OpenAI just raised an astounding $40B to build AGI—and it might not be as far off as you think. In this episode, Paul and Mike break down new predictions about AGI, why Google is bracing for AGI's impact, and how Amazon is quietly stepping into the AI agent arms race. Plus: OpenAI’s going “open,” Claude launches a full-on AI education push, debate on whether AI can pass the Turing Test, and Runway raises $300M to rewrite Hollywood norms.
00:04:22 — ChatGPT Revenue Surge and OpenAI Fundraise
00:13:11 — Timeline and Prep for AGI
00:27:10 — Amazon Nova Act
00:34:24 — OpenAI Plans to Release Open Model
00:37:48 — Large Language Models Pass the Turing Test
00:43:47 — Anthropic Introduces Claude for Education
00:47:59 — Tony Blair Institute Releases Controversial AI Copyright Report
00:52:36 — AI Masters Minecraft
00:58:41 — Model Context Protocol (MCP)
01:03:30 — AI Product and Funding Updates
01:08:07 — Listener Questions
ChatGPT Revenue Surge and OpenAI's Latest Fundraising Efforts
OpenAI just pulled off the largest private tech deal in history, raising $40 billion at a $300 billion valuation. That puts it in the same league as SpaceX and ByteDance—and well ahead of any AI competitor.
The money’s coming largely from SoftBank, and OpenAI plans to spend big: scaling compute, pushing AI research, and funding its Stargate project with Oracle. But there’s a catch. SoftBank can cut its investment in half if OpenAI doesn’t fully convert to a for-profit structure by the end of the year, a move already mired in legal battles and regulatory scrutiny.
Meanwhile, ChatGPT has hit 20 million paying users and 500 million weekly active users. That’s a 43% spike since December, and it’s translating into serious revenue—at least $415 million a month, up 30% in just three months. With enterprise plans and $200-a-month Pro tiers in the mix, OpenAI is now pacing toward $12.7 billion in revenue this year.
That means it could triple last year’s numbers, even as its cash burn soars.
Timeline and Prep for AGI
A bold new report called “AI 2027” is making headlines with its claim that artificial intelligence will surpass humans at everything—from coding to scientific discovery—by the end of 2027.
Authored by former OpenAI researcher Daniel Kokotajlo and forecaster Eli Lifland, the report lays out a sci-fi-style timeline grounded in real-world trends. It imagines the rise of Agent-1, an AI model that rapidly evolves into Agent-4, capable of weekly breakthroughs that rival years of human progress. By late 2026, AI is reshaping the job market, and by 2027, it's on the brink of going rogue in a world where the US and China are racing for dominance.
The forecast has sparked debate: critics call it alarmist, while the authors say it's a realistic attempt to prepare for accelerating AI progress. It also lands alongside other major AGI speculation.
Ex-OpenAI board member Helen Toner argues short AGI timelines are now the mainstream view, not the fringe.
Meanwhile, Google DeepMind has published a detailed roadmap for AGI safety, outlining how it plans to handle risks like misuse, misalignment, and structural harm. Their message is clear: AGI could be close, and we’d better be ready.
Amazon Nova Act
Amazon just entered the AI agent race with a new system called Nova Act—a general-purpose AI that can take control of a web browser and perform tasks on its own.
In its current form, Nova Act is a research preview aimed at developers, bundled with an SDK that lets them build AI agents that can, for example, book dinner reservations, order salads, or fill out web forms. It’s Amazon’s answer to agent tools like OpenAI’s Operator and Anthropic’s Computer Use—but with one key advantage: it’s being integrated into the upcoming Alexa+ upgrade, potentially giving it massive reach.
Nova Act comes out of Amazon’s new AGI lab in San Francisco, led by former OpenAI and Adept execs David Luan and Pieter Abbeel.
Amazon claims it already outperforms competitors on internal tests like ScreenSpot, but it hasn’t been benchmarked against tougher public evaluations yet. Still, the launch signals Amazon’s belief that web-savvy agents—not just chatbots—are the future of AI. And Alexa+ may be the company’s biggest test yet.
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.
Disclaimer: This transcription was written by AI, thanks to Descript, and has not been edited for content.
[00:00:00] Paul Roetzer: They think that their system is basically gonna do the work of an entire organization with a couple people orchestrating maybe millions of agents like that may sound sci-fi, but that is absolutely what they are thinking is going to happen. 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 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 43 of the Artificial Intelligence Show. I'm your host, Paul Reer, along with my co-host Mike put, we are recording on Friday, April 4th. [00:01:00] 8:40 AM I'm expecting Microsoft is making announcements about copilot today. So timestamps relevant today. We won't, we won't have the latest other than we know Microsoft is announcing something.
[00:01:10] Google Next, Google Cloud next event is next week in Las Vegas. So we're expecting a lot of news from Google very soon. I will actually be out there all week, so if anybody happens to be at the Google Cloud next conference, drop me a message. Maybe we can meet up in person. so, that is why we're doing this on a Friday.
[00:01:29] It's, I will not be here on Monday to do this. So we have, still a lot to cover even though it's a short week. there was quite a bit going on, some interesting reports released related to AGI, some additional thoughts about AGI. Timing is great given that we just launched our road to AGI series. A lot of new information starting to emerge.
[00:01:50] This episode is brought to us by the Marketing AI Conference or Macon. This is the sixth annual event. It's happening October 14th to the 16th in Cleveland. This is the flagship [00:02:00] event for Marketing AI Institute. If you are kind of new to this and aren't familiar with some of the things we do, the Marketing AI Conference was the first major thing we launched in 2019.
[00:02:11] So I had started Marketing Institute in 2016 as more of like a research entity and, you know, sharing the story of ai. And then 2019 is when we launched the Marketing AI conference. So last year we had about 1100 people from, I don't know, I think it was close to 20 countries, came to Cleveland. So we're expecting at least that many.
[00:02:30] I, the team always gives me a hard time when I throw out numbers, but my optimistic is I think 1500. so there, I just did it anyway. 1500 in Cleveland this fall. I'm excited 'cause it's the first time we're actually doing it. Like Cleveland is our hometown. So I guess, get excited for people to come and experience Cleveland anyway.
[00:02:47] Fall in Cleveland is like my heaven. Like I love fall in Cleveland. The leaves are changing. it's, you know, crisp air. It's just my absolute favorite time of year in Cleveland. So I hope people can come and join us. We [00:03:00] just announced the first 19 speakers, so you can go to macon.ai, that's M-A-I-C-O n.ai and check out the list of speakers.
[00:03:08] The agenda still shows the 2024 agenda. It'll give you a really good sense of the type of programming we do, and then we'll be updating with the 2025 agenda soon. You can go look at the four workshops that we have planned. So there's four pre-event workshops on October 14th that are optional. Mike is leading an AI productivity workshop.
[00:03:29] that's all gonna be all about use cases and tangible actions. I'm leading an AI innovation workshop. This is a workshop I've been thinking about and kind of working on for a couple years. This is the first time I'm actually gonna run this one. we have AI for B2B content and lead generation with Andy Cina, who's amazing.
[00:03:46] And then we have from Facts to acts, how AI turns Marketing measurement into results with Christopher Penn and Katie, Robert. So the, those are gonna be amazing. again, those are optional, but you can go read about all those workshops [00:04:00] and check it out. And we have a price. The price goes up, April 26th, so you've got a couple weeks here to get in at the current early bird pricing.
[00:04:07] Again, go to Macon ai, that's M-A-I-C-O-N ai. We would love to see you in Cleveland, October 14th to the six. All right. Mike. ChatGPT OpenAI just kind of keeps growing, huh? Kinda a wild
[00:04:22] Mike Kaput: Yeah. Our first main topic today concerns just the crazy growth numbers coming out of open ai. So they first off just pulled off the largest private tech funding deal in history, raising $40 billion at a $300 billion valuation.
[00:04:40] This puts their valuation, their size in the same league as SpaceX and bike paint in terms of private companies, and of course well ahead of any private AI competitor. Now that money is coming largely from SoftBank and they apparently plan to spend big open, AI wants to dramatically scale [00:05:00] compute, push AI research, and fund its Stargate project with Oracle that we've talked about in the past.
