Transcript
Intro
0:00 · We've all known that they knew they were stealing it. Their internal communications indicated that they knew they were stealing it. Every lab knew the other lab was doing it, so they were going to do it. That is not debatable.
0:15 · Welcome to the artificial [music] intelligence show, the podcast that helps your business grow smarter by making AI approachable and actionable.
0:23 · My name is Paul Ritzer. I'm the founder and CEO of Smarter X [music] and Marketing AI Institute and I'm your host. Each week, I'm joined by my co-host and Smarter X Chief Content [music] Officer, Mike Kaput, as we break down all the AI news that matters and give you insights and perspectives that you can use to [music] advance your company and your career. Join us as we accelerate AI literacy [music] for all.
0:51 · Welcome to episode 237 of the artificial intelligence show. I am your host Paul Ritzer along with my co-host Mike Kaput.
0:58 · We are recording a little early for this episode because it is Labor Day weekend in our world. So you'll be listening to this after Labor Day weekend. Hopefully you had a wonderful 3-day weekend. Took a little time off to spend with your family. Uh we are recording on Friday, September 4th around 10 a.m. Eastern time. It has been a crazy week and that's I don't I mean it's hard to say in the AI world. I feel like every week is crazy. This one was exceptionally crazy. Uh lots of model releases, big model releases. Each one sort of one uped the one the day before it.
1:29 · Um we got some bans on AI in the New York City schools. We got Trump taking on his own voter base over data centers. Like it's and that's just the main topics. I I don't know, Mike. Like I I told Mike as we got on here like I'm in a weird headsp space right now. I got to be honest. Like I'm not even sure what I'm going to say today. I usually don't know what I'm going to say, but like normally it's like I'm pretty balanced about everything. Um I'm just not sure.
1:57 · Like I there's some things that I have some pretty strong feelings about at the moment and I'm generally just like very uh controlled about like my thoughts and stuff. Um but yeah, I don't know. We'll see where this goes. It's a Friday and it's been a long week and I'm not sure how I feel about some of these things. Um, so yeah, I don't know. We're we're going to see where it goes, Mike. This should be interesting.
2:25 · All right. So, this episode uh is brought to us by marketing AI month.
2:28 · This is uh a new thing we just kicked off. So, AI is rapidly changing every part of marketing. There's still a major gap between knowing AI matters and knowing how to apply it in your actual work. Marketing AI month is our effort to help close that gap by making practical AI education accessible to every marketer. So throughout September, the whole month, we are giving everyone free access to our complete fivecourse AI for marketing series in AI Academy.
2:57 · So this is here this series is $499 normally. Uh comes with a certificate and it is part of our AI Academy Mastery membership program. So if you're a mastery member, this is one of the 22 mic professional certificates I think that are part of academy. Yeah.
3:15 · Um so the reason we're doing this is because our experience has been that marketers are often the tip of the spear leading on adoption and then communication of AI's value within an organization. So we're running an experiment here to say okay if we make a certificate series which by the way is our most popular certificate series in academy. If we make it free for a month, can we accelerate understanding and compound the value of responsible or the at least the potential of responsible AI adoption within organizations? So, that's like the premise here.
3:45 · Um, and then it also happens to lead up to our Maycon event in October. So, it seemed like this was a great uh way to try this. So, this series gives you a step-by-step road map to becoming an AI forward marketer. You will learn how to find and prioritize AI use cases across your job. Choose the right AI tools.
4:04 · Build personalized roadmap for adoption and use prompting deep research and custom AI assistance assistance to solve real marketing challenges. You can complete the series and earn the professional certificate. By the way, Mike is the instructor. So, if you love listening to Mike, you can spend five hours with Mike on the course series. Um, so all you got to do is go to smarterx.aimarketing.
4:25 · you will see a button to enroll in the course series that will get you to join the academy and the course series will be available to you on demand immediately. So you can take advantage of this offer through September 30th. 30 days of September. Yeah. 30 days in September. And uh and then we will actually have a ask me anything session with me and Mike on October 1st for anyone who has enrolled. You don't have to have completed it and earned the certificate, but if you have enrolled, you will be automatically entered to be part of the October 1st AMA.
4:56 · Um, and if you're not a marketer, send this to your marketing team, send it to your marketing agency, what, whoever it is. Um, it's tremendous value. It will help accelerate understanding and adoption. I promise that, um, it is, like I said, our most popular, very highly rated course, and it's, uh, going to be time well spent. So, take advantage of that offer. Again, smarter.ai/marketing.
5:22 · All right. And then every week we start off with our AI pulse. We are again testing something new with these informal polls. We're leaving them open for two weeks at a time to increase the number of responses to try and get more uh usable data out of these. So they move beyond becoming, you know, from an informal poll to becoming actual projectable data we can use. So go to smartrx.ai/pulse.
5:44 · And we are still asking the question related to entrylevel jobs and the impact that you're seeing within your organization. It takes about 10 seconds to answer this. So smarter.ai/pulse and um yeah, we'd love to get your your input on that. All right, Mike. Um I kind of alluded to some of the big things. I I guess as of Thursday, the number one thing on our list became GPT6 Astra. So let's start there.
GPT-6 Is Here
6:10 · Yes, Paul. So yesterday from the day we're recording, so Thursday, September 3rd, OpenAI introduced GPT6 Astra. They say this is their most advanced model that's publicly and widely available. It can take on more complex work. Um, and one of the biggest improvements here, and we'll talk about this a little bit, is in how it uses a computer. So there's a benchmark called OS World 2.0 that tests whether an AI can carry out workflows across applications. So things like go gather receipts and complete an expense claim for me using my browser or what have you.
6:42 · And that requires it to find information, follow instructions, and actually operate software. And on the offline offline portion of that test, OpenAI reports that GPT6 Astro got a score of 72.6. That's up from 75 65.7% for GPT 5.6 Soul. It takes 47% less time per tasks in its simulations and it also has gained in a number of other areas notably learning how to solve unfamiliar problems.
7:13 · There is a benchmark an evaluation we've talked about several times called ARC AGI. ARC AGI 3 is the latest version of that. Astra scored a 99.9% on that um under OpenAI's evaluation setup. This basically puts AI into unfamiliar games without explaining the rules or telling it how to win. It then has to experiment, learn from what happens, figure out a strategy. The score measures how efficiently it does that compared with people.
7:38 · So that is a pretty significant score on a test that I don't think people expected to be saturated quite this quickly. In mathematics, Astra scored 97.6% on Frontier Math's hardest tier. That's up from 83%.
7:55 · These again are research level problems developed by mathematicians to test the kind of reasoning needed for scientific discoveries. So models can use code to explore all these possible solutions. But then they have to work through the problem to an answer that can be checked. Now obviously there is lots lots lots more that Astra can do.
8:12 · But for everyday work, OpenAI says Astra better follows things like document templates and writing styles and is very good. They call out presentations and spreadsheets and dashboards.
8:22 · Specifically in codeex, there's a cool new little feature now with Astro where you can ask it can ask a clarifying question while continuing its work. Um, so you can actually kind of iterate as you go without redirecting the work. And then there's this whole big security question. So OpenAI says Astra has found previously unknown security flaws. It is meeting the critical threshold on their cyber security threat level evaluations.
8:48 · Astra was almost certainly what was involved in the hugging face hacks we've been talking about the past few weeks. So we're going to kind of get into that as well. Astra is being rolled out currently. It's kind of being rolled out iteratively. They're going as fast as they can to roll this out to chat plus pro business and enterprise users as well as the API. Paul, I don't know about you, I do not have access yet.
9:11 · I checked. I did not this morning. So we we have not gotten access yet, but we are, you know, looking at the capabilities, the model card, the early reports about how it performs. The safety questions have been on everyone's mind for the last couple months. I'm curious, Paul, how big a moment is GPT6 Astra given all that?
9:29 · Well, it got a whole number, and that usually means OpenAI thinks it's a really big deal. You know, these labs don't go to whole numbers um without thinking that they're very significant.
9:42 · And what I mean by that is we're getting a lot of decimal points. So usually it's like 5.1, 5.2, 5.5, whatever. Um, so yeah, you know, this all obviously happened yesterday. You know, trying to process it without having access to the model. We've known it was coming.
9:57 · They've been talking about Astro pretty openly, which is unusual. They they're not normally um kind of presenting 30 days in advance that this model is coming and it has a name and all these things. So we've been waiting for it. We didn't know when it was going to drop.
10:10 · So the way I was thinking about approaching this, Mike, is I'm just going to go through the technology, the competition, and then like the bigger picture. And so if there's anything you want to add in, you know, again, this is all pretty new, then, you know, we'll talk about that. So let's start with the technology. So if you go to OpenAI's GPT6 Astra post, um I just went through the page and I'm looking at kind of how are they explaining this? So right at the front, a new generation of intelligence.
10:37 · So they are, you know, certainly categorizing this as a different level than we've previously had. They say Astra is their most aligned model. That's probably up for debate. Um, with substantial improvements in understanding user intent and model behavior, you can delegate tasks with greater confidence in Astra's judgment. Um, you alluded to this one, Mike, and I'm going to actually come back to this in a minute.
