68 Min Read

[The AI Show Episode 144]: ChatGPT’s New Memory, Shopify CEO’s Leaked “AI First” Memo, Google Cloud Next Releases, o3 and o4-mini Coming Soon & Llama 4’s Rocky Launch

By Claire Prudhomme on April 15, 2025

Podcasts
Featured Image

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

Learn More

Returning from Google Cloud Next, Paul and Mike are back with some major AI updates. They kick things off with ChatGPT’s new memory feature and unpack what that means for your data (and your daily workflows). Then it’s onto Shopify’s leaked memo: no new hires until AI proves it can’t do the job. Databox takes that even further by replacing 80% of its support team with a bot, and actually boosts performance.

Plus, Sam Altman gets grilled at TED2025, Apple’s AI efforts fall flat, and Paul shares what it was like inside the Sphere for Google’s Wizard of Oz AI experience. 

Listen Now

Watch the Video

Timestamps

00:03:38 — ChatGPT Memory

00:13:59 — Shopify CEO’s Leaked “AI First” Memo

00:20:20 — Databox Replaces 80% of Support Team with AI

00:28:25 — Google Cloud Next ‘25 and Google Updates

00:36:44 — OpenAI Will Release o3 and o4-mini After All

00:41:07 — Sam Altman Comments at TED

00:50:15 — OpenAI Pioneers Program

00:53:00 — OpenAI Reduces Model Safety Testing Time

00:57:07 — Llama 4 Release and Controversy

01:00:45 — AI Copyright and Creator Rights

01:05:52 — Anthropic $200 Per Month Subscription

01:08:43 — Behind the Scenes of Apple’s AI Failures

01:12:26 — Writer Releases AI HQ

01:16:07 — Ex-OpenAI CTO’s Startup Making Big Moves

01:20:29 — Deep Research’s Impact on Agencies

01:25:41 — Listener Questions

  • How do you filter out the signal from the noise in generative AI given that the space evolves daily?

Summary:

ChatGPT Memory

ChatGPT can remember all your past conversations, not just the ones you explicitly saved, thanks to a new feature release by OpenAI.

That means it can now reference previous chats to deliver more personalized responses—whether you're writing, brainstorming, or asking for advice.

This builds on last year’s memory feature, but goes further. Now, there are two types of memory: saved memories you ask ChatGPT to keep, and chat history references—insights it passively gathers to make future conversations smoother and more relevant.

It’s a big step toward OpenAI’s long-term vision of AI that grows alongside you, one that “gets to know you over your life,” as CEO Sam Altman put it.

The feature is rolling out now to Pro and Plus users, except in the EU and a few other countries due to tighter AI regulations. You can still opt out entirely or use temporary chat if you prefer not to be remembered.

Shopify CEO’s Leaked “AI First” Memo

At Shopify, the message is now clear: don’t ask for more headcount until you’ve proven AI can’t do the job.

In a memo to staff, CEO Tobi Lütke laid out a new standard: before requesting more people or resources, teams must show that artificial intelligence isn’t a viable option. He called AI usage a “baseline expectation” and said employees are now being evaluated on how well they integrate it into their daily work.

The memo, which was originally internal, was published by Lütke after he heard it was being leaked by the press.

Like many tech companies, Shopify has been trimming costs—and investing heavily in AI. The company’s own tools, like its “Sidekick” chatbot, are already designed to automate merchant support. Now that same logic is being turned inward.

Databox Replaces 80% of Support Team with AI

At analytics software company Databox, AI isn’t just a support tool—it’s become the support team.

Databox CEO Pete Caputa recently posted on Linkedin that the company replaced 80% of its customer support and sales development staff with an AI chatbot… and actually improved results by 40%.

It’s important to note, as Caputa does multiple times, that the company reduced headcount well before adopting AI for this use case. (He says they had 40 people on the sales development and customer support teams, but had to reduce headcount to 8 because the anticipated demand they hired for didn’t materialize.)

After that, the company deployed Intercom’s Finbot, which now resolves about half of all customer chats instantly. That freed up human reps to focus on personalized outreach to high-potential leads, helping drive more revenue.

Finbot’s customer satisfaction scores still lag behind humans—around 71% compared to 95%—but its speed and consistency make up for it. Databox improved those results by feeding it better content, like integration-specific help docs and use-case forums to handle the long tail of customer questions.

Caputa admits the next step is harder: giving bots the ability to log into accounts and troubleshoot like a human would. But 18 months ago, he wouldn’t have believed how far automation could take them.

This week’s episode is 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: I do believe that organizations are going to start to filter their employees and do their annual evaluations, or whenever it's happening, they are doing these evaluations. I do believe that AI literacy and competency are going to become a key filter for who stays and who goes. Welcome to the Artificial Intelligence Show, the podcast that helps your business grow smarter by making AI approachable and actionable.

[00:00:23] My name is Paul Roetzer. I'm the founder and CEO of Smarter X 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.

[00:00:52] Welcome to episode 144 of the Artificial Intelligence Show. I'm your host, Paul Roetzer, along with my co-host Mike Kaput. We are back in Cleveland [00:01:00] after, well, I had a full week in Vegas. Mike, you were, I got to two, two full days or whatever it was. 

[00:01:06] Mike Kaput: Yeah. 

[00:01:07] Paul Roetzer: so we were in Vegas for the Google Cloud next conference last week, as well as a couple of other events and, programming items.

[00:01:16] It was a long time to be in Vegas. It was an incredible event, but, five nights in Vegas is a lot, man. I don't, so yeah, so we were back. It was a crazy week in, in an average week. I think I'll, I, I've said this before, like I'll curate maybe like 30 to 40 links that Mike then goes through and like, you know, picks the things for the show.

[00:01:37] It was well north of 50 last week, and we, I mean, we literally just cut stuff as we were getting ready to come on here. So it, it, it is a lot. And I expect this week is going to maybe be on par with last week. So a ton to cover. We're gonna go kind of rapid fire style through almost all of these.

[00:01:57] There's, there's a couple, we'll linger on a little bit more, but it's [00:02:00] not gonna be the standard, like three main topics. And then the rest there was just so much to get through. We're gonna try and, move through things pretty quickly. But, all right, so before we get started, this episode is brought to us by the Marketing AI Conference, or MAICON.

[00:02:13] You can learn more at MAICON ai. That is MAICON.AI this is our sixth annual in-person conference. It's the flagship event for Marketing AI Institute. It's taking place October 14th to the 16th in Cleveland, Ohio, at the convention center right across from the Rock and Roll Hall of Fame and Cleveland Brown Stadi at least for now.

[00:02:34] Till the rounds may be moving a, a couple years. so we announced the first 19 speakers. There's plenty more announcements coming, but you can, check out the agenda. The preliminary, no, no, we didn't put the full agenda up yet. We still got last year's agenda, I think is the example, right? 

[00:02:47] Mike Kaput: Yeah. But we do have, we have some speakers.

[00:02:49] Okay. 

[00:02:50] Paul Roetzer: Yeah, so you can go to the speaker page, see the speakers, and we'll be updating the agenda soon. You can look at the four workshops. So there's optional, workshops on October [00:03:00] 14th, half day workshops on that day. You can go check those four out again. It is October 14th to the 16th in Cleveland.

[00:03:07] Prices go up on April 26th. You got a little time to take advantage of early bird pricing and you can go again to MAICON AI to learn more about that. I am really looking forward to this year. It's gonna be the best year yet. we have got some pretty cool things planned. Even just like last week I was meeting with the team to talk about.

[00:03:25] The, you know, the exhibit hall and some of the things we're planning on doing there. took a little inspiration from the Google Cloud next event. So, yeah, check that out and we'd love to see you in Cleveland in October. 

[00:03:38] ChatGPT Memory

[00:03:38] Mike Kaput: Alright, Paul, so let's dive into it. Our first topic that has happened in the past week is that chat, GPT can Now, if you so choose, remember and reference all of your past conversations with it, not just the ones you have explicitly saved, thanks to a new feature release from Open ai.

[00:03:59] [00:04:00] So what can happen now here is that chat, GPT can reference previous chats to deliver more personalized responses no matter what. You happen to be prompting chat GPT to do. Now this is different from and builds on top of the memory features released last year and goes a bit further. So. There used to be and still are, those kind of saved memories.

[00:04:22] You can ask Chat GBT to explicitly keep. But now there's also, if you so choose chat history references, which are insights, it passively gathers to make future conversations smoother and more relevant. So this is a big step towards opening AI's long-term vision of AI that grows alongside you and one that as CEO, Sam Altman put it, quote, gets to know you over your life.

[00:04:49] This feature is rolling out now to pro and plus users, except it is not at the moment in the EU and a few other countries due to AI regulations there. And you [00:05:00] can still opt out of this entirely or start a temporary chat session if you prefer for your conversation not to be referenced or remembered. So Paul, this feature is a pretty big deal.

[00:05:13] I mean, some thoughts, I'm wondering if you could kind of respond to here. It seems really, really valuable if you are comfortable with giving chat, GPT, that kind of access to your past chats and conversations. Not to mention, it seems like if this is something you find tons of value in, you are now probably, I would guess, much less likely to leave chat GPT if it has all this context and memory about you.

[00:05:37] But on the other hand, I can also see there are a fair amount of people that appear hesitant here to use this feature. 

[00:05:43] Paul Roetzer: Yeah, I think that the both true like it, it is a very powerful feature. It could be a potentially sticky feature. It's also a feature Google is, has and is going to build on. And if you think about the data, Google has not only, you know, does [00:06:00] it have chat potentially if you use Gemini, but it has emails and search and all these other components that it can know you even more deeply.

[00:06:08] Google Maps. Waymo, like start, you know, mixing in all of these other things, your pixel devices, your nest cams, so Google if again, you allow it. This is the kind of next frontier these companies are competing for, is to take all of your personal data to create truly personalized experiences for you through your ai so they can build truly personalized assistance.

[00:06:35] Apple, if they ever get their selves together, could play in this world too. they are just much more protective of your, private information. So, I I, it's helpful maybe to highlight a couple of things That's a great, or very helpful memory, FAQ, that OpenAI also updated related to this. So I would recommend people read this because I think it's very important people understand what their privacy rights are here and what is maybe automatically turned on and what [00:07:00] you need to turn off.

[00:07:01] So from the FAQ page, it says, chat chip t. Can now remember useful details between chats, making its responses more personalized and relevant as you chat with chat chip t whether you're typing, talking, or asking it to generate an image. This is multimodal. It will remember helpful context from previous conversations, like your preferences and interests, and use that to tailor its responses.

