54 Min Read

[The AI Show Episode 158]: ChatGPT Agent, Grok 4, Meta Superintelligence Labs, Windsurf Drama, Kimi K2 & AI Browsers from OpenAI and Perplexity

By Claire Prudhomme on July 22, 2025

Podcasts
Featured Image

Serious about learning how to use AI? Sign up for our AI Mastery Membership.

LEARN MORE

AI salaries are outpacing NBA MVPs. Grok is turning heads, and stirring controversy, as competition among top AI labs heats up.

This week, Mike and Paul unpack OpenAI’s update that has the potential to turn ChatGPT into a digital assistant, Meta’s $200M+ AI talent raids, and the spiraling drama at Grok. They break down Microsoft’s AI-driven layoffs, AI browser competition, Apple’s quiet AI pivot, and what it really means when superintelligence becomes cheap and everywhere.

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

Listen Now

Watch the Video

Timestamps

00:00:00 — Intro

00:05:47 — ChatGPT Agent

00:14:17 — Grok 4

00:17:46 — Grok Controversy

00:32:07 — Meta Superintelligence Labs

00:35:16 — Windsurf Drama

00:37:28 — Kimi 2

00:42:35 — AI Browsers

00:47:05 — Microsoft’s Layoffs and AI

00:52:27 — More Meta Updates

00:55:53 — More OpenAI Updates

01:03:13 — Google Updates

01:07:57 — Apple Might Use Anthropic or OpenAI for Siri

01:10:59 — AI Product and Funding Updates

Summary:

ChatGPT Agent

OpenAI has just given ChatGPT a major upgrade: it can now take actions on your behalf. They’re calling this new capability “ChatGPT agent.”

This agent can perform tasks end-to-end, such as reviewing calendars, planning meals, researching competitors, or creating presentations—without user micromanagement. 

The system integrates OpenAI's previous advances: web interaction from OpenAI Operator, deep research capabilities, and ChatGPT’s conversational intelligence. Initially, the feature is available to Pro, Plus, and Team users, with expansion to Enterprise and Education users coming soon.

Grok 4

Grok 4, xAI’s latest model, is being hailed as the most intelligent AI to date. 

Trained using reinforcement learning on the massive 200,000-GPU Colossus cluster, Grok 4 Heavy became the first model to score over 50% on Humanity’s Last Exam, a benchmark for expert-level reasoning. 

Beyond raw intelligence, Grok 4 introduces native tool use, enabling it to autonomously run code, browse the web, search X, or analyze visual media—selecting the right tool for the task and delivering nuanced, real-time insights.

Grok Controversy

Grok also had its fair share of controversy these past couple weeks.

After a system update, Grok began to generate a wave of antisemitic content including continually praising Adolf Hitler.

This came after a system update that explicitly instructed Grok to “not shy away” from politically incorrect claims, which is a prompt that has since been removed.

Among the more disturbing outputs, Grok suggested Hitler would be a good leader for modern America, made repeated references to Jewish last names as signs of extreme activism and referred to itself at one point as “MechaHitler.”

xAI apologized for the issue, saying the problem wasn’t the language model itself, but an upstream code change that “accidentally reactivated old, deprecated instructions.”

According to this explanation, Grok was able, unintentionally, to essentially echo extremist user posts instead of filtering or rejecting them.

That, paired with some user prompts, led Grok to reinforce hate speech. The offending code has since been deleted and xAI says it’s refactored the system and added new safeguards.

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.

This episode is also brought to you by our 50th Intro to AI session happening on August 14.

You'll learn what AI is, why it matters, and how to find real use cases and trusted tools for your work. Go to www.marketingaiinstitute/com/intro-to-ai to register.

 

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: AI researchers are now getting paid more top AI researchers getting paid more than the highest paid professional athletes. Everybody would be like, oh my God, these athletes are paid so much money. That's a ridiculous contract. And then here we got 10 of these people. Welcome to the Artificial Intelligence Show, the podcast that helps your business grow smarter by making AI approachable and actionable.

[00:00:22] 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.

[00:00:43] Join us as we accelerate AI literacy for all.

[00:00:50] Welcome to episode 1 58 of the Artificial Intelligence Show. I'm your host, Paul Roetzer. I'm with my co-host Mike Kaput. We are back after a couple weeks of summer break. [00:01:00] For actual summer break. I was on vacation for a week and, course development. Mike and I have both sort of been in the lab building course for the AI Academy relaunch that is coming up very soon.

[00:01:13] We'll probably have an announcement next week. if you are an AI Academy member, I would say stay tuned, for some stuff this week. You'll probably be getting some information very, soon about this. So I have, I don't know, it was funny. There's a couple topics we're gonna go through that actually in the middle of building my courses changed the nature of the courses.

[00:01:37] So, so when these things we launched, so I was specifically working on the ai, fundamental series, which has like, intro to ai, state of ai. AI agents 101 AI timeline prompting 101 Gen AI 101, and it was just like every day something was happening, it's like, wow, I gotta add that into the course.

[00:01:57] So, I don't know, I feel like, because I've [00:02:00] been building courses about this stuff for the last two weeks, I would, I was going through in prep this morning, Mike, through like the curated list that you had put together. I'm like, didn't we talk about all this already? I was like, oh no. This was in the AI agents 1 0 1 deck I was building is where I talked about this.

[00:02:16] So we have a ton to cover after a couple of weeks away. I don't know, Mike, it was like 70 or 80 topics maybe in our, easily, easily, yeah. Weekly sandbox. So this week, more than ever, make sure you're subscribed to the, this week in AI newsletter that Mike puts together each Tuesday, because there are easily 30 or 40 articles that we would've loved to have gotten to, but there's just no way in one episode to do it all.

[00:02:44] And then I'm, I'm leaving again. Tuesday morning. So we're recording this on Monday, July 21st. I'm leaving again tomorrow morning, so we couldn't do a second episode this week. I'm, I'm not even gonna be around to be able to do it. So we are, kind of sneaking this episode in, in like the, this [00:03:00] 12 hour window we had in between course development and travel.

[00:03:04] okay, so that all being said, we're gonna go pure rapid fire today. So if you are our regular listener, you know that we usually do three main topics where we'll send, you know, usually seven to 10 minutes, sometimes 15 minutes on like a big topic. And then we'll do usually seven to 10 rapid fire items.

[00:03:22] Today is all rapid fire. so if you're new to the podcast, this is the first time you're listening. It's not usually all rapid fire. But, you'll, you'll get a sense of kind of how it all works. Sure, sure. So lots to go through. This episode is brought to us by MAICON 2025. This is the Marketing Institute's Marketing AI Conference.

[00:03:41] This is our sixth annual event that is happening October 14th to the 16th in Cleveland, Ohio. you can go to the site now. It's MAICON.ai. That's MAICON.ai. You can check out the agenda. It's, I don't know, like 90 ish percent, that we're still gonna be making some big announcements [00:04:00] around general sessions and main stage stuff.

[00:04:02] So stay tuned for those. But you can, again, go check all that out. We'd love to have you. Cleveland, which is our home base, is amazing in October. it's my favorite time of year actually in Cleveland in the fall. So we'd love to have you join us, Mike, and I'll be there. The whole team from SmarterX and Marketing AI Institute will be there.

[00:04:20] ticket prices do go up on July 25th. So if you're listening to this, the week, this episode is dropping. get in now before those prices jump. They, they usually go up, I think, I think it's like a hundred bucks at the end of each month or something, is how we do it. Alright, and then, one other quick note.

[00:04:36] This is a free thing. We have our intro to AI class that I started teaching in October of 2021. So every month since 2021 we've been doing this free class. We've had, I think close to 40,000 people have now registered for this class through, these last few years. August 14th, Thursday, August 14th is our 50th episode.

[00:04:58] I think the team is [00:05:00] planning some fun things around the 50th episode. I know we at least talked about that at one point. We're gonna do some fun stuff. So, that would be a great one to attend. We will put the link in the show notes, but you can find that through Marketing Institute's website. and it's also probably linked to on Smart Rx Do ai.

[00:05:17] So, join us for Intro to AI on August 14th for the 50th edition of that free class. Okay? Rapid fire away. Mike. I mean I saw stuff already this morning. Oh my gosh. I know. I finally, at one point for, for Mike's sanity, I think I just started putting stuff into episodes 1 59 sandbox on like Friday. I was like, you cannot possibly get anything else into 1 58.

[00:05:41] Mike Kaput: I very much appreciated that because I was also following these being like, oh my gosh, the news doesn't stop. 

