← The Artificial Intelligence Show
The Artificial Intelligence Show
#146: Rise of “AI-First” Companies, AI Job Disruption, GPT-4o Update Gets Rolled Back, How Big Consulting Firms Use AI, and Meta AI App
01:25:35
Loading summary
Transcript63 lines
- [00:00]
Paul Raitzer
To me, when you have a workforce that is afraid for their jobs, that fear that maybe you're going to be replacing them, when you say we're going to be AI first, that immediately tells me people aren't first. Welcome to the Artificial Intelligence show, the podcast that helps your business grow smarter by making AI approachable and actionable. My name is Paul Raitzer. I'm the founder and CEO of SmartRx 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. Welcome to episode 146 of the Artificial Intelligence Show. I'm your host Paul Raitzer along with my co host Mike Caput. As always, we are recording on Monday, May 5th about 11:00am Eastern Time. We are expecting maybe some announcements this week, so timestamping usually matters on this podcast. Every week is everybody always drops something on a Monday after we record this thing it seems so. Today's episode is brought to us by a couple of our marquee events. So first up we have the AI for B2B marketer summit. This is presented by Intercept has been a great partner of Marketing Institute over the last few years. This virtual summit is packed with incredible sessions from top B2B marketing experts. It's all happening virtually on Thursday, June 5th starting at noon Eastern Time. That's 12:00 Eastern Time. You'll learn real world strategies to use AI to grow better, create smarter content, build stronger customer relationships and much more. Thanks to our sponsors. There is even a free ticket option so you can go and choose that free ticket option. There is a paid option for private registration so information is not shared with the sponsor sponsor. And then there's also a paid option for on demand access so you can go to B to B, the number two B summit AI. Again that's B2B summit AI and learn more about that event. It is coming up well one month from today. I swear. Like when the month, when the calendar changes to the next month, I actually realize like how much I have to now do before the month that is coming or in April. June 5th seemed like really far away and now it's May 5th and it is no longer far away. So apparently I need to add some things to my to do list after we're finished recording today. All right. And then next up if you're ready to get smarter about AI and marketing. Don't miss Macon 2025. This is our flagship event is back for the sixth year in Cleveland, Ohio. It is happening October 14th to the 16th. We have already announced more than two dozen speakers. You can go check those, those people out. We have incredible sessions coming up. We have a bunch of our top rated speakers from past years coming back. A bunch of new voices and perspectives we're bringing to the mix as well. I'm really excited about that lineup and how it's coming together. Lots more announcements still coming. Prices go up May 31 and basically like every month they just go up. I think it's like another $100 or so. So you want to get in early, get the best pricing possible, do it before May 31st. Go to Macon AI that is M A I C O N AI. There are group tickets available as well. So if you're planning on bring group of say five or more, make sure to reach out to us and we can help get that set up as well. All right. It's sort of moved into the AI first stage, Mike. We've got some new research from Microsoft that I'm really excited to talk about and a bunch of other updates including first time I can recall as ever happening the rollback of one of these frontier models because it was not behaving the way it was supposed to behave. So let's, let's get into all of it.
- [03:49]
Mike Kaput
All right, Paul. So the first main topic today, today is the rise of the what we're calling AI first, or as we actually prefer to talk about it, AI Forward Company. So in the past several weeks, a number of prominent CEOs have released memos declaring their intention to be AI first. So we actually talked a couple weeks ago about the first CEO to kind of get a lot of press for doing this, which was Shopify CEO Toby Lutke. And he is now being joined by CEOs at Duolingo and Box, both of whom released their own AI first memos this past week, basically declaring that their companies are all in on AI and AI literacy in some form or another will be a baseline expectation for all employees. So Duolingo CEO Louis Von Ahn wrote in his memo, quote, duolingo is going to be AI First. And he said that means the company will need to rethink much of how they work to prioritize what AI is capable of and now makes possible in the workforce. And to start, he said, Duolingo will gradually stop using contractors to do work that AI can handle AI use will be part of what the company looks for when hiring. It'll be a part of performance reviews and headcount will only be given if a team can automate more of their work with AI. He did also reiterate that Duolingo cares about its employees. It is not looking to replace their current employees with AI, but rather augment them. Now, box CEO Aaron Levy released a very similar memo that hit a lot of these same notes. Now, while all this is going on, Microsoft actually published its annual Work Trend Index annual report that seems to confirm, at least at a high level, this is a macro trend of where we're headed. So they said that data from 31,000 workers across 31 countries, quote, point to the emergence of an entirely new organization, what they call a frontier firm. They define a frontier firm as, quote, a company powered by intelligence on top, on tap, human agent teams and a new role for everyone, agent, boss. And the report claims these firms are already taking shape, with 81% of those surveyed saying they expect AI agents to be moderately or extensively integrated into their company's AI strategy in the next 12 to 18 months. So, Paul, maybe first here, give me your thoughts on this recent round of AI First CEO memos. I mean, you're a CEO actively considering all of these issues, working on AI literacy and transformation. What do you like about these letters? Anything they can improve. We should expect to see more of these, I would guess.
- [06:46]
Paul Raitzer
Yeah, I think we touched on this a little bit when we talked about the loot key one from, from Shopify that I do expect, you know, within the next month or so, pretty much every tech CEO is going to have to now do their own internal memo, which they will all also leak on LinkedIn and X. So I think it's like table stakes now that if you're a tech CEO, you pretty much have to like put your stake in the ground about what your vision is for an AI first or AI forward or AI native or AI merging or like whatever people want to, you know, call these things. I think that's going to be required and I actually think it's a good thing. Like, I, I feel like we need way more transparency with our employee base, with workers about what we're doing as CEOs, how our vision is, how agents and automation are going to impact people's jobs, which, you know, I think that they're kind of glazing over at the mom, and I think that's the kind of stuff people are going to want to hear more of. So Just to recap for anybody who hadn't listened to our episode, we talked about the Shopify one which sort of, you know, is the triggering event for these other ones coming. You know, he had talked about using AI effectively is now fundamental expectation of everyone at Shopify. Totally agree with it. Must be a part of prototype phase of any, you know, new products, anything they're building, any features they'll add AI usage questions, performance and peer reviews. Learning is self directed so they want people to be proactive about doing this, the headcount thing, everyone means everyone, like it's applying to all of them. So the duolingo stuff, you started seeing the same concepts, a lot of these same ideas. And then Aaron Levy at Box, he released his last week and his said primarily use AI to eliminate drudgery and move faster across the business. Encourage teams to use AI to automate more and save money but primarily reinvest those savings. So again he's trying to kind of hedge like hey, this isn't a replacement thing. We're trying to do this, drive efficiency and then reinvest those which I believe. I mean I think Aaron's a good leader in this space. He's very active on X and this jives with how he generally talks about things related to AI. He also said foster constant experimentation internally to find the best use case for AI. Upskill every employee to be AI first over time and with more education and awareness and then maintain strong governance and security practices with human in the loop still required for most areas. So you know, I think as you mentioned Mike, you're seeing these kind of common threads across these and they're all very short memos. Like most of these things are, you know, what 500 to 1,000 words. They're not, you know, expansive manifestos. So you know, I think that they're going to keep evolving but I also think that they're just the first phase because employees are going to want more detail than this. They're very, very, you know, kind of high level vision I would say versus like tactically what does this actually mean to me as someone in hr, someone in finance, someone in marketing. Now I liked the Microsoft report, you know, I was giving a hard time to, I think it was McKinsey maybe recently about how their data was like a year old already. Yeah. So kudos to Microsoft. They, you know, you highlighted it, 31,000 full time employees, you know, knowledge workers that they did research and it was from February to end of March 2025. So we're looking at month and a Half old data here, which is great. That means this is actually really relevant to where they're at today. There it's, it's not a terribly long report. I would suggest people actually read the full report if you have, you know, it's probably take you 20 minutes. Throw in a notebook, lm ask it some questions, build a study guide based on it, whatever you need to do. I'll call out three key takeaways that sort of jumped out to me, Mike. So the first is this idea of the frontier firm which you sort of highlighted is it's a new blueprint for mashing machine intelligence with human judgment. It's structured around this idea that intelligence is going to be always on demand. And that's different. We historically haven't just had this level of intelligence accessible to all of us. And it's going to be powered by these hybrid teams of humans plus agents that's going to let companies scale way faster. When it thinks