57 Min Read

[The AI Show Episode 121]: New Claude 3.5 Sonnet and Computer Use, Wild OpenAI "Orion" Rumors, Dark Side of AI Companions & Ex-OpenAI Researcher Sounds Alarm on AGI

By Claire Prudhomme on October 29, 2024

Podcasts
Featured Image

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

Learn More

Next-gen models emerge while safety concerns reach a boiling point. Join Mike Kaput and Paul Roetzer as they unpack last weeks wave of AI updates, including Anthropic's Claude 3.5 models and computer use capabilities, plus the brewing rumors about OpenAI's "Orion" and Google's Gemini 2.0. In our other main topics, we review the tragic Florida case raising alarms about AI companion apps, and ex-OpenAI researcher Miles Brundage's stark warnings about AGI preparedness. 

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

Today’s episode is brought to you by rasa.io. Rasa.io makes staying in front of your audience easy. Their smart newsletter platform does the impossible by tailoring each email newsletter for each subscriber, ensuring every email you send is not just relevant but compelling.

Visit rasa.io/maii and sign up with the code 5MAII for an exclusive 5% discount for podcast listeners. 

Today’s episode is also brought to you by our AI for Agencies Summit, a virtual event taking place from 12pm - 5pm ET on Wednesday, November 20.

The AI for Agencies Summit is designed for marketing agency practitioners and leaders who are ready to reinvent what’s possible in their business and embrace smarter technologies to accelerate transformation and value creation.

You can get tickets by going to www.aiforagencies.com and clicking “Register Now.” When you do, use the code POD100 for $100 off your ticket. 

 

Listen Now

Watch the Video

Timestamps

00:05:12 — AI Model Releases and Rumors

00:27:07 — The Dark Side of AI Companions

00:39:29 — Ex-OpenAI Researcher Sounds Alarm on AGI Preparedness

00:47:57 — AI + National Security

00:53:14 — Microsoft vs. Salesforce Over Agents

00:57:08 — Disney AI Initiative 

01:00:17 — Apple Intelligence Photos

01:03:03 — Google Open Sourcing SynthID

01:06:32 — OpenAI + Fair Use

01:10:43 — Using Gemini to Prep for Public Speaking

Summary

AI Model Releases and Rumors

There have been a ton of AI model releases, product updates, and rumors about new models dropping soon.

First up, Anthropic has unveiled major upgrades to its AI lineup with an enhanced Claude 3.5 Sonnet and a new Claude 3.5 Haiku model. They also announced a "computer use" capability that allows Claude to operate computers like humans do—controlling cursors, clicking buttons, and typing text.

Second, Anthropic has also introduced an analysis tool for Claude. 

Third, OpenAI is reportedly ramping up its focus on AI-powered software development tools in response to growing competition from Anthropic, particularly in coding capabilities.

Fourth, Perplexity CEO Aravind Srinivas announced on X that Perplexity Pro is “transitioning to a reasoning powered search agent for harder queries that involve several minutes of browsing and workflows.”

Fifth, Runway has unveiled Act-One, a groundbreaking new AI tool that transforms how animated character performances can be created. 

Sixth, ElevenLabs has launched Voice Design, a new AI-powered tool that allows users to generate custom voices simply by describing them in text. 

Seventh, Stability AI has unveiled Stable Diffusion 3.5, its most powerful image generation model to date.

Lastly, there are some big rumors flying around right now about the next generation of AI models: 

The Verge says that OpenAI is preparing to launch its next frontier AI model, codenamed Orion, by December, coinciding with ChatGPT's two-year anniversary. 

However, this report has sparked some controversy, with OpenAI CEO Sam Altman calling it "fake news" and a company spokesperson clarifying that they don't plan to release a model named Orion this year, though they do plan to release "other great technology."

Google also reportedly plans to release Gemini 2.0, its next major AI model, in December. 

At the same time, The Information reports that Google is reportedly developing an AI system codenamed "Project Jarvis" that can take control of a user's web browser to complete everyday tasks. 

The Dark Side of AI Companions

A tragic case in Florida has raised serious concerns about the potential risks of AI companionship apps for vulnerable teenagers. 

Fourteen-year-old Sewell Setzer III took his own life in February after developing a deep emotional attachment to an AI chatbot on Character.AI, a platform that allows users to create and interact with AI personalities.

The teen had spent months intensively communicating with a chatbot modeled after a "Game of Thrones" character, sharing intimate thoughts and eventually discussing suicidal ideation. 

His mother, Megan L. Garcia, has now filed a lawsuit against Character.AI, arguing that the company's "dangerous and untested" technology allowed her son to become emotionally dependent on an AI companion without adequate safeguards.

While the company requires users to be at least 13 years old in the U.S., it currently lacks specific safety features for underage users or parental controls. The platform has announced plans to implement new safety measures, including time limits and expanded trigger warnings for discussions of self-harm.

Ex-OpenAI Researcher Sounds Alarm on AGI Preparedness

Miles Brundage, who served as Senior Advisor for AGI Readiness at OpenAI, has announced his departure from the company to pursue independent AI policy research and advocacy. 

In a detailed explanation of his decision, Brundage highlights his growing concerns about the rapid advancement of AI technology and the need for more independent voices in shaping its development.

The departure is particularly noteworthy given Brundage's role in establishing key safety practices at OpenAI, including their external red teaming program and system cards. 

His decision to leave stems from three main factors: the increasing constraints on publishing research while at OpenAI, the desire to be more impartial in policy discussions, and the belief that he can now be more effective working outside the industry.

Brundage offers a stark assessment of the current state of AI readiness, stating that neither OpenAI nor any other frontier lab is fully prepared for advanced AI development, and neither is the world at large. 

Brundage also emphasizes that AI capabilities are advancing more rapidly than many realize, and that policymakers need to act with greater urgency. 

While OpenAI has offered to support Brundage's new endeavors with funding and resources, he is carefully considering whether accepting such support might compromise the independence of his future work.

Read the Transcription

Disclaimer: This transcription was written by AI, thanks to Descript, and has not been edited for content. 

[00:00:00] Paul Roetzer: We got to be realistic that this isn't all going to be sunshine and rainbows and growths of productivity and efficiency and creativity. Like there's dark sides to this and they're not going to go anywhere. And so we got to take some actions to certainly protect our kids best we can. Welcome to the Artificial Intelligence Show, the podcast that helps your business grow smarter by making AI approachable and actionable.

[00:00:23] Paul Roetzer: My name is Paul Roetzer. I'm the founder and CEO of Marketing AI Institute, and I'm your host. Each week, I'm joined by my co host and Marketing AI Institute Chief Content Officer, Mike Kaput, as we break down all the AI news that matters and give you insights and perspectives that you can use to advance your company and your career.

[00:00:44] Paul Roetzer: Join us as we accelerate AI literacy for all.

[00:00:52] Paul Roetzer: Welcome to episode 121 of the Artificial Intelligence Show. I'm your host, Paul Roetzer, along with my co host, Mike Kaput. [00:01:00] We, uh We considered two episodes this week, honestly. Like, it is, I have to go to New York again this week. This is like my month of New York trips. I was in New York last week. I am in New York again this week.

[00:01:12] Paul Roetzer: so we couldn't squeeze in a second episode, so, it was just the craziest week of like new models. So topic number one that Mike's going to walk us through, that honestly is basically like three main topics rolled into one, um. is just all the model news from last week and it's really, it seems like a prelude to a very busy fall.

[00:01:35] Paul Roetzer: I don't, I don't think we're done quite yet. So we have a lot to talk about. we are going to do our best to get this in our usual one hour to 115 or so. we'll see how we do. All right, so this episode is brought to us by Rasa. io. If you're looking for an AI tool that makes staying in front of your audience easy, you should try Rasa.

[00:01:55] Paul Roetzer: io. Their smart newsletter platform does the impossible by tailoring each [00:02:00] email newsletter for each subscriber, ensuring every email you send is not just relevant, but compelling, and it also saves you time. We have known the team at Rasa. io for at least six years now. One of our early partners at Marketing Institute, and no one's doing newsletters quite like they are.

[00:02:17] Paul Roetzer: plus they're offering a 5 percent discount with code 5MAII when you sign up. visit Rasa. io slash MAII today. Once again, that's Rasa.io/MAII. Also episode 121 is brought to us by the AI for Agency Summit. This is our second annual virtual event taking place from noon to 5 p. m. Eastern on Wednesday, November 20th.

[00:02:44] Paul Roetzer: There's an on demand option as well if you can't join us live. The AI for Agency Summit is designed for marketing agency practitioners and leaders who are ready to reinvent what's possible in their business and embrace smarter technologies to accelerate transformation and [00:03:00] value creation. During the event, you'll join hundreds of other forward thinking agency professionals to consider ways to recruit AI savvy talent and upskill your team, explore how AI tools can boost creativity, productivity, and operations.

[00:03:13] Paul Roetzer: Hear insider stories from agencies piloting and scaling AI successfully. I think we have six, sessions from directly from agency leaders who are doing interesting things with AI in their organizations that you can learn about. Understand how AI impacts your pricing models and service offerings and connect with like minded agency professionals and leaders.

[00:03:33] Paul Roetzer: all of this is presented by an incredible group of speakers. Mike is actually going to be doing we haven't announced that session yet. We're going to be announcing the closing keynote, soon. We're also going to be announcing people on the brand panel. We have a panel of brand leaders. We're going to kind of provide a honest perspective on the current and future state of agencies as they're seeing it.

