, then I generally like find him to be, to be the most.

[01:03:24] Believable person around these topics. So he said, I don't think any, and we'll link to this, that this is, you can watch the, or read the transcript. I don't, in this quote, I don't think any of today's systems to me feel self-aware or you know, conscious in any way. Obviously everyone needs to make their own decisions by interacting with these chatbots.

[01:03:42] I think theoretically it's possible. So then Scott Pelley said, but is self-awareness a goal of yours? Dems replied, not explicitly, but it may happen implicitly, these systems might acquire some feeling of self-awareness that is possible. I think it's important [01:04:00] for these systems to understand you self and other, and that's probably the beginning of something like self-awareness.

[01:04:06] So Pelley says, but if a machine becomes self-aware, we may not recognize it. Demis replies, I think there's two reasons we regard each other as conscious. One is that you're exhibiting the behavior of a conscious being very similar to my behavior. The second thing is you're running on the same substrate.

[01:04:26] We're made of the same carbon matter with our squishy brains. Now obviously this with machines, they're running on silicon. So even if they exhibit the same behaviors, and even if they say the same things, it doesn't necessarily mean that this sensation of consciousness that we have is the same thing they will have.

[01:04:45] This is wild. Like you just, so then real quick on the ADE thing, and again, like I, you know, I feel like his writing is, is like bordering on a little bit too much. Like [01:05:00] I feel like they're just like maybe crossing the line slightly into like the hype side or like over exaggerating. But then there's parts of me that think, but maybe he's just right and it looks like hype right now, but he's actually like, right.

[01:05:14] So I try and be very objective with ADE's writings that like. I try and take him at his word, but he, I think he's, he's made some very, exaggerated statements in some of his recent writing. So I would just take that in context of what we're telling you. So I wrote about this in my exec AI insider newsletter yesterday.

[01:05:33] and it basically like, here's the couple of key points. He said in the context of AI interpretability refers to the degree to which a human can understand the reasoning behind AI model's decision. It involves understanding how the model arrives at its output. This is kinda like what interpretability needs.

[01:05:48] essentially it's making AI systems more transparent and understandable that allows a human to grasp what it's, so what I was saying is like. the labs know that the models are getting smarter and more generally capable. They [01:06:00] know if you give them more data, you do reinforcement learning, you give 'em better data, that they get smarter.

[01:06:04] And they know that if you give them time to think, like the reasoning models do that, they tend to reduce hallucinations and again, get smarter on the IQ test. but they can't see the inner workings. It's analogous to like how scientists and doctors don't really know how to explain how the human brain does what it does.

[01:06:22] Or if like the human's doing something weird, like they can't like look in the brain and figure out exactly what's going on. That's kind of how this works with these AI models. And so what Amadei was saying is he thinks that some of the recent research breakthroughs have put them on a path to be able to understand this, to, to do the interpretability, like you were saying, where they can actually look into the model and figure out why it's doing what it's doing, and then that can lead to the building of more interpretable models in the future.

[01:06:49] And then his main thing was calling on other labs like OpenAI and DeepMind. To put more resources into interpretability, which they may be doing. They're just not talking about it as much as . What [01:07:00] anthropic is.

[01:07:03] xAI Funding and Updates

[01:07:03] Mike Kaput: In our next topic, according to a report from Bloomberg, Elon Musk's, XAI Holdings, such as the newly merged entity combining X, formerly Twitter and his AI startup XAI is in talks to raise around $20 billion. So if that closes, it would be the second largest startup funding round ever behind. Only open AI's $40 billion haul earlier this year.

[01:07:28] And this new cash would value this entity at more than 120. Billion. Some of that would be used to pay down debt that Musk took on to privatize Twitter. and obviously plenty of it would be spent on developing X AI's AI chatbot grok. at the same time, grok now has a feature called Grok Studio, which allows grok to generate documents, code reports, and browser games.

[01:07:54] And according to the announcement, Grok Studio will open your content in a separate window, allowing both [01:08:00] you and Grok to collaborate on the content together. Grok users can now attach files from their Google Drive, can now work with documents, spreadsheets, and slides. Yet Paul, despite all this, this is, you know, noteworthy fundraising, cool features they keep shipping you, however, did run a poll on LinkedIn recently that suggests perhaps Grok isn't catching on, at least with some users as the way that some of the other AI tools have.

[01:08:28] Can you maybe tell us a little bit about that? 