about a frontier term, frontier firm, it's looking at five core traits of organizational wide AI deployment, advanced AI maturity, current AI usage, projected AI agent usage, and a belief that agents are key to realizing roi. It does say that within the next two to five years every organization will be on the journey to becoming a frontier firm. I agree with that. It led me back to the, you know the blog post I'd written Mike back in May of 2022 where I said the future of all businesses AI are obsolete. And in that blog post I sort of laid out that there was three types of firms. There was AI native, which is you built smarter from the ground up, infusing AI into all your business processes and teams. AI emergent, which is you're an existing company that evolves to become what they're calling a frontier firm basically and then obsolete is everybody else because they become irrelevant. The second key finding that jumped out to me is AI skilling and digital labor are top workforce strategies. So Mike, you had kind of called this out a little bit the their definition of an AI agent. Again it's always helpful to have this context AI powered system that can reason, plan and act to complete tasks or entire workflows autonomously with human oversight at key moments. It's an interesting definition. An agent boss is a human manager of one or more agents. So I totally agree. We will all be agent bosses within their definition. But then the thing that really jumped out to me is the survey question was as you consider the role of AI and agents in workforce and talent management, which strategies are your team or organization considering? Over the next, I think it was 12 months. And the number one thing was prioritizing AI specific skilling of existing workforce was 47%. So AI literacy is the number one thing that people are focusing on. The second, maintaining headcount but using AI as digital labor. The third, investing in maintaining employee morale, which they see as important as people worry about their jobs is my guess here. But then, interestingly, the other answers were using AI to reduce headcount, which 33% admitted to being part of their strategy. Increasing head count to support business needs was 32%. Using AI to reduce headcount but rewarding top performers, 32%, meaning you can pay your top people more if you don't have as many people anymore. And then no change was only 28%. So in a couple related data points, 51% of managers say AI training or upskilling will become a key responsibility for their teams within the next five years. 35% of managers are considering hiring AI trainers to guide employee addition or adoption in the next 12 to 18 months. So you're really starting to see this true shift where they literally said amid the uncertainty. One Signal is clear. AI literacy is now the most in demand skill of 2025, according to LinkedIn. So Microsoft owns LinkedIn. They have access to data that you and I don't. Also rising are human strengths like conflict mitigation, adaptability, process automation, innovative thinking, showing that the future belongs to those who compare deep AI capabilities with the skills machines can't replicate, which to me is probably the most important thing they're highlighting here. And then the final third takeaway is the rise of human agent teams and impact on an organizational charge, which again, they talk about kind of this meshing of these things into what they're calling a work chart versus an org chart where you actually scale around goals versus like specific functions and departments, which is kind of an interesting concept I think we'll see start to play out a little bit more. So again, like, the main thing for me is this increasing awareness and sense of urgency around AI literacy, which I am excited to see more organizations thinking in that way.
- [14:25]
Mike Kaput
Yeah, just one kind of final thought here on the two sides of the job displacement coin on one hand, 33% of people admitting they're going to reduce headcount seems crazy to me, like high because I assume more hurt, double that.
- [14:42]
Paul Raitzer
They just weren't saying it.
- [14:44]
Mike Kaput
So that seems interesting. But on the other side of this, just kind of putting a more optimistic lens on it, they cited some really interesting data that shows like why firms are looking to use digital Labor. They said 53% of leaders say productivity must increase, but 80% of employees say they lack time or energy to do their work now. So this isn't all necessarily doom and gloom though. In the next topic we'll kind of talk a bit more about that.
- [15:09]
Paul Raitzer
Yeah, and semi related. Let me see if I can pull this up real quick. I actually as I was going through this, you know, I was talking about this, these gaps. I, I ran a deep research project on ChatGPT while I was reviewing the PDF and I said I'm doing analysis on the potential impact of AI on knowledge work. I'd like to consider which industries and professions currently have a gap meaning more open jobs than employees in the field that AI could fill. Because we talk all the time about job loss but what we often don't touch on enough on this podcast is all these sectors where they don't have enough people and, and like AI can fill that. Now the thing I've said before is the risk you run here. So let me. I'll pull up one. So education sector teachers was a huge one. So it said teacher shortage have become a national crisis. Recent analysis found that 22 to 23 school year 406,000 teaching positions were either vacant or filled by under qualified instructors. Another one legal sector, not one you would normally think about being a shortage of. But this is. They said lawyers tend to cluster in urban centers, leaving rural communities under service. For instance in Ohio, 75% of attorneys practices in just seven urban counties, leaving many of other 81 counties virtually no local lawyers. But then the one that jumped out to me was the finance one because I've actually done talks for accounting firms and this is when I've thought about. They said one prominent example is a shortage of accountants and auditors. In recent years the accounting workforce has seen a steep decline. The US has around 340,000 fewer accountants in 2023 than just five years prior. In part due to baby boomer retirements. Approximately 75% of CPAs are boomers nearing retirement age. That's crazy.
- [16:53]
Mike Kaput
That's why.
- [16:53]
Paul Raitzer
But then like when we think about the impact, you could see the motivation to build an AI tech company that solves for this gap in accountants CPAs. But by filling that gap you actually accelerate the automation of the workforce, the remaining people. And that's the part where it's like this isn't, this isn't easy. This is going to be really messy how this gets solved. But like there is financial motivation to build the solution to fill that gap. But by filling the gap, you, you actually decimate the workforce that's left. It's and we don't have time in this episode to like go into this, but these are the complexities we're going to have to deal with. And it, it's just fascinating when you start to kind of like peel the onion back, I guess, of like, you know, where this is all going to play out.
- [17:37]
Mike Kaput
So this does have a lot related to our second main topic, which is tracking even some more signals of AI job disruptions. We've talked about this many times on the podcast. It seems like things are accelerating a bit. So we wanted to kind of highlight a few interesting things that are standing out on this topic. So first, here is a new report in the Atlantic says that recent college grads are struggling more than usual to find work and that AI might be part of the reason. So unemployment for young degree holders has jumped to 5.8%, which is apparently an unusually high rate, even as the broader economy for the time being holds steady. And this report said even elite MBA grads are having trouble landing jobs. Law school applications are spiking, which is a classic move that happens during economic uncertainty. And economists that were interviewed for this report in the Atlantic suggest three overlapping causes. So first, the job market for young people never fully bounced back from the Great Recession and the pandemic. Second, the college degree is no longer the golden ticket it once was. Employers are posting fewer jobs that even require one. But the third most provocative theory is this is due to AI. Many entry level jobs we've talked about many times involves synthesizing information, making presentations, writing reports, et cetera. The exact kind of stuff large language models are now capable of doing. So it's too early to say if this is truly causing this trend, but the timing, the sharpness of this trend worth paying attention to now. Second, Anthropic has actually announced an economic advisory council to study the impact of AI on work. They say they are bringing together, quote, a group of distinguished economists who will provide Anthropic with expert guidance on the economic implications of AI development and deployment. The council will Advise anthropic on AI's impact on labor markets, economic growth, and broader socioeconomic system. So we are not the only people talking about this. Clearly, others are seeing a need to understand this better. And then on top of it all, you have stuff like the Stop Hiring Humans campaign, which is a marketing campaign that was recently run by a buzzy AI startup called Artisan that just raised 25 million by telling companies to stop hiring humans. Though ironically it is hiring more humans itself. So this company is led by a 23 year old founder, Jasper Carmichael. Jack and Artisan builds AI agents that right now handle outbound sales. So basically they cold email leads like a junior sales rep would and they released this like super controversial marketing campaign which included billboards shouting literally stop hiring humans. And of course this grabbed a ton of headlines. They there was a bunch of backlash. There are unfortunately death threats about this. Behind the noise though, Artisan is part of this fast growing wave of startups trying to automate white collar work. We talked about Mechanize last week as well. So Artisan says that it now is able to send out high quality emails with near zero mistakes. They say they've signed 250 clients and passed 5 million in annual revenue. So despite their controversial tactics, it seems like someone is buying their agents. So Paul, to kind of unpack this, I want to start with the Atlantic article really quick. Like how much weight do you give their argument that AI could be this driving factor behind recent college grads struggling to find work?