[00:07:22] The more you use chat GPT, the more useful it becomes. You start to notice improvements over time as it builds a better understanding of what works best for you. You can also teach chat GPT something new by saying it in a chat. For example, quote, remember that I'm a vegetarian. When you recommend a recipe to check what chat GPT remembers, just ask what do you remember about me?

[00:07:42] I don't know if you've done that yet, Mike, but I did it. It's kind of fascinating. it says you're in control of what chat GPT remembers. You can delete individual memories, clear, specific, or all saved memories or turn memory off entirely in your settings. If you'd like to have a chat without using or [00:08:00] updating memory, use the temporary chat.

[00:08:01] As you mentioned, Mike, temporary chats won't reference memories and won't create new memories. If memory is on and you want to fully remove something, chat, GPT remembers, you will need to delete both the saved memories and settings and the chat where you originally shared this information. If this sounds to you like people are just going to not change anything and they are just going to accept whatever OpenAI sets as the standard, you are probably correct.

[00:08:26] I, I would imagine that of their, what now sounds like six or 700 million users. The vast majority are never gonna touch these settings, and they are just going to not even probably know that this is a thing. So, that to me is like, memory is just gonna be a part of the chat experience moving forward. It says chat, GBT Enterprise Workspace Owners can turn memory on or off for all users.

[00:08:49] So if you are a workspace memory or workspace user, in enterprise I, it doesn't say team. I don't know if team doesn't have that option. You don't have a choice here. It's [00:09:00] your admin that has the choice for you. couple other quick interesting notes. They had a, in FAQ it says, does memory remember sensitive information?

[00:09:10] it says memory raises important privacy and safety considerations, especially around what type of information should be remembered and how that information is used. They continued, we're taking steps to reduce bias and have trained chat g GT not to proactively remember sense of information like health details unless you explicitly ask it to.

[00:09:28] We are continuing to prove how the model handles this type of information. You're in control. You can review and delete, save memories. Ask Chad GT what it remembers about you and delete specific conversations. This one is interesting to me because I have had, a health issue recently that I was definitely talking to chat GPT about.

[00:09:46] And when I was asking for, you know, what it remembers about me, none of those details are there. I think it's real important for people to know they are there. they are just choosing not to surface them to you. So what they are basically saying is like, we [00:10:00] remember everything you ask us to, we just classify stuff and extract it.

[00:10:05] When you ask us about memories, basically. So it's how good are they at filtering this sensitive and private information is really the question, not whether or not they know it. and then a couple other quick notes. Do you train your models with memories? It says if, if you have the, improve the model for everyone's setting turned on.

[00:10:23] So if you're not aware of it, there is an option in your settings. I think it's under privacy where you can turn on or off, improve the model for everyone. If it's on, they will use your information to train their models. So it says, if you have improved the model for everyone's setting turned on, we may use content you've shared with chat CBT, including past chats, saved memories and memories from those chats to help improve our models.

[00:10:48] Now a couple quick personal notes here. The first thing I thought about when we first started talking about memory is I have two chat GBT accounts. I have our chat GBT team account for the company, and I have a personal chat GBT [00:11:00] account. I happen to have a pro license Personally, I, I'm generally pretty good about having business related conversations within my business account and my personal conversations and my personal account, except I don't always think to like, look and see which account is active.

[00:11:16] I just go in the app and what everyone happens to be active. I, something I think would be correct if you only have a single CHATT account and you use it for both personal and business reasons. What I learned is you can go in and ask it to segment those memories. 

[00:11:34] Mike Kaput: Hmm. 

[00:11:34] Paul Roetzer: So the way I learned this is I actually gave it the, like, what do you remember about me thing?

[00:11:37] And it sort of like mashed together some stuff. So it gave me, it categorized these as professional interests, business and strategic projects, creative and personal interests, preferences and styles. And so what I found when it first outputted is it was mixing memories of things relevant to me and then things I did for someone else.

[00:11:55] So if I was like demonstrating it to someone, like how to, you know, create a children's book or [00:12:00] something, it thinks I'm writing a children's book. And so all these memories are just like mashed together and it doesn't know the context, but the part that was really weird is when it did this and it mashed together my own stuff and other people's stuff.

[00:12:13] I said to it, your mixing memories of things I've done to help others, like a children's book and my own activities and interests. And it replied. Great point. And thank you for the clarification. Let me separate those more accurately. And it did. It, it was weird. Like it all of a sudden knew everything that wasn't for me, and then everything that was for me.

[00:12:32] So my guess is if you use chat GBT for personal and business, and you say, Hey, split my memories into personal memories and things that I do for business, my guess is it's gonna filter really well. and then it, it replied back. It's like, want me to revise what I remember based on your own projects and pursuits versus what you've done for others?

[00:12:52] I said, yes, and then it said, got it. From now on, I'll make a clear distinction and focus on your own work and interest, unless you specify otherwise. [00:13:00] If you ever want to highlight work you've supported for others, just let me know. Do you want a fresh recap of your core activities and interests? Now this is wild.

[00:13:07] Like it's there, there's more going on here than just like information retrieval. Mm-hmm. There is like an understanding of the context of like, how are you using these things? And again, said it a million times. This is the dumbest form we're ever gonna have. Like I. These things are only gonna get smarter.

[00:13:24] And the people that choose to allow them to have access to these memories, again, not just open the eye, but Google and others, they are going to have very different, more powerful experiences with these chatbots, but it's a very slippery slope also. so yeah, that's just some of context. Check your settings though.

[00:13:44] It's a good reminder to check your, your settings and see what you're allowing to be remembered. 

[00:13:49] Mike Kaput: Yeah, absolutely. Yeah. It's already opened up some really intriguing possibilities, but like you said, it's how comfortable are you going to be taking advantage of this? Yep. 

[00:13:59] Shopify CEO's Leaked "AI First" Memo

[00:13:59] Mike Kaput: Next [00:14:00] up, at Shopify, the CEO is sending a very clear message to staff, which boils down to basically, don't ask for more headcount until you've proven AI can't do the job.

[00:14:11] So in a memo to staff, CEO Toby Lutkey laid out a new standard before requesting more people or resources, teams at Shopify must show that AI isn't a viable option. He called AI usage a baseline expectation and said that employees are now being evaluated on how well they integrate it into their daily work.

[00:14:32] This memo was originally internal, but Lutkey published it after he heard it was being leaked by the press. So like many tech companies out there, Shopify has been trimming its costs and trying to increase its efficiencies and investing heavily in ai. They have AI. Tools that they sell to customers like their sidekick chatbot, and now they are trying to turn that same logic inward to the company's own operations.

[00:14:59] [00:15:00] Now, Paul, some quotes that stood out to me from this memo, he said, quote, what we have learned so far is that using AI well as a skill that needs to be carefully learned by using it a lot. It's just too, unlike everything else. He also said, I've seen many of these people approach implausible tasks, ones we wouldn't even have chosen to tackle before with reflexive and brilliant usage of AI to get 100 x the work done.

[00:15:25] And he also mentioned that using AI effectively is now a fundamental expectation of everyone at Shopify. And everyone means everyone. This applies to all of us, including me and the executive team. And Paul, I noticed that you also had a post about this, which seems pretty interesting to talk about, saying this will be universal across industries by the end of 2025.

[00:15:47] Can you maybe talk to us about that a little bit? 

[00:15:49] Paul Roetzer: There was definitely a lot of noteworthy, items from this post, and it's not long. I mean, it's probably like a thousand words or something. It wasn't a, a crazy long memo, but, yeah. So the [00:16:00] one in particular that jumped out to me is this before asking for more headcount resources that teams must demonstrate why they can't get it done with ai.

[00:16:07] That is the right approach. Like, so anyone who listen to this podcast re knows how pro-human I am in all of this, that, that we have to, reskill and upskill as a, a top priority. We have to try and create opportunities for people, as jobs start to become impacted by ai. But as a business leader, that is fundamentally absolutely what you need to be asking of your team.

[00:16:31] If you want three more customer service reps, first, show me why we can't do what they do with ai. If you want two more BDRs, first, show me why we can't do what, what they do with ai. So before you start adding staff. It's, it's the only responsible thing to do because what happens is if, if we only look at today and we say, okay, let's add those customer, you know, service managers or whatever the role is, and then six months from now we realize, oh wait, that's only like a half FTE now we don't need those people [00:17:00] anymore, then you are in a tough spot.

[00:17:02] So I think that organizations have a responsibility to maintain as many workers as possible and to re-skill and up-skill them. But you also have a responsibility to be looking out, you know, 6, 12, 18 months from now and saying, do we really need to make this hire? because it's way better to not make the hire than to be in a position to cut that role, you know, in six to 12 months.

[00:17:24] So I do think that AI forward companies, the ones that have leaders who understand what these models are capable of today, and what they are going to be capable of in the next six to 12 months. That is, that is absolutely what they should be doing. and I think we're gonna start to see more leaders take a very direct approach to this and be more specific about we are going to require AI usage.

[00:17:44] I think that, and I, I'm saying I think this, I'm, I'm, I can also verify, I've had these conversations with leaders in the last two weeks that are doing this exact thing, which is, AI literacy and competency are going to become a filter for your employment. [00:18:00] Meaning you're not gonna be there if you don't figure this stuff out.

[00:18:03] So if you look across organizations, and I don't care if it's marketing, sales, hr, finance, legal, whatever it is, in the next 12 to 18 months, I do believe that organizations are start to going to start to filter their employees and do their annual evaluations or whenever it's happening, they are doing these evaluations.

[00:18:19] I do believe that AI literacy and competency are going to become a key filter for who stays and who goes. And I think the employees that move forward and prove their ability to drive productivity, efficiency, creativity, and impact on revenue and growth, they are, they are gonna be in the best position to keep their jobs and thrive because revenue per employee numbers are gonna go up.

[00:18:42] Profits in theory, go up and those people stand to, to, to benefit greatly. I've said this a couple times, but like there, there's no better time in history to be building a company from scratch because all these KPIs you would look at like a revenue per, per employee number. So depending on your industry, that number, you know, maybe [00:19:00] it's 300,000, 400,000 per employee target.

[00:19:03] If you're in a software company, it might be six, 700, 800,000. If you're Nvidia, it's like 1 million, 1.2 million per employee. I think those numbers are gonna get completely reset. And I can, I can say this from personal experience of building our company in a more efficient way and how I look to our future, I don't, I don't think that.

[00:19:22] It's, it's unrealistic for service companies, knowledge based businesses to be doing closer to the NVIDIA numbers than to the standard numbers. Mm-hmm. And, but it takes reimagining what those companies look like. And the way you do that is by building an AI literate, AI competent, like AI forward workforce.