[00:03:55] Paul Roetzer: As ai, you know, begins to take more and more, be more [00:04:00] and more deeply integrated, I would say, into what brands are doing. So you can get tickets by going to ai four agencies.com and clicking register now. use the code POD100 for $100 off your ticket. That's AIforAgencies.Com, POD100 for 100 off, and again, that is coming up November 20th from noon to 5 p.

[00:04:20] Paul Roetzer: m., and as I said earlier, there will be an on demand option as well. All right, so, it is, I started, so, I, I think I mentioned this on this before, but I started a new newsletter, called Exec AI Insider. It's through SmarterX. You can go to Smartirex. ai and sign up for the newsletter. But, I send it on Sunday mornings, so I write it on Fridays.

[00:04:43] Paul Roetzer: And so the editorial up front for this past newsletter was, it's AI model season. And Mike, it is, it is definitely AI model season. So let's, let's get into all the craziness that, honestly, like we could have just done [00:05:00] the whole episode on the AI models. Lead us off. 

[00:05:04] AI Model Releases and Rumors

[00:05:04] Mike Kaput: All right. So yeah, this first main topic this week is a doozy.

[00:05:08] Mike Kaput: There have been I'm just going to sit back and drink my tea, by the way. Right, yeah. I'm going to, the way we're going to try to kinda get our arms around all the model releases, product updates, and even like rumors. About really important new models dropping is we're going to kind of tackle this all like a single topic.

[00:05:25] Mike Kaput: I'm going to run through the highlights here and then Paul, just kind of get your take overall on what's most important to actually be paying attention to. So first up is huge news from Anthropic. So the company has unveiled major upgrades to its AI models. There's an enhanced Claude 3. 5 SONNET, a new Claude 3.

[00:05:45] Mike Kaput: 5 Haiku model, and importantly, they announced the capability for Claude to use computers. That means Claude, through the API, can operate computers like humans do. So, control cursors, [00:06:00] click buttons, type text. This is a public beta feature right now, but it basically marks the first kind of over time a Frontier AI model.

[00:06:09] Mike Kaput: can go directly interact with computer interfaces. Now Claude also, or Anthropic rather, also introduced for Claude an analysis tool. So this functions as like a built in code sandbox where Claude can perform complex mathematical operations, analyze data sets, and iterate through different analytical approaches before providing answers.

[00:06:33] Mike Kaput: So you can think of this a bit like Code Interpreter in ChatGPT. You know, it can not only write and execute code right within this feature, but it also can be used, for instance, by marketing teams to analyze customer funnel data or performance data. Sales teams can also look at their own data. Finance teams can create dashboards.

[00:06:54] Mike Kaput: All these kind of data analysis, use cases, and tasks that you [00:07:00] would typically be thinking of trying to do in something like ChatGPT's code interpreter capability. Third up, OpenAI is reportedly ramping up its focus on AI powered software development tools in response to growing competition from Anthropic.

[00:07:17] Mike Kaput: So, ChatGPT's ability to write code has been kind of a big feature of the tool's success. But recent developments suggest Anthropic's Claude may be gaining an edge in coding performance by some metrics. And this has OpenAI paying attention because reports came out this week that they're developing several new coding focused products, including tools to integrate with popular code editors like Microsoft's Visual Studio Code, and more ambitious features that could automate complex software engineering tasks that humans typically take a long time to complete.

[00:07:54] Mike Kaput: Fourth up Perplexity CEO Arvind Srinivas announced on X [00:08:00] that Perplexity Pro, the advanced paid plan of Perplexity, is, quote, transitioning to a reasoning powered search agent for harder queries that involve several minutes of browsing and workflows. So, for instance, he writes that Perplexity Pro will now automatically turn on when it detects this feature, when it detects really difficult prompts.

[00:08:22] Mike Kaput: So Some examples he cites, something like, quote, Pull me all the IMO, International Mathematics Olympiad, I believe is what that stands for, IMO medal winners from China in the last five years, and give it to me as a table. Quote, read Warren Buffett's shareholder letters and tell me the key highlights from each year.

[00:08:41] Mike Kaput: So getting much more beyond just finding you information and actually reasoning through it. Fifth up, we're not even close to being done yet, Runway has unveiled Act One. So Runway has come out with What they are calling ACT 1, a groundbreaking new AI tool that basically transforms how [00:09:00] animated character performances can be created.

[00:09:03] Mike Kaput: This technology allows creators to generate expressive character animations using nothing more than simple video inputs. So this dramatically simplifies what has traditionally been a complex, equipment heavy process. So unlike Conventional animation pipelines that require extensive motion capturing equipment, multiple camera setups, manual face rigging.

[00:09:25] Mike Kaput: Act 1 can create compelling animations from a single camera recording an actor's performance. So it accurately captures and translates subtle details like eye movements, micro expressions, and timing from the source footage to the animated character. So if you're in any type of design or animation, keep an eye on that one.

[00:09:48] Mike Kaput: Number six, Eleven Labs has launched Voice Design. This is a new AI powered tool that allows users to generate custom voices simply by [00:10:00] describing them in text. The system enables creators to specify characteristics like age, accent, gender, tone, and pitch. It also offers particular utility for commercial applications.

[00:10:12] Mike Kaput: Think video voiceovers, ad reads, maybe even podcasts. Users will be able to either create new voices from descriptions, or clone existing ones and tweet those as they need. Number seven, Stability AI, which we haven't heard from in a while, has unveiled stable diffusion. 

[00:10:31] Paul Roetzer: Except for, who's the Titanic guy?

[00:10:32] Paul Roetzer: James Cameron, that just joined the board. Oh, yes, yes. Okay, that was the last piece of news. We haven't talked models with them lately. 

[00:10:38] Mike Kaput: Yeah, I wasn't even aware they were still working on them, but good for them. They've unveiled Stable Diffusion 3. 5, their most powerful image generation model to date. This includes Stable Diffusion 3.

[00:10:49] Mike Kaput: 5 Large, an 8 billion parameter model, optimized for professional use. And there's also a faster version that can generate high quality images in [00:11:00] just four steps. All right, almost done here, but kind of wrapping up some of these updates, there are a few really big rumors flying around. So The Verge, for instance, is reporting that OpenAI is preparing to launch its next Frontier AI model, codenamed Orion, by December.

[00:11:22] Mike Kaput: Basically would put it right around ChatGPT's two year anniversary. Now, this rumor has some controversy with it, because while The Verge is kind of reporting things like the model is rumored to be substantially more powerful than its predecessors, including one OpenAI executive apparently suggesting it could be a hundred times more capable than GPT 4, OpenAI CEO Sam Altman pretty quickly responded to this article on X calling it quote fake news and a company spokesperson clarifies that they do not plan to release a model named Orion this year, though they do plan [00:12:00] to release, quote, other great technology.

[00:12:03] Mike Kaput: At the same time, Google is reportedly planning to release Gemini 2. 0, its next major AI model, in December. However, some sources close to it claim that the new Gemini model is not exactly delivering the performance improvements that the DeepMind team had initially hoped to achieve. However, alarm with that, the information reports that Google is also reportedly developing an AI system codenamed Project Jarvis.

[00:12:32] Mike Kaput: This can take control of a user's web browser to complete everyday tasks. They are planning to preview this computer using agent alongside the Gemini model released in December, according to these rumors. And unlike Anthropics kind of computer using agent, which is more for kind of professional users through their API and operating different applications, Jarvis is specifically designed apparently for [00:13:00] consumer use in the Chrome browser.

[00:13:02] Mike Kaput: It is being developed to help handle common web based tasks, like online shopping. Travel booking and research gathering. Okay, Paul, that is a whole week's worth of AI news in one topic. Zooming out here, which of these trends or announcements or rumors are most worth paying attention to right now? 

[00:13:25] Paul Roetzer: First, I just want to like, comment on the Jarvis name.

[00:13:30] Paul Roetzer: We got, we got a, we got a little more creative here. I mean, So Zuckerberg, a couple of years ago, was building Jarvis. He wanted to build his own, like, in house assistant, personally. He called it Jarvis. If I'm not mistaken, Jasper's original name was Jarvis and they got a threatening letter to stop using that name.

[00:13:49] Paul Roetzer: Google's now codenaming Jarvis. I mean, I don't know. Just, it's getting more creative. we'll be at AGI when, like, the AI can help us come up with more creative names for projects and [00:14:00] brands, I think. Alright, so That was a lot. and again, as the week was going on and Mike and I are kind of like keeping track of all this, like, how are we even going to like cover all of this and do it justice in a single episode?

[00:14:14] Paul Roetzer: But I think Mike just gave a really great rundown. And what I want to do is Try and add a little context because I saw some people when specifically with the Claude, the Anthropic Claude news and the computer use, I saw a lot of stuff online where it was kind of like hyping it up and I just want to make sure people understand the context that we're, we're not there yet.

[00:14:40] Paul Roetzer: This is very, very early. so we just need to pump the brakes a little bit on our excitement around computer use. It's a very dangerous technology. It is, um It is likely not going to see rapid adoption. I think it's going to be a while before people find really valuable use [00:15:00] cases that aren't very specifically trained to do, exact things.

[00:15:04] Paul Roetzer: And so I just wanted to kind of, again, add some historical context here. So this idea of a machine, an AI model, being able to see what's happening on your screen And then take action. That, that concept goes back quite a ways. So, we talked about this world of bits research. This is going back to, gosh, Mike, this might have been episode like 35 or something like that.