[01:08:31] Paul Roetzer: Yeah, I was honestly just curious. So like, I have, I've said on this podcast many times, I'm very active on, x in terms of monitoring the AI space. It's where a lot of the links and resources and things that we find actually comes in from a highly curated list of a couple hundred AI researchers and entrepreneurs and leaders.

[01:08:53] so I see x, Grok all the time. It's embedded in Twitter. It's like they talk about it nonstop. And so [01:09:00] if you're living on X, you could get this impression that Grok is relevant. and so I was just curious, now you could jump over to LinkedIn and obviously people like the, kind of the X elitist, like think LinkedIn is ridiculous and would never go there.

[01:09:17] so it's like almost two different bubbles in a way. But I just went over there and I was like, okay, have, has anybody tried this? Like, I was just curious. So I said, have you tried Grok by XAI, Elon Musk's ChatGPT competitor? It was a, a poll I put up. and so the survey results was, 30% said yes.

[01:09:37] 62% said no, and 8% said never heard of it. So 70% have never tried Grok on LinkedIn. Now this was 18,000 impressions and 1400 votes. So it wasn't an insignificant poll, like it was reasonable. Right Now, there was also 98 comments, and I did not export all these comments and like analyze them. [01:10:00] But Mike, have you looked at the comments of the people that said No?

[01:10:04] I think it's reasonable to assume like 50% of those is because they hate Elon Musk. Yes, it was. There was definitely commentary about like not liking him and interestingly like some European comments, like, I would never buy anything from that guy or using it. So I don't know how much his persona plays into it.

[01:10:25] It's certainly impacting like Tesla sales and stuff right now. Yeah. But I do think more than anything it's, it's just, it's not a mainstream thing right now. It's very much living in that X bubble. Not, I'm not saying anything against the product itself. Like they're making a lot of advancements.

[01:10:43] They're copying other capabilities very quickly. They're a fast follower right now. Yeah. They're not innovating it would look like yet. They're just like, anything OpenAI comes out with like three weeks later it seems like they come out with something like it. but yeah, that's, that's kinda where we're at is 70%, at least on LinkedIn in [01:11:00] my network of, you know, 50,000 people, whatever it is.

[01:11:03] 70% have not tried it or heard of it. 

[01:11:07] Other AI Product Updates

[01:11:07] Mike Kaput: Interesting. Alright Paul, so to kind of wrap us up here, I'm gonna go through a few other quick AI product updates and then get us into our last segment, which is listener questions. So chime in here at any point if any of these product updates jump outta you.

[01:11:21] Otherwise we'll just ease right into that final question here. So in terms of some AI product updates, we alluded to this one before. Google has, released VO two, or isn't the process of releasing VO two in the Gemini, app. So Gemini advanced users can create short, high quality videos with this newest video generation model.

[01:11:45] You write a detailed text prompt describing a scene, and VO two transforms it into an eight second 720 P video clip. At the moment, these are not blurry AI clips. They're built to capture real world physics, [01:12:00] human motion, and fine visual details. So definitely worth checking out if you're interested in video generation, whenever it becomes available to you.

[01:12:10] Next up script has announced an age Agentic AI video editor feature in its video and audio editing platform. CEO Andrew Nas calls it cursor for video in reference to the popular age agentic AI coding assistant. According to the company, you can give this AI co-edit instructions and it will go do the thing right in script.

[01:12:32] So an example might be which they provided. You can give it a screenshot from Wikipedia and say, can you write a script about this? Or tell it to break this video into scenes and add layouts and stock media. Last but not least, a new startup is getting a ton of buzz because it's aiming to replace your CMO or your entire marketing team with ai.

[01:12:54] It's called Icon and it builds itself as the world's first true AI chief marketing [01:13:00] officer. Now this is backed by Founders Fund and execs from labs like Open AI and Cognition. So definitely some heavy hitters behind it. Icon claims to be able to plan, create, and launch thousands of ads end to end, and then learn and improve from real performance data, not just guessing.

[01:13:18] So definitely one to watch if you're able, my one note there 

[01:13:21] Paul Roetzer: is, I didn't dig into this company deeply, but if they think that all a chief marketing officer does is run ads exactly, then they, they might be looking at a quite a limited total addressable market. So I dunno if right name for 

[01:13:36] Mike Kaput: that, that problem.

[01:13:37] Yeah. Yeah. All right. 

[01:13:40] Listener Questions

[01:13:40] Mike Kaput: So Paul, let's wrap up with our. New recurring weekly segment listener questions. Here is the question from our listeners and from our audience. This week. I've been hearing about AI assistance or AI agents. Are these real things or just built out versions of a custom GPT? 