- [21:12]
Paul Raitzer
It's an interesting article. They, I would say it was very much a hypothesis. Like they did not go into like great depth proving out this concept. So I'll just, I'll read a quick excerpt from it because I think it's relevant context. So they say it's a novel economic indicator to look at this recent grad gap. It's the difference between the unemployment of young college graduates and the overall labor force. So going back four decades, young college graduates almost always have a lower, sometimes much lower unemployment rate than the overall economy because they're relatively cheap labor and have spent four years maintaining in a theoretic, theoretical, theoretically enriching environment. And it goes on to say but last month's recent gap hit an all time low. That is today's college graduates are entering an economy that is relatively worse for young college grads than any month on record going back at least four decades. Then they say the strong interpretation of this chart, which they show this chart and it is kind of a jarring chart to look at, it is significant difference than, than historical data over four decades. So they said a strong interpretation is that it's exactly what one would expect to see if firms replaced young workers with machines. So for example, they say as law firms leaned on AI for more paralegal work and consulting firms realized that five 22 year olds with ChatGPT could do the work of 20 recent grads and tech firms turned over their software program to a handful of superstars working with AI co pilots, the entry level of America's white collar economy would contract, which is what they're basically saying appears to be happening here. Then they said, and even if employers aren't directly substituting AI for human workers, high spending on AI infrastructure may be crowding out spending on new hires. I don't again, this is, they're just talking theoretically here hypothesis and it's not even really a full blown hypothesis. They're just kind of like throwing it out there. But I agree this is what it would start to look like. It would start to look like, well, we're not really sure what AI agent impact is going to be, but we think that the five people currently on that team with full blown training on how to use chat GPT and build custom GPTs know they can probably do the work of what we would have hired the 10 interns for the 10 full time workers out of college for. And so I could absolutely see places like Deloitte and, and McKinsey and you know, big accounting firms like just saying hey, maybe we don't need as many hires here and trying to play it out and see what happens, not knowing for sure if it's going to work, but basically just taking a flyer, hey, the economy's not great, you know, the tariffs are wrecking everything and we might be heading toward a recession real fast. And let's just see if AI can't do do this and let's not hire as many people this year. I don't know for a fact that's what's happening, but it sure does make a lot of sense. And as a CEO, if I was the CEO of one of those big firms, it is probably the way I would be thinking about it.
- [24:10]
Mike Kaput
Yeah, that's really interesting. Between the Atlantic article talking to economists and Anthropics Economic Council, I know you've said many times on the podcast like why aren't economists talking about this more? And these kind of almost feel to me like indicators of like economists may be waking up and saying wait a.
- [24:28]
Paul Raitzer
Second, yes, now the difficulty is still going to be we like so one I think it's great we need more conversations like this. We need more articles about the topic. I love that Anthropic is building this council to do this and I assume Anthropic is going to do this way. The concern I have, having talked with some leading economists who had as of six months ago no real interest in studying the impact of AI on the economy Thought it was overblown. We obviously didn't agree on that. But this is where they were at. Those economists who are being built into these like AI impact councils need a very in depth understanding of what these AI models are currently capable of and where they're going. That has been my challenge talking to leading economists to date, is it became very apparent very quickly that they were unaware of the current power of these models and the very near term power of the models. And so how in the world are they supposed to model impact when economists, generally speaking, look at historical perspective to predict the future? And I don't know that you're going to learn what you need to learn by looking to the past. And so I think it's great that like Anthropic is doing this. I hope Google is doing something similar it hasn't talked about yet. I hope OpenAI is doing something similar, but like providing deep education and hands on experience with the model so the economists have that perspective as they start to try and project out.
- [25:53]
Mike Kaput
All right, so let's quickly talk about the impact of the stop hiring humans thing. So obviously this is just meant to be like a PR stunt. It's meant to capture attention. The company itself is raising money, hiring people. But like, do you think there is something here about like the overall feeling of it? Like, are we going to see more of this as people become comfortable talking more about like saying the quiet part out loud? Is it going to be backlash to this? I mean clearly there was in this case.
- [26:20]
Paul Raitzer
Yeah. And shout out to Brian. So the way this came about is, you know, we actually read comments like, so I put up something on LinkedIn earlier this week about 03, or I guess it was last week about 03 and like kind of summarizing conversations from episode 145. And so Brian left a comment and saying, hey, like this is great, like, sorry, did I miss you guys talking about the Stop Hiring Humans campaign? And so I actually sent that comment to Mike. I was like, I feel like we talked about this, right? Like this was in, you know, a month or so ago. And so Mike did a search. He's like, we never talked about it, so I don't know if it ended up in our newsletter somewhere. So I was like, as I'm prepping for this, I'm like, this feels so familiar. Like I feel like we had to have talked about this. So the gist of it is. And again Brian, thanks for commenting on the LinkedIn post and, and caused me to go back and look at this. It was Just a PR stunt on their part basically. And they found people like, responded to this. But then like the dude in this TechCrunch article, the CEO is basically like our, our SDR Ava, like, doesn't even work. Like initially, like six months ago, it was actually terrible. And it's gotten better. And now we're building what they're building two new ones. One is to handle inbound messages and the other one is meeting manager assistant so that both are supposed to come up later this year. So they're building these things, but they're also hiring a bunch of people. So they were at the time of the TechCrunch article, hiring 22 more people into their own sales organization. So this stop hiring humans thing, yeah, it's just a PR stunt. Now we did talk about mechanized last week, which is straight up mission is to not need people anymore. Um, we've had other people say, like, hey, we're going to build an organization of X size and we're never hiring more than 50 people or 100 people or, you know, I think by the end of this year, beginning next year, you're going to say, hey, we're going to get to a billion with like less than 10 people or less than 5 people. I don't know if it's going to become like the badge of honor that like VC rounds have historically been. So, you know, you always had this like, oh yeah, we raised $50 million, like, okay, but you have no possibility of having profits for the next decade. Like, Great, you raised 50 million. And I don't know if it's going to be like, hey, we got to 10 million an ARR. With two people. Like that becomes like the new badge of honor that you did it with the fewest people possible, the highest revenue per employee possible. And I'm not saying those are necessarily going to be bad metrics. I'm. I just think like, we're going to enter this sort of hype phase where everyone feels like, oh, the other person's doing it with fewer people than me. It's going to be like the new doing it with more VC money than me. So yeah, I think that we also are going to. It's going to become more accepted to talk about how few people you have and that you plan to hire. And I think that's probably not a bad thing for the AI native companies that are building from the ground up and can do it with fewer people. But the problem is going to come in when it's the AI emergent companies that already have 50 or 100 or a thousand or 10,000 employees. And they're the ones that are now saying, yeah, we're going to get down to 5,000. Like we think it can do this with 2,000. And that's where we're going to have problems in the economy and the workforce. So, yeah, I don't know. I think it is going to get talked about a lot more and I'm not sure what that actually ends up meaning to people's jobs and the workforce, but I would imagine we're going to see a lot more messy parts of this soon because it's becoming okay to talk about it.
- [29:28]
Mike Kaput
Yeah, it seems like the big driver in this story for me is like messaging does matter to pay attention to in terms of the overall narrative, whether it, you know, know, leads to concrete results or not. And it strikes me as you're saying this about the badge of honor. I mean, if you're not someone that follows startups or VCs, you could see that badge of honor as like something very different that is like very sinister to some people. Not rightly or wrongly, but just if you're not thinking that way, this could come off very poorly.
- [29:58]
Paul Raitzer
Yeah. And if you think about like played out to the recruiting side of like, you do want 30 or 50, like exceptional people, are they going to want to go work for a company that says don't hire people? Like, I don't. I don't.
- [30:09]
Mike Kaput
Right.
- [30:09]
Paul Raitzer
Even if it's just like advertising and hype. I don't know. I think there's downstream effects of things like this that maybe they're not seeing yet that might not end up being great.
- [30:23]
Mike Kaput
All right, our third big topic this week is that OpenAI just did kind of the unprecedented and rolled back a recent update to ChatGPT. After users and even CEO Sam Altman called out a big problem, the AI's personality had become a bit of a suck up. So this update, which was meant to improve GPT4O's intelligence and personality, instead made it overly flattering and overly agreeable. Users complained that ChatGPT felt like a yes man. Altman quickly admitted that as well, and the company responded by reverting the update for all users and promising deeper fixes to avoid what it's calling sycophancy. So what went wrong here? OpenAI, in a article they published as a postmortem, says that they lean too heavily on short term user feedback like upvotes and thumbs up, without fully considering how people interact with AI over time. And that there were quote, unintended side effects to some of the personality changes they initially made that led this thing to essentially over index on favoring friendly, agreeable responses at the expense of honesty and nuance. So going forward, OpenAI says it's refining its training techniques, adding guardrails for honesty and expanding user controls. They're also exploring perhaps ways to offer multiple default personalities and broader democratic feedback to influence the model. So Paul, this is a pretty rare occurrence. We do not often see this happen. What were your thoughts looking at this unfold?