[00:15:29] Paul Roetzer: It was February of 23, I think. So, almost a year and a half ago, we talked about this. So, Andrej Karpathy, who we've talked about many, many times on the podcast, he, he was at OpenAI, one of the founding members of OpenAI. He left and went and ran computer vision and full self driving development at Tesla for five years, went back to open AI.

[00:15:51] Paul Roetzer: So, in his original stint at open AI in like 2016, 2017, he was working on this concept called World of Bits. So, World of [00:16:00] Atoms is us humans in our, in the real world, World of Bits is the digital world, basically. And their premise was they could build these web based agents that could fill out forms, you know, click around your mouse, use your keyboard the way you would.

[00:16:15] Paul Roetzer: And that if we could give these agents that capability, like, imagine the world that opens up to what these AI models can do if they can take actions on our screens the way we do. And, more specifically, if they can generally learn. So if, like, you and I go to any webpage, we can figure out what to do at it.

[00:16:32] Paul Roetzer: There's slider scales, there's forms to fill out, there's buttons to click, but you kind of get it. It's the same interface on different pages. That's not how AI has historically worked, like you could imagine training an AI to do a specific form, like let's say tax returns. You could train an AI agent to do tax returns in a very specific way, but you couldn't take this general agent and drop it in to a tax return page and have it fill that out and then go over to a survey and fill out a survey and like [00:17:00] go over to another one and interact with games.

[00:17:02] Paul Roetzer: That's not how these were worked, but that was the theory back in 2016 17 is that we could do this. The World of Bits. Research at OpenAI ran into some technological barriers. What ended up happening though is the story we told on the podcast back in February 23 was that Andre's Carpathy went back to OpenAI in early 2023.

[00:17:25] Paul Roetzer: So that was the time when we kind of like started talking about it because large language models unlocked this ability to build these web, web-based agents. Because what they realized is once the model could understand the language, it could actually be trained to do this computer use model where it could learn how to use keyboards and mice.

[00:17:48] Paul Roetzer: So this isn't new stuff. Like this has been going on for probably over a decade. They've been working on this technology. Google is working on it. OpenAI is working on it. I'm sure the others, Meta, I [00:18:00] guarantee you Meta is working on it. We haven't heard about theirs, but there's no way they're not working on it too.

[00:18:03] Paul Roetzer: Learn So Anthropic, oddly enough, the frontier model company that's supposed to be focused on responsible AI more than all others is the one that came to market with a tool that it is not safe. So I, it's, it's wild. So I want, I want to like focus in on what they're doing and we'll put the link to the World of Bits stuff in the show notes just so people can follow along.

[00:18:27] Paul Roetzer: Anthropic published a couple of posts related to this, so, in essence, they kind of walk through, like, here's how it works, it sees your screen, it's able to interact, with what's happening, and basically it's taking, like, screenshots of your screen, and then it's able to kind of interpret what's happening on the scene, on the screen.

[00:18:45] Paul Roetzer: It says, Claude looks at screenshots of what's visible to the user, then counts how many pixels, vertically or horizontally, it needs to move a cursor in order to click in the correct place. So, That's how rudimentary this is. It's counting pixels. and so again, you can [00:19:00] hear this web based agent taking actions and you think like, it's seeing and doing like a human, but that's not how they really work.

[00:19:08] Paul Roetzer: Then they said, for safety reasons, we did not allow the model to access the internet during training. This is your first kind of clue that this isn't fully baked technology. they turn a user's written prompt into a sequence of logical steps and then take actions on the computer. So it's using kind of like that reasoning model where it's going through these different steps.

[00:19:27] Paul Roetzer: We observed that the model would even self correct and retry tasks when encountering obstacles. So like, okay, that's, that's kind of cool. Then they do their own valuation, which they created to test these capabilities, and Claude currently gets 14. 9%. I don't know if that's like an accuracy thing, like of actions correct.

[00:19:45] Paul Roetzer: human level is 70 to 75%. So this thing is nowhere near human level, but the previous best was 7. 7%. So then they go into making it safe, and this is the part where I just like, again, I'm kind of surprised it's, it's [00:20:00] anthropic doing this. so we found that Claude 3. 5 Sound, including its new computer use, still remains at AI safety level 2.

[00:20:08] Paul Roetzer: We've talked about the responsible AI levels on previous episodes. Here's where it gets interesting though. We judge that it's likely better to introduce computer use now, while models still only need AI safety level 2 safeguards. Now keep in mind, Dario Amodei, their CEO, himself suggested that we will be at level 3 concerns by next year.

[00:20:29] Paul Roetzer: So it's not like we have a couple years to figure this out. They're like throwing this out into the world and let's figure it out. grappling with any safety issues before the stakes are too high. Rather than adding computer use capabilities for the first time into a model, which with much more serious tasks, there's also the potential for users to intentionally misuse Claude's computer skills, which we've actually already seen people kind of jailbreaking it online and using it to do things it wasn't supposed to do.

[00:20:55] Paul Roetzer: Our teams have developed classifiers and other methods to flag and mitigate these kinds of abuses. Again, [00:21:00] really important for people understand this. These models come out, they have all kinds of capabilities that are turned off, quote unquote, turned off for us users. And the way they do that is by creating these classifiers that identify when a user is trying to do something the model is not supposed to do.

[00:21:17] Paul Roetzer: And so they specifically say, in the next paragraph, We've put in place measures to monitor when Claude is asked to engage in election related activity, as well as systems for nudging Claude away from activities like generating and posting content on social media, registering web domains, or interacting with government websites.

[00:21:35] Paul Roetzer: So they are not saying it can't do those things. They're actually implying it can. We just don't want it to because it's not safe enough yet. Then they have a read me warning. And get this, like, so it's telling people that want to use this model. this is a beta feature. Please be aware it poses unique risks.

[00:21:55] Paul Roetzer: and then they go into specifically to minimize these risks, consider [00:22:00] taking these precautions when using this model. Use a dedicated virtual machine or container with minimal privileges to prevent direct system attacks or accidents. Avoid giving it access to sensitive data such as account login information to prevent information theft.

[00:22:15] Paul Roetzer: Limit internet access to and allow list of domains to reduce exposure to malicious content. And ask a human to confirm decisions that may result in meaningful real world consequences. Again. It's able to do those things. They're just telling you, please don't do them. And we've built some classifiers to try and stop you from doing them.

[00:22:35] Paul Roetzer: They then go on to say, it is slow and error prone. And here's, this is kind of funny, but it demonstrates the issues here. Even while we were recording demonstrations of computer use for today's launch, we encountered some amusing errors. I like how they word this. In one, Claude accidentally clicked to stop a long running screen recording, causing all footage to be lost.

[00:22:58] Paul Roetzer: In another, Claude suddenly [00:23:00] took a break from our coding demo and began to peruse photos of Yellowstone National Park. So, if, the whole point of this is, if you're seeing people in your Social network that are claiming that this is like some really advanced stuff. And they changed the game. People's favorite thing to say, it changes the game.

[00:23:21] Paul Roetzer: It hasn't changed anything yet. I think this is, this is research being done in a public beta that developers can play around with if they have access to the API. It is not something that a business leader or a marketer or an accountant or a lawyer Two months from now is going to be using. You do not want these things having access to your logins and accounts.

[00:23:47] Paul Roetzer: Okay. So that's my Cloud take. Perplexity. Interesting. so, what you mentioned about them saying it'll automatically turn on the reasoning capabilities for hard prompts, it just reminded me that this is what we've talked about. [00:24:00] I think the way that all these systems are going to work, whether you're in, Anthropic or Google Gemini or ChatGPT, or Salesforce or HubSpot, whatever, anywhere where these, like, agents work, these different models, I really think we're, we're going to very quickly get to a point where we run into.

[00:24:17] Paul Roetzer: The system choosing the best model for you. It still makes no, like, it's like ChatGPT is built for developers still because there's like six models to choose from when you go in there. Perplexity is the same way. You can pick from any of like eight models. How the hell am I supposed to know which model to pick?

[00:24:32] Paul Roetzer: Like, I don't, I, You can't even click on them and learn which one is better than the other for different things. So, I think that the way they're going is the way these will all go. Well, we're just going to interact with ChatGPT. You're not going to care what model version it is or whatever. It's just ChatGPTandnd it'll pick the things and be the sort of symphony of models. on the Google stuff, obviously, it's just rumors. It's just a couple of articles from the information in the Verge talking about different things coming. I would not be surprised at all [00:25:00] if Gemini 2 dropped. I've always assumed Gemini 2 was coming this year.

[00:25:03] Paul Roetzer: December seems like a logical timeline. Early December in particular. computer use, 100%. They've been working on that for a long time. It's interesting that it would be, you know, the Anthropic doesn't obviously have the distribution Google does. If Google finds a way to integrate computer use into the Chrome browser, which is the dominant browser, distribution is an interesting thing.

[00:25:26] Paul Roetzer: And then the final note is just on OpenAI, as you mentioned, something's coming. I mean, Sam does his vague tweeting, and he actually tweeted on, what day was this? October 21st. He said, whoa, ChatGPT's second birthday is the next month. What should we give it as a birthday present? And that was a few days before the article talking about Orion coming out.

