If you listened to the podcast last week, you know that MAICON 2023 wrapped up on July 28. Last week’s episode covered MAICON, plus a handful of other topics. Thinking things may slow down for a minute, we’re back to cover even more topics than usual, in what was a busy and groundbreaking week in the world of AI.
Listen or watch below—and see below for show notes and the transcript.
This episode is brought to you by MAICON, our Marketing AI Conference. Main stage recordings are now available for purchase, and a $50 discount code is mentioned at the start of the show.
00:05:45 — Big ChatGPT updates
00:13:44 — HyperWrite launches Agent-1
00:24:06 — Palmyra-Med, a powerful LLM designed for healthcare, is announced
00:28:54 — McKinsey State of AI Report is published
00:38:08 — 3 new things you can do with generative AI in Google search
00:41:45 — Google Assistant gets a makeover
00:43:50 — YouTube tests using AI to summarize videos
00:48:15 — Meta’s AI ‘personas’ might launch next month
00:51:10 — Meta launches AudioCraft
00:55:05 — Partnership on AI creating AI guidelines for newsrooms
00:58:13 — Zoom says its new AI tools aren’t stealing ownership of your content
Huge updates for ChatGPT are announced
OpenAI just announced huge updates for ChatGPT that hold the potential to transform how we use this popular AI tool. The updates include a number of things including: Prompt examples: At the beginning of a new chat, you’ll now see examples to help you get started; Suggested replies: ChatGPT now suggests relevant ways to continue your conversation; GPT-4 by default: When starting a new chat as a Plus user, ChatGPT will remember your previously selected model — no more defaulting back to GPT-3.5; Upload multiple files: You can now ask ChatGPT to analyze data and generate insights across multiple files. This is available with the Code Interpreter beta for all Plus users; Stay logged in: You’ll no longer be logged out every 2 weeks; Keyboard shortcuts: Work faster with shortcuts, like ⌘ (Ctrl) + Shift + ; to copy last code block. Try ⌘ (Ctrl) + / to see the complete list.
Also, missing from the update announcement but spotted by entrepreneur Neal Khosla on Twitter (and confirmed by Marketing AI Institute) it looks like the cap on messages in GPT-4 (within ChatGPT Plus) has gone away.
A new, autonomous AI agent debuts from OthersideAI
Matt Shumer, the CEO of OthersideAI, maker of the popular AI writing tool HyperWrite, just debuted an AI system called “Agent-1,” a breakthrough model that can operate software like a human. Agent-1 will power the company’s Personal Assistant product, which lets you give AI commands that it can then execute autonomously using your web browser. In a demo video posted to Twitter, Shumer showed Agent-1 controlling a Google Cloud dashboard on its own.
Other demo videos of Personal Assistant have shown the tool autonomously planning travel for a user and drafting and sending an email to team members on its own based on a simple command given by the user.
A vertical-specific LLM launches - and is a good example of what’s possible
Writer, a leading AI software tool and friend of Marketing AI Institute, has released a large language model designed specifically for use in healthcare. The model is called Palmyra-Med, and Writer says it has outperformed both GPT-4 and medically trained human test-takers on PubMedQA, the leading benchmark for biomedical question answering.
Unlike a generic model like GPT-4, which knows a little bit about a lot of things and, as such, can be used for a wide variety of tasks, Palmyra-Med is specifically trained on publicly available sets of medical data. Writer clearly hopes to make generative AI much more accessible to healthcare organizations. Historically, adoption in healthcare of LLMs has been limited, given healthcare organizations’ needs for specific medical accuracy from AI tools and robust security and compliance features that many AI systems lack. How will this change AI in an industry like healthcare?
There are many more topics to be discussed, including McKinsey’s annual report, and a peek inside what’s happening over at Zoom.
Disclaimer: This transcription was written by AI, thanks to Descript, and has not been edited for content.
[00:00:00] Paul Roetzer: do I trust the company building it?
[00:00:01] Paul Roetzer: And then am I crossing any ethical lines or legal lines using it? Those are messy questions and they just aren't good answers yet. It's kinda like the generative ai and is it, breaking copyright law, we don't know yet. And so that's the first way I would look at this.
[00:00:16] Paul Roetzer: So it definitely fits under that law of uneven AI distribution of that. Like, I'm not sure I'm willing to give up what would be needed to use this tool because there's lots of unknowns about it.
