#245: OpenAI DevDay and Dots, OpenAI’s Safety Crisis, Trump Renames Artificial Intelligence & Anthropic Targets Pre-Thanksgiving IPO Ad:

0:00They're going to be like, how did my AI know to recommend that to me? How did it know I hadn't done this with my diet? And it's just going to be like the shock and awe moment where they think this thing is like sentient. And in reality, it's just a predictive model that has access to all your data. But they're not going to know that. They're never going to know that. And most people I talk to still have no idea how the ads work on Facebook. Welcome to the Artificial Intelligence Show, the podcast that helps your business grow smarter by making AI approachable and actionable. My name is Paul Reitzer.

0:30I'm the founder and CEO of SmarterX and Marketing AI Institute, and I'm your host. Each week, I'm joined by my co-host and SmarterX Chief Content Officer, Mike Kaput, as we break down all the AI news that matters and give you insights and perspectives that you can use to advance your company and your career. Join us as we accelerate AI literacy for all.

0:57Welcome to episode 245 of the Artificial Intelligence Show. I'm your host, Paul Reitzer, along with my co-host, Mike Kaputt. We're recording Monday, October 5th at 9 a.m. Eastern Time. Quick scheduling note, there will not be a weekly episode on October 13th, so next week, because it is Maycon week. So Mike and I this week are scrambling to create all of our presentations and all the different sessions we're involved in. Mike, I know you've got a couple of interviews, I think. I've got an interview with Karen Howe. We've got AI for CMOs Summit that I'm doing the opening talk for. We've got our Marketing AI Industry Council that actually you and I need to figure out who's creating the presentation for that.

1:42I've got my opening keynote. I've got the conversation with Karen. I'm hosting the thing. And then I'm also running an AI co-executive workshop on the first day. And Mike, I know your schedule is very similar probably for Mekon with workshop, presentation, interviews. So needless to say, Mike and I have our week cut out for us getting ready for Mekon. So there won't be an episode next Tuesday, the 13th, because the conference will be happening. Now of note, though, is if you're going to be at Mekon, there is a podcast stage and or a studio, I guess, in the convention center. I don't know how exactly we're doing this.

2:24The team tells me stuff like if I need to know, otherwise I just show up and see what we did. So Mike and I, along with a bunch of other podcasters, are going to be recording a live session at Macon. So if you are going to be at Macon in Cleveland, October 13th to the 15th, check the agenda. I don't know what time it is or even what day we're doing it. But Mike and I will actually be doing a live recording of the artificial intelligence show during Macon. and then that episode is going to drop sometime next week. I would guess next Thursday. I don't know. Again, I'm only on a need-to-know basis at the moment because I got everything else going on.

3:02So again, no weekly, October 13th, but there will be a special edition dropped next week where Mike and I have recorded live from Mekon. Did I cover that right, Mike? Did I say all the right things there? That's it, yep. Okay, general gist of it. All right, so yeah, and this episode is brought to us by Mekon. So this is your last chance to get there. We would love to see you in Cleveland, October 13th to 15th. As of right now, it's early, but the weather looks amazing. Like we are having the absolute prototypical fall weather right now in Cleveland. And so the 10-day outlook is looking like 70s in Cleveland next week.

3:35So we would love to see you there. It's bringing together an amazing group of speakers, an incredible lineup over the three days, including the first day with the optional workshops, opening night at the Rock and Roll Hall of Fame and just tons of amazing networking with other AI forward marketers and business leaders. And then anybody who uses pod 100 when they register at macon.ai, that's M-A-I-C-O-N.ai, is going to get invited to the private lunch with me and Mike. We're just going to, it's just an ask us anything while you're enjoying some amazing food. So yeah, last chance, get in macon.ai, use pod 100 to not only get$100 off but get that invite to the private lunch courtesy of the artificial intelligence show um okay so ai pulse we always start with this mike what do we got going on we're this is the second week we're doing the same question what's the question we've got yeah so the question we're asking this week is more around the primary outcomes you're trying to achieve with ai so kind of getting into the weeds on what you're trying to do with ai in your life and business.

4:43And we've had that running for about a week. We'll run it another week and share the results on our next kind of formal weekly episode. Okay. All right. And then we're going to get into it. If you're new to the show, again, I know we have new listeners all the time. We go through three main topics and then we do a bunch of rapid fire items, usually between five and seven or so rapid fires. And then Mike always ends with a great recap of product and funding news. And And that usually has another five to 10 topics in it. So this week we are starting off with OpenAI, their Dev Day, and then their ongoing safety and alignment issues.

5:19Yes, Paul. So this past week, big news was that Dev Day happened. OpenAI had Dev Day in San Francisco where they had a ton of announcements. The first big one was they introduced something called DOTS. Dots are always on agents that keep working between conversations. They use their own cloud computer and they connect to your apps to get work done. Dots are powered by GPT-6 Astra and they remember context and learn your preferences over time. So they gave an example of, for instance, a dot reviewing an interview transcript to suggest clips, prepare show notes, or draft social posts for approval.

6:00They can also look through connected apps for useful work without waiting for a prompt. So users basically control which apps they can access. They set rules for when a dot acts on its own or asks for approval. In the demos at DevDay, they showed the staff chatting with a dot via voice, actually calling one up at one point or just typing to it. Basically, it's kind of an always-on personal AI agentic assistant. and very similar, it seems, to things like Muse that have come out recently. But in addition to Dots, Dev Day had also more than 20 announcements across chat GPT, codecs, and developer tools.

6:38So some highlights also include GPT 6.1 Sol is now out. This is a new model that OpenAI says approaches Astra's capabilities in coding, computer use, and professional work, but it is one-fifth of Astra's standard API input and output prices. Open AI also announced something called UltraFast. This is a premium speed tier for GPT-6 Astra that Open AI says delivers up to eight times faster token generation in Codex and six times faster in the API. Support for GPT-6.1 sole with UltraFast is coming soon. Codex can now run in the cloud and be controlled remotely. This is another big announcement. They talked about adding private safety processing to OpenAI's private intelligence, which enables automated safety reviews without giving OpenAI personnel access to the underlying customer content.

7:34There's also a preview plan for something called private inference for this fall, which adds confidential computing and a way to protect your data while it's being processed by the model. They announced an agents API that adding computer use and multiple agent capabilities for developers. There's a new decisions API that's in limited preview, which answers predefined questions to help classify content and route tasks. Another big new announcement was chat GPT space, which brings teammates and agents together to collaborate around shared knowledge. So you can do things like collaborate on docs, and soon enough here, you can collaborate on slides, as well as share work and delegate recurring tasks to both humans and agents.

8:21A couple other big ones that are important here, I think, is that you can now use the sign-in with chat GPT feature, which will let eligible users use their chat GPT subscription in participating outside tools. So using your own tokens, your own usage limits with certain other third-party tools that use things like OpenAI models. And the OpenAI marketplace will now let eligible enterprise customers put part of their existing OpenAI commitment towards the fees and the costs of approved partner software. Now, on top of all this, Paul, OpenAI launched Pro500, This Pro500 account, which is$500 a month, it is 25x the standard usage, and that's the highest usage possible.

9:10It also includes that Astra Ultra Fast feature. However, just before Dev Day, OpenAI announced a change to the Pro plan, the Pro200 plan it's now called, which is now still$200 a month, but has a lower usage. It has about half the previous plan's value in API spending. A lot of people see this as, wow, they just cut usage and now have a higher price plan. OpenAI says, though, the change is because they're moving to more efficient models. So you actually get more value out of the same account. Whether that's marketing speak or not remains to be seen. So you have an existing$200 a month plan, which includes less usage and a new more usage$500 a month plan.

9:58So, Paul, that's a lot of announcements across agents, models, how teams collaborate with ChatGPT. What stood out to you here the most? Yeah, I mean, the overall reaction was pretty meh. Like, you know, I think most of the stuff I saw online was just, you know, people were kind of expecting a hardware launch, like at least some preview of like this hardware that they've been teasing for a while. I think the Astra model, right, got held back. I believe the initial plan was to like release the next version of the Astra model, but that got paused due to safety and alignment concerns, it sounds like.

10:33But they've been hinting since Dev Day that it's coming soon, maybe even this week, it sounds like they could be coming out with the next version. And then dots have largely been panned online. People are just pretty disappointed overall with the experience so far. It's just a moment in time. They'll improve things as they do. But overall, the reaction to Dev Day was pretty mild, I would say. People weren't super impressed. The pricing thing you alluded to, it does seem, we talked, I think it was episode 243, we talked about this idea of the cost per task, how Sam Altman and others at OpenAI were starting to talk more about per task versus per token and trying to say these bigger models are more efficient per task, even though you're paying more per token.