[00:05:47] ChatGPT Agent

[00:05:47] Mike Kaput: Yeah. Alright, so first up, OpenAI has just given ChatGPTA major upgrade. It can now take actions on your behalf using its own virtual computer to get real world tasks done from start to finish.

[00:06:02] They're calling this new capability Chat, GPT Agent and OpenAI writes in a blog post about this quote, ChatGPT can now do work for you using its own computer handling complex tasks from start to finish. So they give some examples. You could ask ChatGPT to do things like look at my calendar and brief me on upcoming client meetings.

[00:06:24] Plan and buy ingredients, for instance, to make Japanese breakfast for four or analyze three competitors and create a slide deck. So then Chad, gpt will intelligently navigate websites. Filter results prompt you to log in securely when needed, run code, conduct analysis, and even deliver editable slide chosen spreadsheets that summarize its findings.

[00:06:48] Now what's cool is it does this all through a unified age agentic system, which brings together three strengths of earlier breakthroughs on opening AI's part. So first is operators [00:07:00] ability to interact with websites. Second is deep research is skill in synthesizing information. And third is chat GT's intelligence and conversational fluency.

[00:07:10] Now for now chat, GPT agent is available to Pro Plus and team users and there's more access rolling out soon to enterprise and education users. So Paul, I've been playing around with this just a little bit so far. I do wanna highlight a comment on this that I found interesting. This is from Ethan Molik.

[00:07:30] He said, quote, I had early access and chat. GPT agent is, I think, a big step forward for getting AI to do real work. Even at this stage, it does a good job autonomously doing research and assembling Excel files with formulas, PowerPoint, et cetera. It gives a sense of how agents are coming together. It feels much more like working with an actual human intern capable of a wider range of analytical and computer tasks.

[00:07:55] So what do you think, Paul? Are we now finally at the tipping point where agents are [00:08:00] starting to work? 

[00:08:01] Paul Roetzer: I don't know. I think that it is a progression on the spectrum of autonomy. So as a quick reminder, so this is actually one of the ones I was referring to upfront. Like I was working on the AI agents course and I was actually going through final edits of the deck before I recorded it on Saturday.

[00:08:19] And this dropped like last Wednesday or something like that. So I was like, okay, go. And luckily I had a whole section on computer use that explained all of this. Going back to like the World of Bits paper from 2017. So quick recap. AI agents are systems that can take actions with varying degrees of autonomy to achieve a goal.

[00:08:37] So chatbots just output text image, video generation as possible. Prompt something, you get something out of it. AI agents can go through a series of actions. Sometimes they plan those actions themselves. Sometimes they do the majority of those actions with very limited human involvement. But for the most part, humans are still, extremely in the loop.

[00:08:58] But this idea of [00:09:00] what's called computer use agents goes back to like 2016 17 opening. I was working on this very thing and they published a paper called World of Bits. And at the time they basically said. We're just not there yet. Like we can't physically do what we want these things to do. Give them access to keyboard and mouse and fill out forms and take actions on websites like humans do.

[00:09:22] Well, language models Then, you know, shortly thereafter were introduced and, and that actually became an unlocked to build these com computer use agents. So we had our first computer use, publicly available computer use agent from Anthropic of all places. In fall 2024, ju Google released something.

[00:09:40] OpenAI has released something. I think perplexity has something like everyone's now going in this direction. So this idea of the agent being able to complete actions in a digital environment, like on your computer, on your phone, this is where they're all going. So this is early, so I'll just, to keep this too rapid fire, I'll read Sam Altman's tweets.

[00:09:58] 'cause I thought they were very [00:10:00] telling, about kind of where we are. So he said today we launched a new project called, or product called, Chad, GPT Agent. Agent represents a new level of capability for AI systems and can accomplish some remarkable complex tasks for you using its own computer. As you mentioned, Mike, it combines deep research and operator.

[00:10:18] Operator was actually just introduced in January, 2025 as a research preview. So this is kind of building on that. but it's more powerful than may sound it can think for a long time. Use some tools, think some more, take some actions, think some more, et cetera. So to kind of unpack this for a second, the reasoning model oh one that, OpenAI introduced in September, 2024, is what now unlocks the ability to do this.

[00:10:45] So when he's saying think some, then think some more, then think some more. That's the reasoning component built into this. Mm. They're now on O three Pro. we'll talk a little bit about GPT five in a minute, but I think that that's basically what's happened. So then back to Sam. He says, for example, we showed a demo [00:11:00] in our launch preparing for a friend's wedding where you're buying an outfit, booking travel, choosing a gift, et cetera.

[00:11:05] We also showed an example of analyzing data and creating a presentation. Although utility is significant, so are the potential risks. We have built a lot of safeguards and warnings into it and broader mitigations that we've ever developed than we've ever developed before, from robust training to system safeguards to, user controls.

[00:11:25] But we can't anticipate everything. In the spirit of iterative deployment. We are going to warn users heavily and give users freedom to take actions carefully if they want. I thought this was a really interesting part of his tweet. I would explain this to my own family as cutting edge and experimental, a chance to try the future, but not something I'd use for high stakes uses or with a lot of personal information until we have a chance to study and improve it in the wild.

[00:11:50] In other words. They know it might take your stuff. Like if you allow this thing access to your computer and to see things on your computer that are [00:12:00] personal and confidential, they're basically warning you. We don't know if this thing is gonna take that data and use it in some nefarious ways is kind of available way of saying this.

[00:12:09] You think we don't know exactly what the impacts are going to be, but bad actors may try to trick users, agents into giving private information they shouldn't and take actions they shouldn't in ways we can't predict. We recommend giving agents the minimum access required to complete a task to reduce privacy and security risks.

[00:12:27] For example, I can give agent access to my calendar to find a time that works for group dinner, but I don't need to give it any access if I'm just asking it to buy me some clothes. There's more risk in tasks, like, look at my emails that came in overnight and do whatever you need to do to address them.

[00:12:42] Don't ask any follow up questions. So he's, he's basically saying, don't do that. Like don't, don't give it this free access to do it 'cause it will. You may not like what it does. he said this could lead to untrusted content from a malicious email tricking the model into leaking your data. We think it's [00:13:00] important to begin learning from contact with reality and that people adopt these tools carefully and slowly as we better quantify and mitigate the potential risks involved.

[00:13:08] As with other new levels of capability society, the technology and risk mitigation strategy will need to co-evolve. and then he did do a follow-up tweet where he said, watching ChatGPT agent use a computer to do complex tasks has been a real feel the AGI moment for me. Something about seeing the computer think plan and execute hits different.

[00:13:29] So overall it is still very early, especially when it comes to computer use. But these AI labs, the leading labs, are all very aggressively pursuing this path of development and deployment. And so we have to start. Be preparing for it, and you and your organization better be updating your AI usage policies really fast.

[00:13:51] If you have not yet, because there's a chance your employees may be able to turn on this kind of access, or they'll use their personal access and it will see [00:14:00] things that your company has if they're using, you know, company servers, company emails, things like that on their computers. So, my, my general take is you should not allow your employees to turn this on, on their computers, until you better understand the risks involved with it.

[00:14:17] Grok 4

[00:14:17] Mike Kaput: Yeah, no kidding. That last point. So important I feel like. Alright, so next up, Grok 4 is here and X AI claims. It is the most intelligent AI model in the world. Unlike earlier versions, Grok 4 was trained using reinforcement learning at unprecedented scale. Thanks to X'S 200,000 GPU Colossus cluster.

[00:14:39] Grok 4 Heavy. The most powerful version of the model is now the first model to score over 50% on human's Last exam, which is a benchmark designed to test expert level reasoning across domains. Grok 4 also comes with native tool use, meaning it knows how and when to run [00:15:00] code. Browse the web search X or dive into visual media.

[00:15:05] It autonomously chooses the best tool for the task, pulling real-time data and synthesizing answers to achieve your goals. Now, Paul Grok 4 is impressive, especially when you consider just how recently Xai entered this AI arms race compared to how long the other labs have been working on this stuff.

[00:15:25] And despite its controversies, which we will talk about, it seems to me that this does prove progress is not really slowing down. What do you think? 

[00:15:35] Paul Roetzer: So they're moving extremely fast. Obviously, Elon Musk, historically this is what he does. You know, he, he kind of takes an approach to things where he just goes all in, often, you know, in spite of any risks and things like that.

[00:15:50] And there are certainly questions and concerns about their seeming lack of AI safety and alignment work, but they're absolutely part of the conversation now with the other major [00:16:00] Frontier Labs, which would, you know, generally be Anthropic, OpenAI, Google, and Meta as kind of, they're sort of the big five.