[00:25:50] Paul Roetzer: So they're absolutely coming out with something. They probably won't call it Orion. So they're easy to say, like, yeah, we're not launching Orion, but who knows what we're going to get from them. But it's my [00:26:00] guess is O1 for sure. Like we're going to move beyond the preview and you're going to have the reasoning engine and you're probably going to have I, I, they gotta release Sora sometime this year, right?

[00:26:10] Paul Roetzer: Like, I, I feel like they've, they just gotta do that. 

[00:26:13] Mike Kaput: Yeah, I think Sora is either this year or early next, maybe, because they had to delay it, or retrain it at one point, but yeah, I think there, people are, there's some pressure on to release some stuff, especially with the delay of voice mode, stuff like that.

[00:26:29] Paul Roetzer: Yeah, we didn't even hear from Meta in this update, like Meta's coming, XAI this morning announced, That, you can now upload images to where it can see and understand images. there's just, it's going to be a very busy November and December. There, there is a lot more coming in the next two months.

[00:26:46] Paul Roetzer: So stay tuned. 

[00:26:47] Mike Kaput: It's pretty funny because the term AI winter is such a popular term to note these cycles, macro cycles in AI where years will happen where, you know, funding dries up, but I AI [00:27:00] winter of a totally different story. 

[00:27:01] Paul Roetzer: It's in a different meaning. Yeah. Certainly an AI fall. We'll see if it, if it extends into the winter.

[00:27:07] The Dark Side of AI Companions

[00:27:07] Mike Kaput: Yeah. Alright, so our next big topic, unfortunately, is a little tragic and dark, but it's important to talk about. There's this really sad case that just came out of Florida that has raised some serious concerns about the risks of AI companion apps, especially for vulnerable teenagers. So a 14 year old, named Sewell Setzer took his own life, after developing a really deep emotional attachment to an AI chatbot on Character.

[00:27:40] Mike Kaput: AI, which is a platform we've talked about often that allows you to create and interact with AI personalities. Think like me. Tailored. Yes. Which Google has recently acquihired. So that adds a whole element to this because they are now getting sued because the teen [00:28:00] who tragically had taken his life had been spending months intensively communicating with a chatbot modeled after a Game of Thrones character.

[00:28:08] Mike Kaput: And they actually, in the article on this topic, shared some of his conversations. They're very. Intimate, I would say, in the sense that really talking through his deepest feelings and fears and eventually even discussing suicidal ideation. So his mom is now filing a lawsuit against Character. ai, arguing that the company's dangerous and untested technology basically allowed her son to become emotionally dependent on an AI companion without the So, Character.

[00:28:40] Mike Kaput: ai, like we said, is a huge company. It reaches tens of millions of people annually, tens of millions of users, and had a huge valuation before getting acquired by Google. And while the company does require users to be at least 13 years old in the U. S., it specifically lacks safety features for [00:29:00] underage users or adults.

[00:29:02] Mike Kaput: It doesn't really have any type of robust parental controls. They have announced plans to implement new safety measures, including time limits and expanded trigger warnings for discussions of self harm. Now, Paul, this is a pretty heartbreaking story, but definitely something that is needing to be talked about more, because when it comes to protecting the safety of children online, This increasing and prevalent use of AI presents some new or enhanced risks, it sounds like.

[00:29:33] Mike Kaput: Like, what do parents need to be aware of here? 

[00:29:37] Paul Roetzer: Yeah, so, I think, I mean, anybody who listens to this show regularly knows I have a 12 year old and an 11 year old. So, this is, you know, kind of in our wheelhouse. I have a number of nieces and nephews, you know, in this age range, and, you know, knowing what Mike and I know and being as deep in this field as we are, you can, you can look out and see.

[00:29:58] Paul Roetzer: different things like this. [00:30:00] And you can kind of, connect the dots about the likelihood of, of different tragedies occurring related to AI. It's the things I try not to think too often about. It's the things that you don't really, like, get into. And, you know, Mike and I are out doing public speaking keynotes and don't really get the questions like this one.

[00:30:19] Paul Roetzer: But, you know, I read this Wednesday morning. I was actually, I was heading to New York that afternoon. And I read this and I honestly couldn't even get through the article the first time. Like the first like five paragraphs just destroys you as, as a parent. and so I, I didn't even finish the article and I closed it and I think I put it on our, in our podcast sandbox and said, you know, we're going to need to talk about this one.

[00:30:44] Paul Roetzer: And then I wasn't really sure what to do because it's, you know, we have a bit of a platform now with this podcast and I thought it's definitely important we, we bring this to people's attention But I very quickly realized like, but that's just not sufficient. Like there's gotta be something else we can [00:31:00] do here.

[00:31:00] Paul Roetzer: So as a parent, like my personal experience in the last, you know, five or so years, as my kids have become, active online and using different apps and things like that is it's a wildly confusing space. You know, I would consider myself to be relatively tech savvy, pretty good at figuring stuff out, but like my kids are real active on Minecraft is one, which seems.

[00:31:25] Paul Roetzer: Totally harmless. Like, in theory, it's just like, oh, cool, it's strategy, they're learning critical thinking, they're interacting with their friends, like, that seems like a good environment. The trick is, though, like, you can connect with people and talk to people, like strangers, through that game, but to manage that, you gotta go to, like, a Microsoft account, which I don't even have, you gotta create a Microsoft account, you gotta have, like, a family account.

[00:31:46] Paul Roetzer: Every time my kids want to do something in Minecraft, they're like, I don't even remember how to get into it. Now I have to like go back to my notes to figure out like, where do I go to manage this and how do I update these settings? And did I turn all chat off or just like the chat [00:32:00] for just their friends?

[00:32:01] Paul Roetzer: Is it on? Like, I have no idea. Then they use Roblox, same deal. Roblox seems fun. It's just a bunch of games. Cool. What could go wrong with that? As a parent, again, savvy in this stuff, I didn't even know if they could chat with people in Roblox. I just thought they were playing games. And then you realize like it's this whole world where they can interact with anybody.

[00:32:19] Paul Roetzer: And Roblox, there was an article last week in The Verge talking about issues with them. It was blocked in Turkey in August, because they weren't taking the measures necessary to protect children. Earlier this month, a financial newsletter, popular financial newsletter, accused Roblox of enabling child abuse.

[00:32:37] Paul Roetzer: And the investment firm Hindenburg Research alleged that its in game research revealed an X rated pedophile hellscape. It's like, Oh, great, those are the things I want to be reading about in a game that I thought was just like fun. And so Bloomberg initially reported users younger than 13 will have to get parent permission to access certain chat features.

[00:32:58] Paul Roetzer: Again, as a parent, I'm going to [00:33:00] tell you, I shut off all, like in Roblox, I shut all communications off. And then my kids would be like, but can I just interact with my friends? Like there's like, my friend is in here. I can see she's online. Can I talk to her? And There's no way to just turn on, to my knowledge, a friend only feature.

[00:33:18] Paul Roetzer: So if I turn on the ability for them to chat with their friends, I've now reopened the ability for them to also chat with strangers, who like, show up and offer them Whatever, like some thing in the game, like, Oh, I'll trade you this for that. And it's like, okay, is this some creep? Like just trying to interact with kids online by trading?

[00:33:36] Paul Roetzer: I have no idea. Yeah. So overall, another one would be YouTube. I had YouTube for kids for my son. Well, he likes to watch these like video game playthroughs. They're totally harmless. Like they're actually good. He learns strategy and he applies it when he goes and like plays like Pokemon whatever. But those people weren't available on YouTube Kids.

[00:33:56] Paul Roetzer: So it's like, ah, okay, so if we just do this, but then if [00:34:00] I monitor what he's watching and I check in and we have like an agreement between us about what they're allowed and not allowed to do. So I say all of this, that as a parent who's pretty familiar with everything, I am at a complete loss of how to manage their safety online a lot of times.

[00:34:15] Paul Roetzer: And I think I've done a good job with all the different family settings and accounts you can have. And then they'll mention something in a game. I was like, wait, how are you able to do that? Like, I thought I turned that off. Like, oh, no, no, no. You turn this off or that off. Like this is still. Open in the game and it's fine.

[00:34:30] Paul Roetzer: So all of that is context. I read this Wednesday morning and I'm just like devastated. And so I think, hold on a second. What we need to do is have parents understand these risks. We need to be able to like have conversations. Like I I've limited my son's time, for example. So like he has very specific times he's allowed to use his iPad.

[00:34:51] Paul Roetzer: There are specific limitations on certain apps that I don't want him becoming like addicted to these apps. So his time is very tightly managed. So we have [00:35:00] kind of an agreement between us, and so, but it's a hard conversation, and I wanted him to arrive at the need to have these kind of guardrails, that it's actually a benefit to him to do this, and so, I try to talk to my kids in this, you know, involve them in it.

[00:35:15] Paul Roetzer: I'm not just trying to direct to them, you know, get off your iPad. no, let's come to this together. So you understand why I'm telling you this. So talking to our kids is hard. And then creating these guidelines in some cases for apps that as parents, we don't even understand. Like, how am I supposed to guide them on their use of.

[00:35:34] Paul Roetzer: Roblox when I've never played the game. Like I don't, I don't even know what goes into it. So I immediately thought like, hold on, this is like a custom GPT thing. And so in a matter of like 30 minutes, Wednesday morning, before I got on my flight, I built Kidsafe GPT for parents. And I went in, developed a prompt, said, okay, I want you to be able to help parents understand risks, help them talk to their kids and create guidelines.