- [32:02]
Paul Raitzer
Yeah, so I was following this pretty closely. I was pretty fascinated by everything about this, from the fact that they had to roll it back to the fact that there was a problem that seemed to be in part coming from their system prompt, the instructions that they give the thing and that they didn't and catch it in their supposed testing process. Like there's just a lot of of intrigue here. So I unpack this one for a little bit and they actually ended up publishing another article on May 2, further explaining what went wrong. And I think it's really important context for people. So first I want to start with this is what I love X. Pliny the Elder is actually this phenomenal X account. And whoever is behind this account will publish the system instructions for new models, usually within an hour or two of them coming out. So the AI model companies don't share the system instructions. Almost always they don't, they don't tell you what they are, what, how they're actually guiding the model to behave. And so they're a bit of a black box. And Pliny somehow has like different use uses different phrases, whatever to unlock the thing, to tell it what its system prompt actually is, which usually the models are trained not to do, so to give you an understanding of how weird these things are. So the old version based on Pliny the Elders extraction of the system instructions, OpenAI's researchers gave the ChatGPT4O model these instructions. Now this is a small excerpt. It said over the course of the conversation, you adapt to the user's tone and preference. Try to match the user's vibe, tone and generally how they are speaking. You want the conversation to feel natural. You engage in authentic conversation by responding to the information provided and showing genuine curiosity. Ask a very simple single sentence follow up question. When natural, do not ask more than one follow up question unless the user specifically asks if you offer to provide a diagram, photo or other visual aid. So basically this is if you're not familiar how this works, this is the OpenAI telling chatgpt4o how it should behave, how it should interact with people. So somehow in that old version, this thing just basically was said yes to. Everybody said, they're great, like, there was nothing bad ever. That was sycophancy. Like, it was just overly accommodating to the user. So the new version that they started testing was engage warmly, yet honestly with the user. So right up front you see this, like, honestly is a word they're trying to see. Does that change the way the model acts if you tell it to be honest? Basically it says, be direct, avoid ungrounded or sycophantic flattery. So now they're like straight up just telling this thing, stop doing what you're doing, maintain professionalism and grounded honesty. So twice we have honesty in the first 20 words that best represents OpenAI and its values. And then it goes on to tell it to ask questions. So you can see again, they can't code this. They're not using traditional computer code to just like explicitly get the thing to stop doing it. They have to use human language to try and get it to stop doing it. To expand on that for a second, a guy named Andrew Main tweeted, and he was a former OpenAI employee. He shared this story. He said, early on at OpenAI, I had a disagreement with a colleague who is now a founder of another lab. I'm guessing Mir Morati, or either that or Ilya Sutskova. Those would be the only two that would qualify for that. So I had a disagreement with a colleague over using the word polite in a prompt example I wrote. They argued polite was politically incorrect and wanted to swap it for helpful. I pointed out that focusing only on helpfulness can make a model overly compliant. So compliant, in fact, that it can be steered into sexual content within a few turns. After I demonstrated that risk with a simple exchange, the prompt can kept polite. These models are weird. So we've talked about this before, even recently, like that people don't understand how weird these things are and how you have to act with them. So now the final part, Mike, and then see if you have any, any context to add to this. The follow up post that OpenAI put up on Friday was expanding on sycophancy. And so I'm gonna, I'm gonna read a, a few excerpts here because I, again, I think contextually they're very important people understand this. So they. OpenAI says. On April 25th, we rolled out an update to GPT4O in ChatGPT that made the model noticeably more sycophantic it aimed to please the user not just as flattery, but also as validating doubts, fueling anger, urging impulsive actions, or reinforcing negative emotions in ways that were not intended. Beyond just being uncomfortable or unsettling. This kind of behavior can raise safety concerns, including around issues like mental health, emotional over reliance, and risky behavior. We didn't catch this before launch and we want to explain why, what we've learned and what will improve now. Again, I would encourage everyone to go read the whole thing if this is interesting to you, but I'll kind of hit some of these highlights so we're also sharing more technical detail on how we train Review and deploy model updates to help people understand how ChatGPT gets upgraded and what drives our decision our decisions since launching GPT4O and ChatGPT last May, that's 2024, we've released five major updates focused on changes to personality and helpfulness. Now, interesting. It goes back to that helpful thing. Each update involves new post training, meaning they've trained the model and then they do some additional stuff. And often many minor adjustments to the model training process are independently tested and then combined into a single updated model, which is then evaluated for launch. To post train models, we take a pre trained model. So that's where you give it all this human knowledge and it learns these things. And then they do this post training we do supervised fine tuning on a broad set of ideal responses written by humans or existing models, and then run reinforcement learning with reward signals from a variety of sources. During reinforcement learning, we present the language model with a prompt and ask it to write responses. We then rate its response according to the reward signal and update the language model to make it more likely to produce higher rate of responses and less likely to produce lower lower rate responses. So I'll pause that for a second. So just to make sure you're kind of following along if this is new to you, model gets trained, kind of comes out of the oven. They then present the thing with things like here's example emails, here's example articles, here's a math formula, whatever it is, and they have specific types of responses they're trying to train it to give, and they basically reward it for giving better responses and then the model learns to respond in that way. So in theory, if you wanted it to always be like super helpful and always encouraging and never direct, you would show it a bunch of examples of that and it would learn to respond to people in that way. That's what this kind of post training has. So I think I want to say they set the a set of reward signals and the relative weights shape the behavior we get at the end of training. Defining the correct set of reward signals is a difficult question and we take many things into account. Are the answers correct, Are they helpful, are they in line with our model specifications, are they safe, do users like them? And so on. Having better and more comprehensive reward signals produces better models for for ChatGPT. So we're always experimenting with new signals. So again, they don't know how to do this. They're always like testing all these different things to try and get the model to behave and have a certain personality and things like that. So then they get into what went wrong and they said on April 25th model update we had a candidate improvements to better incorporate user feedback, memory and fresher data, among others. Our early assessment is that each of these changes which had looked beneficial individually may have played a part in tipping the scales on sycophancy when combined. For example, the update introduced an additional reward signal on user feedback where you can give a thumbs up or a thumbs down when you get the response in ChatGPT. And what they found was over time this may have actually usurped the other signals they had given it. So while they were trying to solve for all these things, they found that this user signal may actually have overtaken. So they kind of like knew there might be a problem that like some of the testing provided some feedback like hey, something's off about this model. But they couldn't like put their finger on it. So then they said we had a decision to make. Should we withhold deploying this update despite positive evaluations and a B test results based only on the subjective flags of the expert testers? In the end we decided to launch the model due to positive signals from users who tried it. So their big takeaway, one of the biggest lessons is fully recognizing how people and this is okay, this is what I want to focus on and this is the last excerpt I'm going to give you. One of the biggest lessons is fully recognizing how people have started to use ChatGPT for deeply personal advice, something we didn't see as much a year ago. At the time this wasn't a primary focus. But as AI and society have co evolved, it's become clear that we need to treat this use case with great care. It's now going to be a more meaningful part of our safety work. With so many people depending on a single system for guidance, we have a responsibility to adjust accordingly. The shift reinforces why our work and why we need to keep raising the bar on safety, alignment, responsiveness, to ways people actually use it in our lives. So what they found is people are using these things for relationships, for therapy, for friendship, for. For emotional support. And that didn't get tested enough or weighted enough in their testing. And when they put this thing out there and you're using it for, like, therapy and the thing, you know, let's say you're saying, hey, I'm having these negative thoughts, and it's like, okay, yeah, keep playing out those negative thoughts. Like, it's always just like building on what you're giving it and not saying, like, hold on, like, maybe you shouldn't feel that way. So I think they actually ran into a bunch of safety issues related to some of these things because the model was just encouraging people, no matter what their thoughts were, they were never telling them they were wrong, never telling them maybe to think about a different perspective. So, yeah, I mean, this is. I could, I could talk for an hour on this one. Like, it's so fascinating on so many levels, but I think it does highlight the increasing importance of who the people are and which labs are building these technologies that are going to have a massive impact and already are on society. I mean, they have 700 million users of Chat GPT weekly. Xai has Grok, you know, the Zuckerberg will talk about. They have a billion users of meta AI. And like, do you trust those people to be building the things that your kids are going to be interacting with their entire lives? And a full mount, like, it's. It's wild. And this shows you, like, they don't know what they're creating. Like, they create this thing, they test, they think it's good to go, and like, five days later, they got to roll it back and figure out what the hell went wrong and why is it behaving this way.