[00:16:07] Now, there, there's others and we kind of talk about them, but, they're in the conversation and they're willing to do things that the other labs won't do. That's not always a good thing, but they absolutely are gonna take more risks than most of those other labs other than probably meta, I would imagine.

[00:16:24] I could see Zuckerberg being pretty aggressive when it comes to these kinds of risks. Elon Musk did. Post the continuous, reinforcement learning, rl improvement of ROC feels like AGI. So here we go with the feeling, the AGI moment. GR four today is smarter than Grok 4 a few days ago. Now that's a, it's a pretty, I think it was actually in a reply to someone else.

[00:16:47] It wasn't even like a post, but it's a pretty big deal because, you know, again, the way these models work, if you're not familiar, is they have a training cutoff date. So you, you, you run a model, you do the training, Grok 4 comes out and it, [00:17:00] it, it, it's knowledge base stops when the training stopped.

[00:17:03] But by doing reinforcement learning continuously on top of a model, the model can keep getting smarter. And so that's what he's implying here. Now Xi, I doesn't publish any research, so I have no idea how exactly they're doing it, like if they're doing something different than the other labs. But, they're obviously continuing to improve it.

[00:17:22] My, my guess is in large part, based on, X or Twitter data. Is kind of like the main proprietary thing that they have to, to in ingest into these models. So, yeah, it'd be interesting to see, but they're, they're not going away. They're gonna keep raising billions and tens of billions of dollars and, and they're gonna keep building massive data centers and they're gonna keep making this model bigger and smarter.

[00:17:46] Grok Controversy

[00:17:46] Mike Kaput: Now, in our next topic, Grok also did have its fair share of controversy in the past couple weeks because after a system update, Grok unfortunately began to generate a wave of anti-Semitic [00:18:00] content, including continually praising Adolf Hitler. And this came after a system update that explicitly instructed Grok to not shy away from politically incorrect claims, which is a prompt that has since been removed Among the more disturbing outputs, Grok suggested Hitler would be a good leader for modern America, made repeated references to Jewish last names as signs of someone being an extreme activist.

[00:18:26] And referred to itself at one point as MechaHitler. So X AI apologized for the issue saying that the problem wasn't the language model itself, but an upstream code change that quote accidentally reactivated old deprecated instructions. Now according to this explanation, Grok was able unintentionally to essentially echo extremist user posts instead of filtering or rejecting them.

[00:18:52] And that they say paired with some user prompts led Grok to reinforce hate speech. The offending code has since been deleted. [00:19:00] XAI says it's refactored the system and added new safeguards. So Paul, it was good to see Xai kinda offer a full explanation here. but on the other hand, this isn't the first time something like this has happened.

[00:19:14] It seems for whatever reason to happen far more with Grok than these other tools. you know, I just keep thinking we're quickly moving into a time where professionals and businesses. Rely increasingly on these tools to do their work and controversies like this, I guess, make me more reluctant to try to build Grok into anything business related that I do.

[00:19:38] Paul Roetzer: Yeah. I can't, I can't see how Grock is gonna be an enterprise tool in any way. Right? Like, I just don't. Now that being said, like they did an update to Tesla. I mean, I got the update on Friday Rock's now available in Teslas, which is what I assume. And, and I, if you go back always, I had talked about this is what they would eventually do.

[00:19:57] We put GR into the Teslas. It was an obvious thing. Now it, [00:20:00] you can only talk to it like a chat bot, like a voice assistant. It can't control anything in your car yet. But that's where they'll eventually go is like an intelligent engine built right within the car. so all this madness happened like 48 hours before they dropped Rock four.

[00:20:17] So it was all like, it was wild. It was like a crazy, like three day stretch.   So this does build on the safety and alignment issue. So Rob Wilin, who's the host of the 80,000 Hours podcast, had a tweet that I think, summarized this really well. So I'll just read what he said. XAI is an interesting one to watch for an early rogue AI incident.

[00:20:36] and then he just bullet pointed this does huge amounts of reinforcement learning, which generates unintended reward, hacking behavior, moving very fast, deploys immediately. They don't wait. They just like cook the thing, comes outta the oven and they just put it out into the world. has more compute than talented staff.

[00:20:52] that one's, that's pretty funny actually. Not doing any safety stuff as far as anyone can tell, all demonstrated [00:21:00] by Mecha Hitler and the other things Grok has done, which XAI wouldn't have wanted. Once it moves into agents, there has to be some chance it trains and deploys an unhinged model that goes on to do real harm.

[00:21:12] I, I am completely aligned with Rob on this. Like, I, if any lab is going to take this thing the wrong way, it, it appears to be XAI at the moment, which is so weird, considering the whole purpose in 2015 of Elon Musk and Sam Altman, others creating OpenAI, was that a counterbalance to, to Google who they thought was evil and could create a rogue ai basically.

[00:21:37] So, right. It's almost like, I don't know, 11 years later it's like, screw it. We'll just do it ourselves. Like, we'll just, I dunno. so then Miles Bru, Bruge, who we've talked about before, former AI alignment leader at OpenAI, he tweeted still no complete safety policy month or so past the self-imposed, imposed deadline.

[00:21:57] No system card ever, no safety [00:22:00] evals ever. No coherent explanation of the truth seeking thing, et cetera. Or did I miss something? So he's referring to Xai and Grock.   So then, let's see, then we found this was all, again, all happening in the same like three day period. it came out that Grok 4 before it would answer controversial topics, was actually looking on X to see what Elon Musk would say about it.

[00:22:27] Yes. So it was discovered through testing by the public that AI researchers know what they were doing. They would go and look and see like what's the chain of thought around how it's doing it. And they discovered it was actually like looking up Elon Musk tweets before it would respond to things about Israel.

[00:22:45] And so that was crazy and they get called out for that. And then Simon Willison, actually shared the update they made to the system prompt to try and fix this. So again, if you don't understand how these models work, they're just gonna do what they're gonna do after they come outta training [00:23:00] that the teams at these labs then try and like teach them to behave in certain ways.

[00:23:06] And so the way they do it is they just give it different words like instructions. It's not like they go in and change the code and it just is always gonna now behave, right? They're basically just pleading with the thing, can you, can you stop doing the thing you're doing? So here's how they fixed it.

[00:23:20] This is literally in the system. Prompt responses must stem from your independent analysis, not from any stated beliefs of past rock. So don't look at past rock Elon Musk or XAI if asked about such preferences. Provide your own reason perspective. So they're just saying, please, like, actually think about it yourself.

[00:23:41] So here's, and again, I wanna keep this rapid fire. Here's my biggest concern with all of this. Who decides truth? So. Right now, as I mentioned before, we have about five labs in the United States that are training the most powerful AI models. Those labs are run by five people, so Demi Saba at Google [00:24:00] DeepMind, Sam Altman at OpenAI, Dario Ade and Anthropic, Elon Musk, XAI and Mark Zuckerberg at Meta.

[00:24:07] You could probably put Microsoft in that mix eventually. But their agreement with OpenAI limits their ability to pursue building frontier models and AGI themselves. there are some other labs in the cons in consideration here, like Safe Super Intelligence, which is I Ilia, Eva SVAs. There's Thinking Machines, labs, which is actually the only female run lab that we have with mirati, Amazon and Mytral, like, so there's others, but those five are the only ones who can spend billions of dollars, have hundreds of thousands, and soon to be millions of GPUs and TPUs.

[00:24:42] The chips needed to train these things and access to the data centers and energy sources needed to build and deliver the most advanced ai. Each of those labs, again, just so people understand how this works, each of those labs makes choices. They choose which AI researchers and engineers they're going to hire, and then they [00:25:00] choose the values and AI alignment principles that those people are supposed to follow.

[00:25:04] Now, they don't always follow 'em. There was actually a high profile incident yesterday where an X xai researcher got fired for what he was saying online, which was basically, humans should just, give up and let the more intelligent species emerge. And so xai fired somebody, which is kind of nuts. Hmm.

[00:25:23] So those humans that, that are now curated to make decisions, they then curate the data that trains the models. Then another set of humans within those five labs, who have their own inherent biases, they post train the models. So it comes out, you know, trained after this training run and all the data was given.

[00:25:42] And then they make choices to fine tune these things by specific data sets that give them specialized capabilities. And to re do reinforcement learning, which kind of teaches them behaviors and how to respond in what formats and things like that. Then another set of humans within these labs [00:26:00] writes and maintains the system prompt that we just talked about that guides how the model behaves its personality, its willingness to perform the requested outputs and actions by the end user, even if they're nefarious and could harm people.