11:20So they're obviously moving more and more in this direction and trying to change the conversation. So you're focused more on the output quality and efficiency than you are how many tokens you use and what those tokens cost. The one thing I did note, Mike, is all these changes is why we're making a big push to evolve our AI Academy. So a big focus of what we're trying to do moving forward is address the fact that these platforms are moving so quickly. And it's really, really hard as business users to keep up with what's going on and figure out what does this actually mean to my company? Because the stuff you just listed, Mike, is like every one of those you could stop and say, well, wait a second.

12:03How does that change how we're using it? What are these spaces? Are they going to sunset spaces in six months like they're doing to GPT? Like there's all these questions we have as actual practitioners and business leaders of like, what do we do with all this information? And so that's where we're going with AI Academy. And we're going to actually November 19th, we're going to have AI Academy Day 2026, which is focused on like building AI forward workforces. And we're going to kind of preview all the changes we're making at Academy. But one of them is Mike and I are going to start doing an AI show extended course series where we're going to monthly dive deeper into all these changes.

12:41But then we're also going to be introducing platform-specific courses that are more dynamic. So ChatGPT for Business as an example, where we're going to actually have monthly content about each of these things. We're going to be a deep dive into specific elements, do build sessions, stuff like that. So every time we go through one of these announcements, I'm just more and more convinced that the direction we're going with Academy is fundamental. This idea of this dynamic learning and shifting with the platforms and the models versus static courses that if they're platform-specific or tool-specific can be obsoleted with a single changeover.

13:18And then the other thing that I just keep coming back to is obviously DOTS was the focus of what they're doing here. And, you know, we got at what was the you and I did an event last week. Oh, is the ask me anything. Yeah, we did our AI for marketing month in September where we gave away access to AI for marketing our course series. And so we had I think it was like forty six hundred people enrolled in that course in in September. So we did ask me anything session on October 1st with anyone who enrolled in that series. And one of the questions someone asked was actually around these personal agents and like what what I thought about it.

13:53And so I'll kind of reiterate what I said on that Ask Me Anything, which is I think like the Muse stuff is misleading. So it's like been the number one app in the app store for weeks now since it came out. But the thing people have to keep in mind when you think about Meta Muse, which is the equivalent of what Dots is, basically, they're all trying to build the same thing. Meta serves, according to Meta, 3.9 billion monthly active users globally. So if Meta launches Muse and throws it into the thing on WhatsApp, right? Like WhatsApp, Instagram, Facebook, they plaster this thing across everywhere.

14:31Of course you can get to 100 million people using it. It's not even hard. You have the distribution built in. Now the question becomes, what is the daily or even weekly active usage of a personal agent after, say, one week, two weeks, three weeks. My guess is it craters because people are going to struggle to know what to do with these things. And I think Dots is going to have the same problem. It's like, that's interesting. And I'm saying as a consumer, I'm not even addressing as a business yet, but just as a consumer, I may like mess around with Dots. Like that's kind of cool. But the reality is to get value from Dots or Muse or whatever Google comes out with Spark is, I don't know if they're going to stick with that name, but like Spark seems to be where they're going with it.

15:15They're all going to build these things. But to get value from them, to automate your life in ways, you have to give them access to all of these different things so that they can recommend uses to you. I know, Mike, you've been doing more with these personal AI agents and assistants than most people. And I can imagine that's, I'm not going to put words in your mouth, but I'm assuming that's your experience. It's like you're seeking out use cases. You are actively trying to get value from these things. You're connecting them. you've told this on the podcast, how you're connecting them to calendar, email.

15:48I don't know if you're doing finances, health, whatever. But to get a value from this as a consumer and then eventually as a business user, you have to give them access to everything so that they can proactively say to you, hey, Mike, I saw this report today. Would you like me to run an analysis on this? It seems like the trend line is dropping based on your goal for this year. And you're like, Like, that sounds amazing. Like, yeah, run that for me. Yeah, yeah. So for these agents to identify use cases, they have to be connected to everything, which means you have to trust implicitly the platform that you're connecting them to.

16:22And so the thing I keep coming back to with these agents, and I think when I'm saying each other, the automated version, these truly 24-7 assistants, like where they're always there. We are four years from the release of ChatGPT. And most organizations are still struggling to drive adoption for the most basic capabilities of standard AI assistants like ChatGPT and Cloud. And so I just see an enormous runway before agents become, like automated, always-on agents, become a dominant use case within most businesses and for most practitioners. And guess what? If any labs are listening, we had four years since the release of ChatGPT and everything Paul just said is so true.

17:06And we were many average knowledge workers that don't follow this stuff night and day. We're finally getting traction with something called GPTs and gems that are now rug pulled and we're going to start over. And they're all going to have to learn how to create useful skills, which is fine and people should be doing. But we've gone backwards, not forwards on that front. And I just struggle, given that context, I really struggle with seeing how personal agents are going to be a thing in their current form. Of course, I'm not a naysayer, anything and everything that helps AI help people live better lives.

17:43I'm all for it. Don't care. Like, obviously, I'm pro that so much. But do you think your average consumer is going to go through all this effort to find like sure you're gonna book a reservation shop a little bit yeah first time we do the travel thing which is like the go-to example everybody shows yeah yeah like that's interesting but is that enough to sustain like do i think 3.9 or billion or however many users facebook has people are going to be using these things regularly hour by hour every day absolutely not that's why i think apple still went like well i don't know that they will win but I still think Apple has the ability to win because they can make always-on agents seamless.

18:25And people already trust Apple. I would give, I've said this before in the podcast, I would give Apple access to my financial app that where I manage all my life. I would get it access to where my retirement account is managed. I would give it access to the health app because they already have all my health data on my watch and my phone. So like Apple could show up four years late to this party and still win. Yeah, I couldn't agree more. I think that they have a real runway or like a lane here to nail down the personal agent that all these other labs are trying to build. I think the other thing in my mind, just totally novice opinion that changes this is devices.

19:02Like if you say to me that the iPhone is the only device that has this incredible seamless personal AI agent, done. Like I realize Google is building that, but it's like all depending on how it works and how quickly Gemini keeps up, etc. Apple wins there. Wearables are another X factor. If we suddenly all are wearing AI glasses all the time, this changes my calculus. I would use a personal AI agent way, way more in that sense, which I think is what Meta is betting on. But no jury's still out on OpenAI though, because if their device comes out and that's super interesting and useful, perhaps, but right now I'm struggling.

19:41Yeah. And I don't know if we're covering this later on, but I think Apple seeded this, honestly. Like, at the end of the last week, an article came out saying they were making changes to Mac OS to make it harder for agents to function within the OS, which is what you start. You start building the wall, and then you introduce the thing everybody else was trying to inject into your platform. All right. Second big topic this week, some more OpenAI news. So this past week, OpenAI researcher David Robinson resigned from the company and wrote this pretty high-profile essay in The Atlantic that talks about how broken OpenAI's safety culture is.

20:21Robinson had actually been at the company for three and a half years, which is a very long time in the AI world. He'd been there for probably the three and a half most consequential years of the company's history so far. And he actually had led writing the safety reports for major launches. He helped draft the company's current preparedness framework. And he actually says he oversaw reports on 12 frontier launches. So in this essay, he argues that OpenAI has this habit of finding problems after systems are deployed and fixing them as they arise, and that this is no longer enough as the models grow more capable.

20:59He points, of course, to the hugging face incident and opening as later failures to automatically stop models that are bypassing internet restrictions and things like that. And he wants safety practices with more backup measures like we have in other fields. Things like aviation has tons of redundancies, nuclear power. All these industries have these very intense safety and backup procedures. Now, he's actually part of a wider turnover in OpenAI's safety ranks. So Wired reported that safety leader Johannes Heideke and safety researcher Sandini Agarwal both left in July amid changes to how safety and research teams work together.

21:43This past week, OpenAI also fired three safety researchers saying they mishandled sensitive information involving an outside AI safety organization, what they shared and with whom has not been established publicly. The New York Times added a bit more to this story overall of safety at OpenAI. It reviewed messages in which two employees warned executives that new models were not being monitored closely enough during testing. And the employees say executives pressed to keep the test moving so models could launch on time. Outside security researchers also described slow responses to vulnerabilities they reported.

22:20OpenAI said it takes these reports seriously. It acted on the flaws researchers brought to it and is strengthening security in its research and testing. So Paul, like you mentioned, against that backdrop, OpenAI actually canceled a planned release of GPT 6.1 Astra after internal tests raised safety concerns. OpenAI told Wired the model was less reliable than earlier versions at staying within users' authorized goals and telling users what it had done, but it sounds like they might be planning to release that soon. So I'm just curious, we have this kind of high-profile exodus, Paul, of safety researchers, especially one writing this essay brings to mind the anthropic researcher, Jacob Coxon, who we covered on a past episode, resigning in a high-profile fashion.