[00:19:43] And if everyone on the team is moving in that same direction and constantly saying, is there a smarter way to do what we're doing, processes, workflows, campaigns, tasks, the compound effect of that is gonna be insane for organizations that get it. So I, yeah, I feel [00:20:00] like there, there wasn't much in that memo I would disagree with honestly.

[00:20:04] Like, I think he, he's saying the stuff that I've been hearing, that most executives have been unwilling to say publicly, 

[00:20:13] Mike Kaput: somewhat related to this is our next topic about the analytics software company data box. So. 

[00:20:20] Databox Replaces 80% of Support Team with AI

[00:20:20] Mike Kaput: Recently, Databox, CEO, Pete Capta recently posted on LinkedIn that the company replaced 80% of its customer support and sales development staff with an AI chat bot and actually improved results by 40%.

[00:20:37] So it's important to note, as CAPTA does multiple times in the post and in the comments that the company reduced their headcount well before adopting AI for this use case, he said he had 40 people on the sales development and customer support teams, but had to reduce headcount to eight because the anticipated demand they had hired for didn't materialize.

[00:20:57] Now, well, after that, the company [00:21:00] then deployed Intercom's Fin Bot, which now resolves about half of all their customer chats instantly. So that freed up human reps to focus on personalized outreach to high potential leads, helping drive more revenue. Now, CAPTA said that Fin Bot's customer satisfaction scores can still lag behind humans, but its speed and consistency make up for it.

[00:21:23] And Databox actually improved its results by feeding it better content, like integration specific help docs and use case forums to handle the long tail of customer questions. Now, he also admits the kind of next evolution of this is a bit harder, giving bots the ability to log into accounts and then troubleshoot things like a human would, could present some roadblocks.

[00:21:46] But he also says 18 months ago, he wouldn't have believed how far they could have gotten already with ai. Now, Paul, you know, Pete, and we're all familiar with Databox having used this software. I'm curious [00:22:00] about your thoughts here. Pete himself said he hesitated at times to share this story. There were definitely some negative comments about his decision on this post as well.

[00:22:11] But also this is a pretty impressive case study of what's possible. 

[00:22:15] Paul Roetzer: Yeah, so Pete and I go back a a long time. So Pete was the architect of the HubSpot partner program. And again, any long time listeners that would know my agency, my former agency that I sold in 2021 was HubSpot's first partner back in 2007.

[00:22:29] So Pete and I go back all the way to the origins of the partner program, at, at HubSpot, and then he moved over to Databox. I don't know how long he's been CEO there, it's been a while. I wanna say it's like six or seven years maybe. So yeah, we have, we stayed connected. Pete's a good friend and, he was not, I would say, and probably admittedly himself, he was not an early adopter of ai.

[00:22:51] Like, I remember pushing Pete back in like 2017, 1890. I was like, dude, you should be building AI into the business intelligence platform. And, you know, here's all the [00:23:00] opportunities. and to his credit, like he came around and now he's like all in on infusing it into their product and obviously now into their business.

[00:23:10] And, you know, I think he was pretty clear that as you highlighted, Mike, these layoffs weren't because of ai. Right. He didn't reduce the staff, but he, as I was just saying in the previous note, he is reducing the number of new hires by using ai. We just won't need as many people going forward. And so that's the right approach.

[00:23:30] but I do think that we're gonna hear a lot more stories like this where the layoffs will have been because of ai. Mm-hmm. So there, there's going to be a lot of instances. Now, again, I know these things are happening. I hear firsthand these things are happening, but people aren't saying publicly yet, this is why it's happening, but they will.

[00:23:52] so you will have layoffs because of ai because you see leaders who look at teams and say, we don't need five people doing [00:24:00] that anymore. You will figure it out with two of you because they know that AI is now capable of assisting these different roles. And again, it's across departments, but it's, it's starting in like marketing is a, is a big one right now.

[00:24:14] Sales is a, a big one. And so what's gonna happen is you're gonna have leaders put constraints on teams and challenge them to achieve new levels of, of new levels of efficiency and productivity with ai. And again, I'm not saying this is the right approach, I'm just telling you it's happening now and it's gonna happen at a much bigger scale as the year goes on.

[00:24:35] And as a business leader, there's no greater way to drive innovation than to create deadlines and restraints. So if, if everything is great and like there's no real limits on budgets and we have as many people as you could possibly need to do whatever you want, people get lazy. and so this is what, and this is like the tech culture.

[00:24:55] This is driven largely in tech. It is constrained resources and then [00:25:00] help people realize what they are actually capable of doing under constrained resources. And so I think given the economy, given a number of other advancements in AI models, I think you're going to see leaders who put constraint resources on their teams and say, you can, you can produce greater with fewer people.

[00:25:18] We, we believe you can now go do it. And they are gonna challenge them to do it. It's gonna be uncomfortable. It, it's not gonna be maybe fun to be like in an AI emergent company that has maybe hundreds of people in your marketing department and you don't think you need as many. You have to make some difficult deso decisions.

[00:25:38] Um. But it, it's gonna happen. And it's, again, it's happening. It's just not being publicly talked about yet. they are what I call quiet AI layoffs, that there are, there are layoffs happening that are not being put underneath that headline, but they will be. 

[00:25:55] Mike Kaput: You know, I think it's really just worth reiterating that this is such a [00:26:00] useful piece of AI related career advice too.

[00:26:03] I would say, just as someone like myself who's trying to navigate this is really, if you can take a step back and get some perspective and put, pretend you are the CEO e of a company, think about like Pete, go to Pete's post to start. 'cause he goes into the hard parts of his decisions that he had to make.

[00:26:21] He's like, well, you know, here's what I was thinking. Here's how I thought about it. Not everyone understands the decisions I have to make. Like, that's a good barometer for how you wanna be thinking about AI in your own career and function, I think. 

[00:26:32] Paul Roetzer: Yeah. And the other one, Mike, that you know, came up in some conversations last week is.

[00:26:38] Take your top players. Take, take, take the people on your team who are figuring this out or listen to podcasts, taking online courses, reading books, like they are doing everything to figure it out. And they are pushing like the limits of chat, GPT and Google Gemini and the prompting, and they are testing out all these new tools.

[00:26:55] Those people are gonna 10 x the value they create on a team. [00:27:00] How in the world can you talk to that person in like an annual review when they are looking around seeing the other people on their team who aren't doing anything? they are not, they don't know how to prompt anything. they are not using Chad Chip T.

[00:27:11] They haven't taken any initiative to learn ai. And now you have someone who's like creating 2 3, 10 x the value. And all of their prompts are the ones that everybody else is using, and they are the ones that are like educating the rest of the team. How do you talk to that person with a straight face and say, yeah, we're gonna keep everybody else around.

[00:27:28] Like, well, eventually they'll figure it out. It's like, no, because now as, as the person on that team who's doing all this work, right, I'm like looking around saying, where can I go, where I'm gonna be really valued and I'm not gonna be pulled down by everybody who's not figuring this out and refuses to like use ai.

[00:27:45] So I think that that's where you're gonna have these AI forward practitioners who want to be around other AI forward practitioners, and they don't want to be kept down compensation wise, career projectory wise, when they are the one who's doing everything that's being [00:28:00] asked of them, or maybe not even being asked yet.

[00:28:02] they are just like the innovator on the team, and they start to feel like they are being constrained by their leadership who maybe doesn't understand AI or by their peers who refuse to use it. So again, there's, there's just basic business fundamentals that tell us things are going to change like they have to, or the best talent is gonna leave and go somewhere where that that talent's appreciated.

[00:28:25] Google Cloud Next ‘25 and Google Updates

[00:28:25] Mike Kaput: Our next topic this week is one you've already alluded to Paul, which was Google Cloud Next 25, just wrapped up this past week in Las Vegas, and some of our team members were there for some, or in your case, all of the event. So we wanted to briefly cover some of the top announcements to come out of the event.

[00:28:43] there were a ton of them. Google literally published a list of 229 announcements in total. So obviously we're not gonna go through all those, but you can find the link in the show notes. But just some interesting highlights here. kinda what's top of mind at Google Cloud [00:29:00] is, a bunch of announcements centered around Gemini 2.5 Pro, which is now available in public preview.

[00:29:06] It's Google's most powerful model to date. It tops the chatbot arena rankings, and it's designed for advanced reasoning and coding among other things. Alongside this we're also announcements about a leaner, faster Gemini 2.5 flash. As well as major upgrades to image, audio, video, and even music generation across some of Google's different models, including Imagine Three, chirp three, VO two, and Lyria.

[00:29:35] Google also announced a handful of important updates to Agent Space, which is its platform that connects your work apps to Google's AI models and agents. So you can use these AI models and agents with all your information and data, and there were a ton of updates about AI infrastructure. Google debuted new GPUs, TPUs, high speed networking and storage, optimized for training [00:30:00] and inference at scale.

[00:30:02] So Paul, this was a really cool event that you and I and some of our team got to experience. I want to kind of get your. View on, were there any big takeaways you had from the event first, and then also I have to have you share your experience on night one before I got there. When Google used AI to recreate the Wizard of Oz in the Las Vegas sphere.

[00:30:27] Paul Roetzer: Yeah. The Wizard of Oz experience was crazy. So I, I'd never been to the sphere. If you, if you aren't familiar with it, look it up. Like you can watch some YouTube videos of it. It's wild. so I know it's a concert hall, but they, they also are creating these other experiences and so, Sarah Kennedy, who, you know, is a friend of mine and who we collaborate with, who's the VP of Global Demand and growth marketing at Google Cloud, she sort of spearheaded this, event and experience and it was just remarkable.

[00:30:57] So what they are doing [00:31:00] is, James Dolan was the CEO of the sphere, had sort of approached Google. About creating this, about re you know, bringing this 1939 film to life in this amazing arena. And so they've been working with the Google team for I think, like two years now to do this. And what they did is they didn't show us the full film.

[00:31:18] It's not ready yet. It comes out August 28th, this year at debuts at the Sphere. But they showed us almost like a documentary of this, this building, of this experience, of taking this film that's basically in a rectangle from 1939 and low resolution and expanding it to fit this massive screen and to create this, you know, multimedia experience with like wind blowing up from the floors and the sheet, the seats shaking when like thunder hits.

[00:31:45] It is so crazy. So there was all these innovations and they talked about, like, they interviewed this one guy from Google DeepMind and they said like, Hey, when this stop project started, like, what did you think was impossible? And he is like, everything, like there, there was nothing we were [00:32:00] doing. That the models at that moment could actually achieve.