[00:00:27] Paul Roetzer: Welcome to the Marketing AI Show, the podcast that helps your business grow smarter by making artificial intelligence approachable and actionable. You'll hear from top authors, entrepreneurs, researchers, and executives as they share case studies, strategies, and technologies that have the power to transform your business and your career.
[00:00:47] Paul Roetzer: My name is Paul Roetzer. I'm the founder of Marketing AI Institute, and I'm your host.
[00:00:57] Paul Roetzer: Welcome to episode 58 of the Marketing AI Show. I'm your host, Paul Roetzer, along with my co-host Mike Kaput. Hello, Mike. What's going on? How's it going, Paul? What's up? Fresh off the golf course. We had the Orange Effect golf outing today. Our friends, Joe and Pam Pulizzi, do amazing work at the Orange Effect Foundation.
[00:01:18] Paul Roetzer: Benefits children in need of speech therapy, services and technology. It's an incredible organization. We've been doing this, I think this was, I'm going to guess like the sixth or seventh year of this event, but it, it actually may go way back longer than that, that years are starting to run together. But we were out there, today, you know, doing that and just, kudos to them.
[00:01:37] Paul Roetzer: They've raised hundreds of thousands of dollars and benefited hundreds of families with that organization. So, Huge fan of what they're doing there. So that was my morning, and then I was getting made fun of because I was ditching early and not partaking in any beverages because I said, I gotta record the podcast today.
[00:01:54] Paul Roetzer: And they're like, come on man, let's go play another a team. I'm like, no, I gotta do it. We gotta stick to the podcast schedule. So here I am. We're recording a little later, than usual in the day. But otherwise we're sticking to our Monday schedule. So we are recording this on Monday, August 7th. So this is coming out on the, on August 8th.
[00:02:14] Paul Roetzer: Give you a little context what's going on. It was an interesting week in ai. We'll get into that in a minute. But first, our, episode today is brought to us by MAICON 2023. So I know a lot of you who listened to the show regularly joined us in Cleveland for MAICON 2023, but I know even more of you weren't able to be there for a variety of reasons.
[00:02:34] Paul Roetzer: Some are international, you know, obviously scheduling, whatever the reason was. We recorded all the main stage sessions, so there was three tracks at this year's conference. And so all the main stage, which included all the featured and keynote general sessions, as well as some of the feature featured breakout sessions.
[00:02:52] Paul Roetzer: So 17 sessions in total were recorded and are now available on demand. So this includes my opening keynote on the state of AI for marketing and business, the fireside chat with Ethan Mollick, which was like the talk of the conference, especially on day one. Just an incredible conversation with Ethan Mollick.
[00:03:10] Paul Roetzer: I. Beyond the obvious was the topic there. Cassie Kozyrkov, the Chief Decision Scientist at Google, whose job does AI automate. Chris Penn did an incredible deep dive on large language models. Dan Slagan talked about the marketing org chart of the future, c m o tomorrow.io. We had panels from CMOs, agency leaders, tech executives on what's next.
[00:03:33] Paul Roetzer: VMware. We had Jessica Hreha and talking about building an internal AI council. We had ethics and AI and Olivia Gambelin. Just amazing talk. So if you missed it and you want to see it, or if you were there and you want to, you know, revisit everything or some of the SA sessions you may have missed, just go to, make con.ai.
[00:03:51] Paul Roetzer: Oh, and we had Mike's, which I think was the highest rated, one of all 45 tools in 45 minutes. The marathon of like AI tools. So that's also in there. We had a few breakout sessions that were really popular, so we recorded those and put them in there as well. So, just go to MAICON.ai, and right at the top you can scroll down.
[00:04:11] Paul Roetzer: There's a register button for next year, and then there's a buy MAICON 2023 on demand. It's 4 99 for the on demand for all 17 of those. And then you can use AI Pod 50 promo code, and that'll get you $50 off. So again, MAICON ai, click on the buy MAICON 2023 on demand button, and you can watch all of those sessions on demand.
[00:04:32] Paul Roetzer: They're live now. They've been, we got 'em up on Friday. Tracy Lewis on our team's. Incredible work. She hustled and got that stuff up there within a week of the show ending. So, check those out. All right, so this week was interesting. It was funny when we were, you know, when, I think it was on Saturday when I realized I was going to be out Monday for this golf outing.