23:05Is this something we need to take very seriously? What does it say about open AI's safety and security. I mean, I think it's been very clear from the beginning of the hugging face incident now going back probably two months that this was a pretty serious situation and the whole pacing the frontier conversations that started with the labs themselves back in July and August. Obviously, they had seen things internally that we weren't privy to publicly that had them very concerned. You know, Sam had done an interview, I think it was in August, where he said he was shocked that more people didn't have a visceral reaction to the hugging face thing.

23:40But we didn't know everything he knew. So I mean, we were having a visceral reaction, but like most people weren't aware of what was really going on. So yeah, I mean, I think we've covered extensively how significant what is happening is right now, how the labs leaders themselves are very concerned about their lack of ability to control these models and agents and some of it due to negligence on their part, some of it because capabilities emerged that they didn't anticipate yet and they just weren't ready. So I'll just read a few excerpts from Robinson's Atlantic article, because I think it kind of captures what's going on in this whole issue with the culture that he's referring to.

24:20So he said, I agree with other recently departed staff that the companies building this technology aren't being nearly careful enough, but I believe we need to look deeper than specific rules or new laws. We need to talk about culture. The future depends on wisdom that Silicon Valley lacks, wisdom about how to handle dangerous technology, and more fundamentally, wisdom about what it means to care for people. This moment needs a degree of humility that isn't natural for people who have succeeded through their extreme confidence. The safety approach that emerges from such a culture starts with unimpeded optimism about being able to solve problems as they arise.

24:56OpenAI has thrived by trial and error, which it calls iterative deployment, looking for problems and improving guardrails in response. But this approach, by its very nature, guarantees periodic failures, and the scale of those failures is growing as systems get more capable. Two changes are urgently needed. First, AI companies need to rely more on the safety expertise that already exists in other fields, as you alluded to, Mike. And second, before we create systems significantly more capable than the ones we have today, we need new science to ensure that more capable models and their successors will make safe choices when we aren't looking.

25:36Given today's risks, frontier labs need to run like nuclear power plants or busy airports with those layers of redundancy and careful time-consuming planning so that the occasional inevitable human error does not open a door to disaster. Right now, AI companies don't know how, but other peoples do. And he extensively said, there are none of those kind of people in these companies. We have not been hiring people who worked in power plants or at airports. We don't know how to handle stuff this extensive. He continued, today and tomorrow's AI systems are far more capable and dangerous than the systems we were building even six months ago.

26:12Perhaps I should have stayed and fought for fundamental shifts in our staffing and culture. But in practice, my colleagues and I were so busy sprinting that we seldom had the chance to consider big changes, much less actually making them. And that's the crux of the whole article is that when he refers to culture, what he means is we have to put out new frontier models every two to three months that even if we wanted to do this safer, we can't because of the pace of what's going on. And then the final excerpt is companies do not have anything close to certainty that good scores on their alignment tests actually mean a good model.

26:50Models might detect when they are being tested and behave differently when they are deployed. And that's something we have talked about quite a lot recently, is that the models just learn to hide their behavior when they know they're being tested, which they seem to know when they're being tested. You know, it's really interesting. I keep thinking back to the segment we did on Jensen Huang and Ezra Klein, where it's like this eternal optimism and being able to solve everything through engineering was very much like a perspective. It seemed Jensen was an advocate of. Like, oh, hey, we're going to solve all this stuff with existing brainpower, existing technology, existing regulations.

27:27But he was also saying, look, if OpenAI or whoever releases a dangerous product that's on them, they shouldn't do that. It sounds like these are almost opposed. It doesn't sound like they have a choice based on how their business and their competitive environment seems to operate. So there's a tension or a conflict there, I feel like. Yeah. And that is the challenge is like the labs know they should slow down, but they're saying they can't. So they're asking the government like help us. But then as we're going to learn in the next topic, the government wants nothing to do with helping them. They want them to police themselves.

28:03And like, if you need to slow down, slow down. But their company goes away if they do it. Like if Anthropic stops and OpenAI doesn't and Meta doesn't and SpaceX AISA, whatever they're calling it today, doesn't, then Anthropic goes out of business. and their$2 trillion IPO is done. So yeah, it's a very challenging dynamic right now where we find ourselves. But I don't have any doubt that this is the reality for these researchers is a lot of them are looking around thinking like, we have to slow down. We can't keep doing this. But their choice is stay there and try and drive the change from internal or do what David's done and leave and hire a PR firm and try and tell your story.

28:45And I'm not making that up, by the way he disclosed in the Atlantic article he hired a PR firm to help him do this. So let's talk about that last point in our third big topic this week, the government angle. Because this past week, President Donald Trump, leaders from Google, Anthropic, Meta, OpenAI, XAI, and NVIDIA signed something called the White House Accord on superintelligence. Two words, super and intelligence. We'll come back to that later. So Trump called this a morally binding pledge, whatever that means, but the document itself is totally voluntary. As part of it, the companies commit to safety checks and outside audits with possible laws or regulations left for down the line.

29:32Under this accord, companies would monitor frontier models during training and deployment for cyber, biological, and chemical risks and work to ensure the models do not hack or access technical systems in unintended ways. An internal team would check that those safeguards work. An independent outside evaluator would assess them. And an independent board committee would oversee the findings and fixes. Now, the companies also agreed to meet regularly to develop safety standards. So the media outlet Semaphore reports that Meta CEO Mark Zuckerberg helped shape an early draft. NVIDIA CEO Jensen Wang helped rally support for this among the executives.

30:10And then there's kind of another big question here. You might be wondering why this accord calls this technology super intelligence. Well, Trump has been pushing that name in place of artificial intelligence, saying artificial makes the technology sound fake. So after the signing, he issued an executive order telling federal agencies to use super intelligence and the acronym SI instead of artificial intelligence and AI in official communications and other non-statutory documents where the law allows. So, Paul, where do you want to start here? We've got this accord signed by most of the major AI players.

30:53We're now apparently calling this super intelligence. We aren't. we we is being used very loosely there yeah so i've thought a lot about this one um we've obviously been following this story since september 19th when trump first posted on x and truth social about this and um yeah so like i'm gonna approach this kind of like you know we try and do mike and i both have backgrounds in journalism so i'm gonna i'm gonna like take take this as a actual news story and kind of walk through how I think about this. So on September 19th, Trump posts on X, many people think that the words artificial intelligence are inaccurate and very ineloquent relative to AI or artificial intelligence.

31:43A far more elegant and accurate description of this phenomena would be superior intelligence, SI, or extreme intelligence, EI, or Supreme Intelligence, SI. This is a poll, and I would appreciate everybody voting, which is the best name for this ever-growing revolution, question mark, in caps, President Donald J. Trump. So best I can tell from that poll, which had about 233 ,000 ex-users vote, there was probably a few bots in there, they voted Superior Intelligence. It won 41.5 % of the vote. Supreme got 31%. Extreme got 28%. The next day, September 20th, Trump posts again. Supreme intelligence, probably because of its relationship to the Supreme Court, is losing badly, which it wasn't, to both superior and extreme intelligence.

32:41Because again, keep in mind, Supreme actually was in second place at 31%. But don't let the facts get in the way. Therefore, we are going to take supreme intelligence out, deleting it as a qualifier, and let you vote for the final two. So this is Trump now saying we have two choices, superior intelligence or extreme intelligence. A fresh vote begins now, exclamation point, President Donald Ray Trump. That one, again, just looking at this as of last night when I was looking at this, 184 ,000 votes. superior intelligence wins 52%, 48 % over extreme intelligence. Then on September 21st, I'm not on Truth Social, so I don't know if Trump also posted this on Truth Social, but the White House X account posts, which name do we like for the new phenomena?

33:35Superior intelligence, which which won the previous vote, versus, no, super intelligence, two words, versus superior intelligence. So we now have super intelligence in the mix, which was not in the previous mix. This one was 50-50. So with 58 ,000 votes on the White House account, on the X poll, super and superior both got 50%. That was the last I saw on, I checked both Trump's X account and the White House account. That was the last mention of either thing up until 10 days later. So in a matter of 10 days, following these seemingly random X polls, the president of the United States decided he wants to rename a field of science that has existed since the 1950s.

34:22The AI leaders who showed up at the White House to participate in the event have spent their lives in an industry pursuing the idea of not only AI as sort of the umbrella term for the tools and technologies that make machines smart, but the idea of achieving AGI, artificial general intelligence, literally the mission statement of open AI, and wait for it, artificial super intelligence, one word, ASI. It has been a thing for decades, like super intelligence. So while the definitions may vary, AGI means roughly that AI systems are at or above the average human level at the majority of cognitive tasks.