[00:32:04] Wow. And so they had to create all these breakthrough innovations, specifically with Gemini vo, which is their video gen model. And imagine, which is the image generation. And the three components they focused on is one called Super Resolution, and then in painting and out painting. So what happens, this is from a Google, blog post that we'll link to.

[00:32:25] It says, using versions of vo Imagine and Gemini specifically tuned for the task. The Google teams and their partners developed an AI-based super resolution tool to turn those tiny celluloid, celluloid yeah frames from 1939 into ultra, ultra high definition imagery that will pop inside the sphere.

[00:32:44] And it does, having seen it. Then the teams perform AI out painting to expand the scope of scenes to both fill the space and fill the gaps. Created by camera cuts and framing limitations. Finally, through performance generation, they are incorporating composites of those frame performances into the expanded [00:33:00] environments together.

[00:33:01] These techniques help achieve the natural gestures, staging, and fine details that conventional CGI struggles to match. Like there was an example where they showed Dorothy in a scene where she talks to the uncle initially when she comes in the door and then she like goes in to the aunt or whatever.

[00:33:15] Well, in the tradi, in the original film, the uncle's off screen has nothing to do with it. Well, in this expanded version, he's there, like they are, they are recreating characters that would've been in a wider shot. they are actually like the AI is creating these characters with Natural Mo. It was so wild to see.

[00:33:33] So, yeah, I, I, there's no documentary about this, like Wall Street Journal had an article about it. You can't really go watch video of it. It was a private event. but just remarkable. And it does show, like, the thing that I took away with it was the, I. The human machine collaboration. Like there's, this wasn't like you just gave it to Gemini and Gemini figured all this stuff out.

[00:33:53] There was dozens of the top minds within Google DeepMind and Google Cloud, within [00:34:00] Google working on this, envisioning this, and then like pushing the limits of the models and in many cases creating entirely new techniques to make this possible. And it does, it's one of those moments where like when I first put on like a vision pro and you're like, oh wow, like this product might not take off, but this is a whole new experience.

[00:34:17] That was what I felt when I was at the sphere. It's like this is totally, this opens up all kinds of incredible possibilities for things that could be done in that kind of environment with ai, working with the human. So. Wizard of Oz thing was nuts. again, gotta check that out. And then the big thing that stood out to me and I sat through a lot of sessions and content, agent space, like we talked about agent space in December, 2024 on episode 1 27 when they first announced it.

[00:34:45] But I think we just kind of mentioned it 'cause there wasn't much information about it. It wasn't actually available. It was just sort of like a preview thing. It was the most impressive thing, maybe outside of the sphere that I saw, last week. So basic premise is like, it's a single space [00:35:00] that has your prompt galleries, your agent galleries enables you to, in a nod code environment, to build agents, to do whatever you want to do.

[00:35:06] Connects to third party software and data. So it basically becomes this like platform where you live and do everything you need to do. It has deep research built in, it has audio overviews built in. You can turn anything into a presentation on the fly. Create a deep research project. Okay, turn this into 10, you know, PowerPoint.

[00:35:24] Slides or Google sheets or Google slides that I wanna turn, you know, use for my presentation. It's got notebook LM baked in. So it's like it, the vision for it is powerful and you could see how it becomes like a control panel basically for a knowledge worker to just like all the tools they need are just living right there.

[00:35:44] it's not available yet though. Like I, you have to go request access. So that was the only frustrating part was like, I don't even know, like when it's gonna be available. How do we get it? Is it, I have no idea. And I was there and I still don't know. So that, that's the only, you know, downside I guess is, it [00:36:00] doesn't really exist that I can tell.

[00:36:02] But I do know people in Google are using it, so it, it is a thing, it's just not a publicly available thing yet for most of us. And then the last thing I'll, I'll note is, I was at a part of a Leader Circle event and had the privilege of sitting there and listening to Sunar, Pacha be interviewed. And he did say that he expects the pace.

[00:36:20] Of the model advancements to continue for at least 12 to 18 months. And he said specifically new major models every three to four months. So when we say these models are gonna keep getting smarter and the velocities there, this is the CEO of alphabet and Google saying like, yes, for at least the next year to year and a half, every three to four months, we are gonna see, you know, massive model improvements, which is just crazy to think about.

[00:36:44] OpenAI Will Release o3 and o4-mini After All

[00:36:44] Mike Kaput: Wow. So our next piece of news here, actually I believe dropped while we are recording our last podcast. So we didn't get to cover it then, but has some relevance probably for the next coming week or two. so [00:37:00] on April 4th, according to a post on X from OpenAI, CEO, Sam Altman OpenAI will in fact be releasing its oh three and oh four mini models after all.

[00:37:12] So Altman said, quote, change of plans. We are going to release O three and oh four mini after all probably in a couple of weeks, and then do GPT five in a few months. There are a bunch of reasons for this, but the most exciting one is that we're going to be able to make GPT five much better than we originally thought.

[00:37:30] We also found it harder than we thought it was going to be to smoothly integrate everything and we want to make sure we have enough capacity to support what we expect to be unprecedented demand. Now it sounds like originally OpenAI was going to shelve a separate launch for these models and instead kind of bake them into GPT five.

[00:37:48] Kind of the goal here is they want kinda a unified system that has voice canvas search all in a single system that can intelligently decide when to think deeply [00:38:00] or give you faster answers. But until that happens, it sounds like we're getting some new models to possibly experiment with very soon. So Paul, what does this kind of pivot mean?

[00:38:12] Like, is GPT five behind schedule not performing as planned or he just changing the approach here? 

[00:38:19] Paul Roetzer: It might be a, who knows, it might be a mix of all the above. my guess is again, that the reasoning models open up a whole new scaling law and they've been trying to figure out since the fall when they released the first reasoning model, you know, where, where this goes and when they are gonna kind of bring it all together.

[00:38:35] so I think it's just, you know, things are fluid internally, and they are GPU constrained, so they, that might be the biggest thing is they are limited by, their ability to do inference with their chips. And I think that the demand for image generation, since it came out a few weeks ago, maybe made that even more challenging, where Sam's basically on Twitter [00:39:00] begging for.

[00:39:01] Chips from somebody, anybody. So that could, that could be, it is, it just might be, they are constrained by compute power to do these things. Now, Sam did tweet on Sunday, April 13th. we have got a lot of good stuff for you, this coming week, kicking it off tomorrow. And then OpenAI just tweeted about an hour ago this we're, it's Monday, April 14th, right now when Mike and I recording this, developers with a handshake emoji, super massive black hole, live stream, 10:00 AM Pacific time.

[00:39:32] So this is one, 1:00 PM Eastern Time. So I ask Grok. So if you haven't done this yet, if you have, in, in Twitter X, there's a little gr symbol, which ironically is a black hole. and if you have a tweet, like a lot of these times, like the ai, people like to vague tweet stuff and you're like, what, what the hell does that mean?

[00:39:51] Like, only you know a small group, you pop any clue that they are talking about. So you can just click the little gr symbol and it'll actually like give you some. Some hints [00:40:00] it may be what it means. Mm. So I asked Rock 'cause I have no idea what the super massive black hole thing is and it said Open AI's cryptic post hints that a major developer focused announcement tying developers with super massive black holes.

[00:40:10] I could figure that part out. Possibly symbolizing a groundbreaking AI tool with immense potential set for livestream at, one piano I talked about, about the supermassive black hole reference aligns with recent astronomical discoveries like the awakening black hole. Yeah, that's nothing to do with this.

[00:40:26] okay. Community Speculation Centers on Quasar Alpha. A rumor opened AI model with a 1 million token context window potentially linked due to its name, names Cosmic Connection and metadata Similarities with open AI's, API suggesting a stealth release for developers. I'd never heard of Quasar Alpha. I dunno if that's making that up or if that's actually a thing.

[00:40:47] Oh, wait, here, I'll hit explain. Quasar Alpha in, in grok and let's see what happens. Um. It's an intriguing AI model that emerged on the scene in early April, 2025, generates significant buzz within the AI and developer communities and then it [00:41:00] goes into a whole bunch of details. So, yeah, I don't know, maybe Quasar Alpha is coming today, whatever that is.

[00:41:07] Sam Altman Comments at TED

[00:41:07] Mike Kaput: So Sam Altman also took this stage at TED this past week and talked about some interesting points about the present and future of open ai. So he actually revealed that chat GT's user base has doubled in recent weeks and now it's used heavily by about 10% of the global population. that growth includes 500 million weekly active users.

[00:41:32] He confirmed OpenAI is working on a powerful open source model, quote near the frontier, responding to pressure from challengers like deep seek. And he even said, Hey, we are late to act on that, but we're going to do it really well now. When asked about safety, he downplayed fears of things like self-aware ai, but did highlight that agent to ai, you know, systems that can take action on their own was quote, the most interesting and consequential safety [00:42:00] challenge that open AI has faced yet.

[00:42:02] And when asked about a GI, he said, quote, if you ask 10 open AI engineers, you'll get 14 different definitions of a GI, whichever you choose. It is clear that we will go way past that. There are points along an unbelievable exponential curve. Now, Paul, I think you're gonna address this, but like first, what's worth paying attention to here?

[00:42:23] This interview had some like nuggets in it, but got kind of funny, kind of awkward because the whole reason we learned that chat GT's user base had doubled in just a few weeks is because interviewer Chris Anderson said. On stage to Altman. Hey, you told me that Backstage and Altman very quickly and annoyingly responded, saying quote, I said that privately.

[00:42:47] Paul Roetzer: Yeah, the interview went downhill pretty fast from there. It, it was honestly one more uncomfortable interviews I've ever watched. So Anderson obviously had very specific things he wanted to get at. Hmm. [00:43:00] The vast majority of them were, were highly annoying to Sam, I would say. and plus Anderson kept bringing up, AI safety people that were on the stage earlier in the show and referencing them.

[00:43:14] And Sam obviously wasn't like a huge fan of that. So, yeah, it, it was, it was just really uncomfortable. Like I, there's a lot, Sam said a lot and I think that there was some points that. Anderson had every right to maybe push a little bit on for sure, like the image generation and the impact of memory and, you know, the dangers of the models and what they are doing for safety.

[00:43:37] And like, these are all valid questions. I just think that at some point you gotta read the room and realize this is going very poorly. Mm-hmm. And he did not back off at all. Like, it was, it was very uncomfortable. And then Sam started getting very irritated and just like throwing questions back at Anderson at the end.