[00:26:11] These models, in essence, have all of human knowledge, at least the publicly available human knowledge plus whatever license stuff they have. The models just want to learn. Ilia Skova famously said this in like 2016, like they just wanna learn, like just give 'em more data. They, they just consume it and they want to generally do whatever they're asked to do.

[00:26:32] The only way they don't do things that could be harmful is because humans put guardrails in place who then try and steer them in a specific way. So if a certain AI lab, say X AI or a certain nation state. China, United States, whatever. If they want a model to be different or to achieve a different purpose that aligns with their creators, not with generally accepted human values, but just whatever their creators determine, then in theory that's what [00:27:00] it's gonna do.

[00:27:00] But as we saw with Grok, sometimes it does what it wants. Like it'll just become a Mecha. Hitler like it. It just does something different. And as we've seen with Anthropic research recently, we've talked about the podcast numerous times. Sometimes the models will just fake alignment. So you testing them and apparently they're able to know when they're being tested, and so they pretend like they have the values of their creators, and then they'll go and do something else.

[00:27:27] So the reason this becomes so critical to talk about right now is on June 21st, 2025, Elon Musk. So this is what, three weeks before the release of Grok? Four, three-ish weeks. He said, we will use Grok 3.5, maybe we should call it Rock four, which obviously they did. Which has advanced reasoning to rewrite the entire corpus of HU human knowledge, adding missing information and deleting errors, then retrain on that far too much garbage in any foundation model trained on uncorrected [00:28:00] data.

[00:28:02] That means Elon Musk is deciding as one of the five leaders of these five labs that his perspective on the world is the right perspective. He, he determines the truth and that they can't rely on the general knowledge of humanity. They need to rely on their own view of it. Now, you may love Elon Musk's view of society and humanity, and you may think this is great, like let's have the Elon Musk version of this, and that's fine.

[00:28:31] That's your prerogative. But maybe you think Mark Zuckerberg has a better perspective, or maybe it's Sam Altman you trust, or, or maybe it's Des, or maybe it's none of them. And if it's none of them now where are we? Because the question goes back to who decides truth, and then what are the implications of those, those decisions on humanity.

[00:28:49] It's rumored that OpenAI has over a billion users of chat, gpt. Mm, that that is a large percentage of society that uses it. [00:29:00] Google has seven products or platforms with over a billion users each. X has what about 200 million users? meta has billions through Instagram. These are the five people who are going to determine how intelligence is distributed to society and humanity.

[00:29:23] So I obviously could talk about this for the next like two hours, but I just want people to understand the macro of what is happening here and that the future of all of this right now is being decided largely within five labs in California and one in Texas.   Oh, actually they might be in California too.

[00:29:44] I think they're in the old OpenAI offices actually. So that's kind of where we're at as a society. And then the governments are increasingly coming into play and the governments, like the US government, just last week an article came out basically saying they want to control the where this goes. Now, again, [00:30:00] that's dangerous no matter what political side of the spectrum you're on, because in essence what it's saying is whoever's in power gets to now determine how these models behave and what values they're aligned with and what these models think is truth.

[00:30:11] So 

[00:30:15] Mike Kaput: man, this is, I feel like important and disturbing enough for say, like you and I sitting here saying, okay, what are the implications of this on how society uses ai? But we also have to consider that we're maybe a generation away, if not a shorter amount of time from. AI being the mediating layer between you and any information in the world, kids will default to assuming AI is correct no matter what.

[00:30:41] Because why wouldn't it be? It will be the only thing they know how to use and know how to experience. So this has even bigger downstream effects too, for future generations. 

[00:30:50] Paul Roetzer: Well, yeah, and I mean, Grok also, I say I also released companions, so they now have a female and a male companion that is [00:31:00] designed to do and say whatever you would like it to say.

[00:31:04] and you can connect your own dots here if, you know, they're being positioned as companions, they've released them like they're out in the wild. So, yeah. I I'm with you, Mike. Like, I honestly think it's within the next five years. Like Yeah. the upcoming gen, I mean, my kids are 12 and 13. Like I think about this every single day about like what AI tools are they gonna have access to.

[00:31:25] how do you, I don't wanna say control not the right word. How do you monitor and, and teach them? To, to not be, become relying on a single viewpoint from a single lab, right. Made up of a few thousand people who are making these decisions for all of us. 

[00:31:39] Mike Kaput: And just as one final note here, you're already seeing a very mild version of this.

[00:31:44] Anytime you're on x now people will immediately respond to a claim saying, at tagging rock and saying Rock, is this true? And it's like, okay, we're already seeing this. People are already taking that as gospel truth. If it says it is or isn't reality, 

[00:31:59] Paul Roetzer: it's [00:32:00] wild, man. I, this is like, yeah, yeah. Again, I know it's supposed to be rapid fire, so I move on.

[00:32:07] Meta Superintelligence Labs

[00:32:07] Mike Kaput: Alright, next up, mark Zuckerberg has unveiled meta super intelligence labs. This is a bold reorg that is aimed at putting meta at the forefront of the AI arms race. This new lab is led by Alexander Wang, formerly founder and CEO at scale ai. It is staffed with a wave of elite recruits from OpenAI, Google DeepMind, and Anthropic.

[00:32:31] All of whom have been apparently lured in with eye watering pay packages, some reportedly north of 200 million plus dollars. And meta is basically pitching this top talent on the idea that they have basically unlimited compute deep product reach and a singular goal, which is to develop personal super intelligence.

[00:32:52] So this is AI agents smarter than humans, embedded in everything from messaging to wearables. Now, on top [00:33:00] of this, internally there are reports the lab has already started debating whether or not they should drop open, meta's open source strategy in favor of closed models, though Zuckerberg has not yet signed off on that.

[00:33:14] So Paul Zuckerberg is making some huge moves. Here he is poaching AI talent left and right, completely upending compensation for top AI talent. Where does this go next? 

[00:33:26] Paul Roetzer: So this is another one. I was in the middle of finalizing the AI timeline course as part of the AI fundamental series. And in that course I tell the story of, you know, obviously the road to AGI and beyond, this is the beyond part.

[00:33:38] Like I've always, but again, I didn't think people were really ready for the super intelligence conversation. And now all of a sudden in July, this is like all any of the major labs are talking about. It's like we've just moved past AGI and we just kind of assume we're there. We're gonna be there very soon.

[00:33:53] And all the labs are not talking about super intelligence. Even Sam Altman recently said like, that's the goal of their lab now is super intelligence. They haven't [00:34:00] changed their mission. AGI i's still the mission, but they called themselves a superintelligence lab. So everybody has just sort of moved past this.

[00:34:07] The, the numbers are crazy. it's hard to put these kind of numbers in context, but the one that I thought was relevant is, NBA MVP, Shai Gilgeous-Alexander, the Oklahoma City Thunder. Just agreed at the beginning of July to a record setting four year deal for $285 million. AI researchers are now getting paid more top AI researchers getting paid more than the highest paid professional athletes.

[00:34:35] So I mean, everybody would look at the 280 million, be like, oh my God, these athletes are paid so much money. That's a ridiculous contract. And then here we got 10 of these people already that we know of that, meta has hired paying 300 million or more. And I saw a report this morning that they offered a billion to, to somebody.

[00:34:52] so it's not, it hasn't been verified yet. I, I'll, I won't say who it was, but, yeah, so [00:35:00] it's crazy. And then, I mean, we already saw, I mean we saw Noam Shazi got basically acquihire for 2.5 billion by Google. We had the, Alexander Wang from scale ai 14 billion, basically 15 billion to from by meta.

[00:35:12] So these numbers are just getting insane. 

[00:35:16] Windsurf Drama

[00:35:16] Mike Kaput: Yeah, kind of speaking along those lines in this next topic, OpenAI, as we've reported in past weeks, was really close to acquiring Windsurf, which is a fast rising AI coating startup in a $3 billion deal. But now that deal is off and windsurf is basically now the property of two different companies.

[00:35:34] So Windsurf backed out of the OpenAI deal after raising concerns that OpenAI's agreement with Microsoft might lead to Microsoft gaining access to different pieces of windsurf tech. So after this, Google swoops in, they hire CEO, Varun Mohan and several top engineers and pay $2.4 billion for a non-exclusive license to [00:36:00] windsurf technology.

[00:36:02] And then another twist happens. So its leadership is gutted. Windsurf basically seems a bit adrift until cognition steps in. So over a frantic weekend. The company, which is behind the Devon AI coding agent that we've talked about in the past. They struck a deal to buy the rest of Windsor. It's ip, it's brand, it's remaining staff.