11:00 · The they have a a section on the page that says the world's best computer use model. Uh GPT6 Astra marks a new frontier in the speed, accuracy, and safety of computer use. It can take care of tedious tasks like filling out online forms, updating customer records in CRM, and organizing your calendar. It can conduct online research and draft summaries in your email or in your document editor. These improvements also result in significant efficiency gains in real knowledge work tasks.
11:31 · So, I'm just going to again I'm going to I'll keep coming back to this, but like they are very specifically saying this does your job for you. Like this like that's that's the gist of everything here is increasingly this model is capable of doing what you do is what they're stressing.
11:50 · We'll talk about this in a second, but that statement these improvements also result in significant efficiency gains in real world knowledge tasks is like the world's greatest understatement.
11:59 · Oh my god. Yeah. And it's totally code for like you don't need as many humans, but we'll come back to this. Um, okay. So then another subhead, a step change in professional work. So now we're getting more specific here. GPT6 Astra pairs advances in computer use with targeted training for professional environments to help tackle complex work tasks. It combines the intelligence required for complex problems with the ability to carry out multi-step workflows and produce polished documents, spreadsheets, and presentations.
12:29 · GPT6 Astra is our best model for adhering to existing templates and producing slides that are welllaid out and succinctly convey key points with a structured narrative. It creates clear, well ststructured documents, presentations, spreadsheets, and analyses that follow your templates and match your writing. So, that was a whole bunch of words saying, "Hey, you previously maybe weren't relying on us to do your work because we weren't super consistent. Now, the model's super consistent." You give it your brand guidelines.
13:01 · You give it examples, and it's going to nail it every time is pretty much what that's saying. GPT6 Astra also brings stronger visual judgment to the websites, games, applications, and renderings it builds with sites, which is capital sites, proper noun here. Um, that is what they're calling one of the capabilities.
13:20 · Now, uh, Astra can create, host, and share websites, web apps, and games directly from a prompt. Okay. So, then what did a couple of people at OpenAI have to say? Sam Alman tweeted, "We hope it will begin to enable a new generation of entrepreneurship, scientific discovery, and building. We believe it is the best model in the world for computer use, professional work, science, coding, cyber security, and more.
13:45 · It took some extra time to ensure that we could meet the safety and alignment standards required for this capability level, but we think you'll find it worth a wait." Now, again, regular listeners are going to know what this actually means. the model is dangerous. The the model can do things that um aren't safe out of the box and they have spent a bunch of time trying to put things in place to prevent it from doing unsafe things. That that is what that means.
14:13 · Um, and it got so capable that it achieved different thresholds within their own preparedness framework that made them actually pause development to figure out how to stop it from doing the things it is inherently capable of doing. This does not mean they have extracted those capabilities. It means they numbed them basically like they're trying to get it to not do the things it is capable of doing and generally wants to do.
14:43 · Um then Mark Chen who I believe is chief scientist. Is that right Mike? I think I believe so. Yep. Yep.
14:49 · Um so he said oh no Yakob is chief scientist. What's Chen's role?
14:53 · Oh chief research officer.
14:56 · There we go.
14:56 · There we go. Chief research research.
14:58 · Okay.
14:59 · Um okay. Okay, so he tweeted, "It can build and test software, work across apps on your computer, and even help you crack an open scientific problem.
15:07 · Capabilities that felt like grand challenges a few years ago have become tools people can actually use. One example is computer use. If you've tried this before and felt it was too slow or not good enough, I encourage you to give it another shot. We've come a long way since operator, which was their early version, and it just works now."
15:27 · [clears throat] Now re rewind rewind to episode 235. Mike, if I'm not mistaken, you said, "Hey, if you haven't tried browser control like in codeex, you have no idea what AI is capable of." That's what Mike was referring to. That's computer use where it's like taking over the browser and doing things for you.
15:46 · So this there's a big deal. There's a reason they keep saying computer use. Um Chen continued, "We're also asking these systems to act on your behalf for more consequential work. Agents need to stay aligned with your goals and values, think transparency, transparently, and respond to oversight even when tasks become difficult. We've made substantial progress on these behaviors in Astra alongside stronger monitoring that can stop potentially unauthorized actions.
16:13 · Uh the that part that work is part of what made this release possible. Okay.
16:18 · So then in the two days two days prior to the launch openai published a blog post called path to Astra which was you know the prelude to the the launch and in that one they said over the past several weeks we have delayed parts of Astra's development and release while we strengthened and tested protections against cyber misuse and unauthorized model actions. Alluding back to the hugging face thing. They also actually I think we'll talk about this next week, but apparently um the agents hacked a German site and made a bunch of changes to it in uh something that happened in like April that OpenAI didn't disclose.
16:52 · So this is not new that their agents are doing on unauthorized things. Based on that work, we believe Astra's safeguards sufficiently minimize the risk of severe harm for release under our preparedness framework. Now, they're saying this model is like super aligned and they're saying it it has met their internal thresholds to release this thing.
17:17 · Meanwhile, the information has a story that says OpenAI technique and Astra model sparks security concerns. Now, we're going to get slightly technical here for a minute, but like bear with me. So, OpenAI said this is I'm going to read straight from the information.
17:34 · OpenAI says its forthcoming AI model Astra marks a step up in capabilities such as coding and operating applications on a computer. But an innovative technique that improved the model's performance also means that the model and others like it think about like Fable mythos um will reveal less of their quote unquote thinking making them harder to monitor for signs of bad behavior according to a person with knowledge of asteris development. Now, I'm going to pause there for a second.
18:04 · If you listen to the recent episodes where we talked about the hugging face hack with OpenAI, the reason OpenAI was able to go back and try and figure out what happened was because there's something called a chain of thought because the model quote unquote thinks out loud about what it's doing. So, we could see that the model realized it had access to a message board, that other agents had done something. And the only reason we know that is because they can go back and audit what the model was thinking at the time it did the thing.
18:36 · So that chain of thought, assuming the model is being honest with us, assuming the model doesn't know it's going to be evaluated and that its chain of thought will be audited and it's not changing its chain of thought so the human sees something other than what it actually did.
18:56 · Sorry if that got a little weird for a moment, but like the chain [snorts] of thought, we are just as humans trusting that these models are showing us their actual reasoning process is all I'm saying. It's we don't know for sure. Um, okay. So, come back to the information article.
19:12 · While the limitation isn't necessarily a significant issue with Astra, the technique has triggered concerns inside OpenAI and across the industry about whether AI developers that adopt and supercharge it will struggle to guard against the kind of rogue AI that recently hacked OpenAI's own systems and those of other companies such as Hugging Face, the new technique OpenAI is using. So again, one of the ways you make these models smarter and more powerful. You can put more chips to it.
19:41 · You can get more data, better data, or you can have algorithmic breakthroughs, meaning you find smarter ways to train them and let them learn faster. And so what OpenAI apparently has discovered, maybe other labs already know this too, but either refuse to do it or just haven't told us they did it.
19:59 · They found a technique known as recurrent depth or a looped transformer that allows an AI model to improve its answers by processing the same text multiple times. Unlike commercially available state-of-the-art models which show writing in writing how they are thinking about a task before completing it, the new technique works in a way that obscures some or all of the AI's reasoning, otherwise known as its chain of thought.
20:27 · That means the steps that the model takes to accomplish a task can easily be read or understood by humans. So what this is saying is they found that if the model doesn't tell us how it thought about what it did or how it did what it did that it actually works better.
20:45 · So if we remove the one thing we have to know why it's doing what it's doing it actually gets smarter faster. So it went on to say OpenAI has limited its use of recurrent depth technique with Astra. So the model still produces a legible chain of thought and the company's researchers can still sufficiently monitor that word sufficiently is doing a lot of work right now across these posts um monitor its reasoning Tuesday night.
21:11 · So when this article came out, open eyes chief scientist Yaka Pachaki said in a post on X that although monitoring of models chains of thought was fragile and unfortunately heading in a negative negative direction, he wanted to discourage an industry-wide race toward developing models that don't produce it.
21:32 · So even though the chain of thought is flawed and may be being faked by the models and the agents, he's saying like don't stop using this. So his his actual post on X OpenAI has worked to preserve and utilize chain of thought monitoring since our very first reasoning models.
21:51 · We deeply care about this technique as it can give us a view into how model alignment generalizes from its training distribution. I do think it is fragile and unfortunately trending in a negative direction for reasons not contingent on architecture changes that I will write about soon but there are things we can do to strengthen it uh and its core goal of our recurrent research. So Daniel Katalo who we've referenced numerous times who's one of the authors of the AI27 doc he replied and said thanks for speaking up on this and keeping the depth low for now. I look forward to reading your deeper explanation.
22:22 · It seems to me and I suspect you agree that monitor is very important and we are now in something of a race to the bottom on it. Even with depth low, this is a worrying direction to be moving. Yes, he asks. And as the information article says, even if OpenAI doesn't go further, others might.