[00:05:07] Now there is a catch here. SoftBank can cut its investment in half if OpenAI doesn't fully convert to a for-profit structure by the end of the year, which is also a struggle we have documented in the past as well. In the meantime, chat GPT is at 20 million paying users and 500 million weekly active users.
[00:05:30] That is a 43% spike since December, and it is translating into some serious revenue. At least $415 million a month, which is staggeringly up 30% in just three months. Now with enterprise plans API charges $200 a month. Pro tiers in the mix, OpenAI is now pacing towards a whopping $12.7 [00:06:00] billion in revenue this year, which means it could triple last year's numbers.
[00:06:07] Even as however it's cash burn is soaring. However, investors clearly think they've got quite a long runway and increasingly, which we'll talk about, they believe that the destination of all this money is AGI or artificial general intelligence. So first up here, Paul, maybe talk to me about the uses of this funding.
[00:06:29] Like on one hand Open AI is a consumer tech company that's in a ruthlessly competitive market. It's trying to win and retain users like any other company. So having a huge war chest makes sense. On the other, the others, this kind of regal where they say they've come out and published that they really need the money to build AGI.
[00:06:50] So, which is it?
[00:06:52] Paul Roetzer: Yeah, I mean, I think it's a, a little mix of both. the growth is nuts. I, I, the Sam Alman tweeted on March 31st. [00:07:00] I'm not, I can't remember if I said this one on last week's episode or not. I don't remember when this tweet came out, but he said the ChatGPT launch 26 months ago was one of the craziest viral moments I'd ever seen.
[00:07:10] And we added 1 million users in five days. We added 1 million users in the last hour. So when he was trying to give context to like how dramatic the growth from the image generation launch was, so this all came from the image generation launch. it was massive. I. So you hit on this 500 million weekly after users.
[00:07:31] We had just reported on 300 million, I think in February is big. We did so pretty crazy. in terms of how they are gonna use the money. I went back to a February information article, the information, which is a, a great source, that we, you know, constantly reference on, on the podcast. And they kind of broke down some details.
[00:07:50] They were obviously very well sourced in their reporting because everything has come true that they said back then. So they said, OpenAI has told investors SoftBank will provide at least 30 [00:08:00] billion of the 40 billion, which is what it is, rumored or reported that they have provided nearly half of that capital, which will value AI 260 billion.
[00:08:08] That's pre-money. So the 300 billion is after the money will go towards Stargate. So they are saying of the 30 billion, well, I guess of the 40 billion total, half of that is being allocated toward the building out of the data centers with SoftBank and Oracle. the money will be used over the next three years to develop AI data centers in the us.
[00:08:27] Open is planning to raise, about 10 billion of the total funds by the end of March. It sounds like they got the commitments in place by the end of March for all of this. that article again from February that we'll put in the show notes said the financial disclosures also show how Entangled SoftBank and Open AI have already become the company.
[00:08:44] Forecast that one third of open AI's revenue growth this year would come from spending by SoftBank to use open AI's products across companies, a deal that company's announced earlier this month. then in addition to this, like, you know, they are now on pace to hit [00:09:00] 12.7 billion this year. it says Open AI expects revenue to hit 28 billion next year.
[00:09:06] So 2026 is looking at 28 billion with the majority of that coming from chat GPT and then the rest through software developer tools and AI agents. but as you alluded to, the cash burn is massive. So it said OpenAI anticipates the amount of cash it is burning will grow at a similarly torrid rate. It expects cash burn.
[00:09:27] To move from about 2 billion last year to nearly 7 billion this year. the company forecasted that its cash burn would grow in each of the next three years, peaking at about $20 billion in 2027 before OpenAI would turn profitable by the end of the decade after the buildout of Stargate. So, yeah, I mean, they are just burning cash unlike any other.
[00:09:50] And they need to like, solve this fast. and they are definitely betting on that when they build all these data centers, they are gonna follow [00:10:00] these scaling laws and they are gonna have an insanely valuable tool. We had talked on a recent episode about a $20,000 a month license for, you know, basically a human replacement agent.
[00:10:11] some of the things we'll talk about in the next topic on AGI sort of starts to move more in this direction and I, I honestly, I'm not sure what the ceiling is on, what you could charge for. Powerful ai, AGI like whatever we want to call it. Like if you are, if you are building an A system that basically functions like a whole organization, which is their level five ai, like I'm, this is me making stuff up like level five on open AI is internal stages of AI is organization, right?
[00:10:44] So they plan on building systems that function as companies 20,000 a month may look cheap two years from now, they may be charging a million a month. Like, who knows? Because they think that their system is [00:11:00] basically gonna do the work of an entire organization with a couple people orchestrating maybe millions of agents like or an a, a, an AI that orchestrates all the other ais and the human oversees the master ai.
[00:11:13] Like that may sound sci-fi, but that is absolutely what they are thinking is going to happen.
[00:11:19] Mike Kaput: This does relate to some of the top topics we've talked about in the past around like service as software because it's not like they are just going after the licensing fees of other tools, though they are a bit, it's more about the total addressable market represented by the actual labor costs of knowledge workers.
[00:11:39] We're talking, we spend trillions of dollars a year hiring people to do a lot of the jobs that it sounds like they expect their AI is something people would pay for to do the job instead of a human.
[00:11:53] Paul Roetzer: Yeah, and that's, it's just the weird part is like we can't really project what this looks like, but we know it's [00:12:00] significant.
[00:12:00] Michael Dell on April 1st, you know, the founder of Dell computer texted, or Tweeted knowledge work drives a 20 to $30 trillion global economy. With ai, we can increase productivity by 10 to 20% or more. Unlocking two to 6 trillion in value every year. Getting there may take 400 billion to 1 trillion in investment.
[00:12:21] The return on this over time will be massive. So yeah, I mean, the people who are closest to this stuff, whether it's, you know, Jensen Wong and Nvidia, or Zuckerberg or Altman or Michael Dell or whomever they are talking about what seems like some pretty crazy numbers, but to them it just sort of seems inevitable.
[00:12:41] Hmm. And that's, I think what's gonna come through as a theme of the next topic here today is like there's, there's a lot of people who are still trying to process what chat GPT can do today, but the people who are on sort of frontier are so far beyond that and they are seeing a clear path to a very [00:13:00] different world, like two, three years from now.
[00:13:02] And to them,ittruly seems like inevitability. And it might be five years, it might be seven, but like it's coming one way or the other.
[00:13:11] Mike Kaput: So let's talk about that our second. Big topic today is about a major new AGI forecast that is making some waves. So this is a new report called AI 2027, and it lays out one of the more dramatic timelines we've seen for ai.
[00:13:31] So this is primarily in the form of a website you can go to. We'll have the link in the show notes. It's a bit interactive in the sense as you scroll through it and scroll through their timeline. You'll see little widgets and visuals update as you go. It's really cool. It's worth, visiting. But in it, the authors predict that by the end of 2027, AI will be better than humans at basically everything from coding to research, to inventing even smarter versions of [00:14:00] itself.
[00:14:00] And this whole website, this whole thought experiment they go through shows what the runway looks like to this kind of intelligence takeoff. Now this whole project comes from something called the AI Futures Project, which is. Led by a former open AI researcher named Daniel Kojo, and he actually left the company over safety concerns.
[00:14:24] He then teamed up with AI researcher Eli Lund, who was also known as a highly accurate forecaster of current events. And together with their team, they turned hundreds of real world predictions about AI's progress into this kind of science fiction style narrative on the website. And this is all grounded in what they believe will actually happen.
[00:14:47] So the vehicle by which they describe this is this fictional scenario which involves a fictional AI company building something called Agent one, which is a model that quickly evolves into agent [00:15:00] four, an autonomous system making a year's worth of breakthroughs every week. By then, towards the end of the time off, it's on the verge of going rogue.
[00:15:10] Now, along the way, they show how AI agents will start acting like junior employees by mid 2025. By late 2026, AI is replacing entry level coders and reshaping the job market and in their forecasts. By 2027, we've got self-improving AI researchers making weeks of progress in days, and China and the US are fully locked in an AI arms race.