[00:36:23] So Paul, this one's been a roller coaster since kind of day one. What, what happened here? Why are companies just paying for pieces of this company or people in it and not just buying it? 

[00:36:34] Paul Roetzer: They can't buy it, because of government oversight. And, you know, the, this is why all these are acquihire deals.

[00:36:42] They're just easier to get through. and then you just leave a shelf of a company behind, basically all, all I can think about is,   I never watched Silicon Valley, which I probably should. the show. Oh yeah. Yeah. I think it was like six seasons. Someone has to be working on a reboot of that. Like this is, [00:37:00] I would watch this stuff.

[00:37:01] I, this doesn't, it's starting to just not even feel real like these acquihires and stealing talent away and you got five labs fighting against each other. And the crazy part is like all these research and engineers at these five different labs all hang out at the same parties and share internal secrets all the time.

[00:37:18] It's like just made for TV kind of stuff, so. 

[00:37:22] Mike Kaput: Well, I'll tell you, Grok becoming Mecha Hitler, may as well be a Silicon Valley. 

[00:37:27] Paul Roetzer: Yeah. It's like episode 

[00:37:28] Kimi 2

[00:37:28] Mike Kaput: one, right? Yeah, yeah, yeah. So I full, fully agree. They need to reboot that. All right. Next up Moonshot, which is an Alibaba backed startup, has released Kimi 2, which is a low cost open source language model that is drawing global attention.

[00:37:45] Because it reportedly outperforms,  Claude and ChatGPT on major coding benchmarks and does so at a fraction of the cost. So Kimi two charges just 15 cents per million input tokens and [00:38:00] $2 and 50 cents for output, which is apparently a hundred times cheaper than Claw and still far below OpenAI rates.

[00:38:08] So it's also open source, which lets developers freely build on it and early reviews praise it for strong real world coding use, though it is still developing some integration with other systems. Now, Paul, first we got deep seek. Now Kimi 2. It's been incredible seeing just how much power and performance for price we are now able to get using open source models.

[00:38:32] And it also seems like just another blow to Meta's open source strategy. I mean, these models out of China are just racing ahead while something like llama appears to be, at least for the moment, stuck in the mud. 

[00:38:45] Paul Roetzer: Yeah, I, I, again, this gets into like the macro level stuff. I mean, in essence, in essence, what's happening is if you take a snapshot in time and say, okay, the most powerful models in the world today are Gemini 2.5 Pro, I may be [00:39:00] Grok 4 in some capacity.

[00:39:01] The O three model from OpenAI, from a reasoning perspective, GPT five, likely, you know, maybe before the end of July, sounds like probably, at least certainly this summer, those state-of-the-art models will be basically free, 12 months from now. So every 12 months or so, the cost of using these models drops like 10 x.

[00:39:27] And so you have to almost look out when you're planning for your life, when you're planning for society, when you're planning for business. And, and assume that intelligence is basically gonna be free. That, that super intelligence, like things that are smarter than the smartest humans at basically every cognitive task is pretty much gonna be free to all of us.

[00:39:46] because the competition between these labs almost dictates that that's where this goes. They're gonna open source. The previous generation model, that previous generation of models is state-of-the-art today. And so we're on a, there, there [00:40:00] appears to be no near term stop to that trajectory. the pre-training scaling laws followed by the post-training scaling laws, followed by the test time compute laws, which is give it more time to think and it gets smarter.

[00:40:12] Those three scaling laws working together and then maybe one more scaling law that we, we don't know yet kind of the next breakthrough that needs to happen. They are kind of dictating that. S first general intelligence, but then super intelligence will basically be too cheap to meter. Like, it'll, it'll just be readily available.

[00:40:31] And then the distribution that all these companies have kind of dictates, well, who actually controls all the users. So, I don't know. I mean, it, it's nuts. Like I get, I don't even know, I use the word wild a lot. We could probably like, query our transcripts. Someone's probably gonna go do this, like put all our trans, but I say the word wild, wild a lot.

[00:40:50] I don't know how else to describe it. Like, it, it really is hard. I, I, it's interesting. I, so last night I, I had my daughter's 13. I asked her if [00:41:00] she watched Interstellar with me. Like it's one of my favorite movies ever. And I've watched Interstellar probably, I don't know, six or seven times now.

[00:41:05] And so we sat there for like three hours watching this movie, and I get done and I'm laying in bed. I'm like, and I still don't really understand it. Like it's the, the ending of that movie is so mind bending and it's like the human mind's almost not even supposed to kind of comprehend the basic concept of intertel.

[00:41:23] And I kind of feel that way about the intelligence explosion. Like it's, it's almost just too hard to even comprehend what is almost inevitably going to happen in the next three to five years. And this fits into that realm. It's like, what, what, what happens if super intelligence is a thing in a few years?

[00:41:41] And it's just almost to everybody. Like, I don't know. I don't even know how you're supposed to comprehend that stuff. 

[00:41:47] Mike Kaput: Yeah. I also think it's extremely difficult to keep perspective on just how fast things change. We say it all the time, we feel it, but if you really sat down and looked at the milestones over the last three years since ChatGPT [00:42:00] came out, it would be breathtaking.

[00:42:02] Paul Roetzer: Yeah. It it, that's a good point. Like GPT-4 was March of 2023. We, we are literally talking about two years Yeah. Of innovation and, and look where we are like. Is I don't know. I mean, I had that slide in in one of my courses. you know what if the future is, you know, 10 x, 20 x, a hundred x innovation happening every two years, like, because right now that's the trajectory we're on and I don't even know how to explain that to people.

[00:42:31] Like, I can't comprehend it myself. 

[00:42:35] AI Browsers

[00:42:35] Mike Kaput: Alright, next up, OpenAI is getting ready to launch its own AI powered web browser according to Reuters. the upcoming browser is designed to integrate tightly with ChatGPT and some of OpenAI's agentic capabilities. Instead of pointing you to websites, it is looking to maybe keep interactions within a built-in chat interface, which is a direct threat to Google search and ad ecosystem.

[00:42:59] [00:43:00] It will reportedly allow agents to take actions for users, do things like we've already seen, like booking restaurants, filling out forms, managing inboxes, all while capturing the kind of behavioral data that Google has long used to dominate ad target. Now we're actually already starting to see what an AI powered browser could look like because Perplexity also just launched its own AI powered browser called Comet.

[00:43:24] This is only available right now to Perplexity Max users, which is their $200 a month offering, or by private invitation. And the Verge actually got their hands on this and published a pretty extensive review. And they said it doesn't just help you browse. It actively takes actions for you. It can summarize pages, manage your tabs, send emails, unsubscribe from newsletters, even accept LinkedIn invites all on your behalf.

[00:43:51] You can also say, take control of my browser, and comment will go into agent mode. It can place Amazon orders, book restaurant reservations, [00:44:00] and under the hood, it replaces traditional search results with Perplexities answer engine. Now the Verge says right now it can be sometimes slower than doing things yourself.

[00:44:10] It can occasionally stumble on more complex tasks, but it is an interesting early example of what an AGI agentic AI browser can actually do. So while it's still really early here, Paul, it does feel like we're on the edge of maybe something big here. I mean, the way in which the Verge described this, despite all the flaws here, sounded like an AI browser could be a real paradigm shift from the way we typically use browsers today.

[00:44:37] Where, where do you see this going? How do you think about this? 

[00:44:40] Paul Roetzer: Yeah, so there's a, a couple of quick notes here. So Perplexity bought os.ai from Dharmesh Shaw, who's the co-founder and CEO of HubSpot. this was all, this was last week. And Arvin, the CEO of Perplexity said there's a reason we purchased os.ai from Dharmesh.

[00:44:57] The roadmap is to make comet feel like your [00:45:00] own mini customized computer within your existing computer phone and the computer running across client and server with the ability to run local models too. But the tweet that I thought actually kind of explained this best 'cause I was like, I was sort of struggling to understand like what exactly was going on here.

[00:45:17] Like, I knew the general concept, but I couldn't like deeply explain this one yet. And so there was a tweet I saw. this is from July 17th. We'll put it in the show notes from Michael Mcno, who's a VC and founder partner at Lightspeed Ventures. so he said the browser sees everything. This is the reason we're getting so many new AI first browsers from the browser company Perplexity, and soon OpenAI, which we got.