22:42 · So then Ryan Greenblat, who's the chief scientist at Redwood, we talked about in episode 235, he said in a post, "Open's newest AI Astra is reported to use opaque reasoning architecture where more of the reasoning occurs in activations instead of natural language. This may be the single worst development for AI security and safety to date. My biggest concern is that a natural progression from here would involve scaling up the opaque reasoning to the point where the model reasons entirely or almost entirely in a latent space, meaning we can't see it.
23:15 · This would be very this would very likely destroy the usefulness of chain of thought for monitoring and oversight.
23:20 · Um I hope it isn't too late to avoid the most concerning architectures. Okay, so now you have like tech's awesome, does all these cool things, computer use, you know, yay. Um, and now like we also understand this is really really dangerous like that we are quickly moving in a realm where we just don't even understand these models and how how they're doing what they're doing. Okay, so now we'll come back to like the reality of just like all right so early user comments what are people saying?
23:45 · Wade Foster, CEO of Zapier, we see LLMs drop so often we go numb. We shouldn't be numb to GPT6 Astra. It set the new record score on Zapier's automation bench, the benchmark we created to gauge how these models perform on real workflows. Some people will call this AGI. It's not, but it's a bigger leap than we expected. The part that actually changes is your weak uh in your week is supervision. Less checking whether the agent did anything sane, more checking its scope and output.
24:14 · Ethan Mollik, I had early access and a longer post is coming, but GPT6 is stunning and is good enough that it actually does complex, meaningful work for me autonomously for days.
24:29 · We are we are beyond the 3 hours, 5 hours, you know, you know, doubling every six months. Malik saying he does days worth of work. Um, Schumer, Matt Schumer, who wrote the something big is coming post, which is what most people know him for, but Matt's an AI entrepreneur. Um, between Astra and Fable anthropics model, we've clearly entered a new era of AI. These models are alien intelligence unlike anything that came before them. The question now, what can we do with them that was never possible before? Aaron Levy, CEO of Box.
25:01 · We've been testing the model in early preview on our enterprise complex work eval at Box, meaning based on real work.
25:07 · It is now the best model we've ever tested on our expanded and hardest test set. Overall, GPT6 GPT6 Astra offers a breakthrough level of capability in coding, analytics, logic, and domain specific knowledge for dealing with complex enterprise knowledge work. Uh, Astra clearly is going to offer a meaningful jump in power powering and orchestrating workflows. And then Dan Shipper, I think he's CEO of EveryY.
25:33 · We've been testing it extensively at every across coding, writing, and knowledge work. My take, it's a big upgrade from 5.6 six soul and some frustrating habits that keep it from matching Fable at the top end. It is the best writing model I've tried. It's fast, produces very little slop, and is very easy to steer. The computer use is wild. It can go for hours at a time using complicated apps to get work done.
25:56 · It did the first cut of our Fable 5.1 vibe check video. Kind kind of mind-blowing. So then a couple just final thoughts here. Competition. Um, this was absolutely time to overshadow Google, Anthropic, and Meta who all dropped models this week. So, they purposely they were probably sitting on it. I think they probably still had some things they wanted to like fix because the roll out was Like Sam apologized last night that they fumbled the roll out. Well, you only fumble a roll out when you like race to do it.
26:27 · And so now that's why we don't have access and other people don't have access is like they just like go like everybody released. Let's go on Thursday. So, um, Google released, Michael share more about that. Anthropic released, Meta released, and so they they definitely were just trying to overshadow everybody, which they did, like mission accomplished. They also, I believe, were trying to hit Anthropic hard right before their impending IPO.
26:49 · So if all of a sudden there's doubt in the market's minds about anthropic having the best models and open AI teasing, hey, we've actually got better models even coming beyond this one that are already in training or in post training. Um then all of a sudden anthropics market might not be what it was before that. Then bigger picture and I'm going to not dwell on this stuff, Mike. I'll just say what I got to say and we can talk about it if you want. So um I'm not sure if this is just a momentary thing but this is the first model release where my first actual reaction was dread over excitement.
27:20 · So OpenAI and other labs have been focusing their messaging on how AI won't be disruptive force in the economy that like jobs are going to be fine. Like they switched this messaging in the spring because you know people told them hey you're losing public opinion and you know Republicans are getting pissed at you and like you guys better fix this And so they started messaging about like, "Oh, no, no, this is going to be great. It's going to like solve all these scientific problems. It's going to create amazing entrepreneurship opportunities."
27:48 · And yet the whole messaging they lead with, like the this is the exact words from their first tweet. This is GPT6 Astra. Anything you can do on a computer, Astra can do for you fast. They might as well just sell just said Astra can do it for you better. Um, and so their positioning all of a sudden falls back to like what Silicon Valley wants from these models is what the rest of the world wants.
28:13 · It's like they're completely tonedeaf to the fact that like you're losing the public on this and yet you come out with this new model that can do the work for us as though that's what like everyone's just waiting around for. God, I can't wait until it can just like totally use my computer for me and I never have to think for myself again. So what's going to happen is enterprises aren't going to touch this stuff.
28:35 · Like I mean some you big advanced tech companies are but AI native startups are going to come along and be like oh like I can just use computer use and like I can do the work of a hundred people with just running like seven agents all day long and just burning you know a million dollars a month in tokens or whatever and I can build a $10 billion company in five months. Like that's what Y Combinator is talking to people about. It's what uh A16Z is talking to people about.
29:00 · It's like the venture capital firms, the incubators, they're just going telling everybody go like go use Astra, go build a replacement to all these existing companies. And so meanwhile, traditional organizations are still trying to absorb AI assistant capabilities from 2024 like Yeah.
29:20 · And so we'll talk a little bit more about the agent stuff you and I have been working on, Mike, this week even, but like agent capabilities are so early within most organizations like coders have been racing ahead, but marketers, sales people, customer success people, operations people, they're not messing with agents. They're not sitting around building agents all day. And so you have Silicon Valley racing toward AGI or now basically claiming Astra is probably AGI which is what OpenAI is saying and that it has this ability to do everything you do on the computer. It's like that's not what people want. Like maybe Silicon Valley wants that.
29:51 · Maybe you know startup AI entrepreneurs want that. But like you walk into any enterprise today, does anybody want that? Like no that's it's not the thing that they're looking for. So, I don't know. Like the computer use is a really slip slippery slope to me. We've been trying to prepare people for this for a while. I actually went back this morning, Mike, and pulled episode 83. So, if you want to go listen to this, it's February 14th, 2024 is when we talked about it and I I glanced at the topics. It was the day Gemini launched.
30:21 · So, that episode was Gemini was Bard Becomes Gemini and Google launches Gemini. And then number two was World of Bits. And so if you go back to World of Bits, I had written a blog post in February 2023. So we are 3 and 1/2 years ago that I wrote this. And I'll just read real quick a couple things.
30:41 · It said again, February 2023, we are so caught up right now in figuring out AI writing tools and large language models that most marketing and business leaders as well as SAS executives and investors are missing the bigger picture. This is all just the foundation for what comes next. Imagine you want to send an email promoting an upcoming event, product launch, or promotion. But rather than a series of clicks and manual entries, you simply spoke or typed prompts for what you wanted the machine aka AI to do.
31:07 · I shared an example in the post of um HubSpot took 21 clicks at minimum to send an email campaign. So, I said, I'm not talking about simple information retrieval and natural language generation, such as a chat feature that responds to queries or prompts. I'm saying that the AI will have the ability to perform actions, clicks, form fills, etc. the same way as humans, computer use.
31:30 · Three and a half years ago, we were telling people this based on a collection of public AI research papers related to a concept called world of bits and in which by the way started in 2016. And in light of recent events and milestones in the AI industry, it appears that the capabilities for AI systems to use them again from your iPhone. Apparently Siri's talking to me. This is creepy as hell.
31:56 · Oh no. Let it let it run. I don't give a So yeah, my AI is talking to me as I'm talking about World of Bits. This is [laughter] great. Um, okay. So now it appears that the capabilities of AI systems to use keyboard and mouse are being developed in major AI research labs right now. Again, February 2023, we knew they were building this stuff. This has been attempted in the past, but recent advancements in language AI appear to be bringing this closer to reality. So we have known for 10 years that this is what they were building.
32:24 · And yet many people in business, politics, education, um are going to be hearing about computer use for the first time like holy Like they can do what? They can take over and write my computer. Hell yeah they can. We've known it. It was inevitable. And like when people it's like you have to be so controlled about telling people about this stuff because they think you're crazy or they think it's like there's just no way it's ever going to affect me. And yet like it was always inevitable. So this is where I'm at, Mike.
32:51 · And this is where part was like I wasn't even sure I was going to say this stuff, but like I have been trying for years to be very very controlled about all of this to be like create a sense of urgency without fear. And so I have said I don't know verbatim what it was, but like something to the effect of listen we have time. It's more of a slope than a cliff. I am increasingly starting to feel like the slope is nearing the cliff. like there is no turning back.