[00:43:57] Like, well, what do you think? And it, it was, it was [00:44:00] super awkward. But anyway, um. On memory one, a couple things that jumped out to me. Sam said one day, you will talk to Chad over the course of your life at, at some point maybe if you want. It'll be listening to you throughout the day and sort of observing what you're doing, and it'll get to know you and it'll become this extension of yourself, this companion thing that just tries to help you be the best you can be.

[00:44:22] This is, not gonna go away. I think that this idea of ever listening, ever watching AI is going to be pushed into society, whether we're ready to it or not, through glasses primarily, I think is the main, the main way. And I just saw this morning that Tim Cook is like all in obsessive now I'm beating meta at like wearables and glasses.

[00:44:44] and that's like the next frontier for Apple. I, my guess is that's what Johnny Ivy and Sam Altman have been working on together as glasses. I think glasses is the most logical. vehicle with which AI will be [00:45:00] delivered. That's just always on, always listening, always recording, always watching and providing context.

[00:45:04] Mike Kaput: Hmm. 

[00:45:04] Paul Roetzer: So I think all these other wearables are just gonna be irrelevant over time. I think glasses is the form factor that makes the most sense. So it's kind of weird, like, honestly thinking about a future where everybody's just got their devices or their glasses, like watching and listening and, remembering everything and being saved into memories.

[00:45:24] And you're, what you do, even if you're choosing not to participate, is being saved in someone else's memory all the time. It's very sci-fi and, you know, black mirrors, but that I'm pretty convinced, like that is an inevitable future. Like by the end of the decade probably. Anderson pushed him really hard on the copyright stuff, the in intellectual property, specifically around like image generation and, Sam needs, I don't know if there's PR people at OpenAI or comms team.

[00:45:53] They need real fast to like get their messaging better. Like I will say like [00:46:00] it's their messaging on AGI and copyright and intellectual property is so poor. Like as someone who PR communications background, it's like they've never had a single meeting to like figuring out their talking points around AGI and copyright and intellectual property.

[00:46:19] And it is the core of everything they are going to do to need societal support for what they are gonna try and do next. And they can't even vocalize their, their position other than on the image generation stuff. Like they don't think artists have rights basically is pretty much all they are, they are saying.

[00:46:37] And that is not gonna, that's not gonna fly. Like you can win the court cases maybe, um. But it's just a very poor approach of like, yeah, we need to find a better way to compensate people then do it like this. You've been work, you've been at this for like, how many years now where you knew this was coming.

[00:46:53] Do something. Stop saying we need a new model. And then on a GI, I just, I'm in disbelief at [00:47:00] their inability to just say what they think it's, 'cause every time Sam gets asked this whole, like, oh, ask, you know, it's the joke internally ask, you know, 10 open AI researchers, you gonna get 14 different answers.

[00:47:11] It's like, it's not funny. Like it's, it is the mission of your organization literally to create a GI and distribute it safely to humanity. What is it, what, what are you actually creating and distributing? I don't comprehend that. They don't have a consistent answer to this. It's wild. and then the safety stuff, and again, I still get, as I'm saying this, I'm getting worked up.

[00:47:32] Like I understand why Chris Anderson was like, pushing on these things. I just thought the interview was not handled great. Um. His views for the future. he said he really believes that society figures out over time with some big mistakes along the way. This is, again, verbatim how to get technology right, and this is going to happen.

[00:47:52] basically he's saying like, these AI advancements, and this is his quote, is like a discovery of fundamental physics that the world now knows [00:48:00] about and it's going to be part of our world, and I think this conversation is important, blah, blah, blah. So he's basically saying like, it's happening with or without us, like, we now know this is possible and we're going to do it.

[00:48:11] And then Anderson pushed him on like people leaving the safety team and Sam basically said like, yeah, we have different views of what safety is now in essence. Mm-hmm. So, I think it's an important interview because Anderson asked the hard questions. Sam didn't have good answers to most of the hard questions, which is why I think he kept pushing him.

[00:48:32] and that itself is sort of maybe illuminative of, if that's a word, um. Of where we are right now with AI is that the people building it don't really know what it is that they are building or like what's gonna happen when they do and they just keep pushing forward this idea that we'll figure it out and like we always do, which I think we will.

[00:48:53] Like, I do think he's right. I think this is technology moves, society moves with it. [00:49:00] Sometimes it's really uncomfortable. I just really opening eye has to get a a, a comms team locked in on how to give them talking points. Because watching these interviews, it's so hard to like see their inability to answer what should be fundamental questions to their brand positioning.

[00:49:18] Mike Kaput: And I never wanna assume malice given how just quickly and chaotic this whole space moves. But it does kind of make you think, and I think you alluded to this, like with some of the copyright and the A GI stuff, it's like really you have the smartest tools and people on the planet. Nobody has thought of this maybe, but at what point is it obfuscating?

[00:49:37] We don't wanna tell you what the real answer is at some point. Yeah. 

[00:49:40] Paul Roetzer: Truly. It's, it's like a, I don't know, like, almost assuming like people are ignorant like that, they are, yeah. I don't know. It, it's just lazy. Like, I can't think of another organization where it's been so important to understand what it is they are doing and they just have a complete lack of ability to declare and it's [00:50:00] not funny.

[00:50:00] Like that's, I think that's what bothers me about it, is they laugh it off. Like, we can't define a GI and you know, yeah. We're, you know, stealing stuff, but like, it's okay that humans have always stole from other humans. Like, no, that's not okay. Like that, that can't be the talking point. 

[00:50:15] OpenAI Pioneers Program

[00:50:15] Mike Kaput: Right. Some other open AI news this past week, they are launching a new initiative called The Pioneers Program.

[00:50:23] And the goal of this is twofold. First, help companies evaluate how well AI actually performs in high stakes industries like finance, healthcare, and law. Then fine tune models specifically for those use cases. So companies in this program will work directly with OpenAI researchers to design domain specific benchmarks or evals that measure what good performance actually looks like in their field.

[00:50:50] And they'll also get help customizing models to reinforcement, fine tuning to basically build expert level AI tailored to narrow tasks in [00:51:00] these industries. So Paul, there's a signup form on the page that we linked to in the show notes people can kind of apply to this program. This sure seems like what we have talked about on past podcasts.

[00:51:12] This need for better ways to evaluate AI's capabilities across real world knowledge work that isn't just related to coding, math, or science. And I don't know, it also strikes me as like a sign that open AI might be gunning for some high value industry verticals with its models and products. 

[00:51:32] Paul Roetzer: Yeah, so on episode 1 41, the Road to a GII, I talked about this, like, you know, moving past pure IQ tests and into the domain or industry specific tests, and, it's what they should be doing.

[00:51:43] It makes total sense. I hope we see more of that. I do think that, if I was building a software company for a specific vertical or industry, I would be like really paranoid right now. If I was the VC firm investing in those companies, [00:52:00] I would be asking some really hard questions of the companies that want funding.

[00:52:03] Because if OpenAI thinks some, industry's big enough and they see like billion dollars here, $10 billion there to go build the legal AI or healthcare AI or finance ai, like we already know, Microsoft's building the financial analyst. Ai, like any market they choose to, they have a, competitive advantage over a.

[00:52:26] Startups that wanna do it. So whether they wanna do it through a venture fund and they wanna fund the building of these and just take an equity stake in some of 'em, or if they want to just build them themselves, it's gonna be pretty powerful. Especially if they are the ones, the proprietary data where they work with people on these evaluations and they know exactly what to build.

[00:52:45] That's gonna be something that the startups don't have if OpenAI chooses not to share it with them. So, yeah, I mean, this could be the next trillion in market value for them is if they, you know, pick off 10 industries and just start building custom solutions for 'em. 

[00:53:00] OpenAI Reduces Model Safety Testing Time

[00:53:00] Mike Kaput: One other piece of OpenAI News this week, according to a new report from the Financial Times.

[00:53:07] Open AI has drastically cut back the time and resources it devotes to evaluating its most powerful AI systems. Now, what used to take months in terms of evaluation and safety testing now takes days and insider say testers were given under a week to assess open AI's upcoming O three model, which is a dramatic shift from the six months of safety checks used for GPT-4.

[00:53:32] The reason according to the FT is competitive pressure. As the race heats up with meta Google XI and others, OpenAI appears to be prioritizing speed to market. Some staff warn that this is a quote recipe for disaster, especially as models get more capable and maybe more dangerous. Now, critics also say OpenAI isn't rigorously testing the exact versions it releases and hasn't [00:54:00] followed through on promises to fine tune advanced models for biosecurity and other high risk scenarios.

[00:54:05] OpenAI, on the other hand, claims its new processes are more efficient and still robust. So Paul, I guess we have to accept that OpenAI may indeed be more efficient and just as robust as before in its safety testing. But it sure seems like the incentive is basically just to release as quickly as possible in order to compete, doesn't it?

[00:54:28] Paul Roetzer: Yeah. So again, this is one of the issues that Anderson pushed Sam pretty hard on. And you know, again, the basic take was we have evolving views of what safety means, but we use our preparedness framework to evaluate these models and we're not that concerned yet. And we have a track record of iterative deployment, which they think is the safest way to do it.

[00:54:48] Meaning we're just keep putting things out, see what we learn, and if they reach a level where we don't think they are safe, we won't put them out. And Anders was like, well, how do we know that? Basically? And then at one point, Sam, 'cause he [00:55:00] got really irritated. I think there was, um. It might have been like the ring of power tweet or something.

[00:55:06] Yeah, yeah. Where, where he was alluding to like Sam being like power hungry and, you know, money hungry. And that was when it was like, you could tell he just was ready to walk off stage. Mm-hmm. And I don't remember if it was right around there, but he's like, listen, like all of us care about safety. Well, all of except maybe one of us, which I can almost guarantee you he is referring to XAI and grok.

[00:55:26] Mm-hmm. He just didn't wanna name him directly. 'cause earlier in the interview they, they brought up Elon Musk and Sam wanted nothing to do with it. so yeah, I think that the concern here is there's no uniform. Um. Agreement on what safety is, what alignment is. Sundar Phai and his interview at Google next that I was referring to earlier, even said we needed more collaboration on safety.

[00:55:52] But it's gonna be hard to do it 'cause you're gonna need to get a lot of egos and a lot of power and a room together to like talk about these things [00:56:00] and agree on it. And it only takes one to push forward and do something outside of the bounds. And then the pressures everybody else to decide, are we gonna all go now and do this or, or do we, you know, gonna hold our stuff back?

[00:56:11] So I do think that by the end of 2025, decisions are going to have to be made within the Frontier Labs to, to hold back models that have been deemed unsafe internally. Whether or not we hear about that publicly, I don't know, but I'm fairly confident that at some point this year, if it hasn't already happened, these models are going to display some capabilities that.