[00:04:52] Paul Roetzer: II sent it to Mike. I was like, okay, we have three options. One, we just don't do an episode this week. Two, we record late Monday, three, we record Tuesday. And, Igotta be honest, there was this part of me post-conference where I was like, let's just bump a week. Let's just not do it. Because I looked at the week and I was like, wow, there wasn't that much that happened.
[00:05:11] Paul Roetzer: Like usually we have like 30 or 40 links in our Zoom message, our sandbox each week we were like 15 or something. Then I started looking through, I was like, no, we gotta talk about this stuff. Like if we don't do it this week, like these things are going to stack up to the next week. And so we've got more, kind of like, of a rapid fire today.
[00:05:27] Paul Roetzer: We're going to go through and do our usual three main topics. But we're going to kind of try and get through, I think we got about 10 or 12 quick hitting topics today. A lot of interesting stuff. So we wanted to, you know, stick to the schedule and hit it with her every Tuesday. So Mike, I'll turn it over to you and let you get us started on the topics for the week.
[00:05:44] Paul Roetzer: Sounds great,
[00:05:45] Mike Kaput: Paul. First up we have some big ChatGPT updates. So OpenAI just announced some significant updates for ChatGPT and these kind of hold some serious potential to transform how we're all using ChatGPT as I assume most of us are. The updates include things like, first up, we now have prompt examples rolling out.
[00:06:08] Mike Kaput: So at the beginning of a new chat, you'll now see examples to help you get started engaging with ChatGPT. It also is going to start giving you suggested replies. So ChatGPT will now suggest relevant ways to continue your conversation based on what you already have asked it or talked to it about.
[00:06:27] Mike Kaput: What's really cool as well is GPT-4 is now selected by default. If you are starting a new chat as a plus user, typically it would default to chat, to GPT-3.5. So now you'll automatically be on GPT-4 by default. You're also able to now. Upload multiple files. So for anyone who doesn't realize, the code interpreter, beta feature for plus users allows you to upload files to actually start analyzing data and generating insights.
[00:06:56] Mike Kaput: And you can now do that with more than one file. You're going to no longer be logged out every single two weeks, which is kind of nice. That was a bit of a frustration for some users. And there are also now some keyboard shortcuts, which you can go to. The link in the show notes to see what those are.
[00:07:15] Mike Kaput: Just bas basically ways to work faster, in ChatGPT. Missing from the latest updates that were documented by OpenAI was something that was spotted on Twitter by entrepreneur Neil sla, and I've confirmed it in my ChatGPT plus account. It looks like there are no longer caps on messages in GPT-4.
[00:07:37] Mike Kaput: It used to be, I think 50 every three hours after they raised the limit. I'm no longer seeing any documentation in ChatGPT plus that tells you there's a limit. So it looks like that may also be gone, as well. So first up, I found it pretty interesting that OpenAI is now offering prompt examples and suggested replies.
[00:07:59] Mike Kaput: So Paul, these seem like hugely valuable features, but do they devalue or kind of remove the need for having prompt engineering
[00:08:07] Paul Roetzer: skills? I don't think so. I mean, we've talked about this as you know, prompting is definitely, gives you a leg up in terms of getting value out of these systems, but they're, it's also probably the greatest friction point to getting value outta these systems.
[00:08:23] Paul Roetzer: And the companies building these models know that. So it's inevitable that they're going to build prompting capabilities into these systems that rely less on the human's ability to give the prompt. Because the better the system at is interpreting or enhancing the prompt it's given, whether it shows you it or it's happening behind the scenes, the quicker you are going to get value outta these systems and the more value you'll get from them.
[00:08:49] Paul Roetzer: So from the beginning, we've said, I mean, going back to December, we probably have podcast, podcast episodes talking about over time your ability to prompt is probably going to become a. Less important. And we've seen that, like, I think I mentioned, runway gen two when I first went into their mobile app.
[00:09:08] Paul Roetzer: II, it seemed like they took it away at some point, but it would, it would take your prompt and it would actually give you like three or four enhancements to your prompt and say, is this better? And most times it was. And so what you realize is they, they're learning a ton of how to properly prompt these systems.