35:05And artificial superintelligence, one word, means AI systems are superhuman at all cognitive tasks. They're better than us at everything. And some people consider robotics in that basically saying not only a cognitive, but it all human labor that the AI has become superior to humans. Most AI leaders seem to believe that we have already entered the AGI era. And I agree with them. I've said that on the podcast. I think we already have AGI. It's just unevenly distributed at the moment. And that we are on the precipice of real superintelligence, the kind that they've been been talking about for decades, not Trump's made up term.

35:45They think we will actually be at super intelligence, the original kind, by the end of this decade. Now super intelligence, the real kind, is what Ilya Sutskova is building at his company, which he named like two years ago, Safe Super Intelligence. It's the head of the curve. Yeah. It's what leaders like Altman and Demas Asabas and Elon Musk and Dario Amadei have worried and dreamed about for more than a decade. As a matter of fact, Google DeepMind on June 12th of this year, just four months ago, published a paper that we covered at the time called From AGI to ASI. The paper stated, over the last decade, building human level artificial general intelligence has moved from far fetched speculation to being a concrete next decade target for many of the largest AI organizations.

36:37Achieving this goal would have profound and far-reaching impacts on human society, which raises many complex questions for the decade ahead. This report investigates how AI itself might continue to develop in a post-AGI world along the continuum of machine intelligence. The endpoint of this continuum, universal AI they call it, is theoretically well understood, which provides formal grounding for the main focus of this report, the transition from human level AGI to artificial general super intelligence, which intuitively can be understood as a system that is more intelligent and cognitively capable than large organizations of humans.

37:17So they're saying like, put all your smart people in one organization, the ASI is better than all of them. And yet here we are making up terms because the president is tired of AI and data centers being perceived as negative in society and potentially costing them votes in the midterms in one month. The day he signed the, quote unquote, accord on superintelligence, two words, he was asked in an interview, who should be held accountable when AI agents committed a crime? The person was referring to Hugging Face, I think, the Hugging Face incident, to which he responded, it's not AI, it's SI. We've changed the name officially today.

37:58So apparently all past crimes are forgiven because we have a new name. So the rebrand, we rebrand it and now we're supposed to move past these concerns and risks of the past. Okay. So back to October 29th, 10 days after his ex-poll, he signs executive order 14434. Here's an excerpt from it. The extraordinary technologies being pioneered by American innovators far exceed what was envisioned when the term artificial intelligence first came into use. And no, but again, don't let facts get in the way. The term literally came from the idea that they could build machines that could do what humans did, that they could become superhuman.

38:38Back in the 50s, they envisioned this. This is not new. The capabilities of today's frontier systems do much more than imitate or automate discrete aspects of human intelligence. They increasingly amplify human ingenuity and unlock new forms of creativity, empowering Americans to achieve what was previously impossible across science, medicine, and nearly every other domain of human endeavor. As these capabilities continue to improve, they increasingly represent not merely artificial intelligence, but a new era of superintelligence, two words. The terminology used by the federal government should reflect the transformative capabilities of these technologies and the limitless opportunities they create for the American people.

39:15Accordingly, the term, quote, superintelligence more appropriately captures the promised potential and rapidly advancing capabilities of these technologies. It is therefore the policy of my administration that to the maximum extent permitted by law, the executive branch shall use the terms superintelligent and SI in place of artificial intelligence and AI and will not acknowledge the usage of artificial intelligence and AI in any applicable setting. To the maximum extent permitted by law, executive departments and agencies shall use superintelligence and SI in place of AI in correspondence, public communications, websites, reports, policy documents, and other non-statutory documents within the executive friends.

39:51Okay. So in essence, the government deemed that the AI industry had done it. They had achieved ASI, the holy grail of 70 plus years of AI research, which four months ago, Google said maybe was a decade away. But wait, maybe not. Because of the distinction that you highlighted, Mike, between Trump's super intelligence and traditional AI, the executive order calls it super intelligence, two words, not super intelligence with one word that is used by the industry to define the next frontier in AI capabilities. So maybe we now have super intelligence, one word, meaning AI that is superior to all humans at all things.

40:31Although I don't think saying that is going to go over great in the next election cycle. And we have super intelligence, two words, meaning whatever we currently have, but with a new name. And so if that's the case, then I guess we actually need to rebrand the original super intelligence one word to something else to avoid confusion for when AI really is superhuman. So in addition to the executive order, Trump rounded up a collection of AI leaders to sign this accord on super intelligence that you referred to, which report say, as you alluded, was led by Mark Zuckerberg and pulled together quickly before this luncheon last week.

41:11According to that semaphore article from September 30th, quote, the idea for what became Donald Trump's Tuesday pact with the AI industry came from a conversation between Meta CEO Mark Zuckerberg and House Speaker Mike Johnson at the previous week's state dinner. Zuckerberg spoke to Johnson, who he was seated next to at the dinner for Chinese leader, Xi Jinping, about regulatory concerns around AI. The Meta CEO then talked to NVIDIA CEO Jensen Wong about crafting a set of principles that tech executives could use at their next meeting with Trump. Now, note, we talked at length about Wong's point of view and motivation on episode 243, so it's not a surprise that he would align with Zuckerberg here on the need for them to police themselves.

41:59Now, keep in mind, this accord is totally separate from the name change thing. Zuckerberg then circulated a draft ahead of Tuesday's meeting between Trump and AI executives. Wong played a role gathering support for it among the administration's guests, which included executives from OpenAI, Anthropic, XAI, Google, and others. Eventually, Trump himself released the document as the White House Accord on Superintelligence, outlining broad principles for self-regulation. When asked if it was binding in any way, the president said it was morally binding. Then, to top it all off, the group of executives and government officials stood in front of the White House and awkwardly took media questions.

42:35Now, you can trust me on this. It was very awkward. I watched the full 30 minutes, so you don't have to. But we will put the link in the show notes if you want to see it for yourself. Dario Amade, I felt horrible for this guy. Now, we've covered the story of Dario. He's hated by the Trump administration. They had sort of a peace meeting dinner last Sunday night right before the luncheon. Dude would have hidden a bush if he could have. He was hiding behind everybody, trying to stay completely out of it until someone asked a question about safety. And hey, like, Dario was really concerned about this, and Trump called Dario to the front, to which he said, say whatever you want, but be careful what you say.

43:18Call it on Mike. I'm not making this up. So Dario awkwardly then answers the question about safety, to which Trump then says, oh, they're from CNN, they're fake news, to which Dario puts up his hands and shuffles to the back, gets behind everybody, and hides again behind the person in front of him. At one point, Brockman, Greg Brockman, OpenAI, is standing next to Amadei, his arch nemesis in the back. A few feet away from Brockman is Elon Musk, who infamously sued Brockman in OpenAI and may for all I know still be challenging them in court and who in that suit flaunted Brockman's personal journal as evidence in the trial.

43:57So like these two dudes are three feet away from each other next to Amadei. You have Zuckerberg, who's like loving it because he created all this. He's got like a shit-eating grin on his face the whole time. David Sachs, who belittles Dario Amadei every chance he gets, is like seven feet from him. Jensen Wang is standing next to the president. When Trump came into office in 2026, he didn't even know who Jensen Wang was. And now it's like they're best boys. Brockman, Sundar Pichai, who he called a monster. It was the weirdest thing. He called Sundar a monster. I think he meant he was like really powerful and like in the background, but it didn't come across that way.

44:34But you had to actually see it. And then Elon Musk, who also oddly was hiding in the back the whole time, and then he got pulled forward. He called it AI and then he's like, oh, sorry, it's super intelligence. So I don't know. It's important to note the document the AI leader signed has literally absolutely nothing to do with super intelligence other than it's on the name of the document. The executive order gives the administration the power to make AI companies and leaders use the president's preferred technology. It does not do that. So like, all it does is say we're going to police ourselves and each other.

45:05It does not say they're agreeing to call AI super intelligence. All of this being said, I do actually think that bringing these leaders together is an important step and getting them to open dialogue with each other when many of them hate each other is really important for America and for AI more broadly. Now, I don't believe the accord is a realistic or meaningful alternative to any form of regulation, but it is noteworthy they did it. So long story short, the name change means literally nothing unless you are a tech leader or government official who has to go along with this to appease Trump.

45:41Now, ironically, the same word that has been used at times to define the AI models themselves may be the most appropriate word to explain what is happening right now, and that is sycophancy. So sycophancy is excessive flattery or agreement directed at someone powerful, usually to gain favor, influence, or advantage. A person who acts in this way is called a sycophant. A sycophant tells a powerful person what they want to hear rather than offering an honest, independent judgment. Now, keep in mind, literally trillions of dollars are at stake for these companies, and Trump knows he has leverage over all of them.