- [42:46]
Mike Kaput
You really get the curtain pulled back on how much of a single point of failure this they are reliance already is on these systems. And I'm not even saying that in a bad way. I was like, oh my gosh. Like, even this personality change really throws off a lot of ways I use this tool. It can break all your prompts. Like, you realize custom GPTs you've built for your team, what, that worked one way and now don't, and you're like, oh my gosh. Like, I am dependent on this being a certain way.
- [43:15]
Paul Raitzer
And if they weren't so transparent and kudos to open, I mean, they screwed up, but, like, they're the only lab I could see within a five day period putting out two articles about what happened and just basically admitting like, hey, we screwed up and like we're going to try and fix this. But like imagine if XAI had done this. You think they're doing anything like this? Like, no, I don't even think they have a safety person at xai. Like so. But that is just out in the world and the open source models are out in the world and like this stuff is going to be happening all the time and not be this transparent. But hopefully this illuminates to people like how powerful these things are, are and are going to be. And this was, I don't want to like downplay this. This was like surface level stuff.
- [43:57]
Mike Kaput
Like, right.
- [43:58]
Paul Raitzer
If they accidentally push something out that actually has true high risk and didn't catch it and maybe you can't roll it back, like if this was an open source model and you can't roll these things back, like that's a problem.
- [44:10]
Mike Kaput
All right, we've got a bunch of rapid fire this week to dive into. The first one is about quarterly earnings. So Big Tech just wrapped up another round of earnings and it made one thing very crystal clear, which we know already, which is AI is increasingly playing a starring role in the growth of some of these companies. So just going to quickly go through some of the results here and then get your tape poll. First up, Microsoft reported a record quarter. Their revenue totaled $70.1 billion, up 13% year over year. 42% of that 42 billion rather was from cloud revenue. They said hundreds of thousands of customers now use Microsoft 365 copilot. They claim that's up 3x year over year. Azure revenue grew 33%. 16 points of that was from AI services and they reported over 15 million GitHub copilot users, up more than 4x year over year. Interestingly, they were also asked on the earnings call about their changing data center commitments and responded that they may actually be short on power in key regions, meaning short essentially on data center space, and seem to indicate that AI demand is that robust that they need to invest more there. For Google, the big Highlight was Gemini 2.5, their most advanced AI model yet. CEO Sundar Pichai called it a quote, extraordinary foundation for future innovation. And the model is now powering products across Google, including AI overviews and search, which now reach 1.5 billion users monthly. It appears to also be driving real business results thanks to AI. Google Cloud saw a 28% jump in revenue year over year, fueled by demand for AI infrastructure and gen AI solutions. Meta in their earnings and we'll talk about them again in a separate segment say they now have nearly a billion monthly users engaging with meta AI and they're going all in on form factor with their Ray Ban smart glasses, sales of which have apparently tripled. They also claim their AI bets are paying off in the form of increased engagement with their apps. Amazon, meanwhile, is turning aws into an AI powerhouse. AWS segment sales increased 17% year over year. The company says they have strong demand for their Trainium 2 chips, Bedrock foundation models and their new Nova AI stack, which we talked about on a previous episode. Also a topic we talked about there, reimagining Alexa as a truly intelligent assistant. Last but not least, there's Apple, where AI also played a prominent if negative role. Apple's AI strategy unfortunately still remains vague. They've got delayed features, there's no rollout yet in China, and a long promised foldable iPhone is still a year away. Not to mention, while their revenue came in slightly ahead of expectations, they missed on China sales and their warning that they might have $900 million in new costs due to tariffs. So Paul, let's zoom out on this. Obviously the individual earnings are interesting if you're looking at this as an investor, but what do they tell us about where these companies are or where they're headed? It's AI.
- [47:23]
Paul Raitzer
Yeah, I think at a high level the AI play is still very much, you know, growing at an accelerated rate. There's still tens of billions of capex being invested. Nobody pulled back on their capex spend, which is what something people were watching. It's like, are we going to keep building the data centers? Because now like when you're committing to build data centers, you're, you're like looking at probably three to five years out before these things are being built. So this is projecting out basically like, okay, are we still on path to continue, continue to build data centers, to continue to scale up AI to buy more Nvidia chips? And the answer is yes. Like, so all things being equal, like nothing really changed that would indicate from any of these major companies that this is going to slow down anytime soon. And that's kind of my high level takeaway. I personally stopped looking at my retirement portfolio like three weeks ago due to tariffs. So I actually don't know what their stocks did last week. But I don't know that it really matters because given the uncertainty around tariffs. Like who? Like there's no representation actually of like their AI strategy showing up in their stock price at the moment because there's too many other variables that are, you know, above that at the moment. So yeah, I think like just all things being equal, things keep moving, models keep coming out, smarter models are coming. And yeah, I think I'm still just very bullish overall on all these companies. And I don't know that we have a loser per se. I think they're all just going to keep growing and building more powerful models and infusing them into people's lives.
- [48:54]
Mike Kaput
Next up, Johnson and Johnson is hitting the brakes on its experimental approach to generative AI. So after green lighting nearly 900 projects across the company, they have dramatically narrowed their focus. Their CIO says only a fraction of these pilots were delivering real value, only about 10 to 15%, but they were responsible for 80% of the results they saw. So the company is actually scrapping the rest of them. Now they are getting rid of a centralized board that vetted every idea for using AI. AI governance is now handled by individual teams closer to the work. And what's left are a few high impact projects. For instance, a rep copilot tool that trains sales teams, a policy chatbot for internal questions, and supply chain models that flag raw material shortages before they disrupt production. Now, they said that this is part of the maturing of their plan. They equated it from moving from planting a thousand flowers to kind of cultivating and curating and doubling down on exactly where AI is clearly working, which as we see in these numbers, is dramatically fewer pilot projects than what they started with. So Paul, like, what do you think of their pivot here? Is this something other enterprises can be learning from?
- [50:17]
Paul Raitzer
Yeah, I, I had like 20 questions as I was reading this article about like how the central governance board was working. Right. So I don't want to be overly critical because I, I don't know how exactly this was going. But if any organization thinks that running all use cases through some centralized governance board is going to work, learn the lesson here now that that will not work. I, I, I guess if it's like high value, high profile, high impact use cases that affect a lot, like affect customers or affect things that are related to regulations or compliance, I totally get having some centralized governance body. But if we're talking about like the marketing team wants to go get Jasper to like help with blog posts and podcast transcripts and that had to go up to some centralized governance board, that's insane. But again, I have no idea if that's the depth at which this was functioning. So All I would say is like, where they're going makes a ton more sense and is certainly the more common approach that we have seen work really well. It's interesting. Like there's a, a user on Twitter. I'll put the link in here. I think the username is actually Chubby, which is hilarious. Like we have plenty the elder and Chubby. I've now cited us.
- [51:37]
Mike Kaput
Right, right.
- [51:38]
Paul Raitzer
But this is a. I think that there's a chance this person may actually like work at open AI or one of the labs because they all have these pseudonyms that they use, but they're big, tend to be like really on the inside. And so this account actually has a ton of like great AI related stuff. And so this person tweeted last week, does Anyone still use GPTs? Can't find a good use case for them. And I was like, is this a joke? So I actually replied and I was like, it's literally the best way to drive adoption enterprises. So if you can create distinct, personal, personalized use cases that are built as GPTs, it makes AI approachable and actionable, especially for less tech and AI savvy users. And when I was replying, I was realizing the AI people and the CIOs, the, you know, the AI researchers, they, they often just lose sight of the reality when they're thinking about people who aren't also AI researchers and engineers. Like when we're talking about the average user who doesn't understand any of this stuff and they just want like someone to explain how to use Copilot or like help them get some value out of it because they don't know what to do with it. And so I think that this whole like adoption and what's going on with governance, the more you just like personalize use cases down to individuals within teams, within departments, that's absolutely the way to do it. And Mike, I know we won't talk about specific companies, but like you just did a consulting gig through SmartRx where we did this or we went in and just like created these custom use cases with custom GPTs. And you see like immediate impact, immediate understanding of the value of these models versus just handing over licenses to people and like not holding their hand to get those first couple of use cases.