[00:09:27] Paul Roetzer: And so right now there's a bit of an art and a science to it, but my assumption has always been, I. That these companies will eventually reduce the importance of the human's ability to prompt. So I think we're starting to see that here. I know you can already see, like even with the suggested replies you talk about, so like Google search generative experience has that, where it recommends follow on questions.
[00:09:54] Paul Roetzer: We're starting to see that where it'll kind of like prompt you of other things to be asking the system and in some ways that's injecting the prompts themselves. But yeah, I think it's obvious that where it's going is either all of them, you'll either put your prompt in and it'll behind the scenes make that prompt better without you seeing what it's doing.
[00:10:14] Paul Roetzer: Or it'll show you suggested alternatives that are just more enriched versions of your prompt and they're often probably going to be better than your. Now the other thing is you will learn how to better prompt by seeing this because you're going to be seeing what they think good prompts look like, and so you'll probably start to become better at it.
[00:10:33] Paul Roetzer: So it's not that prompting as a skill isn't going to be valuable, it's just we won't rely as much on it in the future as we do today to get value from these systems. So it also
[00:10:47] Mike Kaput: seemed to me, given how valuable we found code interpreter to be that kind of uploading multiple files is. Somewhat of a big deal, especially if you're an audience member who hasn't used code interpreter or is new to it.
[00:11:00] Mike Kaput: I mean, this thing is pretty impressive. And I was wondering if maybe you could just kind of walk us through at a high level, zooming out a little bit why this data analysis functionality is so important.
[00:11:10] Paul Roetzer: Yeah. Well, most of us just aren't very good at it. I mean, especially if we're talking, you know, about marketers and business executives.
[00:11:16] Paul Roetzer: A lot of people that listen to this show, it's just not your course competency. So you and I both can do data analysis. It's not what we were trained to do. We've developed those skills through the years, but I always look at it, I'm sure I have a ton of blind spots when it comes to data analysis. Like even when I go into Excel, I can do the basic functions and I can build pivot tables and I can do some stuff, but I'm sure that I know maybe about 10% of the actual capabilities of Excel.
[00:11:44] Paul Roetzer: And so if you just imagine what a data scientist or someone who is, has more advanced data analysis capabilities can do with Excel and some data. I can't do most of that. So just to have that, I mean, in some cases people said it's like almost a PhD level, data analyst that you can call up at any moment, say, here's my data.
[00:12:06] Paul Roetzer: What should I, what should I be asking of this? What, how should I visualize this data? So just imagine having that at your fingertips 24 7 for $20 a month, like is a part of it. So, you know, I think that's the stuff and that's where Ethan MoCo, I mentioned, you know, upfront on the on demand, who is one of our keynotes.
[00:12:26] Paul Roetzer: He's done some of the coolest stuff I've seen in terms of sharing ways. He's applying code interpreter. But I do think that that's going to be one of the most, immediate. Value creators for this whole generative AI movement. As these tools, one code interpreter gets wider use, but two, those same capabilities eventually get baked into Excel, into Google Sheets.
[00:12:50] Paul Roetzer: And all of us, like all knowledge workers have these, like on-demand data analyst capabilities. It's going to be transformative for a lot of knowledge work that's done. So I think we're just starting to see what tools like code interpreter are going to be able to do. It's very early, but it's exciting to me because, like, I love these AI tools that are additive.
[00:13:13] Paul Roetzer: Like there's, there's, if you, I mean, if we had a hundred person company, we wouldn't have a data analyst on staff. Like the, it's not replacing any jobs. It is like, Giving us a superpower we just otherwise wouldn't have. And so I think for a lot of marketers and business people, tools like code, code interpreter are going to be an additive AI thing that enhances what we're capable of without affecting any jobs.
[00:13:38] Paul Roetzer: And that's exciting to me. Like I love those kinds of use cases.
[00:13:43] Mike Kaput: So next up, we have a interesting development in the world of AI agents. Matt Schumer, who's the c e O of Other side ai, which makes the popular AI writing tool hyper right, and who we know well through the institute. He just released an AI system called Agent One.
[00:14:00] Mike Kaput: I. This is a breakthrough model that can operate software like a human. So Agent one is going to power the company's personal assistant product, which we've mentioned before on the podcast, and that basically lets you give AI command that it can then execute autonomously using your web browser. So in a demo video that Matt Schumer posted to Twitter, it showed agent one controlling a Google Cloud dashboard on its own.