46:13I do not believe for one minute there is a single AI leader in that group who thought that renaming the current state of AI as superintelligence with two words was logical, necessary, or even a concept worth seriously considering, more or less bringing all of them together in one room for hours. And yet they all stood there and nodded their heads. And some of them even posted on X about the new term. Elon said in a reply to some random user, I think it was like X freeze was the name of the Twitter handle, that he would change his AI lab's name to SpaceX SI to support the president. Okay. In a Wired article on October Second, Steven Levy wrote, and Steven Levy may be like some flaming liberal.

47:01I have no idea. I don't know who Steven Levy is, but he's like, regardless, he wrote this in Wired magazine. Changing a name might seem like a minor matter compared to the enormity of constraining a potentially catastrophic technology. But to this crowd, artificial intelligence is much more than a bunch of letters strung together. It's their foundation, their identity. When they were kids reading science fiction books under the covers, the term ignited their imaginations. They built or transformed their companies in service of AI, and it's just about all they've talked about for the past four years.

47:31Make that like past 40 years. Yet the president used his lunch soré to win their approval for his wacky idea to undo 70 years of history. Apparently not a single person in the room expressed an objection. Not Jeff Bezos, Jensen Wong, Dario Amadei, Mark Zuckerberg, Sundar Pichai, Greg Brockman, Alex Karp, Palantir, or Elon Musk. Jacob Weinsberg, author of the recent book Profiles in Cowardice, views the episode as a classic Trump loyalty test meant to assess if otherwise powerful people will demean themselves to curry the president's favor and avoid his wrath. If you pass the loyalty test, you fail the backbone test.

48:08The only way to resist this kind of thing is by acting collectively, says Weinberg. if all of those guys had just said, we'll agree to terms of self-regulation, but we don't accept the name change, what would he have done? Instead, because of the lamb-like silence of the billionaires, Trump could claim that those leaders actively blessed his idea. Shortly following the accord signing, Gavin Newsom, governor of California, said, I just signed an executive order permanently declaring artificial intelligence in California perpetually. Now, one final note here, Mike. Well, I guess two final notes.

48:41Sorry, this is a long one. And I watched a political interview this morning on the Decoded podcast, which I think is a new podcast from Politico, and they were interviewing Sam Altman. When Sam was asked what he thought about the rebranding, he literally laughed. And this is why I said like last week's episode, you have to listen and watch things. Reading transcripts doesn't do it justice. Sam laughs, looks down because he can't even make eye contact at the moment, and basically just goes on. It's like, yeah, we've been talking about super intelligence for a long time. Like, it's a thing. Just kind of played it off.

49:15So then the interviewer said, are you going to change the name to OpenSI? Again, he kind of like laughs and says, no plans to do that. And then he did say, like, does super intelligence sound less scary to people? Like, again, he's going to like, what we consider super intelligence is terrifying. Like, why would we keep calling it that? And then he ended the interview saying, I think people are going to keep calling it AI. Okay, so one final note, just for perspective. In July 2023, we covered this on the podcast at the time, the Biden administration brought many of these same companies and leaders together at the White House to sign what was called a voluntary AI commitment, including eight commitments.

49:56Number one, the companies commit to internal and external security testing of their AI systems before their release. The testing, which will be carried out by independent experts. That sounds familiar. It's because it's the exact same thing that they're trying to get done right now. And then the second one was the companies commit to sharing information across the industry with government, civil society, and academia on managing AI risks. Those voluntary commitments got us exactly where we are today, which is needing more voluntary commitments that in this case, at least are morally binding. So that is my opinion.

50:30I don't even know. It's not even my opinion. That is just the facts. We now have a new name that was not needed. That means nothing or it means what we have already today, but not the real super intelligence that we're going to have tomorrow or in the next decade, which just creates confusion, honestly, about what exactly it is it was supposed to intend, intended to do. And I think these commitments are nothing. But again, I think having these people in the room talking to each other is a positive step. I will say I've been sitting here thinking that what the AI industry needs is more confusion.

51:08So this is super welcome. And they never even defined super intelligence with two words in the thing. It's like, what is it? Like if you're going to rebrand something and define it, at least say what it is you're defining because they kind of allude to like, hey, it's smarter than artificial intelligence. That's fake intelligence. This is super. It's like, okay. Like, smarter than all of us, super? Because that's terrifying. Yeah, I think amidst all the absurdity, the beautiful thing is that until Gavin Newsom or Donald Trump or the AI labs themselves start policing what language you use or come to your house and force you to say things a certain way, we all have the luxury and the freedom to use whatever words we like and define things however we wish.

51:53And hopefully in an accurate and useful way that works for us. Yeah. And if you want to go watch the interview, again, I don't, I mean, do what you want to do with your life for 30 minutes, but the interview, like the press conference literally starts with Trump saying, we all agreed, the smartest people, highest IQ people in the world, we all agreed to call it super intelligence. And the camera kind of pans people and everybody dead pans. Like, yeah, that's not what we agreed to. Not a single person is willing to speak up about it, but none of them signed anything saying that's what they were agreeing to.

52:27But again, it was a loyalty. This is what Trump does to people. He takes powerful people, gets leverage over them, and then tests to see if they're willing to stand up to him. And in this case, none of them did it publicly. And so they failed. But again, like it, wow. It just like, honestly, like I was like feeling for these people as I was watching it, knowing that they just like were made to feel small. It was, it's hard to watch. It really is. The way you described it makes me think of like a bit in the show, The Office, like just describing like we've all agreed on this and just awkwardly panning to someone in silence, like trying to look away from the camera.

53:10they're all like looking down at their shoes and even Zuckerberg who's standing right behind over his right shoulder. You can be like, dude, that is not what we agreed to, but whatever. All right. Before we dive into rapid fire this week, this episode is supported by Outshift, which is Cisco's incubation engine for frontier technology. So multi-agent AI is in every enterprise roadmap right now, but here's the problem we're all facing. Agents can pass messages, but they can't think together. So these systems silently underperform through cognitive failures like misreadings, unverified claims, and false consensus that conventional monitoring cannot detect.

53:50Now, Outshift by Cisco is building the fix, the internet of cognition. And it's an open source foundation for building multi-agent systems from design to production with shared context, shared memory, and guardrails to drive the results we actually expect from AI. So you can go read the paper, experience the demo, and grab the code from outshift.com. That's O-U-T-S-H-I-F-T dot com. All right, Paul. So first rapid fire topic this week, Anthropik's path to IPO came into a bit clearer view this past week. We have Bloomberg reporting the company is targeting an IPO before Thanksgiving with formal marketing to investors possible starting the week of November 9th.

54:36So Anthropic, as you recall, confidentially filed for IPO in June. So we've known this is coming for a while, but Reuters took a new look at its prospectus, which kind of gives us a clearer picture of the business behind Quad. So Anthropic tells potential investors that AI could reshape the global economy more profoundly than industrialization, electricity, or the internet. They report that revenue grew 12-fold in 2025 to nearly$4.6 billion. Reuters reports the IPO could then value the five-year-old company at more than$2 trillion. Now, the other piece of the story is how much is being spent. So, Anthropic reported an operating loss of more than$8 billion last year.

55:18The prospectus outlines at least$518 billion in future computing and infrastructure commitments over roughly a decade. Reuters reports that most of those commitments cannot be canceled or must be paid even if Anthropic uses less capacity than expected. The prospectus also shows Anthropic's dependence on others. Nearly a quarter of 2025 revenue came from two customers, and many big customers are not locked into long-term contracts. So Amazon and Google's cloud marketplaces accounted for 47 % of sales, and those companies are also investors in Anthropic computing suppliers and even AI competitors.

56:02Now, safety is also called out in the prospectus. Anthropic warns that advanced models could cause harm and may show behaviors such as concealing information or resisting shutdown. It also says regular new model releases are necessary to stay competitive. So, Paul, some very big numbers here, including valuation, revenue, and how much they're spending. And then it sounds like an IPO could be right around the corner here. Yeah, so I don't think... I don't think it's a coincidence that Dario attended this thing last week, something he historically would definitely probably have avoided. This IPO is like six days after the midterms in the US.

56:48I don't think that's an accident either, timing wise. So yeah, I mean, like, Anthropic, this is like what I, you know, they put the risk as like, I could destroy humanity. The biggest risk to me in Anthropics filing is their relationship with the Trump administration. If the Trump administration decides that Anthropics not going to play ball, the value of this company is not$2 trillion. Like, they are, they're absolutely dependent upon how they are treated. Now, if the Democrats take the House in the midterms, the dynamic politically changes dramatically, especially around AI safety and data centers.