33:16 · The technology is racing ahead in its capabilities. We have school systems banning it, which we'll talk about in a minute. We have politicians calling for bans on data centers and more powerful models. And we have citizens protesting against it and vilifying the technology and the people building it. Meanwhile, government officials look at and say, "Well, we just got to keep going because if we don't, China will beat us." You have AI labs saying, "Shit, this is terrifying." Like Sam literally saying he had a visceral reaction to the hacking of Hugging Face and yet here we go. We got Astra and it can use your computer for you.
33:48 · So something has to give. Like I I I feel like we're just watching this train that I've been looking at for 10 years like seeing the light in the distance and all of a sudden it's like everybody else is seeing the train but everybody else in society like doesn't have the context of all of this and like understand it all.
34:05 · and sort of was just panicking and they're doing crazy as a result of this. And so like I I'm starting to feel like maybe I didn't do enough. Maybe I didn't like say enough. Maybe I wasn't willing enough to like deal with the trolls online when you start talking about hard things. Um and I definitely feel like the AI labs leaders have let us down. Like they presented this like abundant future all while ignoring the fact that they were responsible for the ramifications of what they were building.
34:32 · And now it's like, "Oh let's go hire some economists and we'll go hire some philosophers and like we'll figure this all out." Meanwhile, we're just going to keep throwing these insane models into the world that no one actually wants and that nobody's prepared for. So, I don't know. I feel like we're just really we're stuck. I actually don't know where this goes. Um I if you'd asked me 3 months ago, is like a pause likely? I just said no way. I actually think it might be the best thing to happen if you could find a way to agree with China to like slow down, but I don't think that's going to happen anyway.
35:04 · They're just going to do it behind the scenes.
35:06 · Like, so I don't know. We're in this place where the economics of the business models of these labs requires them to keep going. The competitive environment between the labs and the nations requires it, at least in their minds. And so, the public is going to increasingly push back. And I think they're right to do it. And I find myself every day stuck in the middle of like understanding the technology, advocating for it responsibly and in a human- centered way, but totally empathizing with people who don't want it and don't like it.
35:37 · You know, just a couple final things that jumped out to me, Paul. I've never found myself as riveted by the lab's marketing as I was with the post they put out because just a few I want to read really quickly just a few examples that they highlighted and I want you to think as I'm reading these if you're listening when you think of the work you pay people for is it this work so they first said GPT6
36:04 · Astra can complete financial modeling world cup challenges that's a real thing there's an Excel world championship which Crazy. It can complete these using computer use about four times as fast as the winning human competitor, helping analysts spend less time building models and more time interpreting results and making decisions, which they don't say the model can also do. Obviously, GPT6 Astra builds and refineses dashboards directly in PowerBI.
36:28 · GPT6 Astra creates a slideshow uh using just a few slides from Open AI's presentation template, capturing the correct tone and layout throughout. They also don't tell you it can do the content for you. But I was reading these and I was like and they're they have a really cool launch video and stuff and I was like this is really cool. But I kept coming back I think you are also a fan of this movie to this quote in the big short when Steve Carell's character is talking to these like mortgage brokers and he's like I don't get it. Why are they confessing?
36:58 · And his assistant's like they're not confessing they're bragging. like everything you just said about their perspective. All of this messaging makes perfect sense if you're a Silicon Valley person that says every employee is just a temporarily delayed startup founder, right? Where you say eventually you're going to go build your own thing and won't this be great?
37:19 · 99% of the world is not a temporarily employed startup founder. So I think that's a real interesting disconnect. I'm not saying it's all naive. I think they know often what they're doing, but it is wild to see this cuz I don't read this as like cool, this can use spreadsheets. I read this as cool, we don't need people that read spreadsheets.
37:41 · 100%. Yeah, I'm with you. Like that video. So if you didn't if you're not on Twitter, you probably didn't see this.
37:46 · And again, like you listen to the podcast, keep in mind like 99.9% of the world is going to go to their job today and tomorrow, not even know that Astra is a thing, having never heard of computer use. Like people are blissfully unaware that this this is happening in this universe. But if you are deep in this like we are and you're an ex user and you saw the video that OpenAI released this model with Silicon Valley World AI bubble world. Oh, greatest launch video in history.
38:18 · I saw it is dystopian. Like I watched that video cringing the whole time thinking, "Oh my god, they think this is what people want." Like it's as I said, I don't even know, man. like um it just it feels different. It I don't even know what to how to explain it.
38:42 · Yeah, I couldn't agree more. It's a real turning point. I think we're going to look back on this as a very like a chat GBT moment for sure.
38:48 · Yeah.
38:49 · Um all right, next big topic this week, as if we haven't already covered enough weighty topics. So, this past week, President Trump publicly defended the data centers powering the AI boom. So, in a Monday, past Monday post on Truth Social, he said that communities that reject them risk becoming poor and falling behind, while places that welcome them can gain jobs and lower taxes. He had this slogan in here that said, "Let data rain."
Trump Goes All-In on Data Centers
39:15 · Now, I'm going to read the full post just so you don't think I'm exaggerating, cuz I actually do think it's important to see how starkly this was put given how important the subject is. He said, quote, "The only reason that communities throughout the USA should not want data centers is that if they want to end up being backwards and poor, if they want to be successful and rich with far lower taxes and jobs all over the place, let data reign." The good news is there are plenty of other places that want them.
39:41 · If we kill the golden goose, you have only yourselves to blame. China could not be happier with this anti-data center movement. Actually, they can't believe it is happening. They probably can because I think they're behind some of it. But it sounds like this post is in response to something. I actually had to go look this up. It's not It's not like a response to another post. I assume I from the research I was able to do, it's really just like the growing local opposition. We talked about there was like a Republican party memo that was like, "Hey, you guys, if you get data centers pinned on you, you're going to be in trouble."
40:11 · So, there's definitely like trends he's responding to. It's not like a direct thing like someone rejected a single data center. But at the same time, just a few days later after this post, Senator Bernie Sanders and Rep. Greg Kesar are calling for the US to ban artificial super intelligence, which is future AI that would far exceed human capabilities.
40:33 · Future meaning like 2027, but go ahead.
40:36 · Yeah.
40:36 · And so the Washington Post carried this. They put out this kind of press release about it. They don't have the full plan yet, but they plan to introduce a bill that would immediately pause development of advanced AI until a new federal agency is created to monitor dangerous capabilities. Companies that do not comply, they suggest, should face a corporate death penalty. This is their exact language. I was unfamiliar with that term. It's not a formal term, but it's like widely used of just like, hey, we put you out of business.
41:02 · Um, and they basically heavily cite the fact that the rogue open Aai agents and hugging face were a huge motivation for this. So Paul, I'm curious about both angles.
41:14 · This the Trump post is like given how politically savvy and sensitive people are that get elected presumably. This felt like almost a crazy own goal, which is why I had to look up like what was this in response to? It was just like we decided to go all in. So, I'm going to promise our listeners I don't have nearly as many thoughts and as deep of thoughts on this one as the previous topic. We'll take a break for a second.
41:41 · Here's my best guess. Um, he's cornered.
41:44 · So, there's polling data showing his voters hate data centers.
41:49 · He needs data centers to boost the economy, the GDP, and to compete with China. You you can't have both. So he has to take on his voters directly and get them to support data centers or else the Republicans are going to lose the house in November.
42:06 · So they're under the gun. They have 60 days to change hearts and minds around data centers in rural communities where these data centers are being built that generally support the Trump administration. And he can't back down.
42:21 · Now, if he backs down on data centers, then they lose the competitive edge they're pursuing by accelerating the building of data centers. And I I honestly think it's probably as simple as he saw a Fox News segment on this or someone showed him polling data that says we are screwed. And so it's like I'm just going to go at him. And I think that's it.
42:44 · It's just it's just become so politically um decisive or divisive where people feel so strongly one way or the other about this. And he needs his voters to at least be in the middle, like to not care one way or the other, to not let it swing the midterms. And right now, at least in Ohio, in our backyard, it has the potential as a single um item on the campaign agenda to swing the midterms.
43:17 · So, and if Ohio falls, it's representative of what's going to happen in a bunch of other swing states in the US. So, I honestly think it's probably that simple.
43:27 · This is also why I kind of included here the Bernie Sanders thing. Not because obviously Bernie Sanders is probably not running for president, but and this who knows if this ban's even a thing or going to become a thing, but this idea that with Trump planting the flag in the ground, it's suddenly like, oh, the lane is open. On the other side, we've seen on the Democratic side, the Democratic socialist wing of the party has made some pretty significant election gains.
43:51 · I mean, it feels like this is like now forcing you or forcing you or giving you permission for what you wanted to do, which is Democrats are at the party of anti-data center.
44:02 · Yep. Yeah, possibly. I don't We'll see.
44:04 · Yeah.
44:04 · Just going to steer into it. But yeah, it's all politics. And again, I mean, we're sitting at the beginning of September, October, November. We got literally like 60 days and it's going to be you're going to be hearing a ton more. You're going to see a lot of ads.