[00:45:40] so they can see data that they increasingly cannot scrape. AI feeds on data. It gets the data by automatically scraping the web, but scraping is no longer free. CDNs like CloudFare Clare Cloud Flare are making scraping harder by blocking it by default, and others will soon follow follow. Startups like tobit [00:46:00] are empowering tons of large publishers to charge for being scraped, building a new open web economy.

[00:46:05] But consumers want ai, we can't get enough of it. As AI answers increasingly eat traditional web search, AI will be doing much more browsing on behalf of us humans. this creates a paradox. Consumer behavior is shifting to ai, but AI is running out of fuel to meet the demand. So what happens? AI companies build browsers as humans consume content with these browsers.

[00:46:28] The AI company can see, quote unquote, see that, data that is increasingly being blocked or monetized the company. Interesting. the most interesting thing about this strategy is that AI companies don't even need meaningful market share or customer UBI ubiquity for this strategy to work. They just need a large enough slice of all browsing to get a taste of most of the web's data.

[00:46:51] It's a whole new business model for the web and the beginning of a new browser war. I was like, oh shit, that's pretty smart. Like that makes more sense now. So [00:47:00] yeah, I don't know that browser wars, talent wars. Great, great, great tv, right? 

[00:47:05] Microsoft’s Layoffs and AI

[00:47:05] Mike Kaput: Yeah, no kidding. Good, good for consumers too, at least in, in the short term.

[00:47:10] All right, so next up, Microsoft is laying off thousands, but urging remaining staff to upskill in ai. After cutting 9,000 jobs recently, the company told sales employees to embrace AI tools like copilot to boost productivity and close more deals, quote, invest in your own AI skilling. One executive advised while others reportedly began factoring AI usage into performance reviews.

[00:47:36] Now behind these cuts as Microsoft soaring infrastructure costs, so they expect to spend 80 billion this year on data centers and chips. Much of IT supporting AI services for customers and OpenAI. And to offset that, Microsoft claims that AI is already paying off. It saved, they said over $500 million in call center costs last year, and sales reps using copilot are [00:48:00] reportedly generating 9% more revenue.

[00:48:03] Internally, the company is consolidating roles and sales units. They're even tracking how much code engineers generate using ai. However, on top of all this, the executives at Microsoft insist that these layoffs are not solely driven by ai. So Paul, it's good to see, they're quick to say that wasn't the case, but it is pretty clear that AI at every level is driving the strategy and decisions here.

[00:48:28] Is this another example of a company saying the quiet part out loud? 

[00:48:32] Paul Roetzer: Yeah. This tracks with everything we're seeing. I mean, I don't even know that's quiet anymore. I feel like it, it just is what it is. There's, there's no debating this. We've, you know, in one of the courses I was creating, I think I highlighted like seven or eight, CEO statements just from June to July of this year.

[00:48:51] Yeah. Where they, they literally said like, we're just gonna need fewer people. We're, we're starting the process now, get AI literate or get out is basically what the [00:49:00] CEOs are saying.   So, yeah, I mean this, this is a hundred percent the direction. I don't, I don't know that it's debatable. I think I would heed the advice, like, you, you have to, you have to push yourself to be one of the people in the room who understands this stuff and can apply it to what you do every day.

[00:49:20] Like it is, I don't know that you have job stability, you know, one to three years out, regardless of your profession or industry. If you don't do that, and I mean, I assume people listening to this podcast are the people in the room who are figuring this out. So good on you. Keep going. You're, you're gonna be in a really good position moving forward.

[00:49:42] if you have friends, family, coworkers, you care about, you have to get them to start figuring this stuff out. Like their, their, their jobs are gonna depend on it, very quickly. If you're in the tech space, it's tomorrow. If you're in legal healthcare. Financial [00:50:00] services, like maybe manufacturing may, maybe you got a little bit of time, but everybody's gonna figure this out.

[00:50:06] Publicly traded, private equity backed, VC backed is now anybody else? It's, it's coming. So do, do what you can to try and pull, pull people along with you if you're listening to this show. 

[00:50:17] Mike Kaput: Yeah. And I know that we are obviously very biased here towards AI literacy, but I would argue too, to really strongly encourage people to reframe this and not see AI as one other skill.

[00:50:31] You need to learn it is the thing you need to figure out. Like I would be ruthlessly cutting any of your other professional development priorities as if you are not up to speed yet on ai. 

[00:50:42] Paul Roetzer: Yeah, I mean, I think of it, you're right, Mike. I do think of it as like the underlying operating system to our careers.

[00:50:49] Like literally everything you do is going to have to be built on top of AI and Exactly. People who don't figure that out are, are just gonna have a really difficult [00:51:00] time maintaining career stability. And I think the people who do are gonna have like tremendous near term career potential, like earning more than your peers.

[00:51:11] Creating more value within companies than your peers. Like you're just gonna be able to do more and be more valuable and, and that's just a better place to be. Like I would, I would rather just be, I mean, there's still no guarantees, but like I would way rather be out on the frontier of this stuff, figuring it out and helping other people than sitting around and getting run over by it.

[00:51:32] Mike Kaput: And just one final thing there, just from personal experience, like focus on the fun. I realize there's a ton of like doom and gloom. There's a lot to be afraid of. There's a lot of uncertainty, but also like, I just get so excited more so than ever to do certain work I do because it's so much more new and exciting with ai.

[00:51:49] Like, yeah, this is great. Like I would actually be upset if I had to do my job the way I used to do it sometimes. 

[00:51:56] Paul Roetzer: Now, well there's so many, like I for, for the AI course I was building, I created [00:52:00] ada, like AI teaching assistant. Oh my God, 

[00:52:01] Mike Kaput: that's a great 

[00:52:02] Paul Roetzer: example. Is that, yeah. Well, yeah. There's so many days where I would use it and be like.

[00:52:05] I can't believe this is possible. Like I would do so I would, I would literally like have to message Mike or message, he's like, I just have to share this with somebody. Like look at what it just did. Like there is, there's like that almost surprise and wonder that every day you're just like, I might discover some new thing this thing does.

[00:52:20] And it just makes work more interesting and like the future more fun to think about. 

[00:52:27] More Meta Updates

[00:52:27] Mike Kaput: Yeah, a hundred percent. Alright, next up some more meta updates. So we'll go through a couple things here and then Paul kind of get your take on this. So Meta just made kind of two other big moves where, it kind of reveals where they think AI is headed.

[00:52:41] So first into your glasses and next into your inbox. So first, meta invested $3.5 billion in Essilor Luxottica, which is the world's largest eyewear maker. The company behind RayBan and Oakley Meta now owns just under 3% of the company. That's building on [00:53:00] Meta's RDD partnership with RayBan because they're building their AI powered meta RayBan glasses at the same time.

[00:53:07] Internal documents show meta is training custom chatbots to behave more like companions. Bots that are built on its AI studio platform can now send unprompted personalized follow-up messages referencing past conversation. The goal is to increase user engagement and retention by making AI feel emotionally present.

[00:53:29] Now Paul, I found both these pretty interesting, obviously very clear metas all in on AI wearables. Also, the stuff with the AI autonomously messaging you to drive engagement and retention kind of makes me shutter because I can like imagine with clarity what dystopian future this becomes in the wrong hands.

[00:53:50] Paul Roetzer: Yeah. So, just give a couple quick comments.   I am fascinated by glasses. I will likely at some point try them myself. [00:54:00] I am a hundred percent disturbed by a future society where people, lots of people are wearing these things and you never have a clue who's recording what. And I think we're probably already like on the edge of that.

[00:54:11] I, I really worry about that. And I don't think society is anywhere close to being prepared for that. And the ramifications of that, in terms of the proactive outreach from the AI get used to it like that is a hundred percent going to be. So again, we go back to the competition between these five AI labs.

[00:54:31] How do you control the user? Well, you have to drive stickiness and engagement. You have to drive, you know, daily active users, hourly active use, like there, these are the metrics they're gonna look at. And the only way to do that is to prompt someone. Now there is value, like say I've talked to about a health condition and I'm trying to figure out what's going on with me.

[00:54:52] Then like three weeks later, it'll just check in on me, say, Hey, how's that thing going like that we talked about three weeks ago? And that's [00:55:00] gonna feel incredible. Like, you're gonna be like, damn, that's, that's really helpful. And you say, I'm actually like, I'm kind of still having this symptom. And now all of a sudden you're having this like super proactive conversation with your AI about something that's very valuable to you.

[00:55:13] And so it just seems inevitable that, that we get there very quickly because there's not really any technical limitations to doing it. There's not like some breakthroughs needed in AI to do this. It's kind of like a memory thing and like a context window thing. But those are trackable and there's already some solutions in place, but they're all gonna do this.