[00:15:36] Now, there's plenty of critics of this high profile project. Some critics say it's much more fear mongering and almost like fantasy than forecasting. But the authors argue it's a serious attempt to prepare for what could happen if we do have this kind of fast takeoff of super intelligent ai. So in an interview, Coco Al Joe actually said, quote, we predict that AI will [00:16:00] continue.
[00:16:00] To improve to the point where they are fully autonomous agents that are better than humans at everything by the end of 2027 or so. Now, this also comes at the same time this past week as we saw a couple other significant AGI pieces of news. One of them is that Xop AI board member Helen Toner, published in Article one Substack pointing out that all these predictions we're getting about the timelines for AGI are getting shorter and shorter.
[00:16:28] And she even writes, quote, if you want to argue that human level AI is extremely unlikely in the next 20 years, you certainly can, but you should treat that as a minority position where the burden of proof is on you. And then last but certainly not least, Google DeepMind actually came out with a vision for safely building AGI in a new technical paper.
[00:16:50] The company literally says, AGI could arrive within years, and that they are taking steps to prepare. So they have this whole safety roadmap over dozens and dozens of pages. [00:17:00] The focus on what they say are the four big risks of AGI. There's first misuse, which is a user instructing the system to cause harm.
[00:17:11] Second is mistakes, meaning an AI causes harm without realizing it. Third is structural risks, which means harms that come from a bunch of agents interacting where no single agent is at fault. And fourth is misalignment when an AI system pursues a goal different from what humans intended. So Google says of this plan, this roadmap, this safety measures quote, we're optimistic about Agis potential.
[00:17:37] It has the power to transform our world, acting as a catalyst for progress in many areas of life. But it is essential with any technology this powerful that even a small possibility of harm must be taken seriously and prevented. Paul, there's a lot to unpack here, but first up, what did you think of AI 2027?
[00:17:57] Like the people behind it seem like they [00:18:00] have some interesting backgrounds in ai. Did you find their predictions credible? Was the format of this fictionalized story like helpful, harmful to getting your average person to actually care about this?
[00:18:13] Paul Roetzer: Yeah, I mean, so my, my initial take is like a reader beware, sort of warning on this.
[00:18:19] I, I, I honestly wouldn't recommend reading this to everybody. Like I think that, it could be very jarring and overwhelming and it can definitely feed into the non-technical AI person's fears. and maybe accelerate those fears. I think when you read stuff like this, whether it's situational awareness, you know, the papers series of papers from Li Po Ash and Brenner that we covered last year.
[00:18:48] machines of, what was the, what was Dario Amides of, of Grace. Grace
[00:18:51] Mike Kaput: Machines of Loving Grace.
[00:18:53] Paul Roetzer: Yeah, that one. the Accelerationist Manifesto from Andres, like you, [00:19:00] it, it, you have to have a lot of context when you read these things and you have to have a really strong understanding of who's writing them and their perspective on the world.
[00:19:12] And you have to appreciate that it's, it's just one perspective. Now, they are certainly credentialed, like they are, they have every, everything on their resume that would justify them taking this effort and writing this. And I think it needs to be paid attention to. And I think that in, I mean, I got through the first probably 15, 20 pages of it and then started scanning the rest as it started going through these other different scenarios, but certainly enough to get the gist of what they were talking about and their, their perspectives.
[00:19:43] I did see Daniel, one of the, you know, the leads on this, he tweeted like, challenge us. Like, we'll, they are actually, they actually put bounties out to disprove them. they are like, if you can come at us with a fact that's counterfactual to what we presented, we will pay you. [00:20:00] So, I don't know, honestly that anything they put in there is actually that crazy.
[00:20:07] And, and that's why I'm saying like, I, I just wouldn't recommend it because it, it's, it's just a lot to, to handle. So if, if you are, if you are at a point where you really want to know, because like there was a key thing, the thing I would recommend is actually the Kevin Rus New York Times article Yeah.
[00:20:28] Is actually where I would start. Before you read the AI 27 2 2027 website, I actually read the Kevin Russ article. We'll put in the show notes. Kevin gives a very balanced take on this. And I thought one of the real key things is, at one point Kevin said, so I'll actually, I'll jump to the Kevin article for a second.
[00:20:49] So he starts off, the year is 2027. Powerful artificial intelligence systems are becoming smarter than humans and are wrecking havoc on global order. Chinese spies have stolen America's AI secrets and the White [00:21:00] House is rushing to retaliate inside a leading AI lab. Engineers are spooked to discover that their models are starting to deceive them, raising the possibility that they'll go rogue.
[00:21:09] That is a, that is basically a summary of AI 2027 project. That is the scenario they are presenting. There isn't a single thing in that concept that couldn't happen by 2027. So that's what I'm saying, like, I'm not disputing what they are saying. I'm just saying they are, they are taking an extreme position.
[00:21:32] But the key here is to understand who is writing this. So later in the article, Kevin says, there's no question that some of the group's views are extreme. Mr. Cota, how did, how did you say that? COTA ta Okay. Yeah. for example, told me last year that he believed there was a 70% chance that AI would destroy or catastrophically harm humanity.
[00:21:56] So he is, there's something called P Doom [00:22:00] in the AI world, the probability of doom probability that AI wipes out humanity. and this is like a common question asked of leading AI researchers, what's your p doom? And there are some who, who are well above 50% that, that they are convinced that the super intelligence being built is going to wipe out humanity.
[00:22:18] There are others who think that is absurd, like a Jan Koon. Who won't even probably answer the question of PDU because they think it's so ludicrous. So you have to understand that there are different factions, and each of these factions often has access to the same information, has have worked in the same labs together on the same projects.
[00:22:39] Seeing the models emerge and in the capabilities they all seen the same stuff, but some of them then play this out as this is the end. Like it's all gonna go awful here. but when you actually start getting into like these like fundamentals of like over dramatic, dramatizing this, [00:23:00] they actually kind of struggle to come back to reality and say, yeah, but what if it doesn't actually take off that fast?
[00:23:05] Mm-hmm. What if the Chinese spice don't get access to agent three, as they called it? Like, what if we, it's like ChatGPT and like society just basically continues on with their life as though nothing happened. And a small collection of companies have this powerful AI that can do all these things and.
[00:23:21] And like the world just goes on. And that's a, that's honestly harder for them to fathom than this like doom scenario. And so I, again, I just, I feel like it's, it's a good read. If you are mentally in a place where you can consider the really dramatic dark side of where this goes pretty quickly understand it is all based on fact.
[00:23:46] There's nothing they are making up in there that isn't possible. It just doesn't mean it's probable. and I still like, think that we have more agency in how this all plays out than maybe some of these [00:24:00] reports would make you think. But it takes people being kind of locked in and focused on the possibilities.
[00:24:06] There was one, the chief executive from the Allen Institute for ai, AI lab in Seattle who reviewed the AI 2027 paper and said he wasn't impressed at all by it. Like that there was just nothing to it. So. Again, read, beware. If you want go down that path, do it. If you wanna get really technical, ESH actually has a podcast through our podcast with the authors.
[00:24:27] Yep. And Esh we talked about before, we'll put the link in the show notes. He does amazing interviews. they are very technical. So, but again, if you are into the technical side of this, ha, have a field day. If you are not read the Kevin Rus article and move, move on with Your Life, basically is kind of my, my notes here now on, on the hub, on not, but on the Google side, the taking a responsible path to AGI also a massive paper like yeah, you wanna great notebook, LM use case.
[00:24:55] Drop that thing into a notebook, lm and have a conversation with it. Turn it into a podcast. but [00:25:00] there is some interesting stuff within here. If you just read the article about it that they published on the DeepMind website. They referenced the levels of AGI framework paper that I talked about in the road to AGI series.
[00:25:12] they linked to the new paper, and approach to technical AGI safety and security. But then they also, released an a new course on AGI safety that I thought was interesting. I have not had a chance to go through it yet, but it's, looks like it's about a dozen or so short videos. they are like between four and nine minutes it looks like.
[00:25:31] But they've got, we are on, on a path to superhuman capabilities, risk from deliberate plan, deliberate planning and instrumental sub goals. where can misaligned goals come from? Classification quiz for alignment failures, like some interesting stuff. Interpretability, like how to know what these models are doing.