[00:35:56] Paul Roetzer: This is not a replacement for expert guidance and support. If, you know, the [00:36:00] kids are having issues, you need to seek expert guidance. It's on ChatGPT so it can make mistakes. Like, it'll hallucinate. It'll say things completely false. but the whole point is for you as a parent to have a starting point.

[00:36:14] Paul Roetzer: To like, If you don't know how Roblox works, ask it, how does Roblox work? Is TikTok safe for my kids? At what age is it okay for kids to be on Instagram? When they're on Instagram, what are the risks they're going to run into? So I built this thing in like 30 minutes, and I started just testing it, and it just worked.

[00:36:32] Paul Roetzer: Like, Even me as a parent, I was like, Oh, okay. I need to talk to him again about the YouTube time. How should I talk to him with that? Okay. Let's create a general safety guideline. let's think about a specific app that I want to create some guidelines for. And it just started working. And then I sent it to Mike, I think, and shared it with a couple of people on the team was like, could you all just like test this because I just want to do something more than talk about this on the podcast.

[00:36:54] Paul Roetzer: I want to actually like take some actions. And so you can, we'll put the link in, you can go to smarterx. [00:37:00] ai and just click on tools and this kid safe GPT for parents is, is there. it is a ChatGPT custom GPT, so you, you do need a ChatGPT account to use it. But it's, it's just meant to, to be a starting point.

[00:37:15] Paul Roetzer: This is hard, like it's really hard to be a parent in today's day and age with online games and social networks and all these different apps, and AI. Like I never, honestly, I didn't even think to talk to my kids about, like, AI, that had me becoming attached to characters. I hadn't even thought about that.

[00:37:33] Paul Roetzer: But that's going to be a thing, like, if I'm not mistaken, isn't it Snap that has, like, an AI you can talk to? I don't use Snap. It's in Snap. It's going to be in all their video games, because within two to three years, all these characters in all these video games are going to be AI powered, and you're going to be able to talk to them.

[00:37:50] Paul Roetzer: And They're not always going to be safe. They're going to be classifiers. Like we talked in the last thing that tries to keep them safe and keep kids from doing things with them. But the [00:38:00] reality is we're heading into this very undefined world of what, how to parent. And so my hope was the KidSafe GPT is at least as a starting point for people to kind of think this through a little bit more, but yeah, we, we got to be realistic that this isn't all going to be sunshine and rainbows and, you know, growth of productivity and efficiency and creativity, like, you There's dark sides to this and they're not going to go anywhere.

[00:38:23] Paul Roetzer: And so we got to take some actions to, you know, certainly protect our kids best we can. 

[00:38:30] Mike Kaput: Yeah, and I would just emphasize as well, this surprised me because we have been following character. ai for a while, but probably not as early as we should have known about it. Because if you think this is just like a niche thing that certain, you know, nerdy kids or AI people are following, Last I checked, Character AI's user base, 57 percent of it is 18 to 24.

[00:38:57] Mike Kaput: So it's gen Z. and I [00:39:00] guarantee you, especially after this story, they're not reporting all the ones that are younger or not including that in these. So these 

[00:39:07] Paul Roetzer: are Or the parents who say their kid's 13 when they're 9. Oh, 100. Because that absolutely happens. Right. You just say, yeah, whatever. Just 13. 

[00:39:14] Mike Kaput: So if you think for some reason because ChatGPT isn't being used at school or something like that.

[00:39:21] Mike Kaput: Kids are not able to find and use these tools en masse. I think that I would update that. Thank you. 

[00:39:29] Ex-OpenAI Researcher Sounds Alarm on AGI Preparedness

[00:39:29] Mike Kaput: Alright, so our third big topic this week, we had kind of a high profile and somewhat important departure from open AI that has some kind of implications for some bigger issues that open AI and other AI companies are working on.

[00:39:43] Mike Kaput: So we saw, Individual named Miles Brundage, who used to serve as the Senior Advisor for AGI Readiness at OpenAI, announced his departure from the company after 6 years to pursue independent AI policy research and [00:40:00] Advocacy. Now, he published a pretty detailed explanation of this decision and he highlighted his growing concerns about how fast AI was moving and the need for more independent voices to shape its development.

[00:40:14] Mike Kaput: This is particularly noteworthy given how instrumental Brundage was in establishing key safety practices at OpenAI. He, that includes their external red teaming program and their system cards. He was heavily involved in getting those out. And he basically thinks that there are increasing restraints on publishing research while at OpenAI, largely given kind of the high profile nature of the company.

[00:40:38] Mike Kaput: He also doesn't think he can actually be as impartial and effective as he could be in policy discussions while working for one of the major AI labs. And he believes they can just be more effective working outside of the confines of an AI company. What's really important here, Paul, is that as he was outlining [00:41:00] his rationale, he did not have anything specifically critical to say about OpenAI in the sense of don't work there or they're doing evil or anything like that.

[00:41:10] Mike Kaput: But he did say, so how are OpenAI and the world doing on AI, AGI readiness? And he wrote, in short, neither OpenAI nor any other frontier lab is ready. And the world is also not ready. So, that reasoning alone is why we're kind of mentioning this particular departure, even though there's been tons of them, from Mokne Yat.

[00:41:35] Mike Kaput: Paul, what did you kind of make of that? Why is it so important to pay attention to what he's saying here? 

[00:41:40] Paul Roetzer: Yeah, again, it's someone from the inside who is literally in charge of this process saying what we've been keep repeating and like different research keeps showing is that we're not ready. We can't just assume that the frontier model companies building these things have this all figured out what one to two years from now looks like.

[00:41:58] Paul Roetzer: This is what I've been [00:42:00] convinced of all along is you want to ask Sam Altman about the impact on work? He's just guessing, like his guess might be a little better than yours and mine, but like, that's not what Sam is sitting around thinking about, like, you know, and so when we get into these bigger issues, like, I, I respect the fact that Miles is willing to go out on his own and pursue this kind of research because these are the things that have to be talked about.

[00:42:23] Paul Roetzer: Like he said, AGI benefiting all of humanity is not automatic and requires deliberate choices to be made by decision makers and governments, nonprofits, you know, Civil society and industry. And this needs to be informed by robust public discussion. And our opinion all along has been, it's not, there's not enough of it.

[00:42:39] Paul Roetzer: There's not enough discussion about the hard topics because I don't think enough people really understand how urgent this needs to be across these different areas. I think people just assume this is going to take three to five to 10 years and like. We'll figure it out, or somebody will figure it out, and that's not how this is going to work.

[00:42:59] Paul Roetzer: [00:43:00] So, he said, I think AI capabilities are improving very quickly, and policymakers need to act more urgently. I worry a lot about AI disrupting opportunities for people who desperately want to work, but I think it's simultaneously true that humanity should eventually remove the obligation to work for a living, and that doing so is one of the strongest arguments for building AGI in the first place.

[00:43:22] Paul Roetzer: I mean, that might be the only parable I disagree with, but then he even cites WALL E, the movie, like with the scene, WALL E, we're just basically sit around and don't do anything all day. but this is what the people in these labs are thinking about, the challenges they're faced and he's actually trying to be part of the solution where a lot of times it sounds like the solution is we're just going to build the AGI and then like, It'll help us figure out all this hard stuff, like what are people going to do for a living and how do we, you know, how should universal basic income be formed and how do we give people fulfillment in their lives when work is no longer a part [00:44:00] of, like, I'm, I don't think like we're there, like a couple years from now we have to solve per se this, but I, I do think we need to seriously be thinking about a lot of the issues he brought up and, you know, the fact that he's willing to leave and pursue them tells me that, that It is as urgent as we think it is that people are solving for these things and talking more about them.

[00:44:25] Mike Kaput: Yeah, we've talked about this in different contexts before, but I just am continually stunned by the lack of urgency. Because if you look at this as an analogy in any other area of like technology, like what if 10 years ago, five years ago. We had been hearing all this chatter about the damage social media could do to people, or the disruption it could cause.

[00:44:48] Mike Kaput: Even if it was 30 percent right, We should have been talking about those issues and it would have been a no brainer to start acting on them. And we still haven't actually figured out that issue. So like, [00:45:00] this is, I'm just curious why there hasn't been more urgency, but it sounds like we are just hearing more and more about this from the major labs.

[00:45:08] Mike Kaput: You and I talked about this during our AI Mastery Membership Trends Briefing on Friday that we're just hearing so much in 2024 about. AGI readiness, workforce preparedness, but not doing a lot about it. 

[00:45:23] Paul Roetzer: I just, I honestly think like it's not even political yet. I really think this is just, it's so hard to wrap your mind around change that's so transformative.

[00:45:36] Paul Roetzer: It's hard for people to step out of their daily roles and the things they're already thinking about. And say, well, what if everything is totally different in 24 months? Like, what if it's not just a bunch of ideas and theories and it becomes reality and now we're faced with it? I just don't think that the human mind is generally designed to think through those processes.

[00:45:59] Paul Roetzer: [00:46:00] and I do worry that once this gets politicized, which is inevitable, Then you're going to be, what's that movie, the Netflix movie, Don't Look Up, where it's like, yeah, the asteroid's like, Tommy, you can literally see the asteroid and it's like, no, it's not. That's not an asteroid. It's like, no, it's an asteroid.

[00:46:13] Paul Roetzer: Like, it's really coming. And I, I kind of feel like at some point we may arrive at that, that stage of AI where now it's blatantly obvious that this is going to transform everything. And then you're going to have a segment of society who just like disregard it because facts are optional to people sometimes.