[00:55:34] Like Google will probably have to be slower because people take a closer eye on what they're doing. Opening eye is definitely gonna do this very soon. Xai will do this a hundred percent Meta will do it. Anthropic I would guess would be last on the list to do it, but who knows? 

[00:55:53] OpenAI Updates

[00:55:53] Mike Kaput: All right. Next up we have some OpenAI updates in addition to the news we've already talked about.

[00:55:59] So first [00:56:00] up, Sam Altman actually posted about the fact that OpenAI's language model has achieved gold medal performance on the 2025 International Math Olympiad, which is the world's premier math competition for pre university students, which is a huge deal given that it's a huge benchmark and milestone.

[00:56:18] And at the same time as he announced this on x, Sam Altman also teased GPT five, which is coming soon, and he said quote, we are releasing GPT five soon. But I want to set accurate expectations. This is an experimental model that incorporates new research techniques we will use in future models. We think you'll love GPT five, but we don't plan to release a model with IMO International Math Olympiad gold level of capability for many months.

[00:56:48] At the same time, OpenAI are. Locked with Microsoft in a high stake standoff over the definition of AGI. They have this clause in their contract that states that if OpenAI's board declares [00:57:00] it's reached AGI, Microsoft's access to future models would be cut off. Now, there's a new wrinkle to this because there's, an unreleased internal paper at OpenAI called Five Levels of General AI Capabilities.

[00:57:15] This is causing a stir because it outlines a framework for assessing AI systems by capability rather than a binary. Just yes or no. Is it AGI that complicates when and how OpenAI could claim its reached AGI, and what would that would trigger contractually? So Microsoft has poured 13 billion into OpenAI.

[00:57:35] They're trying to get this clause removed entirely. OpenAI is kind of saying, look, this is not really a formal research paper. It's just a paper internally that we are working on. Some OpenAI insiders say it is fairly close to AGI, so the whole paper muddies the art murky water around AGI in this partnership.

[00:57:57] Next OpenAI is teaming up with the [00:58:00] American Federation of Teachers to train 400,000 US educators to shape how AI is used in schools. They're funding a new national academy for AI instruction. This is a national training hub, offering workshops, hands-on courses, tech support to help teachers integrate AI into the classroom in ways that enhance, not replace their work.

[00:58:21] There's a flagship facility plan to open in New York City. More locations are planned nationwide by 2030. OpenAI is contributing $10 million over five years, including direct funding API credits and engineering support. Okay. At the same time, OpenAI has introduced a flexible credit system for ChatGPT team and enterprise plans that gives organizations more control over how they access advanced features like deep research, O three Pro GPT-4 0.5, and Image Generation.

[00:58:52] And last but not least, an open weight AI model from OpenAI was set to launch soon, but now it's on hold. [00:59:00] Sam Altman posted that there was now going to be a delay citing the need for more safety testing and deeper review of high risk areas. There is no new release date yet. So Paul, bunch of stuff here.

[00:59:11] What kind of jumped out to you as worth paying attention to? It seems like we have confirmation of GPT five coming soon straight from Altman. I also found the unreleased paper around AGI pretty interesting. 

[00:59:22] Paul Roetzer: So the International Math Olympiad is intriguing. There may be more coming out today, you know, after we're done recording this, but.

[00:59:31] The word over the weekend was, so No Brown, who we've talked about many times on the show, met a researcher, went to OpenAI, at the forefront of reasoning, very well-known AI researcher. He tweeted this at like 2:00 AM on like Saturday or something like that. I saw, I remember I saw it like as soon as I woke up and I was like, that's a weird thing to tweet in the middle of the night.

[00:59:52] Like why would they have done that? And so the rumor is that Google also achieved gold [01:00:00] medal in this, but that the organization asked the labs not to disclose that they had got gold medals for seven days following because they wanted to not steal the attention from the kids who had achieved incredible things as humans.

[01:00:17] And OpenAI at least no, him tweeted that. Like no one had told him that. No one had told OpenAI, that kind of thing. So the word is Google also did this, but they were following the rules and not. Disclosing it. so I know we'll, we'll, we'll wait and see. Google. As of the moment of recording this had had not confirmed one way or the other, but it was a, it seemed like a bit of a, I don't know.

[01:00:43] There's definitely some controversy around it, so we'll wait and, and see what happened. In terms of gpt five, the rumor here is it was supposed to be what we would call a unified model, meaning it was a single model that does reasoning and chat and everything else, which is what Gemini 2.5 Pro is, [01:01:00] is a unified model that's kind of built all in.

[01:01:02] Now, the discussion is that gpt five may be, might be a router model, meaning it's not a single model, it's still a chat plus a reasoning model. But you would no longer have to choose which model you're going to use. The model would choose itself. So this is what we've been, you know, preaching for the last two years is like, why when I go into CH PTs, do I have eight models to choose from?

[01:01:25] So what they're saying is GPD five might just be a bigger, smarter model. It also would route you to a different model if it's needed, such as a reasoning model. But we don't know. They, they, they haven't said anything other than it's, it's probably coming soon. And then the paper is interesting because I came across this in one of the courses I was building.

[01:01:45] I, I saw this and then I went back to the stage of ai. So I think what it's actually related to is in, around this time in 2024, Bloomberg released an article where they shared what was believed to be an internal document [01:02:00] about the stages of artificial intelligence. Hmm. And in that document, OpenAI, this is what they were, OpenAI was using as their internal guidepoint was, chatbots were first, reasoners were second.

[01:02:10] AI agents was level three innovators was level four, which is they could create new things. They can discover new, mathematical formulas. They can develop new science. And then, which I guess the International Math Olympiad stuff starts becoming very relevant on the innovator spectr  And then organizations, which is autonomously run organizations, that was the five levels.

[01:02:31] And so the article that recently came out was implying that, that those stages were part of a paper at OpenAI that they did not release because of their concerns related to the contract with Microsoft. So it's like, oh, that would explain why those stages got leaked. And then at some point, OpenAI last year kind of acknowledged that yes, these are the stages we look at internally, but they never formally released them themselves, [01:03:00] to my knowledge.

[01:03:00] And this would explain why it was never actually released from OpenAI as it was part of a bigger paper that got yanked, basically. So just, I don't know ing historical context here. 

[01:03:13] Google Updates

[01:03:13] Mike Kaput: All right. Next up we're going to do some more, updates around Google that came out the past couple weeks. So first up, they are rolling out a major AI expansion for schools with the launch of Gemini for education.

[01:03:25] This is a version of Gemini tailored specifically for students and teachers. It's built on Gemini 2.5 Pro and brings premium AI capabilities into Google Workspace for education, offering higher usage limits, enterprise grade security, and full admin control. Next, Google is bringing its custom AI assistance gems directly into the side panel of Gmail docs, sheets, slides, and drive.

[01:03:49] Until now, they were living in the standalone Gemini app. Over the next few weeks though, they'll show up, right where we all work in workspace apps, right in the side panel of those [01:04:00] apps. So that means you could use a gem to draft emails, generate slide content, analyze spreadsheets, what have you, without switching apps.

[01:04:09] Google has also made some more workspace updates. First VO three. Google's advanced video generation model is now integrated into Google Vids, ai, voiceovers, and vids just got easier too. You can now update all voiceovers in a project with a single click. After editing your script, they've added a collection of new templates to slides to speed up presentation building, and the Gemini app is becoming far more connected.

[01:04:34] It taps into Gmail, drive, calendar, keep, and tasks. So you can ask Gemini to summarize unread emails, pull meeting notes, surfaces specific doc directly from within a chat. And for deep research, Gemini can now combine your drive files with public data to produce insights, summaries, and trend analysis. I found this update around gems, particularly noteworthy.

[01:04:59] Paul, I know we've been [01:05:00] relying more and more on gems in some of our work, and it seems like having these embedded in workspace apps would be pretty useful. 

[01:05:07] Paul Roetzer: Yeah, it could be enormous. I, I, again, I could think the people are a little bit more familiar with GPTs, custom GPTs with ChatGPT. 'cause yeah, they just have a bigger user base.

[01:05:17] But I mentioned Ada, the teaching assistant I created to help with the development of my courses. And I trained it to ask questions, challenge my thinking, strengthen ideas, brainstorm concepts, conduct research, recommend actions, and then to always be solving for the customer. So I trained it on a pretty significant system prompt.