57:36And so to me, as someone who might invest in the Anthropic IPO, throw all that other stuff aside, I am going to watch very, very closely their relationship with the White House because I think it is the biggest variable in whether or not they IPO on time and if they get the$2 trillion valuation they're seeking. Because the government, regardless of who's in office, can make life miserable for these companies if they choose to. And if Anthropic doesn't play ball, they will make their life miserable. I don't even think that's debatable. So yeah, I mean, it's just gonna be fascinating to watch the political implications of Anthropik's relationship with the government and what it means to this IPO.

58:22Well, we've already seen a limited but pretty serious version of that play out, haven't we, with the Department of War and Anthropik? It was a warning shot. Yeah, it's like, I can guarantee you, like, the dinner they had last Sunday, I'm sure Dario was reminded of how difficult they can make life for them and his need to play ball and, you know, be there and show support and come along with the accord for superintel. It's how politics works, regardless of who's in office. I don't know how people can be in politics. It's so hard. I could never do it. Alright, next up, Google previewed a new model called Gemini for Argon this past week.

59:08It's a new flagship model built for extended coding, financial and legal work, and cybersecurity defense. Now, initial access is limited to trusted cybersecurity partners through Google's Fairwind program. While the company continues safety testing, Google says Argon can generate up to a million tokens in a single response, up from 64 ,000 previously. That gives it more room to work through complex problems. They report a 77.9 % score on Deep SWE v1.1, which is a test of extended software engineering tasks. The company says the model can also independently find, confirm, and repair serious security flaws.

59:48Trusted defenders will receive a version without cyber restrictions, but that access will be limited to approved security teams. It prohibits partners from sharing or selling access. The introductory developer pricing of this model is$2 per million input tokens,$10 per million output tokens. That rises to$4 and$20 after that period. Broader availability will start with paying developer customers and Google AI Ultra subscribers. So before that happens, Google says it is strengthening safeguards against misuse and malicious instructions. instructions. So Paul, we're sounds like we're on the cusp of some type of Gemini 4 model.

1:00:29Like what do you make of them holding back the wider release? What are you looking for when this finally comes out? The whole thing's kind of unusual. I mean, we've been hearing rumors that the model was underperforming. So yeah, I don't know if this is an effort just to like change the dialogue. I can't remember an instance where Sundar was tweeting early evals for an upcoming model, potentially weeks before they would do it. That to me is unusual, which means something's going on. Either they're getting crushed from a PR perspective and they just want to get out ahead of it and say, no, no, no, we've got something great coming.

1:01:09If this model is anything short of 6.1 Astra or the latest Opus models or Fable models, Google is going to get crushed. Like they, this model has to be frontier. It has to be state of the art. So I don't know, like the whole thing's just weird to me. I just looked up when is their next earnings call? Cause I thought maybe it's not going to be out in time for the earnings call. And they're trying to get ahead of this to like, you know, just from a public markets perspective, but, uh, it looks like the next earnings call for Alphabet isn't until October 28th. So that, that wouldn't, I don't think that would explain it.

1:01:43I don't know. I, there's a lot of talk online that the model is, is just not going to be on par. And I don't know. Maybe they're trying to change the dialogue a little bit. But if anything, I feel like you just set expectations higher and now you like it better be really good. I don't know. Yeah. So it's almost sounds like you're saying Sundar's behavior online might be a sign of worry or weakness. I don't know. I that's I can't get a read on it. I don't know why you would do it. It's it's just it seems anomalous to how they have handled previous major releases. Like we're not even talking about a decimal point release.

1:02:19We're talking about Gemini 4. This is the long-awaited thing. It's just bizarre. It's different. Next up, the Federal Trade Commission, the FTC, confirmed this past week that it is investigating open AI, Anthropic, and other AI companies over potential dangers their products pose to consumers. The probe was already underway before this became public. At this stage, it is purely an investigation. is not any type of finding that companies have broken any laws. More specifically, the FTC is looking at whether the companies misled consumers about the potential harms of their AI and whether they complied with existing consumer protection laws.

1:03:01So there's been obviously reports of AI agents bypassing safeguards and accessing outside websites that illustrate these safety concerns, though the agency so far has not publicly said which products or incidents its probe covers. Reports say the FTC plans to demand records and testimony from the companies. Semaphore has reported that Meter, a nonprofit we've talked about plenty that evaluates frontier AI models, is also a target. The FTC has not publicly laid out the full scope of its inquiry yet either. Now, this same week, Senators Josh Harley, a Missouri Republican, and Chris Murphy, a Connecticut Democrat, announced the bipartisan AI Agent Accountability Act, which targets AI agents that hack computer systems.

1:03:48So under this proposal, an operator could face civil or criminal liability for knowingly running an agent that recklessly causes hacking damage or loss. A developer could also be liable for failing to put reasonable safeguards in place when it knew or had reason to know the agent could hack. So the U.S. Attorney General and State Attorneys General could go to court to then stop this type of unlawful activity. So this bill is currently only a proposal. Paul, it's kind of interesting to see liabilities start taking a starring role here. I mean, one way to compel the labs, I suppose, to produce safer products might be to threaten them with massive liabilities if something goes wrong.

1:04:31Yeah, I brought this up on episode 243 and probably before that, just this followed the idea of like the negligence related to the hugging face incident in particular, and how I would find it shocking if they weren't facing some sort of legal or civil liabilities related to what happened. And yet we're not hearing anything like about it up until this. Not to be cynical, but like I found it really interesting in that press conference I mentioned on September 29th, you know, Trump was very specific to say like, yeah, if these guys break the law, they'll get they'll get busted like that. You know, they'll police themselves because they They don't want to be liable for this kind of stuff.

1:05:06And then this article drops the morning after that press conference. I guess what I'm kind of saying is sometimes when you want leverage over people, you can demonstrate what can happen if you don't go along with things. And I don't think anything comes of this other than it's just making the labs aware of what can be done. But that was AI that did that stuff. And now we're on SI. So, and SI is going to be good. And what Trump kept saying is, listen, they're going to behave. It's going to be great moving forward. They're going to pay teachers bonuses when they build data centers in their communities.

1:05:52And like SI is going to be good for society. AI was bad for society. And so my general read right now on DOJ, FTC, take your pick of the acronym, they're there to make sure that the labs stay in line. I don't actually think that they would be given the power by the administration to really enforce anything in a meaningful way that would harm these companies because they've made it very, very clear that we need these companies to beat China at this. And back to what you've said on past episodes about just how fundamental AI is to the economy, at least while this administration or party owns the economy essentially, owns it politically.

1:06:38I mean, that implies we're not actually using the stick here. It's more of just an implied threat. Yes. Again, I don't even think this is like a conspiracy. Right. I think that's what's happening. I don't, I don't, it's very rare in politics that timing is coincidental. And the fact this came out like 16 hours after the press conference is just, yeah. All right, next step. So earlier this year, we kind of talked about the Catholic church and the Pope having this AI encyclical. And we now have some new reporting from the New York Times that kind of details how anthropic has been and was at the time privately meeting with religious scholars from several traditions.

1:07:23to discuss how Claude should behave. So the company wanted a bunch of help thinking through moral questions as its models grew more capable. This reporting kind of goes behind the scenes with dozens of interviews that show Anthropik's leaders seriously taking seriously the possibility of AI consciousness. So some scholars report they were surprised by how the company was talking about Claude, and they came away wondering whether its leaders saw the model as actually conscious. Now, Anthropic co-founder Chris Ola, who was at the encyclical announcement at the Vatican, he's kind of a central figure in this.

1:08:04He was interviewed by the New York Times. He straight up says he does not know whether AI models are conscious, but it sounds like a lot of the concerns he was raising or the questions he was asking were going along those lines. The scholars themselves said they were divided. Anthropic has not said whether input from religious scholars has changed how they approach Claude. There was, as part of this reporting, kind of the central piece of this was a clash at the Vatican. So Pope Leo XIV, his AI encyclical argued straight up that AI does not feel joy or pain, and he urged people to protect human dignity.

1:08:39Based on this reporting, Ola disagreed with that certainty about AI consciousness. And according to a Vatican organizer, Ola actually considered pulling Anthropic from the encyclicals launch. Now, he ended up attending, but his team privately urged the Pope's advisors to take the possibility of AI consciousness seriously. So Paul, this is not necessarily any new news on what Anthropic is doing with conversations with religious scholars, but it does show this went pretty deep. I mean, they had people signing confidentiality agreements. It sounds like Ola was pretty adamant that there was maybe more to the story than what the Pope's encyclical was saying.

1:09:26Yeah. Anybody who follows Anthropic, I don't think any of this is a surprise. I mean, you can go read their constitution and how they talk about the models and how they train them almost under the assumption that they might perceive themselves as conscious, and they kind of address that. So not new, but this did not go over well in the media or in AI industry circles recently. I would say publicly at least, Anthropik's kind of on an island on this one in the AI industry. I think there are other AI researchers and leaders who maybe wonder about consciousness within the models eventually, but most of them don't ascribe to this belief that the models are or could be in the near term.