[00:56:38] hit sort of the red zone of these preparedness frameworks and they have to hold back and figure out what to do about it. I don't think it's far off at all. And that's just from listening to interviews with like Dario Ade and Yeah. And others is these things are moving really fast and maybe that's part of the GPT five thing is like, maybe it did [00:57:00] meet some threshold where they have to hold it back and put in some more safety standards.

[00:57:04] I don't know. 

[00:57:07] Llama 4 Release and Controversy

[00:57:07] Mike Kaput: Meta has started launching models in its LAMA four family, which are the latest versions of its large language model. But these are arriving a little later than expected and under a bit more pressure than expected. So the headlining models right now are LAMA four Scout and LAMA four Maverick.

[00:57:26] So these are two open source models that can process text, images, video, and audio. They use a mixture of experts architecture activating only select parts of the model per task, which boosts performance while cutting down on Compute Scout. Fits on a single GPU and supports a 10 million token context.

[00:57:45] Window Maverick. Meanwhile, apparently outperforms GPT-4 O on many benchmarks, though a little more on that in a second. And it's optimized for tasks like image understanding, coding, and multilingual reasoning. Coming after [00:58:00] these models is what they are calling LAMA for Behemoth A still training 2 trillion parameter giant model that meta says will outperform GPT-4 0.5 in STEM tasks.

[00:58:11] But it is not out yet, though it is serving as a teacher model to boost Scout and Maverick through distillation. Now this launch has a bunch of controversy already around it because earlier last week, meta began to post that LAMA four Maverick had jumped to the number two spot on the popular LM Arena leaderboard.

[00:58:34] But some users, including prominent AI voice, Ethan Molik, started to report that the winning model was different from the version of the model released to users. He actually posted quote, the LAMA form model that won in LM Arena is different than the released version. I have been comparing the answers from arena to the release model.

[00:58:53] They aren't close. And then the Verge reported that in some fine print meta acknowledged that the version of [00:59:00] Maverick tested on El Marina isn't the same as what's available to the public. According to Meta's materials. They deployed an experimental chat version of Maverick to the leaderboard that was specifically optimized for Conversationality.

[00:59:14] El Marina then responded, basically saying, yeah, it seems like meta released a model to our leaderboard that was more customized to human preferences, which play a huge role in the LM Arena rankings. So basically a model that others couldn't use but was designed to rise higher in this specific leaderboard.

[00:59:36] L Marina actually said, look, as a result of this, we're updating our leaderboard policies to reinforce our commitment to fair reproducible evaluation. So this confusion doesn't occur in the future. Alright, Paul. So it certainly seems like meta might have ruined their kind of fanfare here by basically gaming.

[00:59:54] A leaderboard. why would they risk doing this when it was like bound to be [01:00:00] discovered? 

[01:00:00] Paul Roetzer: I don't, I don't know. Plus they released on a Saturday. Yeah. Which was weird to start. Like it was, I knew they, like, they were just maybe getting out ahead of like the Google announcements. But I, I, as my initial reaction, I was, I was like, must not be very good.

[01:00:13] Like, you're, you're almost like trying to just get it out there and you don't want a bunch of fanfare around it. Yeah. I mean, lm arena tweeted meta's interpretation of our policy did not match what we expect from model providers. That was a very PR way. 

[01:00:26] Mike Kaput: Yeah. Diplomatic, I guess cheated, figured, 

[01:00:28] Paul Roetzer: yeah.

[01:00:30] I don't know. It wasn't a good week for meta. They, they are, and they are getting, they are not having fun in court over the copyright lawsuit either. So, or, or the, efforts by the government to break them up. It's. It's a tough covid for meta at the moment. And yeah, this didn't help things. 

[01:00:45] AI Copyright and Creator Rights

[01:00:45] Mike Kaput: So kind of related to some of the copyright discussion, there's been some really heated discussion in AI circles this past week related to copyright.

[01:00:54] it was kicked off predictably by a controversial post from Jack Dorsey, co-founder of [01:01:00] Twitter, and founder of the Financial Services Company Block, formerly Square. On April 11th, he posted on x the following quote, delete all IP law. Elon Musk quickly replied, quote, I agree. Basically just like that to of texts, influential figures appear to have called for a total tear down of IP protections.

[01:01:22] It's not exactly clear yet what has prompted this, but this comes as AI companies including open ai. Are under legal fire for allegedly scraping copyright material to train their models. As you alluded to, meta is getting on blast for the same thing. And we have also talked about many times how most, if not all of the model providers have done the same things.

[01:01:44] basically Ed Newton Rack, so we talk about all the time, founder of fairly trained called this a War on Creators. a writer Lincoln Mitchell put it bluntly, none of Jack LAN's companies will exist without IP law. And Dorsey doubled down arguing that this [01:02:00] current copyright system favors gatekeepers over artists.

[01:02:04] And this comes obviously as all these other copyright things we have talked about have come to pass, where OpenAI drew some heat for its studio Ghibli kind of viral moments where people were having a lot of fun using its model to generate studio Ghibli style images. But people also were up in arms about the fact that that style was probably stolen or used without permission.

[01:02:28] We have controversy about reports that meta used copyright books to train its model. And on top of all this, a coalition of major publishers, including the New York Times, Washington Post, and Vox Media, is calling on the US government to stop AI theft. So they are launching a sweeping ad campaign this past week that accuses AI companies of using their content without permission or payment to train their models.

[01:02:55] So Paul, this isn't a new debate. People have been [01:03:00] suing over this for a while, but it does seem like it escalated pretty quickly with some tech leaders now feeling comfortable saying we should just delete IP Law. 

[01:03:10] Paul Roetzer: Yeah, I don't know. I mean it's certainly a provocative tweet. There's probably a bunch of context behind it.

[01:03:17] I, you know, I, it's like anything else when you want to argue, I. Point you like, take this extreme position and forget the nuance of the fact that you own copyrights yourself. Your company probably had patents on its initial technology that prevented other people from doing it for years. Like you've made your billions on the top of IP law and now it's convenient to just want to get rid of all IP law because you're a billionaire and it's inconvenient for you to have to, you know, pay people for their creative works.

[01:03:49] So, I don't know. It's like I've said many times on the show, I have a hard time with extreme positions on anything. I don't care what it is. and so these extreme [01:04:00] positions where people pretend like there's no nuance to the conversation just bother me when obviously there's lots of nuance to this conversation.

[01:04:08] And then, you know, Elon Musk, whatever, like yeah, it's obviously he's stealing everything he can possibly steal to build rock, and it's a extreme annoyance to have to potentially face lawsuits over it. So they are just gonna, the way I look at this is like, they are gonna just keep taking it. And now, like we have seen with image generation from Grok first, and then, you know, shortly after from four oh oh image generation with OpenAI, they just don't care anymore.

[01:04:36] Like, they are, they are, they are just all in and they assume it'll work out in the courts somehow, or there's gonna find some model to pay people back in some class action lawsuits and be done with it. But we are, we are in an accelerated phase of IP theft, and that is not going to stop unless the courts somehow stop it.

[01:04:56] And I just don't see that happening. And I, I, and I have to be [01:05:00] honest, like, you know, we were using the image generation thing while I was in Vegas. I was hanging out with a, a buddy of mine and we were like, you know, I took like a family photo and you can turn it into the Simpsons and South Park and anime and Muppets and like Pixar, like you, you can use all these names now and it's fun and it's like really cool to create it.

[01:05:19] But there is that part of me that's living this, like in this gray area of, but like, they are not paying for these. Like I know if you're not allowed to technically do this, they are just doing it. Hmm. But as a user, it's great. Like, it's nice to have the ability to do these things. but as someone who likes, you know, studies the space, I also sometimes are like, man, I don't, I kind of feel guilty about creating these things with these copyrighted, you know, images, but Right.

[01:05:48] Yeah. It's a weird space to be in. 

[01:05:52] Anthropic $200 Per Month Subscription

[01:05:52] Mike Kaput: Our next topic is that Anthropic is offering a new clawed Max plan with two pricing levels. And this is kind of aimed [01:06:00] at power users. There's a hundred dollars per month for expanded usage, or $200 for full access, which basically matches the cost of open AI's, top tier chat GPT Pro License.

[01:06:11] So what you get in return is priority access to new features, including a voice mode that's launching soon. And importantly, significantly higher usage limits, which people have been demanding for a very long time. Philanthropic says the demand for this tier has been building for over a year, especially from professionals in finance, media marketing, and code heavy fields who rely on Claude to scale up their work.

[01:06:35] And just a quick reminder here, 200 per month seems like a lot for your average business user, but it seems like there's demand for this. According to some reporting for the information back in January, they actually estimated that chat GT Pro licenses are the 200 buck a month ones could be generating as much as $25 million per month for open ai.

[01:06:59] Paul, should [01:07:00] we expect, to see more demand and, people buying these licenses from philanthropic open ai? 

[01:07:08] Paul Roetzer: I don't, I don't know who's paying 200 a month. I mean, maybe it's like developers and stuff, but again, I can only provide the context of like, my conversations with big enterprises and, and, you know, leaders and I don't know anybody using Claude.

[01:07:21] Hmm. Like, it's, again, it's, it, I think it's a great product. I still have a subscription to Claude. I do still test it at times. I just don't, I don't think I have a good context of their market and if they are seeing this as like a competitive product to, to, to like the $200 model from OpenAI.

[01:07:40] Yeah. And I just, I feel like it's just gonna be a really tough battle. Like, I feel like OpenAI has escape velocity with their user base. And I, I, again, I dunno, maybe Anthropic is being like super smart about the verticals they are going after the use cases. I don't know, maybe, maybe it might be good to like just get somebody from philanthropic on at some point [01:08:00] and like hear what it is they are doing, what their, what their market looks like.

[01:08:03] Because viewing them as a direct competitor to OpenAI is seeming more and more unlikely. Mm-hmm. Just given the usage rates at OpenAI. But I could be wrong, but again, I go, I just go back to like, they don't have their own proprietary data. They don't have distribution anywhere. Like two of the critical things that would tell me they, they are set up to like remain a key competitor, aren't there?

[01:08:28] And, and so I do just continue to wonder about their long-term viability as like a major player in this market. But they, they keep growing. I mean, it's, they are incredible growth and looks like a good company from all things, but I, I, whether it competes or not is, is just hard to say at this point. 