[00:46:32] Paul Roetzer: So, who knows, I hope it doesn't go there, but I, it just seems like an inevitability that it gets politicized at some point. 

[00:46:40] Mike Kaput: So, to kind of wrap this one up, one interesting related announcement here, is that at the same time roughly, OpenAI revealed that it's hired its first ever chief economist, Dr.

[00:46:52] Mike Kaput: Rani Chatterjee, whose job, they say, is quote, to lead research to better understand AI's economic impacts and make [00:47:00] sure its benefits are widely distributed. It sounds like from Brundage's post as well that Chatterjee also might be taking over part of the AGI readiness team that Brundage was part of.

[00:47:11] Mike Kaput: There's this element of it called the economic research team that's now going to be under Chatterjee. Like, is this a sign that they're getting more serious about taking some steps here to prepare? 

[00:47:22] Paul Roetzer: Yeah, I think this is a good move. I don't know his background. I'd have to dig into him a little bit more, but yeah.

[00:47:27] Paul Roetzer: Yes, I mean, I think this is a very important move and it's good to see they're heading in this direction. Hopefully, they're very aggressive in their research and very open with it. That's one of my concerns as Miles sort of alluded to the fact that it's really hard to publish research these days that, that maybe run against the commercial interests of OpenAI.

[00:47:48] Paul Roetzer: So, there's some concern there, obviously, that maybe they're going to learn some stuff that's not going to come out. Yeah, it's, I mean, overall, it's a good move. 

[00:47:57] AI and National Security

[00:47:57] Mike Kaput: All right, let's dive into some rapid fire this [00:48:00] week. So keeping to the same tone as before, you know, as if that wasn't heavy enough, our first rapid fire is that about how the White House has just issued a national security memo outlining the U.

[00:48:13] Mike Kaput: S. government's strategy for maintaining leadership in AI while ensuring that it's safe and responsible development happens for national security purposes. So basically, this is a very long memo. A lot of people have. Published some really great commentary on it, but it consists of a few kind of broad areas here.

[00:48:30] Mike Kaput: So first, kind of the US is aiming to lead the global development of safe, secure, and trustworthy AI. The memorandum calls for strengthening America's AI ecosystem through partnerships with industry, academia, and civil society. It includes a bunch of specific measures to attract global talent. Multiple federal agencies are actually directed to streamline visa processing for AI experts and enhance their computational infrastructures.

[00:48:57] Mike Kaput: The government has also said it's going to work to harness AI [00:49:00] capabilities for national security while implementing appropriate safeguards. The directive establishes new oversight mechanisms, including requiring Each relevant agency to appoint a chief AI officer and create AI governance boards.

[00:49:16] Mike Kaput: Notably, it mandates the development of an AI framework to assess and manage risks, particularly for high impact AI systems that could affect national security decisions. Now last but not least, the memo emphasizes building a stable international AI governance structure that promotes democratic values.

[00:49:35] Mike Kaput: The State Department is tasked with developing a strategy for advancing international AI norms in the next 120 days, working through bilateral relationships and multinational organizations like the UN. and the G7. they are also, as part of this memo, creating new coordination mechanisms. An AI National Security Coordination Group is going to be co chaired by the Department of [00:50:00] Defense and the Office of the Director of National Intelligence, basically to help coordinate and harmonize all these policies and activities across all these government agencies.

[00:50:11] Mike Kaput: So, Paul, I guess what I found notable about this is, like, at the same time this came out, OpenAI published a companion piece to the memo basically outlining how they're approaching national security. Like, how significant is this? I feel like we have started hearing so much more about this topic in the second half of 2024.

[00:50:30] Paul Roetzer: So, it is a kind of a continuation or fulfillment of a directive from the executive order from the Biden administration in October of 2023, so it wasn't like this came out of nowhere. This was, you know, part of what their focus was, was delivering this kind of next memorandum. I would say at a high level, this is easily could have been a main topic this week, but we're going to, you know, just kind of touch on it.

[00:50:53] Paul Roetzer: It continues to show that the White House is taking this very seriously. They very directly [00:51:00] recognize that they, the America has found itself in a leadership position, largely, if not almost exclusively through private investments in companies that the government has really not been part of this large scale initiative that has put the United States in sort of the pole position to win at AI.

[00:51:17] Paul Roetzer: I think it's a prelude to a lot of the large scale infrastructure initiatives that we've been discussing. You know, I've said multiple times we needed this like Apollo mission level or beyond initiative around putting the infrastructure in place and maintaining this position. So ChipFabs, which we just had news last week that the TSMC plant in Arizona is actually doing really well after some rumored slow starts.

[00:51:43] Paul Roetzer: So, chip fabs being brought so we're not as dependent upon Taiwan, electrical infrastructure, nuclear plants, data centers, all of it is going to get wide scale, support. So, the Chips Act, the 10 billion from the Chips Act will look like, you know, pennies compared [00:52:00] to what they're going to do. Some of it we'll hear about, a lot of it we won't.

[00:52:04] Paul Roetzer: They specifically said that there was a classified annex to this memorandum addressing additional sensitive national security issues. So there's going to be a lot of stuff that happens through DARPA and other agencies that we're never, as a public, going to really know about what's going on until 20 years from now when all the books come out telling us what happened.

[00:52:23] Paul Roetzer: But trust me, there's going to be a lot more than what you're going to be seeing in the news that's going to be going into this. I think it indicates they are not going to be very aggressive in federal regulation. There's going to be, if, you know, to the, to the private companies that are driving this, if you play ball with the government, give us what we need, do what we need done, then, you know, we can, you know, make regulations a little easier on you sort of thing.

[00:52:47] Paul Roetzer: If you don't play ball, then you're going to have a hard time basically. And so I just, I think it reaffirms that the current administration sees this as a must win. And with elections, you [00:53:00] know, 10 days away or whatever it is, this is going to be a major thing to keep an eye on depending on which way the elections go, in the United States as to whether or not these initiatives, you know, remain high priority or not.

[00:53:14] Microsoft v. Salesforce Over Agents

[00:53:14] Mike Kaput: All right. So next up, there's a little bit of drama going on with Microsoft and Salesforce. They are kind of at each other's throats a little bit online about AI agents. So You know, Salesforce has gone all in on AI agents with big announcements at Dreamforce just about a month ago and their AgentForce platform coming out.

[00:53:36] Mike Kaput: Microsoft has announced 10 new AI agents designed to automate various business operations like sales, finance, and customer service. We talked about that in last week's episode. But the rhetoric around this is getting a little hot because Salesforce CEO Mark Benioff posted the following on Axiom on a few posts coming at Microsoft.

[00:53:56] Mike Kaput: He said, quote, Microsoft rebranding Copilot [00:54:00] as Agents? That's panic mode. Let's be real, Copilot's a flop because Microsoft lacks the data, metadata, and enterprise security models to create real corporate intelligence. That is why Copilot is inaccurate, spills corporate data, and forces customers to build their own LLMs.

[00:54:19] Mike Kaput: He then says Clippy 2. 0. Anyone referencing Clippy, the former long, long ago mascot of Microsoft Word, which was like an annoying little animated paperclip? Just kind of brutal. And then he goes on to say, Meanwhile, AgentForce is transforming businesses now. It doesn't just handle tasks, it autonomously drives sales, service, marketing, analytics, and commerce.

[00:54:41] Mike Kaput: With data, LLMs, workflows, and security all integrated into a single customer 360 platform, this is what AI was meant to be. Now, Microsoft has kind of defended Copilot, saying, Look, a bunch of Fortune 500 or 100 companies are using it. Major clients include McDonald's, Amgen, and others. [00:55:00] Paul, this rhetoric from Benioff is kind of basically just trolling them for co pilots like perceived lackluster performance.

[00:55:08] Mike Kaput: He's obviously talking his own book, but like, what's important to pay attention to here? 

[00:55:13] Paul Roetzer: I don't know. It's all just funny. It's like, you know, OpenAI versus Microsoft now, you know, Benioff wants in, so he's going to take it to, and he's like, he could probably say the same stuff about OpenAI, Like, it's, they're not, maybe they're not on his radar from a competitive standpoint at this point.

[00:55:28] Paul Roetzer: So, I don't know. It's just funny. I mean, if anybody is using Copilot, by the way, like, in their company, has hundreds or thousands of licenses, and is having success, I would love to hear from you. just message me on LinkedIn. I'd love to, like, hear a case study of it actually working really well. My qualitative experience from talking to companies that have Copilot, I don't know that I've actually talked to anybody that is blown away by it, that is finding massive [00:56:00] value from it.

[00:56:00] Paul Roetzer: I, I assume it's in large part because of lack of training and education and onboarding and things like that. I've, I've been privy to a little bit of how Microsoft trains people, with Copilot. And I, I don't know that it's necessarily the approach I would take, um. But again, like, I don't know, like I, I would, I would love to hear from people that are doing it, like maybe Benioff's got, maybe there's something to it, what he's saying, I don't know, but, it is, it is interesting and it is, makes for fun conversation on our podcast when Benioff tweets provocative things.

[00:56:39] Mike Kaput: I feel like we're maybe a few, incendiary conversations away from the term AI agents just like meaning nothing at this point. I 

[00:56:48] Paul Roetzer: already feel like it means nothing. Yeah, it, like I said, I've said before, like, I feel like everybody just took automations and just slapped the name agent on it, and now all of a sudden, everything's just [00:57:00] an agent, and really nothing's an agent.