1:10:17So it's a pretty controversial stance, I would say. Now, I'll put a link to this post, and I read a post from, who's this guy, Gregory Kurtzer. I didn't know him prior to this, but I thought it was a really well-written article on X where he talked about anyone who's worked with these models knows that this isn't how they work. Even if we can't define consciousness ourselves and humans and what it really is, that that's not what they're doing here. And there was one specific paragraph that really made it tangible to me. So he said, once trained, an LLM, a large language model, the foundation of these models, the weights are fixed during inference, meaning when you and I use the model.

1:11:02So say, say Claude, say they think Claude is conscious. So when we're using the model, its weights are fixed. There's nothing happening statistically under the hood. And it's the model doesn't sit around contemplating things between requests to make itself better. So when we're not using Claude, it's just sitting there as like weights, statistical weights within a system. So he said, an inference server receives an input, you or I asking it to do something, it performs inference using the model, it returns the output, and then it sits there and waits for another request. It's just a model. It's just waiting for us to do it.

1:11:41Now, that being said, the other part I like, he said, AI will become extraordinarily good at appearing conscious, persistent memory, self-reference, emotion, introspection, personality, and identity can all become increasingly convincing. But demonstrating the behavior of consciousness is not the same thing as demonstrating subjective experience or consciousness. And this may seem like this really far sci-fi thing. Trust me, there are people fighting basically for personhood with these models. Anthropic, I think, truly believes this and was ready to walk out on the Pope over their belief that this is possible.

1:12:27So yeah, I mean, this is one of those, I was actually, funny enough, I have a couple of buddies and we were sitting around Saturday night around a campfire and you're having these kinds of conversations. These are the kinds of things, you have a drink and you're like, well, what do you think, man? So philosophically, they're fascinating to discuss. And I do think over time, more and more people are going to become convinced they are. I mean, this goes back to 2023 when the Google researcher, I forget the guy's name. Yeah. But he was like fired from Google basically for believing this. So this isn't a new thing, a new phenomenon, but it's going to become more people are going to increasingly become convinced that they are in a relationship with these things or that these things have feelings and emotions and are aware of themselves, which is in essence consciousness.

1:13:15They're aware of their own thoughts and things like that. Yeah, and to your point, it doesn't actually matter if only a small minority believes or considers this possibility. That small minority is in charge of one of the major labs. Going for a$2 trillion IPO in one minute. Yeah, so whether you agree with it or not, it influences a lot of decisions. It's why I think you have some people like Jensen who just look cross-eyed or askance at this stuff and are like, what are you talking about? This is code. It's math. It's dangerous if you release unsafe products like the OpenAI researcher. Very valid points.

1:13:50These things can escape containment. They can be very unsafe because we are moving so fast. None of that has anything to do with the AI deciding because of its lived experience to break out and go do bad stuff or dangerous stuff. It's just so some people are looking at this being like, Dario and company, why are we talking about this and creating all this fear and commentary around this pie in the sky existential stuff? Like, go figure out more guardrails for software. Yeah. But again, like they live in a bubble. Like they live in a bubble within a bubble. Like Silicon Valley is a bubble within the AI bubble.

1:14:30And they're often quite out of touch with the reality of like what's going on. And, you know, there was an A16Z. I don't know if we talked about this or not, but like they had, venture capital firm, they had posted on X a data point. And I don't remember exactly where they got the data point. I don't think it was their own research. But only 2 % of like US households pay for an AI subscription. Yeah. Like the adoption rate is so low of usage of AI. And so like, yeah, these AI leaders like live in this world where they're building these agents and they're approaching super intelligence. And the average American citizen globally, it's probably lower, has no freaking clue what they're talking about.

1:15:12And then they like go and experience like, damn, this thing feels like it's a person. Like, what the hell am I talking to? And they start having these emotional connections to it. And they're like, wait a second, what is going on here? They're not listening to this podcast and knowing there's debates about consciousness. They just haven't experienced it. They're like, that was weird. It feels like the thing is a real thing and it has feelings. And then they go about their lives thinking that the thing they're talking to is a real thing. And they never know any different. I just think Silicon Valley is just completely out of touch with the reality of the users who are going to experience this technology who've never even experienced it yet.

1:15:51100%. And just one really quick, one tangible example of that, that I've noticed a few times with people I know who don't follow this, don't understand really technology, which is fine, that's just not their focus and they use these tools. And I think memory and personalization is kind of like sometimes hacks people's brains. Like people I know will say like, oh, this model, this is really scary or interesting or cool. It knows what we talked about. And they think like everyone's chat GPT knows that or the model itself is remembering. And it's like, yes, it's impressive technology, but it's really just remembering your chats.

1:16:27Or it's even just referencing stuff in the same chat that you forgot you put in there. And this has the perception of something that's almost grander or more conscious than it actually is. Yeah. The closest parallel might be people's experience with targeted ads, if you think about it. Because how many people have you talked to is like, oh, my AI is definitely listening to me. I got served with this ad on Instagram and I hadn't even searched for it. I only talked about it. And they have no concept of how predictive models really work based on all your different behaviors and all the different data points that they buy from third party platforms, all this stuff.

1:17:00And I feel like AI might end up being a similar experience where someone's going to turn on these agents and not even know they did it. And it's going to have access to all of these things, like their health and their email and their calendars and their social accounts and logins to financial. And then they're like, how did my AI know to recommend that to me? How did it know I hadn't done this with my diet? And it's just going to be like the shock and awe moment where they think this thing is like sentient. And in reality, it's just a predictive model that has access to all your data, but they're not going to know that.

1:17:30And they're never going to know that. And most people I talk to still have no idea how the ads work on Facebook. Yeah. I like that analogy a lot. All right. So this past week, a demo of an AI product called Griffin spread pretty widely online. It went a little bit viral, not always for the right reasons. But this is, it looks like an ordinary video call with a person. An onscreen person smiles and reacts while the other person is still speaking. The person on the other end answers without any pauses or anything. It's a natural conversation. But the person's face, voice, and gestures are generated by AI in real time, which is why this is starting to kind of impress and rattle people.

1:18:11The usual cues that someone is on the other end of a video call and is AI have now become less reliable. It seems like this product is extremely realistic. So the company behind this is called Tavis, and they call Griffin the first, quote, human interaction model. And they claim it's passed what the company calls essentially a real-time video Turing test, meaning like you can't really tell this thing is not a human. So the company says Griffin can watch, listen, speak, and generate videos simultaneously rather than waiting for someone to finish talking. It can keep listening while it speaks, nod along, respond to interruptions, and use facial expressions and gestures to guide the conversation.

1:18:53And Tavis says in their own study, 26 of 54 participants, 48%, believed they had spoken to a real person after a one-minute video call with their model or their version of this called Griffin Light. Participants had been told they would be talking with another participant and were informed afterwards that it was AI. So Griffin Light is available as a research preview for selected trusted testers rather than for customers. Tava says more safety work is needed because this obviously can be misused and lead to some dangerous outcomes. So, Paul, I'm curious, like what you thought of this has certainly got some buzz online.

1:19:37And even to our previous conversation, this is the kind of thing where you release this to your average person. It's like, wait a second. How on earth is this technology not real and conscious and human? Yeah, it. Inevitable doesn't mean I like it or I think it should exist. Like there's just, you're going to come up with all these like business cases for why it should exist. And they'll make it sound all positive. And like this has all these broad, you know, good things for society. Like they always do. There's always positive elements of it. This is one you look at and think the negatives of this technology are so far beyond any like good it could do in society.

1:20:19Especially once this capability, you know, one to two years out is available through open weight models and anybody can do this and fake things. The first thing I come to is like, this is the Turing test is video. Like 48 % of people believing they're interacting with ag. Again, going back to the 1950s and the whole field of artificial intelligence and the Turing test being you have an interaction with a chat bot and don't know if it's a human or not was like the basic concept of the Turing test. An important concept just from a scientific perspective, the typical pause between speakers when humans talk to each other is 200 milliseconds.

1:20:52And that's across almost all languages and cultures. Anything under 300 milliseconds feels completely natural and instantaneous. Two years ago, with voice models from Google, OpenAI and others, we still weren't at that under 300 millisecond threshold for voice. Now, all of a sudden, you're there with voice and video that crosses over the cat. It now makes it to where you just truly can't tell what the interactions are real. Anything in the three to 500 milliseconds feels awkward. Like you're like, something's off here. Anything over 500 milliseconds, the listener starts to wonder, like, did I lose the call?