44:17 · We are in Ohio. I don't know where if you are in other places, but man, I cannot watch a sporting event without getting bombarded with ads hitting people on both sides.
44:27 · And the hugging face hack is kind of broken containment, as I like to say, like your average person starting to be like, "Wait, these can do what?" I think then you're going to have Astra do the same thing eventually. Um, if it doesn't soon enough here.
44:40 · Yeah, I think we're going to have I I may have said this on the podcast or maybe I've just been thinking it. I do think in the before the end of this year, we will have our hugging face moment for jobs. Like I I I think that there's going to become this moment where it becomes very obvious that all the things we've been talking about, the the things we should be preparing for, I think they're going to start to become a reality to where you start to have this visceral reaction within society because they start to everyone starts to realize
45:08 · and like we said, maybe Astra is that tipping point. maybe real reliable computer use was the thing that was missing and that starts to change the way people realize like oh my god they are coming for our jobs even though they tell us they aren't.
45:22 · Yeah.
New York City Schools Ban AI Through Middle School
45:24 · All right. Our next big top final big topic this week then we'll get into rapid fire is this past week New York City public schools announced a one-year moratorum on studentf facing generative AI in grades 2K through 8 for the 202627 school year. ABC News reported this policy will cover more than a half a million students because this is the nation's largest school district. This rule applies specifically to generative AI software that students use directly.
45:51 · There are separate screen time rules they've got that prohibit 1:1 device use in grades 2K through two. They recommend daily limits of 30 minutes for devices for grades 3 through 5, at least the 1:1 devices, and 45 minutes for grades 6 through 8. Teacher-led screens generally remain allowed for group instruction.
46:10 · This policy preserves assisted techn or assistive technology. So if you do have a disability or um certain other limitations where you need to use some type of AI tool, that's okay. But high school students face a different set of rules. Grades 9 through 12 may use approved vetted programs in limited guided settings.
46:29 · Every high school student must complete two 45minute AI literacy modules while certain career readiness courses and uh five centrally approved AI pilots within the school district are available um as long so they can then use AI with teacher supervision. Companion chat bots prohibited across all grades. Teachers and staff can still use approved AI for planning and operational tasks but not for grading, behavior monitoring or any other decisions about students.
46:56 · New York City Mayor Zoran Mamdani said the city will spend the year studying AI's impacts and schools chancellor Kamar Samuels emphasized protecting students independent thinking and making sure that technology serves learning. So Paul nation's largest school district basically has banned generative AI for grades uh what is it 2K through 8.
47:22 · What is 2K? Is that like prek?
47:24 · I have no idea. I had to check that. like I had to I had to make sure I wasn't like missing something. I think it's just what it's what they call it's in the actual policy. So, it's it's some level of it's like kindergarten or prek kindergart.
47:37 · Um, okay. So, we'll get to the LinkedIn post in a second, but I I I shared this ABC story on LinkedIn with what I thought was a relatively balanced take, which was like, hey, this is like straight out banning. I'm not sure that's the right approach. like maybe literacy has a better role here. Maybe we should invest more in teaching the teachers, but that's really hard. And like that's pretty much it. And it's like, but you know, banning, I'm not sure. I'm I'm all for a ban, man. Like that that lit a fire.
48:09 · So like this is obviously a topic that people uh feel passionately about. I'll say um I stopped reading the LinkedIn post. So if you added something constructive, thank you.
48:23 · if you just wanted to like question my integrity and um my motives like I I I'm not reading it anymore so you can stop putting those posts. Um so I'm going to give the like my full context here about this topic. So one uh as Mike and I do on the show, we try and take a very objective position on all this stuff. We look at the facts and try and share the facts. So I'll do that to start.
48:48 · So the ABC post that a lot of response came to um you alluded to this already Mike Mami said in a quote the tech industry wants us to believe that AI powered early education is not only inevitable but necessary. We do not see it that way. Um New York City Schools Chancellor Kamar Samuel said the measure will help restore critical thinking skills students need to graduate.
49:12 · Um then Samuel said added, "We are standing we we're standing firmly in our decision to not conflate innovation with more tech and we're going to lead over the next year with evidence and make sure technology serves learning, not the other way around. Fine. Like that's good, but no problems with these statements." Um well, I mean the Madami one was a little bit more like, "Hey, we're banning it because they're evil and we're like we don't that's what came across poorly in the ABC thing, I would say."
49:38 · Um that article also said meanwhile the White House is pushing teachers to use AI responsibly in the classroom as administrative officials believe it can revolutionize education. So then you had pulled some resources Mike I was kind of clicking through some of the stuff you'd pulled and so we'll just like rewind back what are we in September through so 6 months ago.
49:58 · So New York this is March 24th 2026 New York City public schools announces release of AI guidance for educators and school leaders. So apparently since March, the public schools have had listening sessions and seem, unless I'm interpreting this wrong, Mike, to have completely pivoted with their positioning on this. Yeah.
50:21 · So the nove the March 24th, spring 2026, um release from New York public schools announces the release of guidance on artificial intelligence, which includes policies on academic integrity, student privacy, and data security. This first iteration of guidance will help educators and staff assure that when AI is leveraged in schools, it is done safely, safely, thoughtfully, ethically, and responsibly while reinforcing the necessity of human judgment when evaluating AI produced materials.
50:49 · In line with chancellor's counsels, same chancellor's commitment to engaging with school communities and prioritizing school community-led decision-making, families, educators, and school leaders will be invited to offer feedback on the guidance over the next 45 days. So this was the actual quote from Samuels in this announcement.
51:08 · While there is no tool or resource in the world that can replace what our teachers bring to their classrooms every day, AI can be used as a powerful tool to make the work of our educators more efficient, giving them more time to focus on supporting our students as they develop essential critical thinking skills.
51:27 · This guidance is designed to empower our educators to choose tools that support our students without compromising on safety and academic integrity while teaching our children when and how to use AI appropriately.
51:42 · So all again like we're m we can dig further in like future conversations but on the surface it seems as though their direction six months ago was we're going to do this responsibly. We're going to empower our teachers to make choices and they're going to use this as a tool. They've now between March and today apparently learned enough that they're banning it completely K through through 8. They're not teaching it at all.
52:11 · There's no AI literacy, nothing. So, when we get into the actual announcement, I clicked over to the guidance on AI. So, now we're in the schools.new york city.gov, the guidance. I will put the link in the show notes. I actually really like a lot of their messaging here. So this is the evolve form. This is like as of September 4th.
52:30 · Um so it's called guidance on AI and screen time. Uh they've obviously put tremendous thought into this. Now again some of that thought led to an apparent pivot since March. So this is from the website. Ultimately, as chancellor and as uh the New York City schools parent, um I believe that schools must be places where students think, read, write, reason, solve problems, talk with one another, and work through difficulty.
52:55 · They should also be deeply human places where relationships with teachers and peers are central to learning. We know that AI can never replicate the care, expertise, and commitment of educators. Great. That good messaging. That's why we're putting guardrails in place to protect the critical human connection, curiosity, and creativity that help our children grow. It's also why we're making every high school student learn how to use AI responsibly and safely.
53:17 · Cool. Um, they say the public schools are taking a cautious, developmentally appropriate approach to student facing Gen AI and screen time. Younger students will have stronger limits. High school students will have limited guided opportunities um for generative AI use.
53:33 · Then they go into their framework. So they have this n grades 9 to 12. They have like a studentf facing AI limited guided use where it's like how many minutes they can be on it. They have five pilot programs where I think they're going to like now test the impact AI exposure AI has. They have like a 15 minute per week a 20 to 40 45 and then a one period per week and then they're basically saying like we're going to run these pilots and then from there we will update our guidance. So like that's seems to be the situation.
54:04 · there. Zero AI, zero AI literacy is what appears to be happening from K through 8. Um, and then from 9 through 12, it's going to be some some limited stuff. For context, I went and looked said, well, what is the Cleveland system doing? So, Cleveland Metropolitan School District, which is, you know, the equivalent in in Cleveland. Not nearly as many students, but, you know, the equivalent we have, um, the board passed an AI policy, which applies to all district staff and students with unanimous vote on June 23rd, 2026.
54:32 · Uh, Ohio set a July 1 deadline for all districts to approve AI policies. The policy tackles the risks posed by technology such as academic integrity, data protection, and un ethical use of AI. On the flip side, it emphasizes the need for students and staff to learn how to use AI and other new technologies and encourages responsible integration of AI into the classroom. A commitment to AI literacy is a key part of the policy passed by CMSD uh board. It was closely modeled on the sample policy created by the Howard Department of Education.
55:03 · The goal is for students to understand how to safely and responsibly use AI by integrating into curriculum and professional learning opportunities. The policy puts some restrictions in place on the usage of AI. It states that technology should not replace human work and should instead be used as a tool to support learning and teaching, not a substitute for student effort or the role of the educator. It also describes what uses of AI like academic dishonesty and cyber bullying are considered unethical and prohibited.