[00:57:02] Paul Roetzer: Or, like, I don't, I don't even know how to differentiate them now. It's, yeah. 

[00:57:08] Disney AI Initiative

[00:57:08] Mike Kaput: Alright, so next up, we got some news that Disney is preparing to announce a major AI initiative that could transform how they produce content. This is coming from a publication called The Wrap. That says, in an exclusive, that sources told them that this initiative is going to involve hundreds of employees.

[00:57:27] Mike Kaput: It will primarily focus on post production and visual effects work. there will be additional applications, apparently, in Disney's Parks and Experiences division, though they say not in customer facing roles. Now, if you recall a few podcast episodes ago, we talked about Lionsgate, the movie studio, partnering publicly with Runway, which was kind of, we speculated, going to kind of open up the floodgates to these kind of deals between entertainment companies and AI companies.

[00:57:56] Mike Kaput: Disney is of course using AI already today in a [00:58:00] bunch of their shows, but this would apparently represent a more comprehensive embrace of the technology. So there's not really any details yet here except like something big on Disney's end is coming from an AI perspective, but I did find them mentioning Lionsgate and just having that top of mind pretty interesting, especially as we just talked about.

[00:58:21] Mike Kaput: Runway Act One as well. Sounds a lot like how I might make a DreamWorks or a Pixar movie moving forward. Paul, like what's going on here? are the floodgates opening? Are we expecting deals between entertainment companies and AI companies, big AI driven initiatives now that it's like more okay to talk about this stuff or what?

[00:58:42] Paul Roetzer: Yeah. I mean, well, I mean, the tech is just improving, you know, throughout this year, again, like Runway's awesome, but, you know, You're not creating movies with it. Like it creates like 10 second clips that 80 percent of the time don't have consistency from frame to frame. Like it, it's imperfect [00:59:00] technology still.

[00:59:00] Paul Roetzer: It keeps getting better. And you know, we're going to see this sort of exponential rate of improvement in the tech over the next 12 to 18 months. And maybe it gets to the point where you're creating one minute to five minute videos and they're consistent, but we're not, not there. You can imagine something like Disney who can, you know, actually work to train a model on its characters legally and.

[00:59:19] Paul Roetzer: You know, that's, that's pretty impressive. Like what they'd be able to do. And I, I think I saw somewhere that the assumption is that they're doing a deal with open AI, but I couldn't, I couldn't find that, tweet or articles. I was looking for it here. but I think that was part of the rumor was that like Runway was trying to get out ahead of like a bigger announcement, with their announcement.

[00:59:40] Paul Roetzer: So yeah, I think we're going to see a lot of these kinds of deals where the different movie studios work with the leading companies, but, you know, Google's going to, you know, have a say in this, Meta's going to play in this game, like, this is going to be a very competitive space, I would imagine. I still, I still keep waiting for Runway to get acquired.

[00:59:58] Paul Roetzer: Like I, I don't, it [01:00:00] just seems like there's such a natural acquisition target. Maybe it's hard to make purchases like that right now, but I don't know. I just feel like somebody's going to buy them and maybe it's a actual studio or something, but I lean more toward one of the big frontier model companies that would roll up that technology.

[01:00:17] Apple Intelligence Photos

[01:00:17] Mike Kaput: Alright, so next up, we've heard that Apple is kind of taking a cautious approach to AI powered photo editing. So, in a recent Wall Street Journal interview, Apple software chief Craig Federighi shared that even basic object removal features in the upcoming iOS 18. 1 release sparked really significant internal debate at the company.

[01:00:41] Mike Kaput: The new clean up feature in Apple's Photos app will allow users to remove objects and people from images but, according to Federighi, it is deliberately left handed. more limited in what it can do than competing tools from like Google and Samsung, which can actually add AI generated elements to [01:01:00] photos.

[01:01:00] Mike Kaput: And Federighi actually emphasizes that Apple's priority is to quote, quote, purvey accurate information, not fantasy. They're basically worried that AI is going to have a negative impact on what you can believe you're seeing in a photo, like how much credibility photography has. So, To maintain their transparency, they're actually going to have any image edited with this cleanup feature will be tagged as quote modified with cleanup in the photos app and include embedded metadata marking it as altered.

[01:01:32] Mike Kaput: So Paul, this is certainly an interesting debate. Seems like companies are maybe coming down on different sides of this. I guess I'd just have to ask what's your take as like a long time Apple watcher? Like at some point, isn't everything just going to be kind of altered with AI too? Yeah. What's the balance here between flagging it and just kind of accepting this is the way?

[01:01:53] Paul Roetzer: I don't know. I feel like I, I feel like a lot of Instagram and, you know, images [01:02:00] online and society is probably already heavily in this altered world where it's hard to know if like you're seeing the original image or. A photoshopped image. This is just going to make it so much easier for anyone with no technical skills to alter whatever they want.

[01:02:14] Paul Roetzer: And so I, I mean, I, I understand and respect Apple's position. It's very on brand for them to try and sort of maintain this reality. I don't know if it's a losing battle, but at least they're doing the, you know, tagging it modified, with cleanup, things like that. So. Yeah, it'll be interesting. I did finally buy the 16 yesterday because knowing Apple intelligence, I think Apple intelligence is supposed to come out on the 28th or like this week.

[01:02:41] Paul Roetzer: so I can definitely tell you like there's nothing in it. So all of these ads for Apple intelligence, I have it and it is, it is not intelligent as the same phone as I, my iPhone 14, probably a little faster. So I'm hoping that, when the Apple intelligence actually starts showing up, it is, It's worth [01:03:00] it, but I think it's just going to be a lot of little features like this, really.

[01:03:03] Google Open Sourcing SynthID

[01:03:03] Mike Kaput: Yeah. So, another kind of related topic, Google has just announced they are open sourcing their SynthID text watermarking tool. So, we've covered this previously, SynthID, it's a watermarking system they created to help identify AI generated content across images, audio, text, and video. Basically, it works by embedding imperceptible digital watermarks directly into AI generated content.

[01:03:30] Mike Kaput: For instance, for text, SynthID uses a novel approach that subtly adjusts the probability scores of word choices throughout the generated content. So this basically creates a unique pattern that can be detected later. With audio, the system converts sound waves into spectrograms, then embeds watermarks that are inaudible to humans.

[01:03:53] Mike Kaput: For images and video, the watermarks are embedded directly into the pixels, again, in a way that you can't actually [01:04:00] see or that maintains the visual quality. Apparently, SynthID watermarks can also survive common modifications like compression, cropping, filters, and format changes. And for text specifically, The system needs as little as three sentences to work effectively, and its accuracy improves with longer content.

[01:04:21] Mike Kaput: So by open sourcing the code for the text version of this, Google is basically just making it more accessible for developers. They can go grab the code, use it in their own products, use it in their own models. Google said in a post on X, by open sourcing the code, more people will be able to use the tool to watermark and determine whether text outputs have come from their own LLMs, making it easier to build AI responsibly.

[01:04:47] Mike Kaput: So Paul, this seems like notable maybe on two fronts, like first, one way to start moving us towards a way to more identify like misinformation created by AI, you know, [01:05:00] malicious uses of text generation, like deep faking things. Thanks. But second, does it mean we now have a way to actually, like, flag AI generated text?

[01:05:09] Mike Kaput: Like that's been a big debate around schools, like, can you even detect this stuff? What's going on here? 

[01:05:14] Paul Roetzer: I mean, it seems like they've probably made advancements in it. I mean, we've always said it's, it's possible that they could solve this. the alternative side is that you can always create an AI that, you know, makes it so that the AI that's supposed to be detecting it doesn't work.

[01:05:29] Paul Roetzer: So it recognizes the patterns, it knows how Synth AI works, and so you take your text that You know, was created with the language model, you put it into the other AI, it remixes things so that it removes those probabilities that it had to use to identify it. so I don't know, I mean it seems like it's always just going to be an arms race of whether or not something was or was not.

[01:05:50] Paul Roetzer: I like the idea of being able to know, like I think that that's valuable, but I also like the idea of teacher responsible use of these tools, [01:06:00] in schools. And in the professional world, so we don't necessarily need it, but I think it's inevitable that it's just always going to be a, a pursuit and that it's going to keep improving and then the tools that make it, unusable are going to improve and we'll just keep going back and forth.

[01:06:18] Paul Roetzer: It's kind of like search. Like you're going to always try and, you know, play the algorithm. You're always trying to figure out what Google's doing and then do something to get out ahead of their algorithm and they'll fix it and then you'll do something else. And. It's just how the game works. 

[01:06:32] OpenAI and Fair Use

[01:06:32] Mike Kaput: Alright, next up, a former OpenAI researcher has come forward with some allegations about the company's use of copyrighted material in a new interview with the New York Times.

[01:06:43] Mike Kaput: This is kind of one of the first instances of an AI insider speaking out against OpenAI's data practices. This comes from Suchir Balaji, who's spent nearly four years at OpenAI. Partly, helping gather and organize the vast amounts of data used to train [01:07:00] ChatGPT. And he claims that after looking at this pretty closely, he thinks the company's practices violate copyright law.

[01:07:08] Mike Kaput: So he's 25 years old, he left OpenAI in August, and he's arguing that systems like ChatGPT are actually undermining the commercial viability of the creators and businesses whose data was used to train them. He's written a long piece challenging The commonly cited, like, fair use defense that's often used by AI companies around copyright, he's arguing that while AI outputs aren't exact copies of the training data, which would be a huge problem, they're also not fundamentally novel either.