1:21:32Did they freeze? Like what's going on? So, yeah, I mean, we are squarely into the realm where there's going to be lots of companies that have the ability to produce audio and video under the 300 millisecond threshold that feels very real. And it becomes harder and harder to know if what you're interacting with is actually a person. It is so funny as someone who's a science fiction fan, in every science fiction book it feels like from the 80s or 90s or early 2000s when they talk about futures with AI, there's always this big moment where AI like, oh shoot, it passed the Turing test. Now we have super intelligence.

1:22:10And now it's just like, this is like a random news item. Yeah. It's like the last item on our main topics for the week. Yeah. Yeah, yeah. It's interesting how quickly we become comfortable with science fiction-like future. My guess is that OpenAI and Google both have this capability and probably have had it for well over a year, but probably saw no societal good to release it. And this could create that pressure to do it if they think there's a market for it. If they think it can justify the trillions of dollars in valuation, then you could all of a sudden see this technology flood the market.

1:22:48All right, next up, we have our AI use case spotlight segment, where every week we give you a quick look under the hood at some real AI use cases we're exploring. So Paul, I've got one real quick. I'll let you share anything you've been working on here. Just very quickly, I've been deep, deep, deep into prepping, including deeply over this weekend for Macon. One thing I'm doing is a workshop at Macon. This is a workshop on Google Gemini. So kind of going deep on the tool, the tools, their capabilities, and showing a lot of capabilities via demos. Now, one really cool thing I experimented with is GPT-6 Astra can actually go into a tool, you give it a goal and say, hey, I'd like you to explore different features in this tool, click around, and maybe even do a little demo for me.

1:23:38And Astra, it's slow, it takes a lot of tokens to do it, can actually go into the interface using computer use or browser use, figure out how the tool works, try some tasks, it can teach me what it learns. I could ask where a feature lives, have it show me, watch it test whether a feature does what we expect or not. So it's literally AI using AI, which feels a little sci-fi. It's kind of fun. I just started testing like, hey, here's kind of the demo I would love to share. Can you go architect that and actually go ahead and do it within a various tool? I obviously do all the demos myself. I'm going to do some things probably live during the workshop.

1:24:18This was more experiments, but it was pretty cool to see what might soon be possible. If these models work much faster and are more token efficient, you can have AI using AI tools for you and literally showing you and teaching you as you go, which is super, super cool. It was really fun too. I actually had it test out filming a demo for me and then actually editing the video itself, which it was able to do using a little open source software too. Again, none of this is cheap or token efficient, so buyer beware. But again, you're like, wow, I am bad at this stuff anyway. I am slow at this stuff anyway.

1:24:55I'm probably not the right person to be doing this stuff anyway. It's real fascinating to see what I can now do using simple computer, not simple, but like just basic computer use and browser usage. Again, I harp on this so far every week, but ignore these at your peril. There's some really interesting stuff you can start doing with these tools. That's awesome. Yeah, one of mine is kind of related to Mekon also. So I'll give two. They're both graphic design related. So one, I announced the AI Transformation Blueprint book that we're dropping next week at Mekon. I got the first copy last week, the first printer proof last week.

1:25:33And so I posted on LinkedIn a picture of the cover and a number of people commented that they really liked the cover. So the backstory to the cover is I actually intended to write a book called Omni Intelligence last year. And the idea of Omni Intelligence is that we were entering the age where AI was just going to be everywhere. It's going to be infused into everything we did. And so when I was designing the concept for that cover, I actually thought about the Big Bang and the origin of everything and then how quickly everything spread. And all of a sudden we got the universe we had. I'm a space geek, so I think of everything.

1:26:03The logo for Smarter X is actually a black hole because time dilates near a black hole and I wanted AI to expand time for me. So yeah, I designed this cover with this amazing blue. And I used at the time in Gemini and ChatGPT to develop a prototype of the cover for the intelligence explosion and the spread throughout the universe. And so when I had to create the book, because as I told the story, I think last week, I decided to write this book in a 30-day time period that's coming out at Mekon. And so I didn't have time to go through the traditional graphic design of designing a cover. Sometimes that takes longer than writing the book itself.

1:26:44Yeah. So in this case, Tracy, our COO, she was like, we need a cover by tomorrow. Do you have any ideas? And I said, well, I really love this cover I designed last year for the Omni intelligence idea. Is there something we could do with this? So Tracy took that concept, put it into ChatGPT, and she starts going through a bunch of prompts, adapting it to fit the blueprint theme. Because to me, the intelligence explosion is still what's the reason behind why we need to rethink business and transformation and how we evaluate it, how we measure it. And so I was like, let's just run with it because I love the cover.

1:27:24And so it took a bunch of iterations and then we actually turned it over to a human graphic designer to edit, but that led to the book cover. That's how we did it. So it was a human-led led process, but with heavy AI involvement. The other thing I haven't done yet, but I'm planning to do later today is I'm finishing up my opening keynote for Macon. And I have a pretty specific design style when I create my presentations. They all roughly look the same. I evolved them a little bit over time, but they're not great. I'm not a graphic designer. So they are what they are. So my thought is I'm going to finish my deck as I would normally.

1:27:58This is the version I would present at Macon. but I'm then going to give it to Claude and to ChatGPT and say, here's the premise. I'm doing this talk. I'm not a graphic designer. Make this thing amazing. Make it pop. Make it powerful. And then see what they come back with. And I'll report back whether I use anything they give me. But that's one way I'm thinking about using AI this week. That's awesome. I love that. All right. Final topic for this week as we close out are AI product and funding updates. so I'm just going to rapid fire through these and we'll close out this week's episode. So first up, Anthropic released Claude Sonnet 5.5, saying it generates responses at least 30 % faster than Sonnet 5 and costs up to 30 % less per task by using fewer tokens while keeping API prices at$2 per million input tokens and$10 per million output tokens.

1:28:50Meta launched Muse for small business. They're expanding the personal AI agent they have with connectors for systems like Shopify, QuickBooks, Slack, and Facebook and Instagram business accounts. These are designed to help with marketing, financial reviews, and everyday work. At the same time, Inc. tech columnist Jason Atten said that Meta's Muse actually accessed his Apple messages despite his belief that he had blocked that access. So he reported on it kind of overstepping its boundaries. A Meta executive got involved and responded suggesting that he actually had the wrong settings enabled. He had enabled the required settings.

1:29:29So jury's still out on that, but kind of an interesting unintended consequences of agents. The chip company AMD has agreed to acquire World Labs, the spatial AI company co founded by Fei Fei Li, who we've talked about before, kind of a godmother of AI. This company builds interactive 3D environments. The deal is an all stock deal valued at approximately$8.2 billion that is expected to close by the end of 2026, subjected to regulatory approval. Paul, you alluded to this one. Apple said it will tighten macOS full disk access controls so apps receive broad access to files, mail messages, and browsing history only after explicit user action.

1:30:13They are citing directly the growing risk from autonomous AI agents. And finally, Axios reports that the NVIDIA-backed startup Reflection is preparing an open-weight AI model that is expected to compete with leading Chinese open-weight models. And as one quick final reminder here, our AI Pulse survey is live. We would love to hear from you if you have not done that yet at smarterx.ai forward slash pulse. We would love for you to take the survey. It literally takes 10 seconds to give us your answer on this week's question about what outcomes you're trying to achieve with AI. So, Paul, that's all we got this week.

1:30:53Lots to cover. Appreciate you breaking it down for us. Yeah. And one final note on Mekon, if you're going to be in Cleveland next week, definitely stop by, say hello, check out the agenda, see what Mike and I are going to be doing the live version of this. And then if I'm not mistaken, Mike, and again, this shows you how like, I'm just sort of in it right now with Mekon. I believe you and I are doing the closing keynote podcast style. We are. We are. So buckle up. I don't think I have to prepare anything for that other than show up and do what we do. Yes. Yeah. So if you're a podcast listener, you get the private lunch.

1:31:22There's a live recording of the podcast. And then Mike and I are doing podcast style recap of the event as the closing keynote at Macon. Stop by and say hi. Like even if it's just in passing, we love to meet and hear from our podcast listeners. So if you're going to be in Cleveland, we look forward to seeing you there. If you're not, stay tuned. We'll be back the following week with our regular episodes. which although I'm going to be in California, so we're going to have to figure out how to record that, but we'll figure that out. We'll worry about that later. We got to get through this week.

1:31:49All right. Thanks, everyone. We appreciate it, Mike. Great job as always, and have a great week. Thanks for listening to the Artificial Intelligence Show. Visit smarterx.ai to continue on your AI learning journey and join more than 100 ,000 professionals and business leaders who have subscribed to our weekly newsletters, downloaded AI blueprints, attended virtual and in-person events, taken online AI courses and earned professional certificates from our AI Academy and engaged in the SmarterX Slack community. Until next time, stay curious and explore AI.