[00:14:26] Mike Kaput: Other demo videos in the past have personal assistant have shown that tool autonomously planning travel for a user and drafting and sending an email to team members based on its own, on its own, based on a simple command given by the user saying, Hey, send a congratulatory email to the team that's a couple of sentences and send it to this address, and AI agents will then go ahead and do that now.
[00:14:51] Mike Kaput: This isn't the first autonomous AI agent we've seen out there. We've talked a bit about auto GPT in the past, which made a lot of waves, but this does seem like an interesting step forward. We're now seeing this technology begin to get deployed into actual products like personal assistant from Matt and the team.
[00:15:10] Mike Kaput: Do you think we're about to see this rise in autonomous AI agents for
[00:15:13] Paul Roetzer: commercial use? I don't think there's any doubt. We've talked about this many times on the show previously, that this is one of the next phases of AI is the AI agents, the agents that can take actions on your behalf, like mouse clicks and filling out forms.
[00:15:29] Paul Roetzer: It is the inevitable next phase of this. Everyone's been working on this for years. I think it's really important to give a caveat here that, to, to Matt's credit and Jason and the, you know, the team there, I. They, they're very clear, at least on the site, that this is an alpha, it's not even a beta.
[00:15:49] Paul Roetzer: And they, I like that they, it's Alpha V 0.01, like it's not even V one. So they're very, very clear that this is extremely early. In some ways I kind of liken it to quote unquote full self-driving Teslas. So I've had a Tesla model Ss, I'm on my second one. I've had five years. Anyone's read the book.
[00:16:14] Paul Roetzer: You know, Iwrote about my experience with Tesla's, you know, full self-driving and autopilot and things like that. So even like the early days of what's now become full sail driving was autopilot. It, it's an insanely misleading name for what the driver did in a Tesla. There was nothing about it that was autopilot.
[00:16:36] Paul Roetzer: Copilot would've been a great name for what Tesla was building. Because you could use it in some circumstances to drive you. But you always had to have your hands on the wheel and you couldn't take them off and you couldn't not pay attention. And it was really glorified cruise control for a really long time.
[00:17:00] Paul Roetzer: Then earlier this year they came out with full self-driving. I don't know what version is, like 11 or something like that. And now it could actually do like city streets and stop signs and turns and all these things and lights, and it is quite honestly a terrifying experience. So Ihave it. I paid, you can do with theirs.
[00:17:19] Paul Roetzer: It's a, I'll get back to the personal assistant in a second as to why I'm. Pay two, like $200 a month to have this full self-driving capability rather than paying like the 15,000 upfront for it. So I was taking a trip to Milwaukee in April and I said, I'm going to pay the 200 bucks. Let me, let me see how it performs.
[00:17:36] Paul Roetzer: And it has definitely gotten better. Mind you, this is years after we had the first autopilots. But I don't trust it to change lanes. It makes really crazy decisions in some, situations that, you know, it probably hasn't seen much training data. So even though Tesla's claim to have full self-driving, my experience is they are not.
[00:18:00] Paul Roetzer: And you need to know that going in or else you will get killed or kill someone like it, it can go, it can lose control pretty quickly in some environments. Now some people claim it's like 99% of the time there's no problems with it, and they may have a more advanced version than I've seen. But you really have to keep your hands on the wheel and you have to pay attention at all times.
[00:18:23] Paul Roetzer: So I would say where we're at with personal assistance, including hyper rights, is the very early days of what autopilot was like back in 2018, where it's really cool. Sometimes it does something. You're like, how is it even doing that? Like that's pretty incredible. I've never seen a car, or in this case a personal assistant, buy me a pizza online.
[00:18:44] Paul Roetzer: That's amazing. Just know that we're extremely early and it is not a truly automated personal assistant. And then to use the metaphor of the car, you have to keep your hands on the wheel. Like you have to be aware of what it's doing, what it has access to, where it's going to go next, how it's going to perform these tasks.
[00:19:04] Paul Roetzer: And so I think for people who want to be on the frontier, like truly out in the frontier, this is the kind of thing you probably want to experiment with because it gives you a sense of where this technology is going to go. I just wouldn't use it and say, Hey, I'm going to go with the full self-driving personal assistant.