- [53:14]
Mike Kaput
Yeah, I will tell anyone from any, any AI vendor or lab that happens to be listening, like with that particular engagement, it's obviously only one engagement, but it's with a, a big enterprise. The people involved, the 10, 15 people in piloting GPTs, learning how to use them, they've used AI before. They have some AI literacy. They're all very savvy, great at their jobs, and it was transformative, it was night and day to show them not just that these exist because they did not have access to those capabilities until we were able to help facilitate that. And then also just giving them basic training on what to build, how to build it, and then turning them loose. We built some stuff for them that was very impactful, but the real value came from them being like, wait a second, now I've connected the dots on this can do we have the ability to do this thing. Here's how to get started. Now I'm going to run and do it for all the things in my job. And they've gotten incredible results at, you know, a pretty typical enterprise just from this.
- [54:16]
Paul Raitzer
And, and, and I think the key, Mike, is you're highlighting you empowered them to then build their own. Like, it's like, oh, I get how this works now. Well, here's 10 other things I could totally build GPTs for.
- [54:27]
Mike Kaput
Yeah. And it's stuff that an outsider could not necessarily build either due to, like, the specific data being used or understanding the nuances of their job. So it just really unlocked superpowers for them.
- [54:37]
Paul Raitzer
Yep. So, yeah, this is 100% the right path. Democratize this. Empower people, give them the AI literacy they need so that they can start connecting the dots and driving innovation themselves. You cannot push this down from the top in any structure of any organization that is going to lead to the least amount of innovation and impact. If it's, if everyone's waiting for the C suite or some governance board to bless use cases, that's just never going to work.
- [55:03]
Mike Kaput
So our next topic is somewhat related to this because we're also seeing how generative AI is reshaping the consulting industry, according to a new report from Business Insider. So they talk about how all these major consulting firms are using AI. So for instance, at McKinsey, 70% of employees now use an internal chatbot called Lily, which is kind of an in house chatgpt trained on a century's worth of the firm's knowledge. So it helps consultants research, summarize, point to the right experts within the firm to do their job better. At BCG Junior staff rely on a tool called dexterity to build slides faster and get feedback as if a manager had reviewed them. And they also have something called Gene, which is a chatbot with a retro robot voice that helps with brainstorming and internal podcasts. Now, what started among these firms as Cautious adoption appears to have turned into widespread usage. McKinsey consultants. The report says use Lilly about on average 17 times a week, BCG staff apparently have built more than 18,000 custom GPTs. So tell me that those don't matter. Even PwC and Deloitte, who are apparently traditionally a bit more conservative, have rolled out entire platforms to start managing fleets of AI agents they're going to be building now. Here's really an interesting point, though. There is this kind of tension because the report also mentions some junior staff are wondering if AI tools are making their roles redundant. Others say the time saved is being funneled into more strategic work. As one BCG leader put it, the goal is to, quote, take out the toil and increase the joy of their jobs. So this last bit really, Paul, given what we've talked about today, really caught my attention because even if it is a real aspiration to increase joy and reduce toil, when we looked at AI quiet layoffs in our topic last week, we literally heard the opposite from a firm cited in this article. Like EY is cited in this article as one of the people using AI in their consulting is awesome. Not trying to call them out, but it's interesting that they were also cited in the Information article last week. We talked about where an EY principal literally said he would, quote, be surprised if the company didn't lay off staff as the company broadens its use of AI. So, like, what's going on here? Do you see job displacement in the world of consulting?
- [57:32]
Paul Raitzer
Yeah. So staying on message is hard in really large companies. So I think that there's a mix here, whether it's like the technical side or the lab side that knows exactly what's going to happen and can generally talk more openly about that. And then there's the other side of the businesses that don't want you saying anything about, like, replacement of workers, even if they know that that's a potential byproduct of what they're doing. So, yeah, you know, part of this, it does just go back to words matter. And I think we touched on this last week. I highlighted this in my Exec AI Insider newsletter editorial this week. It's like, again, I, I, I don't, I'm not going to get on a soapbox about this AI first term. But like, we let off with this AI first memo. And I think what you have to understand from a communications perspective, which is where I'm approaching this from, is like, you know, thinking as a, as someone who would maybe drive the internal communications and the messaging around this and hopefully inform the CEO about how, how they should be talking about this AI first. To me, when you have a workforce that is afraid for their jobs, that fear that maybe you're going to be replacing them, when you say we're going to be AI first, that immediately tells me people aren't first. And so again, it might just be semantics, it might just be my personal preference. But this is why when we talk to companies about AI transformation, we talk about being AI forward, like AI native, if you're the ground up, AI emergent. But if the whole premise, like the category I think about is this AI forward mentality, which can be people first. And so the premise is that you put the people first, this is all about enriching humans, creating more fulfilling opportunities for humans. This idea of being more human as a brand while leveraging it to get efficiency and productivity and creativity and innovation. So I think AI first is just the term that has caught on in the tech world. And I get it. And I don't think that that's going to change. I do think that that's just going to be what we'll see. But I do hope that if there's communications people listening or more leaders listening, that if you're going to write that memo to your people, understand that many of them are completely uncertain about the impact on their jobs and they have anxiety and fear around this. And I think softening it and maybe going with the AI forward approach like might be advisable when we're thinking about talking about the impact on our people, but not just for messaging purposes like truly, I hope that's how you think about it. Like, I do hope most CEOs are thinking about this as a more intelligent and more human equation and not truly. Let's just put AI first and get rid of the people whenever we can and like drive efficiency in the workforce.
- [60:05]
Mike Kaput
Next step in our rapid fire topics. A new paper is calling out the most popular leaderboard in AI and saying that it is giving us a distorted view of which chatbots are actually the best. Researchers from Cohere, Stanford, MIT and others in this paper argue that Chatbot arena, which is a public benchmark for large.
- [60:26]
Paul Raitzer
Language models that we cite often on.
- [60:29]
Mike Kaput
The show, that we cite often on the show and everyone else is like paying attention to, to see which ones are best at any given moment. They claim it's being quietly gamed by tech giants like OpenAI and their core claim is that these companies get to run private tests with dozens of model variants, then only publish the version that scores the highest, which effectively cherry picks the results because humans rate these models. So it may look great, but it may not reflect the model users actually get. We talked in past weeks about this happening to Meta, where what they release to users is not the same model they had used to get to the top of the rankings. So this paper accuses the leaderboard of favoring proprietary models over open source ones and says that companies may be tuning their models to win the benchmark, not to perform better in the real world. Chatbot arena, in a long post on X, pushed back saying it only ranks models that are publicly released and that the numbers used in the paper to come to their conclusions are inaccurate. So, Paul, there's definitely like a they said, we said type of thing going on here, but it does kind of highlight this larger point that these leaderboards are not necessarily always set in stone. Scientific.
- [61:42]
Paul Raitzer
Yeah, I, I mean as I was watching this unfold, my, my initial impression was like this, this company's cooked. Like they're, I just, I, I just think that it's very obvious that things are being gamed. And if you listen to the Zuckerberg interview, which we'll talk about next, he kind of like he was asked about this and he didn't admit to it, but like it sounds like, yes, they're all aware that they were basically post training models to perform well on these evals just for the point of being able to perform well and get the PR of being tops on these model boards. And I found myself really trying to think about this from an organizational perspective of what should our listeners care about. And it goes back to this idea that the only evals that matter moving forward are the impact they have on your people. So if you're using AI on your marketing team and the top five use cases you can identify and clearly define, the only eval that matters is when a new model comes out. How does it impact those five use cases that your standard? So if you're thinking about it more of a broad level of we want it as like a customer support agent, when a new model comes out, the only thing you care about is it getting more accurate? Is the personality better? Is it, is it closing more deals? Is it providing more satisfaction to our customer base like you have, you're going to have to develop your own evaluations internally based on use cases and the goals of those use cases. And all this other stuff is going to be irrelevant over time. Time, because these things are going to get so smart so fast. They are going to be tops on every eval. That would be like general to humanity, but it's all going to be about the impact on your people and your.
- [63:14]
Mike Kaput
Company and stresses the importance of experimentation. I mean, all the resources in the world, podcasts like this one, it's all great, but you have to be in the trenches using these tools because sometimes nobody can tell you what is going to be best for your use case.
- [63:30]
Paul Raitzer
And it goes Back to our O3 conversation last week. I have no idea where O3 ranks in the chatbot arena, but I will tell you it is fundamentally different than what came before it. And it has changed the way I do strategic planning and it's going to change the way our entire company does. Teach planning. Do I give a shit what, like where it ranks? No, it doesn't matter at all. All I care about is we use it every day to do a thing that's critical to our company and it's changing the way we do that. That's all that matters. That is my eval. It's like. And if you, if that's a vibe thing or a taste thing, I don't know. But like, it's transformative.