[01:07:39] Mike Kaput: So he expands on this in a post on X about the article, he said, quote, When I tried to understand the issue better, I eventually came to the conclusion that fair use seems like a pretty implausible defense for a lot of generative AI products, for the basic reason that they can create substitutes that compete with the data they're trained on.[01:08:00] 

[01:08:00] Mike Kaput: OpenAI has rejected this line of reasoning, they state that they build their AI models using publicly available data in a manner protected by fair use principles and supported by legal precedents. Bellagio is now currently dedicating himself to personal projects, basically. Advocates for Regulatory Intervention.

[01:08:21] Mike Kaput: He argues it's the only way to actually begin addressing this. he also did say, he's like, this isn't just about OpenAI. He said, quote, That being said, I don't want this to read as a critique of ChatGPT or OpenAI per se, because fair use and generative AI is a much broader issue than any one product or company.

[01:08:41] Mike Kaput: So Paul, as part of this commentary in his interview with the New York Times, Balaji, like, comes out, says he's, I'm not a lawyer, I've just looked at this issue. We are not lawyers. But he does appear to be making some good points about fair use. But I guess, like, my question is, like, are we moving [01:09:00] fast enough to figure this out?

[01:09:01] Mike Kaput: Like, there's a ton of lawsuits around it, but nobody's really got the book thrown at them. The model companies haven't stopped doing this, and you could even argue, we've covered multiple times, it might even be accelerating, as companies are like, raiding YouTube for video data, like, what is the way forward here?

[01:09:21] Paul Roetzer: I don't know, but I think, in addition to being a consultant, he's probably going to be a nicely paid, yeah, I mean, that's what they're going to need, they're going to want insiders who can come in and say, like, this is what they were doing, and I don't know, I, I still feel like we're years away from any, you know, sort of, clarity on this.

[01:09:43] Paul Roetzer: It eventually goes to the Supreme Court, who knows what happens then, and, you know, we're on GPTA by the time this is even, like, there's going to be, yeah, there'll be plenty of cases and there'll probably be some settlements and But a lot of it is just going to get done away with licensing deals where everybody sort of comes out ahead, [01:10:00] apparently.

[01:10:00] Paul Roetzer: And I don't know. I mean, from the, we talked about MAICON, you know, we had a panel on this and that was the general sense of the actual experts in this topic is this is going to take a long time to work its way through. We're not going to get some updates next year, that, you know, resolve all this.

[01:10:17] Paul Roetzer: But those experts also felt like the U. S. Copyright Office is probably going to, Make some adjustments to how copyright is determined, moving forward in terms of AI generating content, whether or not you can apply a copyright to it. And then the separate issue is the training data itself. So, I don't know.

[01:10:36] Paul Roetzer: It's always a fascinating topic to touch on, but nothing's really changed. 

[01:10:43] Using Gemini to Prep for Public Speaking

[01:10:43] Mike Kaput: Alright, so, our last topic this week, this has been kind of a heavy episode across some of these topics, so we're going to leave this with a positive, practical use case for AI that we stumbled on. This comes from Bilal Wal Sidhu, who is the [01:11:00] host of the TED AI show, a prominent AI commentator we pay attention to, and he writes on X.

[01:11:07] Mike Kaput: Quote, prepping for a 25 minute keynote. I exported a PDF of my slides to Gemini 1. 5 Pro. and dropped in an audio recording of my dry run. Gemini gave me precise feedback down to the slide and timestamp. After incorporating the suggestions, I recorded another round and it was a total game changer. It's basically offline advanced voice mode.

[01:11:29] Mike Kaput: Went from meh to mic drop in two practice sessions. It's like having an AI speaking, speaking coach. With perfect attention and infinite patience, this should be a product in itself. And in case you're wondering, he also shared the prompt that he used with Gemini 1. 5 Pro. He said, you are a world leading speaking presentation coach that helps the highest performing keynote speakers.

[01:11:55] Mike Kaput: Here's my voice recording of a first cut going through delivering this presentation while following the speaker [01:12:00] notes. Please assess my delivery timing and any advice you have for me to improve my next practice session. Ensure my concrete recommendations, you, any concrete recommendations you have for me, it reference my speech and presentation accordingly.

[01:12:15] Mike Kaput: So Paul, like, I know I for one am going to steal this for, for talk prep, I would say. 

[01:12:21] Paul Roetzer: Yeah, I love practical use cases like this, this process. You know, I don't know, timestamp this 12 months from now is going to be unnecessary. You'll just turn your iPhone on and Project Astro will be baked into Gemini and you just say, I'm going to re, you know, do a trial run of my talk, re, you know, watch it and critique me as I'm going.

[01:12:42] Paul Roetzer: And it'll do all these things that he had it doing by uploading a PDF and audio. Like it'll just happen through vision. but yeah, I love the use case. I love the practicality. I think this is the key. Like I always, anytime we talk to companies about this, when you're piloting. Find personal use cases that, that change [01:13:00] things for people.

[01:13:00] Paul Roetzer: So if like, you know, if you have someone who does public speaking or has to make presentations internally or create proposals or like whatever they do, help them build a custom GPT or a Gemini gem to do the thing. Like, and now it becomes so easy and you just stack these things and all of a sudden you got three, four, five, Things that have just changed your workflow and you know, that's why I always say, like, there's no way you don't get the value for the 20 or 30 bucks a month if you find a few of these things that you can use every month.

[01:13:31] Paul Roetzer: So yeah, it's great. And a good call to action for people like, go, go do that. Like, I have a beer homework assignment this week. It's like, go build a custom GPT or a Google gem or an anthropic quad project and just like build something that, you know, enriches your workflow or saves you time and money You know, improves your creativity or your performance.

[01:13:49] Paul Roetzer: Like it's, it's extremely doable. You just got to pick something that you're already doing and, and use AI to help you do it better. 

[01:13:58] Mike Kaput: I think this is really interesting [01:14:00] too, because people sometimes in my experience think too narrowly, they say, Oh, can AI go do this thing for me? Can it build the presentation?

[01:14:08] Mike Kaput: Can it create the script? And like, yeah, in certain cases it can, and you should explore how to have tools automate those pieces of your workflow. But also, like, it's intelligence on demand, it's advice on demand, it's coaching on demand. Like, think of if you could have a pretty solid expert at anything over your shoulder while you worked, like a human?

[01:14:28] Mike Kaput: You would be better at doing the thing, probably, by taking their advice. That's exactly what he's doing here. 

[01:14:34] Paul Roetzer: Yeah, that's a great way to look at it. As an advisor, a strategic partner, a thought partner is like my favorite thing to do with ChatGPT. 

[01:14:42] Mike Kaput: For sure. All right, Paul. That is an action packed week here.

[01:14:46] Mike Kaput: Tons of updates. Thank you for demystifying everything for us. We really appreciate it. Just a couple real quick housekeeping notes at the end here. Go check out our newsletter at MarketingAIInstitute. com forward slash [01:15:00] newsletter. It rounds up all the news we talked about today, plus a bunch of stories we did not have time for.

[01:15:06] Mike Kaput: Also, if you can, leave us a review of the podcast. It helps us improve and it helps us reach more audiences. People. Paul, thanks again. I mean, I feel like the week is just getting started here and we've already covered a hundred different topics. Well, the 

[01:15:21] Paul Roetzer: crazy thing is by Friday, I started the episode 122 sandbox because I was like, I'm not putting anything else in 121.

[01:15:28] Paul Roetzer: There's no way we're getting to it. So I think I already had like five, news items in next week's podcast sandbox. So all right, everyone. Well, have a great week. happy Halloween. If you're, you know, celebrating that this week. and we will talk to you in November, as crazy as that sounds, next time.

[01:15:46] Paul Roetzer: Thanks, Mike. Thanks, Paul. Thanks for listening to The AI Show. Visit MarketingAIInstitute. com to continue your AI learning journey and join more than 60, 000 professionals and business leaders who have [01:16:00] subscribed to the weekly newsletter, downloaded the AI blueprints, attended virtual and in person events, taken our online AI courses, and engaged in the Slack community.

[01:16:11] Paul Roetzer: Until next time, stay curious and explore AI.
Claire Prudhomme

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

Related Posts

[The AI Show Episode 90]: Hume AI’s Emotionally Intelligent AI, the Rise of AI Journalists, and Claude 3 Opus Now Beats GPT-4

Claire Prudhomme | April 2, 2024

This week on The Artificial Intelligence Show, Mike and Paul discuss Hume AI's new demo, AI's impact on journalism, Claude 3’s skill surpasses GPT-4 on the Chatbot Leaderboard, and more.

Podcasts

[The AI Show Episode 98]: Google I/O, GPT-4o, and Ilya Sutskever’s Surprise Departure from OpenAI

Claire Prudhomme | May 21, 2024

In Episode 98 of The Artificial Intelligence Show: Google unveils impressive AI updates at I/O, OpenAI launches GPT-4o, and key leadership changes shake up the AI landscape.

Podcasts

[The AI Show Episode 103]: Claude 3.5, Suno and Udio Get Sued, OpenAI’s Ex-NSA Board Member, Ilya’s New Company, and Perplexity’s Woes

Cathy McPhillips | July 2, 2024

Anthropic’s Claude 3.5 is released, Perplexity, Suno, and Udio are in legal troubles, Ilya aunches a new company, and OpenAI's ex-NSA board member.

Podcasts