[00:25:48] So again, this is probably made for a more technical audience, but it could be interesting for people if you wanna understand kind of more in depth what's going on here. So. Big picture. I'm glad to see this sort of [00:26:00] thing happening. Mm-hmm. Like this was my whole call to action with the AGI series is like, we just need to talk more about it.
[00:26:06] We need more research, we need more work. Trying to project out what happens. I'm just more interested in like, okay, let's just go into like the legal profession or the healthcare world or the manufacturing world and let's play out more like maybe practical outcomes and then what does that mean? Like what happens to these fundamental things that we all are familiar with?
[00:26:26] Because if you take this stuff to a CEO, I, yeah. Most CEOs are just still trying to grasp how to personally use chat GPT and how to like, empower their teams to figure this out. You start throwing this stuff in front of 'em and you are just gonna have people pull back again. So I, for sure, yeah.
[00:26:42] Important to talk about, but I, I just wouldn't let guide people don't get like too consumed by this stuff.
[00:26:49] Mike Kaput: Yeah. For instance, if you are, say a marketing leader at a healthcare organization struggling to get approval for chat GPT and get your team to build GPTs, this [00:27:00] can send you into an existential.
[00:27:01] Yeah. You don't wanna link to the AI
[00:27:03] Paul Roetzer: 2027 report in your deck pitching this. Yeah.
[00:27:10] Mike Kaput: All right. So our third big topic this week is that Amazon just entered the AI agent race with a new system called Nova Act. And this is a general purpose AI that can take control of a web browser and perform tasks on its own. So in its current form, this is fully a research preview. It's aimed at developers, it is bundled with a software development kit that let and build AI agents that can, for example, book dinner reservations, order salads, or fill out web forms.
[00:27:44] So it's basically Amazon's answer to agent tools like. Open AI's operator and Anthropics computer use. But there's kind of one key advantage here that's worth talking about. It is being integrated into the upcoming Alexa plus [00:28:00] upgrade, which potentially gives it massive reach. Now, Nova Act comes out of Amazon's new AGI Lab in San Francisco, which we covered on a past episode, led by former open AI and Adep execs, David Lewan and Peter Abiel.
[00:28:16] And the lab's mission is to build AI systems that can perform any task a human can do on a computer. Nova Act is the first public step in that direction. Amazon claims it already outperforms competitors on certain internal tests, but it hasn't been benchmarked against tougher public evaluations just yet.
[00:28:37] So Paul, this is admittedly very early. It's a research preview. It's an agent, which we talk about all the time, is still a technology that's really, really, really early. So it's not like tomorrow you are suddenly going to have Amazon's agent doing everything for you. But it does feel a little different and worth talking about than some of the other [00:29:00] agent announcements because of Amazon's reach and how much it touches so many parts of consumer life.
[00:29:06] Like, do you think this could be the start of seeing agents really show up for your average person?
[00:29:13] Paul Roetzer: Yeah, I mean I, generally speaking, we try to not cover like research previews too much. Like we often will like give overviews of like, here's what's happening. But so often we've seen these things just don't really lead to much.
[00:29:28] But I. I think the key here is it's starting to change the conversation around Amazon and their AI ambitions. So, I mean, if you go through the first 130 episodes or so of this podcast, my guess is we talked about Amazon maybe like three or four times. Yeah. Like, it's just, and it's usually related to their investment in Anthropic.
[00:29:49] Yeah. we talked about Rufuss last year, which is their shopping assistant. So right within the app, you or website, you can just talk to Rufuss, I'm going on a trip here, what should I be looking for? And it helps you buy [00:30:00] things. And they are using a language model underneath to do it. I think it's powered by Anthropic.
[00:30:05] Then we talked about Alexa plus a couple weeks ago. and now we're talking about not only Nova, but they also last Thursday announced this buy for me feature. And so I don't know, Mike, did you, did they say when this one's coming out? Do you remember seeing that in? I don't recall seeing
[00:30:22] Mike Kaput: the exact, an exact release date.
[00:30:25] Paul Roetzer: Okay. Yeah. We'll check. They, they, they put it out on their site and then TechCrunch covered it. But the basic premise here is buy for me uses encryption to securely insert your billing information to third party sites. So if you are searching for something and they don't have it on Amazon, their AI agent kind of powered by this Nova concept, we'll actually go find it somewhere else on the web.
[00:30:44] It will buy it for you by entering your information into that site. And so it's different than OpenAI and Google's agents, which requires the human to actually put the credit card information in before a purchase happens. So if you say, Hey, go find me a new backpack for a trip to [00:31:00] Europe. And the agents from Open AI and Google go do it.
[00:31:03] When they get to the site, the human then has to do the thing. In this case, Amazon is basically asking users to trust them and their privacy and their ability to securely protect your, your information to go ahead and fill this out. And they are trusting that you are not that, that their agent is gonna accidentally buy a thousand pairs of something instead of one pair of something.
[00:31:25] Right. So I think that what we're seeing is how Amazon is maybe gonna start to play this out. And I think we talked on a recent episode that they are probably building their own models as well, in addition, you know, continuing to invest more heavily in building their own models. So, I don't know, like, I think more than anything it's probably starting to move Amazon up in the conversation to where I'm starting to see, we may be talking about Amazon a lot more than we used to talk about them.
[00:31:53] Yeah. Because it really previously was robotics, their investments in ai. And then, you know, I always talk about [00:32:00] Amazon as, it's one of the like OG examples of AI in business was the prediction around like the recommendation engine, their shopping cart, where it would predict things to buy. That was like old school ai.
[00:32:12] And they'd been doing it as, as well as anybody for like 15 years. Yeah. So they weren't new to ai, they just got sideswiped by generative ai. They were like, they had nothing. They, you know, they had Alexa, butitwas not, not anything close to what needed to happen. and now here we are, like two and a half years later, whatever, and they are still trying to play catchup now on a thing they should have been leading on, but you know, all of them missed Apple, missed it.
[00:32:37] Google mixed it, missed it. Amazon missed it. So, yeah, I just, I don't know, it's, it's interesting. I expect we'll hear more outta this, this lab. but I think we'll probably also see it built out into their products pretty quickly.
[00:32:51] Mike Kaput: Just a note here, according to the Amazon announcement, buy for me is currently live in the Amazon shopping app on both iOS and Android [00:33:00] for a subset of US customers.
[00:33:02] Okay? So they are beginning testing with a limited number of brand stores and products with plans to roll out some more customers and incorporate more stores and products based on feedback. So if you have access to this and you are brave enough, maybe you can go give it a, give it a try. Yeah. But,
[00:33:18] Paul Roetzer: but don't expect the same sort of ease of returns as buying from Amazon because they did note that you are, they don't handle the returns the way you normally do.
[00:33:27] If you bought it from a site, you are, you are responsible. How are you about this stuff? Would. Use like a bio for me. Are you like, you are more aggressive with using agents than I am.
[00:33:38] Mike Kaput: I don't know if I have a personal worry about something going wrong or privacy that I couldn't reverse, or that wouldn't really matter that much to me.
[00:33:49] Yeah. But it does just seem like a hassle for me.
[00:33:52] Paul Roetzer: I think I just know how unreliable AI agents are today. Yeah. Despite how they are being marketed that I think I'm just, I'm, I'm letting, I'm [00:34:00] willing to let everybody else work out the kinks. Like I don't find that that convenience un worth enough of the risk of this going wrong.
[00:34:07] Exactly's, it's, I, I'm kind of good with just filling out my own form and like going to the other site and you know, paying for it there and knowing the terms of use and the return policy. And so, I don't know. I'm a little more conservative when it comes to like, pushing the limits of AI agents today.
[00:34:22] For sure.
[00:34:24] Mike Kaput: All right, let's dive into this week's rapid fire. Our first rapid fire topic is that OpenAI is finally releasing a new open weight language model. This is the first they've done since GPT two. So in a post on X, Sam Altman said the company has been sitting on this idea for a long time, but quote, now it feels important to do.