[01:08:43] Behind the Scenes of Apple’s AI Failures

[01:08:43] Mike Kaput: Next up, we have some more updates on the pretty serious roadblocks that Apple is running into as it tries to build a smarter Siri with advanced ai.

[01:08:54] So last year, apple promised a smarter AI powered Siri, but according to a new report, [01:09:00] from the information internally, the team appears to not even have been able to agree on the basics. They bounced between models, scrapped a privacy first approach, and cycled through leadership changes, all while rivals like OpenAI raced ahead.

[01:09:16] The chaos led to delays, staff departures, and ultimately the embarrassing admission that the upgraded Siri wouldn't even ship until 2026. Now this article details behind the scenes, kind of the fallout as well being swift apples stripped. Its AI Chief John gm, Andrea of responsibility for Siri handed control to software head Craig Federighe and Vision Pro exec like Rockwell.

[01:09:42] That team is now pushing to rebuild Sir future, possibly even opening the door to using open source or third party AI models. it's basically been kind of this hot potato at Apple that's been passed around between teams without a lot of real [01:10:00] progress. Now it's back to Feder Rigi, who's known for his execution, but many see it according to this reporting.

[01:10:06] As a last chance to bring Apple's AI assistant up to speed. Now, Paul, we have talked, talked about this topic at length. This article though, really seems to pull back the curtain on what's going on at Apple. Can this be salvaged at all in your opinion? I. 

[01:10:24] Paul Roetzer: I don't know. It's Apple. they are incredible company.

[01:10:26] Obviously. They've seem to, outside of, you know, recent ups and downs related to tariffs and other things, like they, their stock price doesn't seem to really be impacted by their slow move into ai. It's almost like the market is just, I, I've said this before on the podcast, it's almost like it'll be a surprise and an uplift if they ever figure out the AI thing, but they are so strong in product and distribution that it's just kind of like, people are kind of right.

[01:10:52] I was like, yeah, okay. The story still sucks. And, you know, they, they didn't figure out, vision Pro really didn't unlock the market we thought they were going to, and [01:11:00] yet they just keep humming along, as one of the, you know, biggest companies in the world. I will say though, this, this article was fascinating.

[01:11:07] Like the amount of infighting and indecision I is hard to fathom for, for such a, a meaningful strategic direction around, like, around ai. so it was definitely the most. Information I've seen about what happened there and why it happened. And it's hard to like look at it and realize like they moved so slowly.

[01:11:30] Mm-hmm. And now you kind of know why for the most part. And then, and again, like this is just one source, but of all the publications that Mike and I read and follow to do this podcast every week, the information consistently has stories often one to two months in advance. And you'll see this in the Wall Street Journal in like two months saying, fighting it Apple like causes delay.

[01:11:55] It's like, yeah, the information had that two months ago. You see this all the time. So they [01:12:00] are very well sourced and it's a, it tends to be a very factual, in credible publication. So I do believe that this is probably pretty close to what actually was happening internally. And it's, it's just crazy to look at and realize there's still a ways away.

[01:12:15] Like they haven't solved this in like, now we have gonna fix it in three months. Like, no, it's. 2026 is kind of when they are thinking they are gonna right the ship and that's wild. 

[01:12:26] Writer Releases AI HQ

[01:12:26] Mike Kaput: Our next topic is that the generative AI platform writer we have talked about a bunch on the podcast, has announced something called a I hq, which is designed to help companies build, deploy, and manage AI agents at scale.

[01:12:41] And basically at the core of the A I HQ is something called agent Builder, which is a low-code environment where IT and business teams co-create agents using, using visual tools. These agents can then be launched across departments from finance to hr, via writers', new agent library, [01:13:00] and personalized home dashboards.

[01:13:02] Apparently, according Tory, major players like Uber and Franklin Templeton are already using these tools to transform support content sales pipelines, and financial reporting. So Paul, this is just kind of a preliminary product announcement, but it sounds like writer, which is a platform we're pretty familiar with, is now all in on agents.

[01:13:24] Should we be expecting this to kind of become the norm or the direction that all these kind of third party gen AI startups go? It, 

[01:13:33] Paul Roetzer: it sure seems that way that all the SaaS companies are like what used to be apps or templates are just becoming agents. It's like the new terminology and then the, they, you know, rightfully so, like they, they do have new capabilities that are being built into these, so they are not just these deterministic templates.

[01:13:48] so yeah, I definitely think it's the direction. Within their announcement, they said there's over a hundred prebuilt agents, so I just, I clicked through that. I'm just looking at that now and give people an example, like, so within [01:14:00] marketing you've got a product detail page, copy agent, a retail product intelligence, brief agent, case study, agent blog, post outline agent.

[01:14:09] So it's a lot of like the tasks that you would do. Then they've got it broken into finance. There's like a tax research agent. HR has a job description. Agent sales has a earnings call summary agent. So it's it's task driven. It's, industry driven. So it's looking at different interest saying what are the common tasks in this industry?

[01:14:29] And then it's building those and then allowing you to build your own. It's not unlike agent space. We talk about Google, where I'm gonna be able to go in and I think theirs is a low code environment. I think a lot of these are gonna be no code to where an average knowledge worker with no coding ability is gonna be able to just go in and build things to, Hey, I go through these 10 steps every, you know, Monday morning, build an agent to do those 10 steps for you.

[01:14:51] And I do think that by this time next year, this is all gonna be very real. I still think agents are very early and there's probably a little, a little bit of [01:15:00] hype and maybe they don't deliver quite what you think they are going to, but I do think by like end of this year, you know, spring of next year at anybody in marketing, sales, service, whatever.

[01:15:10] You're gonna be able to just go in and say, Hey, I go through these 10 steps, optimize these 10 steps, Maria, and build me an agent that does this for me. And then like, have it send me an email every morning. I think that is very much going to be reality really, that you're gonna be able to just automate, through prompting with your AI assistant, just build agents to do repetitive processes and it's gonna make work.

[01:15:32] Wonderful. I I really think like anything you can imagine that has a repetitive process, you're gonna be able to build an agent to help you do it and to do it way faster than you, you did it before. 

[01:15:43] Mike Kaput: And like some of the examples you just mentioned with writer, even though you know the company is called writer, it's definitely moving beyond just kind of marketing content or sales content, right.

[01:15:53] Hr, finance, this is touching every single function. 

[01:15:56] Paul Roetzer: Yep. Yeah. There's a healthcare and life sciences, retail, [01:16:00] and you can tell it's early, like retail and consumer goods has two agents like yeah. It's all gonna get built out and verticalized and, yeah. 

[01:16:07] Ex-OpenAI CTO’s Startup Making Big Moves

[01:16:07] Mike Kaput: Next up, Mirati, former CTO of OpenAI is back in the spotlight because her new AI startup called Thinking Machines Lab is reportedly seeking a $2 billion seed round.

[01:16:21] That would make it one of the largest early stage raises in tech history, which is also notable because a company doesn't have a product or revenue and only recently emerged from stealth. It does have a pretty stacked roster of AI talent. Advisors now include Bob McGrew, open AI's former head of research, and Alec Radford, one of the key minds behind the original GPT models and open AI's dolly image generation model.

[01:16:48] Now, TI says the goal is to build AI that's more customizable, more general, and more understandable than what's out there today. They also have on the team ex OpenAI scientist, John Schulman and Barrett Soft as [01:17:00] CTO. Um. Paul, this is a pretty staggering seed round, but like not many details, which we have also seen with IA startup.

[01:17:08] Like what's the bet here that investors are making that makes these numbers make sense? 

[01:17:13] Paul Roetzer: You know, I think I saw in the last couple days, like Ilia just raised another 2 billion at like a $32 billion valuation. and I'm, I believe Google and Nvidia were in on that, investment round, if I'm not mistaken.

[01:17:24] Yeah. So I don't know, like there's literally tens of billions in value in startups that we don't know what they do. Mm-hmm. So when you look at this and you look at safe super intel intelligence, Familiia, we're talking about raising billions with no public knowledge of how they are gonna differentiate from OpenAI and others.

[01:17:46] Right. And I find that fascinating because there has to be something there, there has to be some unique approach to algorithms. There has to be some unique approach to like training that's more efficient, like. It's [01:18:00] gotta be something. It, it, it is not, let's go compete with Google and OpenAI and Xai and philanthropic and spend a billion dollars on a training run that cannot be what this is.

[01:18:11] because they know they can't keep up like that is, I think that ship has sailed. I think we now know who the frontier model companies will be. Maybe one more shows up or something. But you basically have four or five that can spend the billions to do the massive training runs that, you know, Sam talks about, and Google talks about these, these aren't them.

[01:18:31] I don't, I don't think these are frontier model companies. I think these are something different. And I don't know exactly what it is. There's some spot I would make some bets on a couple of things, but I don't know what they are, or what their market is like. It's so, you know, I was saying earlier like, what is the philanthropics market?

[01:18:48] Like, what is their total addressable market? What do they look at as, you know, the verticals or the industries they are gonna go after? and I think the same thing with these is like, what could they possibly be going after? Because at a $2 billion round, [01:19:00] like if they are, like, if you're ilio, I don't know what the valu what was the valuation on this?

[01:19:04] Of the $2 billion? C 

[01:19:05] Mike Kaput: This is 10 billion, I believe. Total. Okay. 

[01:19:08] Paul Roetzer: So, so at a 10, so I am not an expert on like investing rounds, but I've, I, I've spent some time on the, on the topic. So if you're investing at a $10 billion valuation at a seed round, the investors are gonna be looking for, at at least a 10 x return, like far, probably far greater than that.

[01:19:26] So they are saying like, this is a hundred billion dollars company out of the gate. Like that the market they see right now is like a hundred billion dollars market for this company. Now, that's not outta the realm of possibility. If you see Ilias and it's already worth 32 billion and you see opening, I just raised 40 billion, right?

[01:19:43] So, I mean, there's, there's huge markets out there to be won, but a $10 million, $10 billion seed round is absurd. Like the growth of that company must be so massive. So it's like, what could you possibly be bringing to market that is that vast in its market [01:20:00] potential that you're getting a $10 billion valuation before there's anything to show?

[01:20:04] Right. It's 

[01:20:05] Mike Kaput: crazy, right? Yeah. Especially when we're already saying who is paying $200 a month philanthropic. Right. Which is Right. My, I mean, something we understand 

[01:20:14] Paul Roetzer: realistically, like as an investor, you have to probably be looking this as a trillion dollar bet. Like you, you're, you're guessing that this company has the potential to be a trillion dollar company if you're putting that kind of valuation at a seed round with no products, no revenue, nothing.