[00:19:20] Paul Roetzer: It's going to start booking my travel for me and completing my forms and sending my emails. We're not there yet. We're early days of this. Now, I will also say the caveat to that is early days might mean six months from now we have full self-driving personal assistance. Like, I don't know when we're going to get that digitally.
[00:19:40] Paul Roetzer: It's taken a long time for cars, but in the terms of these personal assistance, it could very well start happening rather rapidly and that becomes very disruptive to a lot of the way we do work, a lot of the way we find information. It becomes disruptive to the companies building their sites, knowing that agents can now come to those sites and take actions on humans' behalf like.
[00:20:04] Paul Roetzer: It's a, it's a whole new frontier that hasn't been explored much. I haven't seen great research papers yet on what does this look like when these things work. And so I think that's not only do you have to be cautious of the tools don't really work the way you might hope they work, yet, but we also have to think about, there's a lot of lot more concerns that start emerging when agents can take actions, especially when you can string agents together.
[00:20:31] Paul Roetzer: and so we need a lot more research on safety and alignment of these things. we haven't seen much of that yet. So what
[00:20:41] Mike Kaput: jumped out to me looking into Agent one and some of these other autonomous AI agents is a piece that you wrote a while ago called The Law of Uneven AI Distribution. And you basically were making this argument that you're going to get benefits from AI proportional to the amount of, control your comfortable seating to some of these systems or data that you're willing to share.
[00:21:04] Mike Kaput: So in the case of an autonomous AI agent, I would immediately kind of jump to, well, wow, do I want to give this control to my of my browser? Do you want to maybe talk us through a little bit of that? 'cause it seems like a really stark example Yeah. Of what you had theorized.
[00:21:18] Paul Roetzer: So, Iknow Matt personally, and we've known them for years, and Itrust them.
[00:21:25] Paul Roetzer: So my, honestly, like if this was anyone I didn't know or a company I wasn't familiar with, I wouldn't even consider. Getting access to this tool, like because you have one, you have to drill into the terms of use. What are they getting access to? What data are they keeping? How are they using that data? So there's all these immediate questions that come to my mind.
[00:21:44] Paul Roetzer: Am I breaking the terms of use on sites that it's going to go to? Like is even allowed to go to LinkedIn and curate, you know, lead for me? 'cause it can do that. Am I allowed to be doing that? Like, is that, does it violate LinkedIn's terms? I don't know the answers to these questions. And so my first thing would be, do I trust the company building it?
[00:22:05] Paul Roetzer: And then am I crossing any ethical lines or legal lines using it? And those are, those are messy questions and they just aren't good answers yet. It's kinda like the generative ai and is it, you know, breaking copyright law, we don't know yet. And so that's the first way I would look at this.
[00:22:22] Paul Roetzer: So it definitely fits under that law of uneven AI distribution of that. Like, I'm not sure I'm willing to give up what would be needed to use this tool because there's lots of unknowns about it. But I would definitely exercise caution because you're going to see a flood of this kind of technology, the rest of 2023 and into early 2024.
[00:22:44] Paul Roetzer: And you have to develop your own systems for evaluating the companies and the people because you have to trust them. We talked about this like rewind ai. I think it was a, a podcast episode or two ago where I was like, seems like a really good dude. A lot of people I know know the guy invested in the company, trust the guy.
[00:23:03] Paul Roetzer: I'm just not doing it. Like it, it just crosses a line for me personally, of access. I'm not comfortable giving up to, to get the benefit of that technology. And for me right now, personal assistance would fall into that realm outside of, you know, the obvious, like voice assistance and things. But these kinds of assistance, I'm not there yet where I'm comfortable giving up something to a company, even if it's one I trust that could have access to literally every click everything you're doing.
[00:23:32] Paul Roetzer: And I know like Google and Apple and, you know, they probably have that kind of data, but it's just, it's a different ball game for me when it's a startup company or a, you know, an early company. So yeah, I think these are the kinds of messy things we're going to have to work through individually and as companies that this tech is going to be available to us.
[00:23:51] Paul Roetzer: Are we allowed to use it? Are we willing to give up what's required to use it if we are allowed? Yeah, I, it it's a good example of the kinds of things we're going to have to be really be thinking about in the months and years ahead.