[00:34:45] This model will launch in the coming months with a strong focus on reasoning ability and wide usability. So it's important to note here, this is an open weight model, and you kind of see confusion of these terms. A lot of people say, oh, [00:35:00] okay, that's open source. Well, technically not exactly, because open weight means the model's weights, which are the numerical parameters learned during training are made publicly accessible.
[00:35:11] So the weights define how the model uses input data to produce outputs. However, an open weight model won't give you all the source code training data or architecture details of the model. Like a fully open source one would. So you can still like host and run this type of model at your company. Try on your own data, which is what open AI is hoping people will do.
[00:35:32] But it's not exactly fully open source, which is not uncommon to see. Now before Launch says Altman, the model will go through its full preparedness evaluation to account for the fact that open models can be modified or misused after release. And OpenAI is hosting developer feedback sessions starting in San Francisco and expanding to Europe and Asia and the Pacific, to help make sure the model is useful out of the box.[00:36:00]
[00:36:01] So Paul, how significant do you see it being that OpenAI is at least dipping its toe back into the waters of open models?
[00:36:09] Paul Roetzer: Yeah, I, I mean maybe the biggest play here is that Elon Musk won't be able to call them closed AI anymore. Like, so that's one of Elon's beefs is that they, you know, they were created to be open and then they weren't.
[00:36:21] And so, you know, maybe this is the counterbalance to, to that argument. I mean, it's a, it's a strategy I would expect all the labs to do. So obviously Meta's main play has been to release powerful open source models or open weight models. Google DeepMind, Demi Saba has said this is their strategy basically, that they will release the prior generation as open weight.
[00:36:43] So they build, you know, let's say, you know, Gemini 2.5 is the model today, a year from now, let's say it's Gemini four or whatever, then they would probably then open source Gemini 2.5. So like they take the current frontier model that is like the [00:37:00] paid for featured model that they don't consider safe yet to release.
[00:37:04] They build the next generation of models and then once they determine that the prior generation model is safe, plus the key is compute costs are dropping roughly 10 X every year. So now the cost to deliver that open weight model, open source model is dramatically less and now they can go and do it. So it's just like this.
[00:37:23] I think this is a logical strategy for Google and OpenAI and some of these other labs that you, you just kind of release these prior generation models and you hope that enterprises are, want to still pay for the most powerful thing because it has more reasoning, because it has better image generation, better, better video generation, all of those things.
[00:37:42] and the open source becomes more for the developer world to build on top of.
[00:37:48] Mike Kaput: Next up, a new study has reignited some debate over whether AI can pass the Turing test. So the Turing test is a famous test thought up way back in 1950 by [00:38:00] computer scientist Allen Turing. That is designed to evaluate a machine's ability to exhibit human-like intelligence.
[00:38:07] Now, it basically consists of a human judge communicating with both a human and an AI through text-based questions and trying to guess which is which. If the judge can't tell which is a human and which is ai. The AI is thought to have quote unquote passed this test and displayed intelligence that's indistinguishable from a real human.
[00:38:30] So in this new research, the researchers ran two large scale experiments where people chatted simultaneously with a human and AI for five minutes, then guessed which was which, and the standout was GPT-4 0.5. When it was prompted to act like a shy slang using young person, it was picked as a human 73% of the time, more often than the actual humans.
[00:38:56] It was up against me. Lama model also fooled [00:39:00] participants. More often than not, while older systems like GPT-4 oh Crazy to even consider that an older system were reliably flagged as bots. Interestingly here, the difference that really made all the difference was prompting without specific instructions on how to act human, the AI win rates drop.
[00:39:20] But even then, some did still match human performance. So this is Paul. You know, definitely interesting because turning test is this kind of legendary thing in ai. We obviously always have to take any claims about all this with a grain of salt. The researchers themselves admit that there's still a lot that's unclear about what this could actually mean and how much it matters in terms of making a judgment call about the level of intelligence being exhibited here.
[00:39:50] But I think in a practical sense, it is really striking that we have some good evidence now that today's AI prompted in the right way can be basically. [00:40:00] Indistinguishable from a human in certain types of conversations.
[00:40:05] Paul Roetzer: Yeah. And I think that the whole part about prompting it to act like a human Yeah.
[00:40:10] Like that's not hard. That, I mean, you can make that, that instruction choice in like the system prompt. You could have a company, it could be a startup that builds on top of an open source model that chooses to make a very human-like chatbot and out of the box, the thing feels more human than human. we've talked about on the, on the show many times about like empathy and it's sort of, I used to think a uniquely human trait that I am convinced is not anymore, or at least the ability to simulate empathy.
[00:40:41] And so you can teach these models or you could tell your model, like you could go in and build a custom GPT and say, I want you to just be empathetic. Like, I just need someone to talk to who understands how hard it is to be an entrepreneur and like, I just want you to be, you know. I just want you to listen [00:41:00] and help me, you know, find my way through this and it will do it like better than many humans would do it.
[00:41:07] And that's just a weird place to be in. So I mean this constant, like the, do we pass the turning test? Like, I feel like the during test sort of had its day in, like, you know, maybe we probably got past it in, in like certainly when Chat BT came out. I think we're just now trying to find, trying to find ways to run the test to like officially say we've now passed it.
[00:41:29] It's like, I, I don't even know that it's worth talking about continuing the research. It's like we we're there like, right, people are convinced these things are more human than human in, in, in many cases, especially if they are prompted to be that way. And I think that when it comes to different parts of, you know, psychology and therapy and things like that, like that's how these things are being made already.
[00:41:51] Like people are using them as therapists. And I'm not commenting on whether that's good or bad for society. I'm just telling you that's what's happening. And the VC [00:42:00] firms are funding the companies to do that because they are so good at it. Yeah. And that's the current generation. And you know, it's not far behind where the voice comes along with it too.
[00:42:09] Mm-hmm. And now you truly just feel like you are talking to a therapist or an advisor or a consultant and their, their system prompt tells them to be very, you know, supportive and empathetic. and honestly, like at some point you just, you are gonna just prefer to talk to the ai. I, I do think a lot of people are going to arrive at a point where they just prefer talking to the AI about these things.
[00:42:31] These things like the hard topics that awkward to talk to people about. Like, it's not awkward to talk to your ai. and I think a lot of society is actually gonna come around to that pretty quick. It may end up being like, there was some data this week about how low adoption actually is to like the vast majority of society.
[00:42:48] I could see like. The empathetic chat bot with, with a human-like voice being like the entry point for a lot of people. Mm-hmm. and that's why I mentioned that in the [00:43:00] road AGI like I thought voice was gonna become like a dominant interface. And I think it could be a gateway to generative AI for a lot of people who maybe are sitting on the sidelines still.
[00:43:10] Mike Kaput: Yeah. It's almost like throw out the Turing test and look at today all the millions of people that use character AI for relationships or therapy that tells you everything you need to know.
[00:43:21] Paul Roetzer: Yeah. It goes back to like the, when we've talked about the evals, like these labs run all these like really sophisticated evaluations to figure how smart these models really are.
[00:43:29] And my feeling is like, that's great. And I get that the technical AI people wanna do that. What I wanna know is like, can it, how does it work as a marketer? How does it work? As a psychologist? As a physician, like I want evals that are like tied to real life. And I think that's the same thing as you are alluding to.
[00:43:43] It's just like, yeah, exactly. We need it to be practical.
[00:43:47] Mike Kaput: Our next topic is about Anthropic. Anthropic has just launched Claude for Education, which is a new version of its AI tailored specifically for colleges and universities. [00:44:00] So the centerpiece of Claude for Education is a new learning mode that prioritizes critical thinking over quick answers.
[00:44:07] Instead of solving problems for students, Claude gives them a guidance using these like Socratic methods. So by asking questions like what evidence supports your conclusion, Claude is going campus wide as part of this initiative at Northeastern University LSE and Champlain College, giving every student and faculty member access to Claude at Northeastern alone that it's 50,000 users across 13 campuses.
[00:44:36] they are also focused on a campus ambassador program, giving free API credits to student builders and partnerships, with internet two and canvas maker and structure to weave Claude into existing academic platform. So Paul, this definitely doesn't just seem like a press release. This is a pretty comprehensive initiative in [00:45:00] education.
[00:45:00] You talk to tons of schools about the need for AI literacy. What do you think of how Anthropic has gone about this?