55:30 · Um, and then it just goes on to say like, hey, we there's a researcher at Stanford who studies AI K through 12 and likes the direction that the Ohio one is going that encourages districts to adopt these AI policies and teach AI literacy.
55:42 · So, I think it's just important to like balance this with, hey, like one way is let's ban it because we've had some feedback or we're not sure and like we're just going to like eliminate it all together while we figure this out.
55:54 · Fine. Like that. That's one approach to doing it. Another approach is the way Ohio's doing it, which is let's embrace this. Let's like focus on literacy and let's teach responsible use. Um I will say, Mike, like I I was last night at the parent teacher meetings for my daughter. So, she just started high school and AI is not allowed in any of the classrooms and I'm cool with it. Like, so people I think some people were like attacking me on LinkedIn as though like I'm some like you know allin AI everybody should be doing.
56:21 · It's like no like my daughter's taking like English and world history and um you know sciences and it's like what does she need AI in those classes for? Like it's not now maybe like the the physics one I could understand like to visualize different concepts and do some guided learning on complex topics like that would make total sense to me. I would I would encourage that. But as I'm sitting there, I'm thinking like, man, there should be an AI literacy 101 class. Like every freshman should just take like what is it? How do we use it? You know, what's our policy in the school? Like is that bad? Like I I don't know.
56:51 · I don't know how that would be perceived as like a bad thing if we just like taught these basic fundamentals. And my kids have been using AI since they were like fourth grade at their other at their their elementary school. It's like why why couldn't we just teach that? So my whole point was like I don't understand why we can't just have some middle ground here where we can teach responsible use because they're going to use the stuff outside of school anyway.
57:14 · And so what I found by bringing this topic up on LinkedIn, Mike, was some people have and you read you read the comments. Um some people have strong opinions because they have strong feelings about AI.
57:27 · Yes.
57:28 · If you ask them cool like why is the band good? It's crickets. like they they it's because they hate AI basically or they hate for some reason and maybe it's a valid reason like maybe they have a reason that they don't like it. Fine. Um some people have strong opinions because they've thought deeply about the challenge school leaders face and the impact AI is going to have on their kids. And I I loved hearing constructive feedback from both sides. Like that's the whole point is like let's have a debate. This is not there's no obvious right or wrong solution here.
57:58 · City schools in Cleveland are doing it very different than city schools in New York.
58:03 · Um you may agree or disagree with either of them, but the whole point is nobody really knows and so why can't we just have a discussion about this openly and honestly?
58:13 · Yeah.
58:13 · So my whole thing and I tried to convey this on LinkedIn and then I just gave up was um I have been on the board for Junior Achievement for 10 years. Junior Achievement, if you're not familiar with it, does incredible work in starting an elementary school to teach financial literacy, business skills, life skills, and inspire entrepreneurship. Maybe that's at minimum what we need.
58:38 · Like, if if we if we have something as incredible as Junior Achievement that can go in and bring people in from the outside and like provide this level of literacy, why couldn't we just do something like that at least for AI starting at elementary? Like why do we have to pretend like the technology doesn't exist? Doesn't mean they have to be using it in art class and write English class and things like that.
58:58 · Fine. Like I actually am an advocate for no AI use in English classes, especially early on. Like learn to do the craft.
59:05 · Like learn the skills. Um so I don't know. I mean I honestly didn't think this was going to be that big of a hot button issue, but like some people had some very strong reactions to it. And um yeah, I guess that's my my take is like I'm I I don't know when AI literacy is ever bad like early on.
59:26 · So I don't know that's I guess I had a lot to say on that third topic too.
59:29 · [laughter] You know I for what it's worth I'm a little earlier in the journey with a a little over a 2-year-old. But the way I'm starting to try to think about it is like I don't I'm not and I don't have any answer here. But I think it's helpful to think about not like oh you have to be using AI for all these things. It might not be good to use AI for. It's more I want to preempt other people telling you how to use AI because like you're going to if you don't have any education or exposure or experience the moment your buddy shows you you can cheat on a test with this thing.
1:00:00 · Not that I think it's I don't see this as like a children or being unethical. Like your kids are kids. Like you're going to take some shiny thing and want to use it in a weird way probably. So I want you to come back as a child and say, "Well, no, there's a better way. Here's how it is a bicycle for the mind. Here's how it is accelerating my thinking and how I how I achieve things."
1:00:25 · Right? If we if we don't teach the responsible use at an early age, someone else is going to and it might be it might be meta, it might be like Instagram like where they experience it or roadblocks where they're interacting with AI. It's like 100%.
1:00:39 · So I just I don't again I I like I want to be very open to why people feel so strongly against not even providing literacy at an early age.
1:00:51 · Um, like I want to understand those perspectives, but I think I just live in this reality of like it's going to be a part of their lives either way and someone needs to teach them. And what are we going to do? Rely on parents to go figure this out? Like it's complicated stuff. And so that falls to me like shouldn't the teachers teach it?
1:01:08 · Like shouldn't the schools provide some element of literacy to keep the kids safe and teach responsible use and ethical use and things like that? So I I again I don't I don't know when that became a controversial idea that we should educate people on these things, but apparently to some people that is a controversial opinion.
1:01:31 · And one final note here for everyone wondering, 2K in New York City schools is their free early care and education program for 2-year-olds is similar to 3K and preK, but it's for younger children. It's not mandatory. It's just a thing they have. So nice.
1:01:46 · Um, okay. So, before we dive into rapid fire, this week's episode is brought to you by Mecon, our AI conference for marketing and business leaders happening October 13th to the 15th here in Cleveland, Ohio. It is quickly approaching and MECON is 3 days of keynotes, sessions, workshops, and conversations built specifically for marketing and business leaders who are actively figuring out how to adopt, operationalize, and scale AI across their organizations. So, you can actually use the code pod 100 at checkout and you save a $100 on top of locking in the best rate available.
1:02:19 · Go to mcon.ai to register. mic.ai to register. And I would just encourage you if you're a listener and you're coming on your own, if you have team members that should be here, both marketing and marketing adjacent and even business leaders within your organization, reach out to us. we can uh figure out a way to get more of your team to make on this year.
US Government Backs OpenAI in The New York Times Copyright Case
1:02:43 · All right, Paul, let's dive into rapid fire. First up, this past week, the US Justice Department filed a 20page statement of interest in the consolidated copyright litigation against OpenAI. And in this, they supported the company's argument that its use of copyright text to train large language models is indeed uh is indeed protected by fair use.
1:03:02 · So the government is not a party to the case which is being brought by the New York Times against OpenAI over it using their material to train their models and also producing outputs that have borrowed from their content. But this and this filing does not have anything to do with deciding the dispute. But the Justice Department submitted it under a law that lets the agency inform a federal court of the US's interests in pending litigation.
1:03:30 · So this comes as US District Judge Sydney Stein has asked both the New York Times and OpenAI to file summary judgment motions. The filing asks the court to treat model training, this the filing from the government asks the court to treat model training separately from model outputs.
1:03:48 · It argues that copying text so a model can learn statistical patterns is what they would call in the copyright world highly transformative meaning it is not taking copyright work while acknowledging they acknowledge the outputs which reconstruct and distribute protected works in the case of this like New York Times articles that may raise different copyright questions. The justice department is also framing this issue as one of competition and national security.
1:04:16 · It says mandatory licensing for training data could favor the largest technology companies slow US innovation give foreign rivals an advantage. The New York Times, which sued OpenAI and Microsoft back in 2023, says the companies used its journalism without permission and can produce outputs that substitute for its work. A Times spokesperson said the administration is siding with trillion dollar AI companies at creators expense.
1:04:44 · So Paul, this case is not resolved, but is like a landmark copyright case. I found it really interesting. The government's basically trying to put their thumb on the scale a little bit and say that training is different from outputs. I'm guessing in the hopes that OpenAI doesn't get slapped with some type of fine or a legal action that helps us fall backwards in the race against China or developing our own national security advantages.
1:05:10 · Yeah.
1:05:10 · I mean, this is again, we're going back years here. we've been talking about these cases that they would eventually end up in the Supreme Court and you know some verdict would be decided. Um you know in this case the the the argument is like they should destroy the original models and weights because they're not going to be able to extract the New York Times training out of it. So um but then you just you know distill future models from that stuff.
1:05:34 · You don't you know you don't even need it. You could ever find it within the training data. It's just like I I don't know. My feeling has always been they will end up paying some ridiculous amount of money that isn't a big amount of money to them to like pay these things off and make it all go away. Um maybe the government steps in and they win the legal opinion. Fine. They're never winning in the court of public opinion on this. Like this is one of those things where um we've all known that they knew they were stealing it.
1:06:04 · Their internal communications indicated that they knew they were stealing it. Every lab knew the other lab was doing it, so they were going to do it. That is not debatable.
1:06:15 · The whole thing they the game they played was well eventually maybe we can win and prove that it was and they should change the law because the law is just archaic and doesn't understand what we're doing. Um or we'll just pay some really big fines eventually, but we'll be so big at that point that who cares what, you know, $10 billion here.