[01:14:00] Paul Roetzer: I feel like we could a answer a different version of this question every week.

[01:14:04] Mike Kaput: A hundred percent. 

[01:14:05] Paul Roetzer: Absolutely. The hot topic, when I go do talks like, it's, everybody asks me about AI agents. It comes up on our intro to AI class. Like it's all everybody's talking about. and part of the confusion is like everyone's just labeling everything AI agents now. . So things that were previously templates or apps or, I don't know, workflows, like whatever.

[01:14:25] They're just, they're just being called agents. So it's a super confusing space. I, I empathize with anybody trying to understand whether agents are real or not. basic premise is an AI agents like a system that can take actions to achieve a goal, and sometimes those actions are defined by the human.

[01:14:44] you know, there's a rules given to this agent to do something. but oftentimes there's some element of it that it chooses its own path. There's, an ability for it to write its own rules and figure out a direction. So the example I like to give is [01:15:00] deep research. Open AI or, or, you know, Google both have their deep research product.

[01:15:05] you give a prompt. It figures out what tools to use. it figures out what sites to visit. It summarizes those sites, it comes back, it creates a report for you, it builds its research plan like that is agentic. Like that, that is a good example of a real AI agent. and I think a prelude to much more technology like that, that can actually build and perform its own plan.

[01:15:30] And the human is there to kind of give it the goal and then like review and vet the quality of the product and things like that. So that's what an agent is. I think what's happening right now is a lot of tech companies are just piggybacking off of the AI agent terminology and calling everything agents, but in reality it just means like if, if there's a project you need to do, if there's a, you know, something, an an activity you need to perform that might require 5, 10, 20 steps.

[01:15:57] These things are increasingly gonna be able to do [01:16:00] those and even plan for what to do, versus you having to go in and say, I want you to do this, then this, then this, then this, then this. That's just automation. Like if you can, if you just go in and define the 10 steps and then you, you know, set up through make or Zapier or whatever, like a way to do it, that's just automating something.

[01:16:16] Mike Kaput: Yeah. 

[01:16:16] Paul Roetzer: With an AI agent, there's some level of intelligence happening. There's the thing, there's doing some planning or, decisioning on its own and then creating the output. So it's confusing, but it basically just means there's a series of actions that are taken by the AI system to achieve an outcome or a goal.

[01:16:34] Mike Kaput: Cool. Paul, that is a ton going on in the past. Yeah. Couple weeks in ai. I really appreciate you as always breaking everything down for us. 

[01:16:43] Paul Roetzer: I thought today was for sure gonna be an hour and a half for him. We, it looks like we knocked this out about 15. So, nice, nice work, curating and organizing it all.

[01:16:50] 'cause there was no less than like 75 links. Oh gosh. I know. In the sandbox this week. So Mike does an incredible job every Sunday night of pulling this all [01:17:00] together and getting it organized for us to talk about. And it was a, it was a lot of work this week. So thanks everyone for listening and giving us the grace of a week off to enjoy spring break.

[01:17:09] And, we're, we're back now for the foreseeable future. I don't, I don't think we have any known days off coming off, so we will be back every week, as originally planned. And, we'll talk to you. I guess it'll be may next time we talk to you, so. Oh yeah, boy. All right. Thanks, Mike. I'm off to Boston. All right.

[01:17:25] All right. Later. 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 earn professional certificates from our AI Academy and engaged in the Marketing AI Institute Slack community.

[01:17:51] Until next time, stay curious and explore ai.
Claire Prudhomme

Claire Prudhomme is the Marketing Manager of Media and Content at the Marketing AI Institute. With a background in content marketing, video production and a deep interest in AI public policy, Claire brings a broad skill set to her role. Claire combines her skills, passion for storytelling, and dedication to lifelong learning to drive the Marketing AI Institute's mission forward.

Recent Posts

The White House Just Made AI Literacy a National Priority. Now What?

Mike Kaput | April 29, 2025

In a surprise executive order, President Donald Trump has made AI education a core national initiative—starting in kindergarten.

AI Layoffs Are Already Here. But Don't Expect Companies to Always Admit It

Mike Kaput | April 29, 2025

AI-driven job disruption isn’t coming. It’s already here—and many executives are quietly betting their future on it.

OpenAI's New o3 Model May Be the Closest We've Come to AGI

Mike Kaput | April 29, 2025

OpenAI just dropped two game-changing AI models—o3 and o4-mini—and if you're paying attention, you can feel the seismic shift happening under our feet.