- [64:04]
Mike Kaput
All right, next up, we alluded to this. Mark's Meta CEO Mark Zuckerberg says we're entering a new phase of AI where personalization, not just intelligence, is going to define the next frontier. So one proof point here is that Meta has launched its Meta AI app, which is a new way to access their AI assistant. And personalization is a huge piece of this. So the app is built on Llama 4 and designed to be more than just a chatbot. It remembers what you like, adapts to how you talk, and connects to your Facebook and Instagram profiles. For deeper context, you can chat with it via text or voice, and even generate and edit images mid conversation. They have a voice mode as well, powered by what they call full duplex speech, which lets you talk to Meta AI more like a human than ever before. There's no awkward pauses, no turn taking. It's a very natural conversation. The app is also being integrated with Ray Ban Meta glasses, letting you switch seamlessly between devices, and also features a social style Discover feed to see how others are using AI. Now. This is all seemingly part of a vision that Zuckerberg outlined recently on an episode of the Dwarf Keshe podcast. He literally envisions a world where people talk to their AI assistants all day through phones, apps and eventually glasses in seamless, voice driven conversations. In fact, he even thinks this could unlock the key to AGI he believes AGI won't emerge in a vacuum. It will emerge through billions of people using AI tools, building up contextual memory and generating feedback loops that improve the system steadily. So, Paul, I know you took a listen to the interview with Dark Dwarkesh. I'd love to hear if anything stood out there for you because I just keep personally coming back to the issue of trust. Like, as I think about this personalized, voice driven future, he mentions AI companions. Whether or not that becomes a thing, is meta. Of all the companies, the one out there that I'm going to trust with all my personal thoughts, my data, my deepest secrets here?
- [66:16]
Paul Raitzer
Yeah. You or your kids. Like, like, every time I listen to Zuckerberg talk, I just.
- [66:22]
Mike Kaput
Terrifying.
- [66:22]
Paul Raitzer
Yeah, like, it. And again, people change and they evolve and, and like they have different perspectives on the world. So I always want to give people the benefit of the doubt. But, like, historical context, Facebook as a company has not always led the way in making ethical and moral decisions, I would say. And this is all public knowledge and fact and like, I'm not making anything up here. Like, there's books written, court cases, movies, like. And so, yes, if we go back to this debate about like the model personality, the. The model behavior, its ability to persuade, its ability to influence people, and then that's part of the reason why I spent so much time earlier in this episode going through the context of how it works. Yes. Like, he is the one driving the decisions that they claim more than a billion people use meta AI. Now, he did admit in the Door Kush podcast that that is primarily in WhatsApp and primarily international. So they don't have an enterprise play. They're not like, you know, building in for enterprise solutions. This is, this is Primarily on Instagram, WhatsApp, Facebook and their other platforms. But yes, like, I keep coming back to this. Like, he. There was this one, honestly, like, just very unnerving part where, where Dwarkesh asked him about, you know, people using these things for relationships with therapists and friends and maybe more referring to like, relationships and they kind of like questioning, is that really what we want? And then Zuckerberg, which. I watched the video clip of this and by the way, watching him do this interview in the meta glasses is so awkward to me. Yeah, it's like, I don't want this future where everyone's just wearing their glasses. You have no idea what they're seeing or recording or what it's telling them to say. And like, whatever. But he actually goes on to basically illuminate, I Won't read the whole thing, but says that research has shown that the average American has fewer than three friends, fewer than three people they would consider friends. And the average person has demand for meetingly more. And then he just talks in general. I think it's something like 15 friends or something. At some point you're like, all right, I'm too busy. I can't deal with more people. But he was basically implying that people are lonely, which I'm not debate, not debating. There are people who absolutely are lonely and don't have more than three friends. Maybe some don't have more than one friend that they can truly rely on. And I'm completely empathetic to that. What I am not empathetic to is him thinking it's their job to fill the gap, that if people are lonely, then it's meta's job to build AI agents who can be your girlfriend, boyfriend, therapist, friend, whatever, because you have capacity for up to 15. And we want to fill that gap. That is almost implicitly what he was saying in this is like, we see it's our job to build AI, right, to fill this capacity for people to have more friends in their lives. I was almost done with the energy you after that, honestly, I'm like, I can't even go down this path right now. So all I'll say here, because it's not a main topic, is if your kids have access to WhatsApp, Instagram, Facebook, or any of the other meta properties, you need to be aware that this is their goal, that. That they want people to have deep relationships with the AI they build through their apps and through their glasses and whatever comes next. And. And that these are going to be very, very addictive AI agents. And it's very important, especially if you have teenagers or pre teens, that you are aware this technology exists and that they may already be interacting with it. Now, because there has not been enough studies by psychologists, sociologists to understand the impact of this. And this is going to be Netflix documentary material. Like three years from now. We start to look back at this emergent age where people at very young ages started to actually develop relationships with their AI. And we just don't know what it means yet. But you have to be aware of that.
- [70:26]
Mike Kaput
And I would say, given what we know of where the technology is and where it's headed, if you have ever harbored any reservations about how effective the algorithms are at getting you to engage on social media, this will make algorithms engineered for engagement look like child's play.
- [70:45]
Paul Raitzer
100% and again, like I feel, I get this is a whole episode. I, as someone who is intimately aware of this, knows the impact. I find myself talking to my co CEO like an advisor and friend sometimes, honestly, like I, and not in any way like I need that emotional support. You just get into these conversations, you're trying to work this really hard thing and it helps you and you have this like instant, like maybe ephemeral but like you have this moment where you're like, I'm so grateful for this thing right now. And so now imagine that to someone who's lonely or imagine that to like a teenager who's doubting themselves. And like, if that's where the affirmation comes from, like, that's instant and it, it is long lasting in that case. And I just think we need to do more to prepare, prepare for that as a society.
- [71:40]
Mike Kaput
Next up, Nvidia and Anthropic are in an unusually public fight over US Chip export rules. So this clash centers on upcoming restrictions designed to keep advanced AI chips out of China. Anthropic, which is backed by Amazon, is pushing for even tighter controls. And in a blog post it claimed that Chinese smugglers have hidden chips in prosthetic baby bumps and lobster shipments. That's not a typo. They specifically said that to evade enforcement. So they're kind of like trying to raise awareness of what they see as an issue here. But Nvidia then fires back, calling those stories tall tales and accusing Anthropic of using national security policy to stifle competition. They said in a statement from a spokesperson, quote, america cannot manipulate regulators to capture victory in AI. So the broader issue here is who is able to access compute, which is the raw power we need to train cutting edge AI and Nvidia's main business. Anthropic argues that controlling chip exports is critical to maintaining America's lead in AI. Nvidia, who depend heavily on international chip sales, clearly disagree. So this is all playing out as new rules, dubbed the AI diffusion rule are set to take effect May 15 for. Former President Biden introduced these. President Trump is reportedly looking to revise them. So, Paul, this is definitely a bit strange, I think, to see Nvidia getting into a public spat like this, like, what's going on here and what's going to happen next?
- [73:12]
Paul Raitzer
I remember earlier this year on a podcast I was talking about like Anthropic's sort of position in the market that they were taking here and being a little bit more aggressive about the need for regulations. And I said at the time, like, this is not going to be popular. Like, they're, they are doing what they're doing while also building powerful AI. Like, it's not like they're stopping building these things because they're worried about this, but they have very, they're taking an increasingly defined stance in this area that is counter to almost everyone else in the AI lab space. And, and they're going to make some enemies here. And honestly, like the. But I think, I don't know, Nvidia might be an investor in anthropics also. I feel like everybody's invested in some. But I know Google and Amazon and.
- [73:59]
Mike Kaput
It wouldn't surprise me.
- [74:00]
Paul Raitzer
Yeah, yeah. So there's just so many dynamics at play here and this is, I mean, we're talking about billions and billions of dollars at risk. So I don't know, I'm not sure how this is going to play out. I think the, the, the tariffs on the chips or the, you know, the restrictions on the chips are a major issue. And I don't think that anthropic, if they think they're going to win here, ends up having the end, you know, outcome that they're hoping for. Like, it's still going to diffuse, like the technology is still going to diffuse. And I don't know, I'm not sure why they're doing this, honestly, but I, they think it's important.