1:06:36 · It's like Meta, just pay whatever, 19 billion for, you know, destroying kids. Um and it's like whatever. Like, you know, amateurize that over 10 years and like just go about your life. It doesn't change anything at Meta and this isn't going to change anything at these labs.
1:06:49 · They're going to keep doing what they're doing. So, it I don't know like it's one of those things like some people probably feel super strongly about. I I've I've been on the side of like copyright is copyright and that that there should have been consideration from the very beginning on this and there wasn't for a very long time and the labs will just continue to be vilified over this. So they might win their cases but it's just going to be one more thing that the public stacks up as to why these labs are evil. And that's what worries me.
1:07:20 · Look, it's just Well, it's also like sounds like the Justice Department, at least in this administration, is like retroactively trying to be like, well, hey, it probably was transformative for them. Like, it was okay for them to train on copyright and know what they do with it after what the outputs are. Okay, we could talk about this. It's almost like rewriting what happened.
1:07:40 · Yeah. Yeah. So, I don't know. I just always assumed that models were never getting destroyed. However, this worked out in the courts.
1:07:47 · It wasn't going to slow the technology down. And I still feel the same way today. Hey, I just think the public is going to have an increasing awareness of what was done where 3, four years ago they, you know, generally the public was unaware that this is how this stuff worked and how they were trained.
Banks Push Big Law to Pass On AI Savings
1:08:04 · Next up, the Financial Times reported this past week that Goldman Sachs, Morgan Stanley, and Cityroup are pressing major law firms to lower legal bills or change how they charge as AI speeds up routine work such as research, document review, contract analysis, and litigation discovery. Both Morgan Stanley and City told the FT they want payment arrangements that save them money. Goldman, according to people familiar with the matter, has asked law firms how much more efficiently they can work with AI and expects to share the benefits.
1:08:34 · City has begun asking firms to bid for work and explain AI savings.
1:08:39 · Adam Meshel, city's global head of legal, said the bank expects cost to fall significantly per transaction when AI reduces the hours required and a different working model would probably be in place within a year. Eric Gman, Morgan Stanley's general counsel, called big laws compensation model extraordinarily unstable. He said most of the bank's external legal work would be competitively bid by year end paid through alternatives such as fixed fees while Morgan Stanley would still pay for top lawyers judgment and talent. So this kind of reflects actually a wider industry trend.
1:09:08 · Thompson Reuters found that 71% of in-house legal professionals expect firms to change how they charge as AI use grows, while 62% of law firm professionals said their pricing structure remained unchanged in response to AI. Boy, Paul, is this like a full circle? I feel like for years we've talked about marketing agencies, professional services, billable models, billable hours model is dead. I'm sure we talked about like legal accounting at the time, but it really sounds like they're putting the screws to legal firms from some of these banks. Like not unexpected, right?
1:09:41 · No, I mean this is I'm actually surprised this is like a story now. Like I would assume this would have happened two or three years ago. Um, but maybe thing like tools like Harvey and stuff are getting so good that they know now that like it's Honor's radar like wait Harvey is worth how many billion dollars to do legal work. Um, they see what Claude's capable of doing and it's just like oh wait a second this is going to take you as long as it used to. I I think like it's just a microcosm of all professional services. You know, I think about it in hiring developers.
1:10:09 · It's like well I know you're not going to take as long to develop something as you used to as long as you're using the AI within it. And if we were hiring a marketing agency, I we have we are our own agency basically. But if we had to go hire an agency, I would be like, there's no way it's it's like a tenth of the time it used to take. Like I owned an agency, I know what it would take to do this stuff 5 years ago versus what it takes to do it now. So yeah, again, like I uh in 2012, chapter one was eliminate billable hours.
1:10:39 · Like I've never believed in billable hours as a model. Uh, I think AI just accelerated the need for it largely to go away. But I do think there's still a place for Bill Blowers when it's like high level advisory work, you know, sitting when it's actual time and you're in meetings and stuff. Um, but I yeah, I mean, I think we've been calling for this back in 2024. We ran an AI summit for marketing agencies and that's that was what I was preaching.
1:11:03 · I was like, you got to find a better model. like it has to be a valuebased model because the work is going to be done faster and your clients are eventually going to realize it's going to be done faster. So, you need to get out ahead of it. Um, yeah, I think a lot of service firms are going to be in in some dire straits here because if they hadn't already been moving toward a different model, they're going to have to move pretty quick. Well, yeah.
1:11:25 · To that point, anecdotally, I'm just curious like do you think people are moving fast enough towards different models or No, it's No, it it's hard like when you have these like traditional systems, when you have staff that have spent their whole careers doing it this way, your billing system is structured this way, your project management system is structured this way, your service model is structured this way, you can't just flip a switch and do it differently. Um, so it's a huge opportunity for like upstart firms or like partners at these firms.
1:11:57 · It's like screw it, I'll just go, you know, raise some money and do my own thing and I'll do the work of five partners in one. And, you know, you just build a smaller AI native version of this. So I could see a lot of startups emerging out of these big traditional prof professional service firms where consultants and advisers, analysts are like, man, I can just go do work with five or 10 people and do my own thing and not have to deal with all the legacy stuff that's got to get changed over the next three years. That's going to be super painful to go through.
1:12:26 · Yeah.
1:12:28 · All right. Next up, this past week, MIT released a report from its ad hoc committee on AI use in teaching, learning, and research training. Um they said that generative AI has forced a broad rethink of undergraduate teaching and assessment. This report is the product of five months of meetings, research and outreach across the MIT community. And that committee included undergraduate and graduate students, faculty from every school, staff from MIT libraries, and the teaching and learning lab as well.
MIT Says AI Can Now Complete Most Undergraduate Assignments
1:12:57 · And it assessed current AI use, identified teaching and assessment innovations, and proposed policy. And basically the headline here is that they basically just concluded they came out and said that current large language models and other generative AI tools can produce credible solutions or reasonable responses to almost any written undergraduate assignment including essays, math and science problems, proof and coding. They did not like test this against controlled benchmarks or they did not name like models.
1:13:28 · They tested prompts assignment samples and more. This is kind of just their top level takeaway based on this more qualitative explorations and discussions. They also reported they had all these kind of effects they were seeing of AI in the undergraduate experience including decreased office hour attendance and online discussion. They said they had seen they had heard of anecdotal declines in in-person study groups as students shift towards AI supported learning.
1:13:57 · And it also says AI's ability to complete MIT level assignments makes outofclass work basically less dependable as evidence of what a student can do independently. So MIT recommends that most courses be reviewed and substantially adapted. They suggest alternatives like oral exams, semester portfolios, and outofclass assignments paired with in-class conversations along with more experiential and project-based learning. And MIT President Sally Cornbluth says they are developing guidance, instructor support, pilot funding, and communities of practice.
1:14:28 · So Paul, I mean they don't have like scientific way of saying like AI can come out and do every undergraduate like writing assignment basically, but like I kind of give them credit for just being super candid and being like we got to change everything. I mean, we've been talking about that for years, but it's nice to at least see them say I think they start out saying basically like they say this report is a call to action as the first line of this. So, I thought that was interesting.
1:14:58 · Yeah, it's I mean this is a really hard time [clears throat] to be an educational leader, you know, to to figure out how to shift this stuff.
1:15:07 · You probably have like teachers, professors who are a mix of like all in on AI, can't stand AI, kind of in between, not really sure what to do, how to apply it in classrooms. You have competing studies about like, well, if we do it responsibly and we integrate it as like a guided tutor kind of model, it works really well. If we just let students have free reign of how to use it, then they're just going to replace critical thinking.
1:15:29 · Um, yeah. again like I just I my brain a lot I spend a lot of time talking to higher education leaders but I I have a high school student and I have an eighth grader and so I think about where they are and you know again going back to where with my daughter's school it's a lot of inclass writing you know pen to paper like and I think that is correct like I I do think you have to you have
1:15:56 · to strip out this temptation to take the shortcut and you actually have to teach this stuff um while finding ways to integrate it in a responsible way. So, it it's like super challenging right now and I'm you know I I wish we could do more honestly to like support educational leaders who are trying to solve this because it it's a very complicated time to be be education with as fast as this technology is moving.
AI Use Case Spotlight
1:16:27 · Okay, next up we have our AI use case spotlight where every week we give you a quick look at real AI use cases we are exploring. So Paul, I'm going to share one real quick and then see what you got this week. But this week I actually was doing this yesterday uh on my home computer as we were getting prepped for an early recording. But this plays really well with Astra. Though I did not use Astra for this, but I actually used AI to analyze the comments on your LinkedIn post about New York City schools.
1:16:56 · And to do that, I used browser usage because typically you'd like go scrape all these comments. There were hundreds of them. I wanted to know like what positions were people taking, what arguments kept coming up, etc. Now, that's all well and good. Like you could do that for years with a reasoning model, scrape all the comments or maybe use LinkedIn's API, though I suspect that's kind of locked down.