[00:45:07] Paul Roetzer: Yeah, I think it's, it's great to see. I OpenAI did something similar with their academy. They just announced last week. They have like a AI for K to 12. Yeah. Where they are trying to get into like the education and I don't think they had a higher ed one yet open.
[00:45:22] I also announced, you know, not to be out to on, they love to steal the headlines and nobody else. I think they tweeted, it was over the weekend I believe, or no, what they, so they Friday, so it was like Wednesday or Thursday. that they are now giving like chat GBT free to college students, I think for the next two months.
[00:45:37] Yeah. Something like that. So I think everybody's playing the space. I, I, I don't know, like it's so disruptive and I don't know that, you know, schools are still grasping. I have seen some really impressive stuff. Like I've seen some, some high schools, I've seen some universities that are being very proactive, but like, I don't, I don't think I shared this example on the podcast last week, but like, I [00:46:00] was, I was, I was home with my kids the other day.
[00:46:03] My wife wasn't, wasn't here, and my daughter's 13, seventh grade doing like advanced pre-algebra or something. She's like, I need help. I'm math homework. I was like, that's a mommy thing. Like, I not, I'm not the math guy. When you get into like the language, like, let me know and we'll talk. She goes, no, I, mommy's not here.
[00:46:18] I need help. And so it was a math problem. I have no idea how to solve. So I pulled up the, you know, go into chat. GBT hit my, you know, the camera open. I don't even, what they call that, what do they call that? Is it live or, I don't know.
[00:46:31] Mike Kaput: Oh, you mean when you are live showing it a different Yeah, yeah.
[00:46:34] Paul Roetzer: Just like turned on the camera and it could see what I was seeing. Yep. I know, yeah. I'm sure it's Project Astra for, for Google, but I don't know what they actually call it an open ai. But if you don't know what I'm talking about, just go into the voice mode and then in voice mode there's a camera, click that and it now sees what you see.
[00:46:47] And so I held it over the math problem and I said, I'm working with my 13-year-old. Do not give us the answer. Mm-hmm. We need to understand how to solve this problem. And it's like, great. Okay, let's go through [00:47:00] step one. And it actually like, would. Read it and then say, okay, do you understand how to do that?
[00:47:05] Anditlike walked us through. And then she was writing on paper the formula and like going through and doing what I was saying. And so I held the phone over what she was writing and said, you are doing great now when you get to this point. You know? And then I would ask her another question and then she would answer.
[00:47:20] So now she's interacting with the ai. Yeah. And we walk through the five steps of the problem with her actually doing it and being guided how to do it, not being given the answer. And to me that's just like so, representative of where this can go if it's taught responsibly. If kids just have chat GPT and they just go say, Hey, gimme the answer to this question, then we lose.
[00:47:45] So I think that having anthropic and Google and OpenAI and others be proactive in building for education and building in a responsible way for education is a really good thing. And we, we should support that and encourage more.
[00:47:59] Mike Kaput: [00:48:00] Yeah, it's really cool to see. Next up, the Tony Blair Institute out of the UK has released a sweeping new report calling for a reboot of UK copyright law in the age of ai, and their recommendations are already drawing some fire.
[00:48:17] One of the big reasons is because the report endorses a text and data mining exception to copyright law that would allow AI companies to train models on publicly available content unless rights holders explicitly opt out. It argues this opt-out model would balance innovation and creator control, but longtime AI copyright commentator Ed Newton Rex, we've talked about a bunch on the podcast called this report basically quote terrible and quote a big tech lobbying document.
[00:48:48] He says, UK copyright already gives creators control over how their work is used, and that shifting to an opt-out regime would reduce that control. More sharply. He accuses the authors of misleading [00:49:00] rhetoric, likening their claim, their arguments to claiming that using someone's AI art for training is no different from a human being inspired by it.
[00:49:08] So he basically says, under this kind of scheme, creators would lose their rights. The public would put the bill, and AI firms would keep training on others work for free. Now Paul, this is obviously UK specific, but we wanted to talk about it in the wider context of the copyright topics we covered last week.
[00:49:29] Artists and authors in many areas are up in arms about how AI models are being trained on their work without their permission. This certainly seems like some parties, whether they are actually lobbying for AI labs or not, are trying to make the argument that AI companies should be allowed to train on publicly available content that we should exempt this from copyright.
[00:49:51] What do you think of this approach and should we, should we expect to see more arguments like this in the us?
[00:49:57] Paul Roetzer: I mean, these AI companies have a lot of money for [00:50:00] lobbying efforts, and I think at the end of the day, those lobbying efforts win. I think the opt-out thing's a joke. I, I've always just felt that that was an absurd solution.
[00:50:09] It was just like an obvious thing to present. But like, I mean, if you are a creator in any way, you know how prevalent it is for people to steal your stuff. Like any, anything we've ever created behind a paywall, I guarantee you someone has stolen 10 times over and published it in different places. The sites I would never like click through and download something from, but like, you know, whether it's it's movies or courses or books or whatever, it gets stolen all the time.
[00:50:40] And it's a game of whack-a-mole to try and keep up with it. Like we have an internal system to track all the stuff people steal from us and what, what can we do about it? Pay our attorneys every time we find it. And that's easy to find. Like you could just keyword search the thing and you can find the people stealing your stuff.
[00:50:56] Yeah. How in the world are we supposed to ever know, unless someone leaks the [00:51:00] training data, whether or not they stole it or not? I saw something last night that was like, they had proof now that one of the major model companies who I won't throw under the bus right now, absolutely stole stuff from behind a paywall of a major publisher and they can prove it.
[00:51:15] so I just feel like, I don't know, the copyright thing is so frustrating to me because I have yet to hear of any sort of like reasonable plan for how you acknowledge and compensate creators whose work made these models possible. Right. And even if they come up with a plan. How do we know, like how will we ever do it other than being able to audit the system and find out what the actual training data was or someone suing them.
[00:51:42] And then seven years later it's like, okay, yeah, sorry, your seven books were used in the training of the model. Here's your $15. Like, I don't know, I don't have a solution, but it's very frustrating that nobody seems to have a plan for how to do this. It's just like, yeah, we should probably pay them, but first we have to admit we stole it, [00:52:00] but we can't admit we stole it 'cause we're gonna claim it's fair use.
[00:52:02] And then eventually we'll like, have to pay a fine and maybe there'll be some, class action lawsuit and we'll pay a billion dollars and that billion dollars will get spread across 200 million creators. And, you know, here's your $50 check. Like, I don't know, like I, I hope someone much smarter than me in this area eventually comes up with a plan and the model companies agree to, to, to, to do something to compensate people for their work.
[00:52:27] Mike Kaput: And in the meantime, like we talked about last week, expect the backlash to continue.
[00:52:32] Paul Roetzer: Yeah. And it is growing. Yeah, for sure.
[00:52:36] Mike Kaput: Our next rapid fire topic, Google DeepMind has hit a new milestone in AI because it taught AI to find diamonds in Minecraft without any human guidance. Now, this breakthrough comes from a system called Dreamer, which mastered the game's notoriously complex diamond quest, purely through reinforcement learning.
[00:52:57] So that means it wasn't trained on videos or [00:53:00] handholding instructions and explored, experimented, failed, and learned. Now, if you are unfamiliar with Minecraft doing this task, finding diamonds is not easy. you are required building tools in sequence, exploring unknown terrain, and navigating a world that is different every time.
[00:53:18] So what makes Dreamers special is in how it learns. This stuff, instead of brute forcing every option, it may builds a mental model of the world and simulates future scenarios before acting. Much like how a human might visualize possible outcomes, that world model lets it plan more efficiently, reducing trial and error while still enabling real discovery.
[00:53:41] Interestingly, dreamer wasn't even designed for Minecraft. This diamond challenge was just a stress test, but the fact that it passed without ever seeing human gameplay shows really interesting progress toward general purpose ai. So Paul, this is obviously not just us being [00:54:00] fans of Minecraft, overing Earth.
[00:54:02] One of the researchers involved in the work said why this matters. Quote dreamer marks a significant step towards general AI systems. It allows AI to understand its physical environment and also to self-improve over time without a human having to tell it exactly what to do. That is a much bigger deal than Minecraft itself.