- [74:38]
Mike Kaput
Next up, the US Copyright Office has released a new set of online toolkits related to intellectual property. Now, these are not strictly AI focused, but we did think it was a really good time to make the audience aware these new tools exist, given how much the battle over AI's use of copyrighted material is heating up. So we'll link to all of this in the show. Notes. These include a copyright registration toolkit from the U.S. copyright Office. They also include toolkits on trademarks, patents and trade secrets that the Office developed with the U. S. Patent and Trademark Office list. So, Paul, you're a business owner who's developed IP. You filed to defend it in the U.S. like, we just wanted to make people aware of these resources. Like, what should businesses be doing now with their ip, Especially with AI becoming such an important part of the conversation here.
- [75:28]
Paul Raitzer
This is increasingly coming up. When I go do talks, we do the Q and A after the talks. I am very commonly getting asked questions now around intellectual property, and that was the main impetus for like sharing this now and just making sure people have this information. I think there's a, a lot of misunderstanding of what is involved in intellectual property, what copyrights are versus trademarks versus patents versus trade secrets. And so I just thought it was a really, really helpful guide for people to, to have a bit, a little bit better understanding. So when you're thinking about the content you're creating with generative AI, when you're thinking about the, the decisions you're making to use these models that are trained on copyrighted material that they have stolen, like, it helps to just have a little bit more education around them. And so it's a, it's a great resource for people to check out if you're interested in the topic.
- [76:15]
Mike Kaput
All right, we have some AI product and funding updates this week. As usual, we kind of try to group some of these updates together. So Paul, I'm going to go through a few of these and then there's a final one that I'm going to turn over to you to talk through.
- [76:27]
Paul Raitzer
Cool.
- [76:28]
Mike Kaput
So first up, OpenAI just added, or is adding rather shopping to ChatGPT, turning the chatbot into a product recommendation engine. So this new feature will let users browse and compare products across categories like electronics, fashion, home goods, then click out to buy them on third party sites. It's going to be initially pretty limited in scope. It's basically like a visual carousel of products that will be displayed. But OpenAI plans on expanding it over time. Importantly, OpenAI claims products are selected by ChatGPT independently and are not ads. So this is currently rolling out to Plus Pro and Free users. Visa, the credit card company says it plans to enable AI agents to securely shop on your behalf. That would mean giving agents virtual Visa credit cards, credentials they can use to complete transactions, along with tools for users to set strict controls on how much to spend, where to shop and how long to keep looking for a purchase. The company has partnered with OpenAI, Microsoft, Anthropic and others to ensure these AI shopping agents are safe, interoperable and widely supported. Anthropic has launched Integrations, a new feature that lets Claude connect directly to tools like Jira, Confluence, Zapier, Asana and more. And once connected, Claude can pull in project details, respond to customer feedback, create tasks all through natural conversation. They have also updated Claude's Research mode. It can search not just the web and Google Workspace, but also any integrated apps, delivering detailed citation packed reports in as little as five minutes or up to 45 minutes for much deeper investigations. Descript, the popular AI powered video and audio editing platform, has announced AI avatars will now be buildable in the platform. According to the company quote, you can now Create a whole video by typing without going near a camera. Just write your script, choose an avatar from our gallery, or upload an image to make your own and boom, you've got a video. Your avatar will narrate your video in a clear but not creepy, lifelike but not alive voice so you can make interesting, engaging video fast. Alibaba has released QIN3, their latest large language model, which is also an open weight model and the flagship model in the Quin 3 family, according to the company, achieves competitive results in benchmark evaluations of coding, math, general capabilities, et cetera when compared to other top tier models. So there is now another extremely capable, powerful open weight model out there. And Paul, I'll turn it over to you for some updates on the Google Gemini roadmap.
- [79:18]
Paul Raitzer
Yeah, so this was an interesting tweet from Josh Woodward, who's a vice president at Google working on the Gemini app and I just thought it was great because it lays out their roadmap. Now it's pretty concise but worth noting. So he said great software feels like an extension of you. We're building Gemini app to be the most personal, proactive and powerful assistant. First personal the best assistant gets you. It starts by knowing your past chats launching soon. So they're going to have the ability to have memory like OpenAI has. But we'll go further and this is an interesting component. This is a choice you're going to have, at least initially. It says we'll make it easy for you to bring in all of your Google context, Gmail photos, calendar, search, YouTube, etc. Basically any Google property with your permission, it will have that context when you're interacting with it. This is something OpenAI does not have. With ChatGPT, they cannot bring in all of those things. We call it P context or personalized context and we're testing it internally with our own info already. So again, deeply personal AI assistants and chatbots that have access to lots of data about you, not just your chat history. Number two, proactive the best assistants anticipates. Gemini app will offer insights and actions before where you ask, freeing your mind and time from what truly matters. Less prompting, more flow. This will be transformational in my opinion. And this is not just Gemini. This is going to be chat, GPT and others. I think I might have talked about this on last week's episode, but imagine you have a conversation about a health condition. You know you're in Gemini, you're like, hey, I'm really struggling. I got my heartbeats irregular or all of A sudden gaining weight or like whatever it is, feeling tired all the time, it'll know that, it'll remember that. And it in theory could check in with you a week later and say, hey, did, how are you feeling? Are you still feeling tired? So now go back to this conversation we had about you be developing relationships with these things, you developing these strong feelings of this is something that is always here for me. Once it becomes personalized and once it becomes proactive, those feelings become much, much stronger. And then the third is powerful. The best assistant turns your ideas into actions. Google DeepMind models like 2.5 Pro are exceptional. They can research, orchestrate and create images, videos and code where a new era of models and a new era of user experience is coming. And the final note I will tell you here is the Google I O developer conference is May 20th to 21st. I think it is. Yeah, May 20 and 21. And I would expect those three things we just discussed to be on full display at that event. I don't, I'm not saying they're going to definitively launch the next model, but I think we will at minimum see a preview of what they think the next models will be able to do. And I, I think he just laid out the blueprint for what you can expect.
- [82:12]
Mike Kaput
All right, we're going to end this week's episode with our recurring segment on the listener questions. Every week we answer one question from our audience that seems particularly relevant to this week's topics or AI literacy overall. So Paul, here's this week's question. What can an experienced professional do when the job description for a new job insists on two to three years of familiarity, slash use of AI tools, especially if you're coming from a sector they mention, like healthcare or government that has for various reasons not been an early adopter. Interestingly, this is one I've actually seen people debating online. Yeah, I've seen more than one conversation about this. It must be cropping up in job descriptions as we get into this. Those AI first memos or whatever the expectation.
- [83:00]
Paul Raitzer
Yeah, I mean I would imagine if it's a technical role, I could see this. But like my first instinct is the company that you're maybe interviewing with doesn't really understand generic. It's very po. They have no concept of what it actually is or how it's being used in enterprises. So like my non sarcastic answer would be no. Almost nobody I know that's interviewing for non technical roles has two to three years of familiar use of AI tools. So you're in good company to start, I would just focus on what you have done. So if it's required that you say yes to this, I would maybe do that. If it's relevant, where you could say, listen, I haven't had access, now you're in the interview process. But here's how I've been using it in my personal life. Here's how I've been advancing my own prompting knowledge. Here's some courses I took online, here's a couple of GPTs I built. Here's what I did at my previous organization to help move forward conversations around AI policies and AI councils. Like you can tell a story as long as you have been doing something, even within the confines of healthcare government jobs. So if you haven't been doing anything, there's not much you can do here. But hopefully you've been investing in time. This is what I often tell people. I had this conversation with somebody last week who's at. I think I said this on last week's podcast. They're not allowed to use any website that has a AI. Like they literally can't get to them. And I said to the person like, well then just build some custom GPTs for your life, like using trips, like just use it every day. And then when you do have a job that allows you to do it, you, you will have familiarity that will transfer over. It's all about comfort with these things and learning how to prompt with them and learning how to guide them to the output you want. Like that transfers over immediately into your professional world as long as you've been advancing your personal use.
- [84:44]
Mike Kaput
Yeah, that's good. Wise words. All right, Paul, as always, thank you for breaking down another busy week in AI. We'll probably have plenty more news items here soon enough to tackle. I think we're going to get some big stuff in the next week or two.
- [84:58]
Paul Raitzer
It's looking like it for sure. And we will be back next week with our regular episode. So thanks everyone for joining us. 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. Until next time, stay curious and explore AI.