[00:24:05] Mike Kaput: The next up, a leading AI software tool called writer writer.com. A friend of marketing AI Institute we've talked about several times.
[00:24:13] Mike Kaput: They actually just released a large language model designed specifically for use in healthcare, and that model is called Paul Myra, med and Writer says it has outperformed both GPT-4 and medically trained human test takers on PubMed qa, which is a leading benchmark for biomedical question answering.
[00:24:33] Mike Kaput: So unlike a generic model like GPT-4, which knows a little bit about a lot of things, and as such can be used for this wide variety of different tasks, Palm Myra Med is specifically trained on publicly available sets of medical data. Generic LLMs are impressive and have a lot of different use cases, but the issue is there of pretty limited utility in domains like healthcare, which requires extreme accuracy and domain specific knowledge.
[00:25:02] Mike Kaput: Now, Palmyra Med aims to solve this need, and by doing so, writer clearly hopes to make generative AI much more accessible to healthcare organizations. You know, we've kind of seen firsthand. Adoption in healthcare organizations of LLMs has been pretty limited to date, given how much they have a need for specific medical accuracy from these tools.
[00:25:24] Mike Kaput: And they have all these really robust security and compliance concerns that many AI systems and products out there are unable to address. Now, Paul, the reason I mention this is because even if our listeners aren't in healthcare, this seems significant. We're starting to see these leading AI companies start tailoring really powerful generic LLMs to specific areas of domain expertise.
[00:25:48] Mike Kaput: Do you see that as a trend that's kind of starting to take off here?
[00:25:52] Paul Roetzer: Definitely. One of the hypotheses we've talked about is this verticalization of LLMs makes a ton of sense that you could have legal GPT healthcare, GPT, Manufacturing energy like that you would eventually start building into the verticals and then, it, it's just faster to value for the companies in those spaces.
[00:26:15] Paul Roetzer: So I think it makes a ton of sense and we're probably going to see a lot more of this approach. Plus, I think it's these, these organizations, especially in healthcare, they have so many other considerations beyond how the large language model works and what it's capable of. They have to deal with regulations, privacy, compliance, all of these things.
[00:26:35] Paul Roetzer: So even if you go to like the page from writer, it says, you know, we're deeply focused on secure enterprise grade AI platform. You don't store share or use your data in the model training, and we're compliant with SOC two, type two, G D P R, hipaa and P c I, like you're talking their language. So like we've talked with healthcare companies, we know that in the back of their mind, even if they're in marketing, they're thinking about the regulations, compliance issues that they're going to have to move any of this forward.
[00:27:01] Paul Roetzer: So the fact that you can talk to a company that obviously understands the complexity of your industry and appears to have built something that's tailored to you, I could see this being a really valuable messaging play in their marketing and writer's marketing to be able to talk at this level to people in the healthcare space.
[00:27:22] Paul Roetzer: Whereas, you know, ChatGPT, they're not going to touch it. And like in terms of scale, because they don't know what's going to happen with the data and you know, they're going to have all kinds of issues with security and compliance. So, yeah, Imean, I think this is smart and I could definitely see a lot more of this.
[00:27:37] Paul Roetzer: As these models try and find niches within different verticals. So yeah, for, for writer it makes a lot of sense. So we
[00:27:46] Mike Kaput: know that a lot of these enterprise businesses with, you know, concerns around accuracy, privacy, data security, they're getting tasked to essentially figure out AI by the board. C e o leadership team is a good first step.
[00:28:01] Mike Kaput: Looking for kind of a domain specific l l m, like this one.
[00:28:05] Paul Roetzer: It could be, I just don't know how many are out there yet. I think for more organizations it's going to be fine tuning models, whether they're open sourced or proprietary models that they can fine tune on their data and keep everything secure and accurate.
[00:28:20] Paul Roetzer: It seems like there's more options to go that direction than there are to go vertical. So if you're, you know, a law firm or whatever, Idon't know that anybody's really built those verticalized models yet. They're going to probably come, but right now it seems like so much energy and resources and honestly, compute power is going to building these horizontal models that can just be used for general tasks.
[00:28:41] Paul Roetzer: So I think, you know, to train these kind of vertical solutions is, again, a, a logical play. I just, I don't think it's very widespread right now.