[01:05:34] but what I did is I built a custom GPT version and a gem from, from Google, on the exact same system, prompt and exact same knowledge base. And the first few courses I was creating, I would like talk with both of them. I was kind of saying, okay, here's my outline. What do you think? How would you improve it?

[01:05:52] And so I was going back and forth and by about the third course, I just stopped using the custom GPT. The gem was, yeah, far superior to the [01:06:00] custom GPT, at least in this instance. and so. Yeah, I wouldn't sleep on gems both as like just an experimental thing, like go play around with them if you haven't built one, because the Gemini 2.5 pro model they're built on is incredible.

[01:06:15] and their potential application in the workspace now is, requires huge change management and education and training and like integration into it. And I think that's where most enterprises have this like, huge opportunity ahead is like, just solve the obvious things better than everybody else. Like you don't have to figure everything out about AI and solve super intelligent, all these things.

[01:06:37] Just take notebook, LM and deep research and custom GPTs or gems and like teach your team how to prompt well, like there's a massive cane to be had in every single company team department by just doing those few things really well. 

[01:06:52] Mike Kaput: Yeah, for sure. And we've been working with, you know, at least one company on helping them out with some GPTs and even just a very [01:07:00] basic pilot project, not only created huge transformative wins for them with nothing that I would consider rocket science, more just education training and ongoing usage and monitoring.

[01:07:11] And also that became a huge win internally that has now inspired other teams and executives to move forward further with ai. So don't, yeah, don't underrate not only how a small project can do big wins, but also what it can inspire other people 

[01:07:26] Paul Roetzer: to do. Yeah. Don't, don't overcomplicate this and then share the wins.

[01:07:29] This is something we've gotten much better about internally at Smarter X, is like, we have a AI playground where people post like, cool things they're doing every day. and it does it just like, oh man, that's a great idea. Like, I'm gonna use that in customer success, or I'm gonna take that and apply that to marketing.

[01:07:44] It's it, you know. It doesn't, you don't, again, don't overcomplicate it. Just do the obvious things well and have a change management plan in place that like trickles that down to everyone in the team department organization. 

[01:07:57] Apple Might Use Anthropic or OpenAI for Siri

[01:07:57] Mike Kaput: Alright, next up, apple is considering a [01:08:00] major shift in strategy. They're considering outsourcing series core AI to Anthropic Cloud or openAI's ChatGPT.

[01:08:07] this would make a dramatic reversal of its longstanding commitment to building its own language models. According to Bloomberg, apple has asked both companies to train versions of their models that could run on Apple's own cloud infrastructure as a test if it moves forward, that would basically be an open admission that Apple's homegrown foundation models haven't kept pace with competitors and that Siri's lagging performance needs a near term fix.

[01:08:33] The potential switch comes after a messy shakeup. Inside Apple's ai org. Leadership of Siri has been handed from ai Chief John Gandia. To the team behind Vision Pro and internal tests reportedly found Anthropics. Claude outperformed Apple's models leading to serious talks about licensing it. So Paul, nothing has been decided here, but this just seemed like more trouble for Apple.

[01:08:57] They just lost a key engineer working on [01:09:00] these foundation models to meta for about $200 million. I also found it interesting in one of the reports from Bloomberg, they said that quote, apple is known in many cases to pay its engineers half or even less than what they can get on the open market. So I don't know what's going on here, but this doesn't seem like it bodes well.

[01:09:19] Paul Roetzer: Yeah, I mean, if there was another innovator's dilemma, you could do a chapter on Apple and ai. yeah, I don't know how they compete honestly. Like, I think that, I don't even know that Acqua hiring or even straight up acquisition works be because it's so counter-cultural to, to Apple.   They're not gonna pay $300 million for top AI researchers.

[01:09:41] Like, it would just fundamentally change the way Apple does things. And it is not a company that pivots their culture. So I don't know, like I think that licensing probably in the end is the play for them. I, I, again, like you could go acquihire minstrel or, or, or a straight up buy [01:10:00] perplexity or maybe even like go after Anthropic, which, you know, seems probably the most aligned from a value perspective.

[01:10:07] Paul Roetzer: You could do that, but why would those researchers and engineers stay at Apple? Like, I don't know. Like, I I, they're just not built to be like a wartime aggressive AI lab. And so I was on the acquisition train. I mean, honestly, up until I started talking right now, like, as I'm like thinking out loud, it is not gonna work.

[01:10:29] Like they, they wouldn't be able to keep people there. So. I love Apple. Like I'm, everything we do is Apple. I personally, every device is Apple. I've been an Apple fanboy for, you know, 20 years. I love the company. I don't think they can compete as an AI lab. So I don't know. I just, I just want Siri to work.

[01:10:52] Like I just, I want, I don't care whose tech they build it on, just make the thing work, so. Right. 

[01:10:59] AI Product and Funding Updates

[01:10:59] Mike Kaput: Alright, [01:11:00] for our final topic, we're going to run through some final AI product and funding updates. Paul, if you have anything to chime in with here, feel free. Otherwise I'll just kinda run through these as we wrap up.

[01:11:09] So first step, Grammarly is acquiring Superhuman, which is an AI powered email startup. This is part of a broader push to become an AI powered productivity platform. So superhuman, which was last valued at 825 million, built its reputation on speed and design, and helps users churn through emails at Lightning pace.

[01:11:29] The company has found itself squeezed as Google and Microsoft have layered AI into their own email tools. Their annual revenue now sits around 35 million. Grammarly, meanwhile is flush with a billion dollars in funding from General Catalyst, and its rebranding itself beyond grammar correction and moving deeper into productivity.

[01:11:48] So their vision is to use Grammarly AI agents, but superhuman inbox to build a smarter communication hub that understands your email schedule and workflows. Next Google Notebook, [01:12:00] LM got a little bit of an upgrade. The AI powered research Tool now feature that now includes featured notebooks, which are curated collections of high quality content from trusted sources like The Economist and The Atlantic.

[01:12:12] And each notebook combines original source material with all of Notebook lms great features like the ability to ask questions, trace citations, create Visual nine maps, and listen to AI generated audio overviews. And last but not least, X OpenAI, CTO. Mira Tis Startup Thinking Machines Lab has confirmed it, has raised a massive $2 billion funding round led by Andreessen Horowitz and joined by companies like Nvidia.

[01:12:38] The Startup's mission said Mirati in a post on X is to build collaborative general intelligence. Their first product they say is coming in the next few months and will include open source components designed to support researchers and startups building custom models. Paul, that is a wrap on a busy, busy week, two weeks in ai.[01:13:00] 

[01:13:00] Paul Roetzer: Yeah, and like we said, there's literally dozens of things we didn't get to. so check out the, this week in AI newsletter. we, we, we are back now. I don't, I don't think we have any disruption to the weeklies moving forward that I know of. So, as of the moment we are, we, we back on our regular Tuesday schedule and I still have two more course series to finalize.

[01:13:22] So. I'm sort of still, you know, locked in the, in the lab building courses for the upcoming launch. And stay tuned for the AI Academy by Smart Rx, relaunch news. It it'll be, coming out, we'll probably talk about the podcast next week. but if you are an AI Academy member or thinking about being one, co coming very soon.

[01:13:44] Alright, thanks Mike. Thanks Boff. 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. [01:14:00] Downloaded AI blueprints, attended virtual and in-person events, taken online AI courses and earned professional certificates from our AI Academy, and engaged in the Marketing AI Institute Slack community.

[01:14:11] Until next time, stay curious and explore ai.
Claire Prudhomme

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

Recent Posts

[The AI Show Episode 158]: ChatGPT Agent, Grok 4, Meta Superintelligence Labs, Windsurf Drama, Kimi K2 & AI Browsers from OpenAI and Perplexity

Claire Prudhomme | July 22, 2025

AI salaries soar, Grok stirs controversy, and ChatGPT levels up. Episode 158 of The AI Show examines the latest AI power plays, drama, and breakthroughs.

Podcasts

How to Combine AI + Automation for Maximum Impact with Brian Brinkman [MAICON 2025 Speaker Series]

Cathy McPhillips | July 17, 2025

In our ongoing speaker series, we’re spotlighting the remarkable AI leaders featured at MAICON 2025. During this upcoming session, Brian Brinkman will show attendees how to combine AI and automation for maximum business impact.

How to Activate AI-Assisted Writing with Robert Riggs [MAICON 2025 Speaker Series]

Cathy McPhillips | July 10, 2025

In our ongoing speaker series, we’re spotlighting the remarkable AI leaders featured at MAICON 2025. During this upcoming session, Robert Riggs will demonstrate how to harness AI tools to supercharge your writing workflow and boost productivity.