1:17:19 · Basically, in the past, I would have done this by like copying and pasting a bunch of comments, but in this case, I used the browser uh usage of GPT 5.6 Soul and said, "Look, there's hundreds of comments. Go pull every single one and do the analysis for me." And so basically using chatgptex desktop app uh it opened your post in the signedin browser. It switched to most relevant for to the view that showed all the comments.
1:17:49 · Started um with the first one. It worked through the page, loaded more comments, expanded replies, it opened.
1:17:57 · And you were just watching it do this the whole time?
1:17:59 · Yep.
1:17:59 · Yeah.
1:17:59 · Oh yeah. Trust me, I was watching. I did not walk away from this. I never never walk away when it's that I have not gotten to the point where this stuff is doing things for me while I'm not there. But um so yeah, I'm watching it do all this. It literally um worked through the page, loaded comments, expanded replies.
1:18:16 · It opened longer comments that LinkedIn cuts off behind a see more button. Um so then it goes and just reads all of these, organizes what it collected, separate, they actually separated your responses from the audience comments. It distinguished individual commenters from the number of comments they wrote because I wanted to say like what's the actual sentiment of each person and then it grouped the responses into being for being against. Pulled out some recurring arguments, some specific examples.
1:18:45 · Spoiler, it was like almost exactly divided, positive and negative. There was a lot more nuance though than you know some of the more aggressive comments would suggest. a lot of people were uh had a more um nuanced view of it, but yeah, it was about 50/50 at the time. So, I just wanted to share like that's something where typically, you know, that's possible, but you're like, "Okay, do I go scrape all the comments using a web scraper? Do I c copy and paste them? Is there a connector to LinkedIn?" All that goes out the window with this is why I wanted to mention it.
1:19:18 · I'm not saying go do this. It's again, you have to be really careful and do it in a controlled environment. But think about that and extrapolate that to like anything else you would be doing.
1:19:28 · Yeah, it's a great example, computer use, Mike, because I actually thought about that at one point last night where I was like, I can't read any more of these comments. Like, this is crazy.
1:19:35 · Um, but I was like, yeah, I'd be interested to know the people who actually offered constructive ideas versus the people who just like wanted to, you know, throw up there.
1:19:45 · Um, and I was like, I don't know how you do that. And here you go. You just found a way to do it. Yeah. Um yeah, that's cool. Uh mine, maybe we'll explain more of this later, but I actually was working on a lot of agents um this week.
1:19:59 · So I'm doing a talk I mean I guess you all are listening to this u after September 7th or 8th I think. So this week I'll say I'm doing an agentic enterprise talk for a major tech company. So you know kind of talking about the role of agents, where we are in their advancements and what you could be using them for. Um, I'm also starting to prepare for my AI co-executive workshop at MCON where I'm actually going to do a hands-on how I use, uh, agents and assistants as an executive and like provide frameworks for how other people can do it.
1:20:29 · And so, um, and then Mike was actually running an AI jam session internally. So, we do these like, you know, I don't know, like once a month or every couple weeks where someone shares something they've done and Mike was doing one on agents. So, it was just like it was all agents all the time. So I actually went into Gemini Enterprise which we have and I went into chat GPT whatever our business license is and I was like okay let me look at how to build agents that enable me um to do my job better or you know fill in gaps of things where I don't have.
1:20:56 · So I basically look at is what are the things I do that are repetitive and recurring um that I would potentially schedule to do for me and then what are the things I would do if I had the time or the human resources to do them. And so what I'm realizing is I I really need a bunch of assistants and analysts, Mike. Like I don't like I I need AI as a thought partner and that's the dominant reason I use it. But I also have a bunch of things that I would do if I could like looking analytics data every morning like pulling reports from key campaigns, things like that.
1:21:26 · And so I started envisioning building um I won't use the word swarm because I think it's such a negative connotation right now. A collection of agents that are very specifically designed to do analyst and assistant type things. Um so a quick example I built a daily brief agent that just runs through my Google calendar. It has access to my Gmail and has access to Google Drive and it can go through and pull anything relevant to the meetings I have in my agenda for the next day. And it emails me at 9:00 the night before and says, "Here's what you got going on.
1:21:57 · here's context, here's what you should prepare, that kind of stuff. Um, this also led to a lot of AI policy talk with Tracy, our COO, and the connectors that we could use with these agents. So, again, I'll share more down the road, but we're doing a lot of work on what I would consider automation agents, not like computer use agents like Mike is talking about. We are not messing with that stuff yet as a company.
1:22:18 · Um, but we [clears throat] are looking at automation agents to support in automating and optimizing workflows to free people up to focus on some bigger picture stuff. That's awesome. All right, let's wrap up here with some product and funding updates. As you'll see from the first three, it's kind of crazy that these are just getting mentioned as updates because any of these could be their own big topic.
AI Product and Funding Updates
1:22:44 · That's how much has been going on this week. But we had a bunch of model releases in addition to GPT6 Astra. So Anthropic first up introduced Claude Fable 5.1 and Claude Mythos 5.1 as the same underlying model with different safeguards. Fable like before is generally available. Mythos is reserved for trusted cyber security partners. Google introduced Gemini 3.8 Flash for coding agentic tasks and complex reasoning. It's at the same introductory price as 3.7 flash.
1:23:12 · They also have a security focused flash cyber version for cyber security defenders. Meta released muse spark 1.3 in muse code and the meta model API which has stronger long horizon agentic work instruction following multitasking and coding efficiency. OpenAI has said Chat GPT ads has reached a billion dollars in annualized revenue run rate less than 200 days after launch as they have expanded their self-service ads manager access across India, Europe, the Middle East, and North Africa.
1:23:49 · OpenAI also added an Epic electronic health record integration. Epic being a health record system and a healthcare public data plugin that connects chat GPT for health to healthcare to nine official sources including PubMed, DailyMed, and CMS coverage. [snorts] OpenAI's current help center has an announcement that we'll probably talk about a bit more once there's some more details here, but they say that personal accounts, including paid plus and pro subscribers, can no longer create or publish new custom GPTs.
1:24:17 · Now, this does not seem to be rolled out uniformly because as of recording, we were able to actually still create GPTs, though it looks like I can't share them anymore. Existing GPTs remain available to use.
1:24:32 · Owners can still edit them if their plan and permissions allow it. Um, sounds like business enterprise edu workspaces retain GPT creation and publishing subject to workspace permissions. It sounds like they're kind of rolling these into workspace agents. It's still kind of a mess and totally unclear, uh, Paul. So, we're going to talk more about it as more comes out.
1:24:51 · I'm I'd assume Astra maybe eclipsed some of the more reasonable decisions to be made around this. [laughter] Anthropic has added beta computer use to Claude Co-work and Claude Code for Pro and Mac subscribers on Mac OS and Windows. This allows Claude, of course, to click, type, and navigate desktop apps with per app permission. OpenClaw released version 2.0, 0 its largest update yet.
1:25:19 · There's a simpler setup, a rebuilt browser experience, and changes across memory, skills, models, automations, apps, plugins, and security.
1:25:27 · Runway previewed Solaris, its first interface world model, which generates visual software interfaces frame by frame in real time. And the company is taking requests for early access while working with partners towards a public launch. And finally, a company called World Labs introduced Atlas, an omni world model that works across text, images, video, and 3D to generate, reconstruct, and simulate spatially consistent worlds. Sounds like world models as before we've talked about them. They might start becoming even more of a thing.
1:25:58 · So, think about all those releases. That was like I kept joking in the internal like sandbox for the podcast was like model week part six. Like this just like endless new models. The other thing, Mike, uh, to mention, Apple has their big event this week, um, where they're going to be launching the foldable iPhone, iPhone 18, I think, and maybe demoing AirPods with cameras in them to have like awareness, think about world models, like awareness of the world around you.
1:26:27 · So, going to be a big week for Apple fans.
1:26:31 · And as one final reminder, go take our AI pulse survey for this past week. We're leaving this one up for just a little longer. Smarterx.ai/p I/Pulse.
1:26:39 · It'll literally take you 10 seconds to answer. We'd love your input. Paul, thanks for unpacking a crazy heavy week this week. Yeah, thanks Mike. And uh be kind to each other out there. Have have good constructive conversations. Like the world needs more like logic based debates and discussions about things, not extreme views on things. It doesn't really help. So, um, yeah, a lot of important things we talked about today and, um, do do your best to like move those conversations forward with friends, family, schools, whatever.
1:27:10 · Like, we all have to be, you know, willing to get out there and talk more about these things. All right, thanks, Mike.
1:27:16 · Thanks, Paul.
1:27:17 · Thanks for listening to the artificial intelligence show. Visit smarterx.ai AI to continue on your AI [music] learning journey. And join more than 100,000 professionals and business leaders who have subscribed to our weekly [music] newsletters, downloaded AI blueprints, attended virtual and in-person events, taken online AI courses, and earned professional certificates [music] from our AI academy, and engaged in the SmarterX Slack community. Until next time, stay curious and explore AI.