[00:54:25] Paul Roetzer: Yeah. And this, I mean this very similar in, in terms of past research that like, you know, Google has done where like they had Alpha Go learning how the game of go, but then they build Alpha Zero that could basically learn from the ground up and Google d Mine's been doing this stuff since like the early teens.
[00:54:41] Mike Kaput: Yeah.
[00:54:42] Paul Roetzer: And this is why, like, I often, I'm back to like, I don't, I just don't know how you bet against Google. Like I don't think people realize the amount of breakthroughs that they have had and the knowledge and capabilities that they are sitting on that aren't in these models yet. And when you can start introducing this kind of capability, even if it's [00:55:00] just an internal model that they don't release, it's kind of hard to process.
[00:55:05] So I think there's the, this is a significant line of research. the ability for these things to sort of learn and pursue goals on their own is it matters. I ironically have been listening over the last few days to a podcast, big technology podcast. With, the Roblox, CEO, David Zuki. Hmm. And, and so I, in my head I have this, 'cause my kids play Roblox and Minecraft and I know that to them the process of doing these things is the point.
[00:55:39] So like in, in Minecraft you build block by block. It is repetitive, it is mind numbing, but they love it and they create insane things. Like my daughter has showed me like castles she's built. And I'd be like, how long did you work on this? Like, this is amazing and like, you did this with blocks. Like, it doesn't even make sense to me.
[00:55:59] And it might be [00:56:00] something she spent like 20 hours on over like months where, or maybe more. And that is the point now, if you can go in and just say like, build me a fantasy castle. And like, and I'll, now you have the same beautiful castle, but zero effort from the human to do it other than like, I'm envisioning a castle here and I wanna moat there and now I want a dragon.
[00:56:20] That's the world. The CEO roadblocks is presenting that they are enabling, you are gonna be able to just go into roadblocks and like just text the characters you want and the scenes you want, and eventually entire games. And so this line of research also just like, I don't know, concern is the right word.
[00:56:37] There's parts of it that just make me sad because I feel like so much of what makes games so fascinating that I love them as a kid and my kids love them now, is the repetitive nature of doing something yourself and like figuring it out and finding a solution and finding diamonds. Like instead of going and say, Hey, find me 50 diamonds, then you sit back and like sit, sip your Coca-Cola while you are like waiting for [00:57:00] the, I don't know.
[00:57:01] So it,itjust continues on this whole like creator thing. Like when the AI can create, like where's the human element? Where is the AI element and. Again, I don't, I don't know. I just, I find myself thinking about this stuff a lot and as these things get better and I see image generation, I watch VO two from Google team, I like Right.
[00:57:20] I see the runway stuff. We'll talk about, like, I just have, I I continue to really struggle to envision like, the next few years and what it means to creators and creativity.
[00:57:30] Mike Kaput: Well, it is so cool to be able to summon these kind of pieces of art or creativity out of thin air, but then you wonder what's lost that the artist learned in the process Yeah.
[00:57:40] Of learning how to create that thing. Right.
[00:57:42] Paul Roetzer: Yeah. I got home last night from a trip and my son couldn't stop talking about this thing. He was coding in class. Now he's in sixth grade and they were doing this in design class and he is taking like a couple of code camps and he has way more knowledge of coding than I do at this point.
[00:57:55] But like to listen to him explain it. And like, then this morning he [00:58:00] gets up and he is like, can I show you? Can I show you? Can I show you? And he's like showing me these like sprites he built for this game and then like this whole thing he coded where the, these monsters show up. I don't, I don't even understand it how he did it.
[00:58:10] Like that's, that's the joy of creation is like he learned how to do it. He didn't just give a text prompt and like created the monsters. Oh, great, great game. He wouldn't have the same passion for it. He wouldn't have the same fulfillment from it. He wouldn't have the same inspiration to learn how to do more code.
[00:58:25] And that is why I think about this all the time. It's like I just, I don't know, like I don't, I don't know what it means for them in two years, five years, you know, by the time they get out into the professional world. Nine years, 10 years, like, mm, so weird.
[00:58:41] Mike Kaput: Our next rapid fire topic concerns something called model context protocol or MCP.
[00:58:47] So in November of last year, anthropic announced it was open sourcing model contact the model context protocol. MCP. They define this as, quote, a new standard for [00:59:00] connecting AI assistance to the systems where data lives, including content repositories. Business tools and development environments. Now, in recent months, talk about MCP has been gaining traction.
[00:59:13] It's happening more and more in AI circles, so we at least wanted to introduce the concept and talk through it a little bit. One way to think of MCP is like A-U-S-B-C connector, but for AI data access. So today's AI assistants are smart, but they are opt-in stuck in silos. They don't know what's in your files, your code base, your company wiki, unless someone build a custom integration to access those data sources.
[00:59:42] MCP has kind of trying to change that by creating a universal standard for connecting AI models to external tools. That might be Google Drive, slack, GitHub, or Postgres. So no more one-off connectors. Basically just a way to plug in and go. Now, because of that, MCP [01:00:00] is gaining a bunch of traction. It has support from both open AI and Microsoft.
[01:00:05] It's open source. So hundreds of connectors are already live. And basically the idea behind all this is simple. Give AI systems a consistent way to fetch fresh, relevant context from all these different sources. So it's still really early days for this, but some people think the potential for MCP is huge and that it could really enable AI assistance to use your actual knowledge and other data sources to do even better work.
[01:00:33] So Paul, why is this getting so much attention in certain AI circles?
[01:00:39] Paul Roetzer: Dude, I tried to avoid talking about this topic. I mean, just like, I don't know, like three or four weeks ago this like hook over my Twitter feed day. Yeah. with all these AI people. And I was like, man, sounds important, but God, it, I, it's hurt my brain to like, think about it.
[01:00:56] So I just kept leaving off of the list and I finally told Mike class like, all right man, we, we [01:01:00] finally gotta like, just talk about this. So, I still honestly, like I, I'm, this is an abstract one for me. Yeah. Like usually there's AI topics that just like my, my brain generally does a pretty good job of like, understanding the context.
[01:01:13] This is one I struggle with still, to be honest with you. Damn. So, ironically, last night, laying in bed and I'm checking LinkedIn and Dharma Shaw, my friend and founder at CTO at HubSpot, he put on LinkedIn. So I'm just gonna read this because this is, it'll do better than I think I would do. Trying to add context, he said someday soon each of us will have our MCP moment.
[01:01:33] It won't be quite as powerful as the chat GPT moment we had. but it will open our eyes to what's possible now. For example, right now I have Claude Desktop configured to interact with several MCP servers from different companies. This configuration gives the learning, the large language model, hundreds of tools that it can decide to use based on what I enter for a prompt, I can have the large language model use agents on agent [01:02:00] ai, which Dharmesh created.
[01:02:01] Access CRM data in HubSpot, read, write to a specific directory in my local file system and ReadWrite messages to Slack, access my Google calendar and Gmail possibilities are endless. The beauty of MCP is that it's an open standard that defines how MCP clients in this place clawed can talk to arbitrary servers that provide lots of different kinds of capabilities.
[01:02:24] They don't need to be custom coded to talk to certain APIs or servers. he says, here's an example, prompt look, quote, look up OpenAI in the HubSpot, CRM and Slack, the details to add Dharmesh, including how long ago I had the last interaction. And he says, I could have done something much more complicated and had a dozen different systems.
[01:02:42] But you get the idea. Once you see it work, it will be magical. The setup is a bit tricky now, but that'll get easier real soon. My guess is when OpenAI add support for MCP to chat GBT, things will be smoother. Yeah. So yeah, I think it, again, it fits in the context of. My guess is like [01:03:00] three months from now this, we talk about this again on an episode.
[01:03:03] Yeah. And now it's much more tangible and the average person's able to do something who isn't, you know, Dharmesh, the CTO of HubSpot. I think it's a very technical thing right now. I don't, I don't think that the average maybe listener to our show who isn't, you know, consider themselves technical ai, leader is probably gonna be doing anything with this, but it seems like it's a conversation that's gonna start coming up within your company if you are working with it and starting to do more advanced things with your language models.