[01:20:29] Deep Research’s Impact on Agencies

[01:20:29] Mike Kaput: Wow. So according to a new report from Digiday agencies, marketing agencies, ad agencies, et cetera, are increasingly adopting deep research tools from OpenAI, Google Perplexity, among others. That we have talked about quite often on the podcast. So just as a reminder, these don't just, you know, do a little research for you.

[01:20:52] They actually autonomously scan tons of different sources on the web or what data you give it to produce. In-depth research [01:21:00] reports on any topic that at least we found rival or surpass the work of humans and agencies are apparently starting to take that concept even further according to JI jy. So they are integrating deep research capabilities into their proprietary data sets and then using what deep research AI tools produce to then do even more.

[01:21:22] So one example they cite is Havas, which has integrated deep research tools within their broader data platform. One exec there calls this the long dreamed of quote, planning buddy. They've always wanted, and then they are taking the insights that deep research produces. I. Turning it into interactive tools like custom gpt that simulate consumer behavior as kind of digital twins of different audience segments or client types.

[01:21:49] And they are also using tools to find, synthesize, and upload info from a range of external and internal sources to get better insights into their customers, clients and [01:22:00] markets. Digiday also mentions one tool outside of the big AI labs that's gaining some traction for its workflows specific to marketing agencies.

[01:22:09] It's called Waldo, and it claims to be able to automate complex research tasks that agencies do every day. So Paul, the reason we wanted to mention this is, you know, we have talked a lot about how agencies need to start evolving based on what AI can do today, especially the deep research tools that made this much more obvious that how agencies worked in the past probably cannot be how they work in the future.

[01:22:35] When I read this, this definitely sounded at least like a step in the right direction. 

[01:22:40] Paul Roetzer: Yeah. It, it's cool to see it. I am, I do think that there are a lot of agencies that are figuring this out. I, I've talked with some recently some bigger firms that are doing some pretty cool stuff, like the digital twins idea, where you're just creating these segments and you're running simulated campaigns and Yeah.

[01:22:55] I just, I think the way this all works out looks so much different a year or [01:23:00] two from now. And I go back to, you know, the thing that originally drew me to AI back in 2012 when I was running an agency was this concept of a marketing intelligence engine where I could put all the data into it and I could actually use an ai.

[01:23:12] This is way before Gen ai, obviously, or in the, just the machine learning era of ai. Deep learning was just becoming a thing, just being proven that it could, it had the potential to become what it is today. But I had this vision to be able to. Automate marketing strategy by feeding data and running simulations and then making predictions about what campaigns would work and, you know, generate ROI.

[01:23:34] So that was what drew me to AI 14 years ago. So to see these sorts of things, it's almost like the early iterations that lead to that intelligence engine I envisioned long ago. I still have not seen someone build that, but I think this is the kind of stuff that starts to get us much closer to where AI is truly infused at the strategic level.

[01:23:54] and this idea of building personas and simulations and digital twins like that, that [01:24:00] that's the stuff that can really on or start to unlock it, where you can have a million customers represented and then you can run simulations against those million customers to predict the performance of a campaign or tagline, things like that.

[01:24:12] Hell open AI could do this. They could build a model that, runs like simulations of how people respond to their messaging. Yeah. Here's 10 ways we're gonna explain a GI. How would a million people respond to that? Go do that. It's, you could probably find a way to do that. Borrow a few GPUs to run some test messaging against.

[01:24:32] Mike Kaput: Yeah. What I also just love about this is how Havas is like not only using the deep research tools apparently, but taking that output and like the custom GPT for simulated audiences is awesome. But also, presumably it's like, how easy would it be for an agency to run deep research reports on any part of their business or their clients.

[01:24:50] Drop a bunch of that into a custom GPT that anyone on the team, even those who are as AI savvy, can then start using. Seems like an interesting idea. 

[01:24:59] Paul Roetzer: [01:25:00] Yeah. And imagine like, you know, we just talk about the writer and their AI agents or agents based with Google. Like imagine an agency or it could be a brand, but whatever.

[01:25:08] And as a marketer, I just go in and say, Hey, I'm trying to plan for a campaign to launch a new product. I. What GPTs do we have that would help me do this? Mm-hmm. And like an AI agent you're interfacing with pulls the three custom gpt that people have built and then recommends which ones will help you and how they'll help you.

[01:25:25] And here's the rationale, and then asks you, would you like me to start running simulations for you? Would you like me to? That's the future workflow. Yeah. And I don't think I'm saying far future. I think I'm like 12 months future probably to be able to start doing those kinds of things. 

[01:25:41] Listener Questions

[01:25:41] Mike Kaput: Alright, Paul, our last topic today is our recurring segment on listener questions.

[01:25:46] So each week we take a question from our listeners and our audience and try to answer it to create some more value for everybody. So this week's question, someone asked, how do you filter out the signal [01:26:00] from the noise in generative ai given that the space evolves daily, to which I would also add, given all the hype and craziness that we have to follow day in, day out.

[01:26:10] Paul Roetzer: Yeah. So, um. I'll tell you how I do it, and then I'll offer like a general guidance. So the way I largely, stay sane and filter this is I actually have notifications set up on, on X. So I have a highly, highly curated list that I've built over the last 10 years of AI researchers, influencers, authors, entrepreneurs, media, who I trust, who I well, who I feel like have some inside knowledge that I don't have.

[01:26:40] And that by curating that knowledge, I can create a better picture of what's actually happening in ai, that if some announcement is made, I can go to that feed that are purely notifications. This isn't just a list, this is like notifications of, of a list, like filtered further down. If something was talked about, something major happened, I can [01:27:00] immediately get 10 different perspectives on that announcement.

[01:27:02] I don't take one person's announce, like perspective, like as the truth. I look around and say, okay, what are other people that I follow? What are they saying about this? And I often will put contrarian people into those notifications too. It's like, I wanna know what the safety people are saying versus the EAC people.

[01:27:18] The accelerationist. Like, I want both perspectives. So for me to do what we do, where I have to talk about this every, you know, Tuesday, I need to take in as many perspective as possible and be as objective as possible. And the way I do that is largely through highly filtered notifications on Twitter, and then I listen to a ton of podcasts.

[01:27:37] Mm-hmm. and then I watch a lot, read a lot of research reports and articles, things like that. So basically it starts with filtering people that are your, that are your influencers. Who are the people's voices I trust, and then what are they sharing, what are they talking about? That's how I do it. So I've, I've done it for, you know, over a decade in ai.

[01:27:55] if you're just getting started though, I think you find a [01:28:00] couple of voices that you trust. If, hopefully we're one of those, you know, hopefully this podcast, each Tuesday is part of your process. I know for some people I talk with, this is their process. Like they don't have time to do everything, you know, research this space, so they just listen to this once a week, and that gives them some peace of mind that they are at least aware of the key things that happen.

[01:28:20] And then I think if you want to go down further, it's like, okay, now you go find your newsletters and you find, you know, your people to follow on Twitter, LinkedIn, and like, you know, there's standard ways to research it. But, that's, that's kind of how I do it, is that I find the trusted voices and then I follow the things that they talk about and the other people that they follow.

[01:28:36] I, how about you, Mike? How do you, you know, keep, keep up to date? 

[01:28:39] Mike Kaput: Yeah, definitely similar. I certainly don't have as built out of a system as you do for the alerts and especially people posting about it. But also I would say what's been really helpful for me, aside from. Trying to stay on top of the news from trusted Voices is also taking some time based on what you know or what you're learning to map out the vision of where you think [01:29:00] this is going.

[01:29:00] It doesn't have to be like crazy in depth, but like, where do we think in the next 12 to 24 months, your job, your function, your livelihood could be impacted. And that helps me then say, okay, am I looking at something that is noise or something that's interesting but not immediately useful? it gives me a good filter to say, okay, based on where we're going, have I made progress towards evolving my career or expertise towards where I think the puck is kind of sliding to?

[01:29:31] or am I just spending too much stuff, you know, reading certain things that are, are interesting but not going to help me move the needle. 

[01:29:39] Paul Roetzer: Yeah. And I just, I wouldn't get overwhelmed. Yeah. I mean, if you're, again, if you're just trying to figure this stuff out. Pick one or two things you can ease into it and like figure it all over time.

[01:29:47] But there's no replacement for experimentation. Like that's the one thing I would tell people is like, yeah, you can listen to podcasts all day long. Take course, whatever you wanna do, but until you start using one of these tools daily, Chacha, bt, Gemini, Claude, whatever you prefer, [01:30:00] that's the best way to learn right now is just experiment with things.

[01:30:03] Try 'em out, figure it out for yourself, what's going on. 

[01:30:06] Mike Kaput: Alright, Paul, that's a pretty packed week in ai. Just a quick housekeeping reminder here. If you haven't checked out the marketing AI newsletter and marketing AI Institute newsletter, please check that out@marketingaiinstitute.com slash newsletter.

[01:30:20] We cover all of today's stories plus all the ones we didn't get to. And Paul, thank you so much for curating and unpacking and demystifying everything for us today. 

[01:30:31] Paul Roetzer: Absolutely. And quick show note, no episode on. April 22nd, I'll be on spring break with my family. And so there, there will be no episode next week.

[01:30:41] So we will be back, I guess that would be what, April 29th with, the next episode. but newsletters will go out in between. I'll, I'll probably still send my exec AI newsletter from Smart x.ai, so you can go to Smart X ai and subscribe to the newsletter there. That's every Sunday I drop that newsletter with editorial [01:31:00] and a preview of what we're gonna talk about on the blog, or on the podcast.

[01:31:03] And then Mike, authors the newsletter from the institute that has kind of a recap of everything and links to the week. So, check both those out. And in the meantime, if anybody else go on spring break, enjoy. I know I'm gonna try and unplug and enjoy and we'll be back with you at the end of April.

[01:31:21] Thanks for listening to the Artificial Intelligence Show. Visit smarter x.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.

[01:31:46] 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

Gemini 2.5 Pro, AI Agents, and a Legendary Wizard of Oz Makeover: The Biggest Reveals from Google Cloud Next '25

Mike Kaput | April 15, 2025

Google Cloud just wrapped its Next ‘25 event in Las Vegas, unveiling a jaw-dropping 229 announcements.

Shopify’s CEO Just Issued a Bold AI Ultimatum to His Entire Team

Mike Kaput | April 15, 2025

Shopify’s CEO Tobi Lütke just gave his company a bold new directive on AI.

ChatGPT Will Now Remember Everything You Tell It

Mike Kaput | April 15, 2025

OpenAI just released a new memory feature for ChatGPT that allows it to recall everything you’ve ever discussed with it.