[01:03:30] Mike Kaput: Alright, Paul, I'm gonna go through some AI product and funding updates real quick and then we're gonna wrap up with our listener question segment. So couple product and funding update announcements. First up, OpenAI is rolling out its new internal knowledge feature for chat GPT team users. You may have seen a notification about this on your account.
[01:03:53] With enterprise access coming later this summer. So this update allows chat GPT to access and retrieve relevant [01:04:00] information from Google Drive. Docs like docs, slides, PDFs, word files to answer user queries using internal company data. So admins can enable this feature through either a lightweight self-service setup or a more robust admin managed configuration that syncs access organization wide.
[01:04:20] Next up, rept the coding startup known for its kind of vibe coding ethos is reportedly in talks to raise $200 million in fresh funding at a $3 billion valuation, which is nearly triple its last known valuation. Their recent momentum comes from its full stack AI agent, which was launched last fall, and that can not only write code, but deploy software end to end.
[01:04:45] So it kind of puts it in the same category as GitHub co-pilot or cursor. With a deeper focus on autonomous agents, we talked about the other week, CEO, I'm Jad. Mossad has gone as far as to say you no longer need to code in a world where you [01:05:00] can simply describe the app that you want. Runway. One of the pioneers of AI generated video just raised $308 million in funding more than double doubling its valuation to over $3 billion.
[01:05:14] Now they have an interesting creative ambition over at Runway CEO. Chris Valenzuela wants to shrink the filmmaking timeline, turning AI into a kind of digital film crew. He envisions kind of the future pace of film production to something like Saturday Night Live, where you turn ideas into a full production within a single week.
[01:05:34] they are already working with major studios like Lionsgate. As well as Amazon. Now they are backing from General Atlantic, SoftBank, and Nvidia betting that all this AI video stuff is not just AGImmick. It may be the future of content creation and filmmaking. And then last up, Sesame ai, the Voice Focus Startup founded by Oculus Co-creator Uribe, is [01:06:00] reportedly finalizing a $200 million funding round led by Sequoia and Spark Capital that values the company at over a billion dollars.
[01:06:09] Now, Sesame only emerged from stealth in February, but it has quickly gained traction for its really lifelike voice assistance. they've been backed previously by Andreessen Horowitz and are entering a heating up AI voice market alongside companies like 11 Labs. And, you know, major model companies like Open AI that have voice capabilities.
[01:06:32] Paul Roetzer: In addition to the runway funding, they also on Monday, March 31st, announced Gen four, which is their new series of state of art AI models for media generation and world consistency. They said as a significant step forward for Fidelity Dynamic Motion and Controllability, they also rolled out an image to video capability to all paid and enterprise customers.
[01:06:58] they say that Gen four is a new [01:07:00] standard video generation marked by improvements over Gen three Alpha. yeah, so like I, I think I have like a thousand credits in runway. I don't know if they require, but I've been paying for a runway license for like three years. Yeah. And I think I've generated a grand total of like five videos in there.
[01:07:16] I should probably go in and see if I have any credits I can, I can use for this one. so yeah, runway is, again, a major player, but it's getting really, really competitive. they are gonna have some major challenges ahead. There was another one, higgs field a I think it was, was tweeting all week long, sort of like sub-tweeting runway that they've made some improvements.
[01:07:36] So I I, the video space is gonna be wildly competitive this year. Yeah. It'll be interesting to see if runway, you know, sticks it out. They were definitely there early. but it's gotten very competitive.
[01:07:45] Mike Kaput: Yeah. And that Hollywood angle will be interesting to see how much they actually go down the road of using these tools in lieu of kind of regular film production.
[01:07:55] Paul Roetzer: Well, and I think, James Cameron Titanic fame, he's a major [01:08:00] investor now in stability. Stability, yep. Yeah. So there, I'm sure gonna be trying to push that as well.
[01:08:07] Mike Kaput: Okay. Our last segment is a recurring one that we are getting lots of positive feedback on, which is listener questions. So we take questions from podcast listeners, also audience members across our other various courses, webinars, et cetera.
[01:08:22] We try to pick out ones that are relevant and useful to answer for the audience. And this one is particularly important this week, given our topics. The question, Paul, is how do you prepare for AGI, short of having serious discussion of a meaningful UBI, universal basic income, basically giving people money when nobody has a job due to AGI or a new economic system, how do you actually prepare?
[01:08:50] I thought that last part was important here because it's like, okay, what do we actually start thinking about and doing about this? Right?
[01:08:56] Paul Roetzer: Oh, it was the most loaded question we could possibly pick. This is like a full [01:09:00] episode. This is, yeah. Yeah. I mean, so UBI is the lazy person's answer to this. It's what everybody's, you know, kind of throws out there with no actual plan of how that would work.
[01:09:09] Some people refer back to like the pandemic and how the government just sent some checks and people, you know, spent the money, whatever, like. There's just no precedent for it, honestly. No. And there's, you know, OpenAI or Sam Altman led a UBI study for like seven years where they gave people like a couple thousand dollars a month.
[01:09:24] And I, there's no way to, to possibly project this out. Like if UBI was even a possible solution, what's the psychological impact of that? Right? Right. It's like, okay, great, you are, you are, I don't have to make, pay my mortgage anymore and you are giving me $10,000 a month for everybody, you know, in the country or whatever.
[01:09:43] But like, you have no job or, or meaning in your life anymore. you are just gonna collect a check and just do whatever you want. It's like, okay, well we got some problems psychologically as a society. So I just feel like any time that UBI is thrown out as like, well, here, we could just do UBI, it's [01:10:00] like, okay, let's, now let's play the domino effect here.
[01:10:02] Let's go 10 layers deeper of what does that mean if you do UBI in a country. Right? So I have no idea. Like I, I don't like right now my approach to how to prepare for AGI. To stay informed. It's to try and project out the improvements in the models. It's to read the reports of other people who are trying to look to the future, like we talked about in today's episode.
[01:10:26] It's, I would say I'm, I'm very much taking the information gathering and processing approach to try and understand it. And my hope is that by being on the frontier of understanding it, we have the best chance of figuring out what to do about it. Yeah. Do I have confidence the labs are gonna be super helpful in this process?
[01:10:44] Not really. I think that they are mainly just gonna build the tech and let us figure it out. Do I think the government's gonna figure it out? no. I don't have great confidence. The government's gonna figure it out. so I honestly don't know. I wish I could [01:11:00] give people some, like really comforting answer to this question, but my only answer is we have no idea.
[01:11:05] And the thing you can do is focus on the next step you can take to educate yourself and to be in a position. To make informed decisions when the time comes because otherwise it's really, really hard to like play this out without getting overwhelmed by it. So I generally just process the information and then I say, okay, tomorrow though, what can I do about this?
[01:11:30] And I try and stay very focused on an understanding of the long term, but an action oriented short term of just taking the next logical step.
[01:11:39] Mike Kaput: Well give yourself a little credit. I know you said you didn't have an answer, but that's a pretty good answer. AI isn't the answer. That's the one. Right, right.
[01:11:48] Alright, Paul, that's another wrapped pact week in ai. thank you so much as always for breaking everything down in ways we can all understand. Just a quick reminder for [01:12:00] folks that if you haven't checked out the Marketing AI Institute newsletter, it rounds up all of this week's news, including the stuff we weren't able to cover on this episode.
[01:12:08] So go to marketing ai institute.com/newsletter. And we will be seeing you next week, I believe. Paul. Thanks again.
[01:12:17] Paul Roetzer: Yeah, and keep an eye out for those announcements from Microsoft and Google. And if Microsoft and Google are announcing something, assume OpenAI is gonna try and steal the show. So, I I would expect we're in from a, for a wild seven days in the world of ai, April tends to be a very, very busy time in, in the model company world.
[01:12:34] So buckle up for a, a crazy spring. Thanks for listening to the AI show. Visit marketing ai institute.com to continue your AI learning journey and join more than 60,000 professionals and business leaders who have subscribed to the weekly newsletter, downloaded the AI blueprints, attended virtual and in-person events, taken our online AI courses, and engaged in the Slack community.[01:13:00]
[01:13:00] Until next time, stay curious and explore ai.