[00:28:50] Mike Kaput: All right, let's jump into a number of rapid fire topics. So first up, we actually just had McKinsey publish the findings of their annual McKinsey Global Survey on the state of ai.
[00:29:01] Mike Kaput: So they do this, they've done this many years running now. And this year's findings seem to confirm the explosive growth of generative AI tools. So McKinsey, according to their research, found that one third of respondents to this survey, of which there were nearly 1700, say they're using generative AI regularly and at least one business function.
[00:29:22] Mike Kaput: And that's less than a year after many of these tools debuted. More than a quarter of respondents from companies using ai. Also said that generative AI is already on their board's agendas. Another 40% of respondents say their organizations plan to increase their investment in ai, thanks to all these advancements in generative ai.
[00:29:43] Mike Kaput: Now, McKinsey actually also looked at the anticipated effect of generative AI adoption on the number of employees by business function that the respondents expect to have over the next three years. They actually looked at marketing and sales specifically in addition to a number of other areas, and in marketing and sales specifically, 39% of respondents anticipated a decrease in the number of employees.
[00:30:12] Mike Kaput: 33% said there'd be little or no change, and 17% expected an increase. Now, again, this is out of, in some cases, nearly 1700 participants, 900 of whom said their organizations had adopted AI already in at least one function. Now, there's a ton of findings in this research. We can't call out everything here, but I think the ones I, that jumped out to me that I just read, I think they tell an interesting story that we've been talking about more and more, which is we're seeing this lightning fast adoption of generative ai, and at the same time, we're seeing some more signals here that the impact of the tech on jobs could be pretty uncertain.
[00:30:54] Mike Kaput: Now, from my read of this, McKinsey didn't necessarily paint this doom and gloom picture, but it did seem pretty significant to me that 39% of respondents in marketing and sales say they anticipate a decrease in headcount. How did you view these findings?
[00:31:13] Paul Roetzer: I think that I. Whenever I read these studies, the first thing I honestly do is go through and look at how it was done and when it was done.
[00:31:22] Paul Roetzer: Like who were the people that they asked and when did they ask them? And so in this report it says it was an online survey fielded from April 11th to 21st, 2023. So the first thing that jumped out at me, some of these numbers seem really low, in terms of impact, in terms of, you know, how many had it on their board agendas based on the conversations we're having.
[00:31:44] Paul Roetzer: Some of that just seemed. Like they could have done this survey last year. And then I was, so in the context though, April, again, we've gone through this timeline so many times, but Chad, GPT comes out end of November. Then you have the holiday months, then we get back to work, you know, in January, mid-January, late January, all of a sudden everybody's trying to figure out generative ai.
[00:32:04] Paul Roetzer: GPT-4 hits in March. Google announces, you know, that Palm or whatever's coming to Google Workspace. So a lot happened in like March, middle of March was when everything exploded. So this survey was like 30 days after that. So all I'm saying is good data. You have to look at all of these reports and not assume that anything is, this is exactly how it is.
[00:32:29] Paul Roetzer: That this is definitive. In terms of representation of what's going on, you have to consider who it was and the timing in which it was conducted. We're going to be releasing our state of. Marketing our report findings in a couple weeks here. And I'll give the same caveat to that. You have to consider who's taking the survey.
[00:32:46] Paul Roetzer: When did we field the survey? So I think it's really interesting data. When it comes to impact on jobs. This is the one I I struggle with the most. Ifeel like I've read every research report that's come out on the AI impact of jobs. And going back to in my second book in 2014, you helped me with this research.
[00:33:07] Paul Roetzer: We did a section on ai, so the book was about tech, talent and strategy, marketing, tech, talent strategy. And we had a section on ai and I cited a University of Oxford study on the automation of jobs. And they looked at, if I remember correctly, it was like 315 different jobs and which ones had the highest likelihood of being automated by.
[00:33:29] Paul Roetzer: Ai. This was 2014, it was 2013 data. So we have been looking at this. What is AI going to automate for over a decade? Like, I've been researching this topic, I, I've yet to come across anyone who seems to be able to confidently tell me what is going to happen. Like, no, no. Literally no one seems to actually comprehend it.
[00:33:51] Paul Roetzer: And what I've, the flaws I often find, and again, like I don't want to like just ignore the fact we're talking about the McKinsey study. Great study. Go read it. It's good context. But what I'm seeing consistently happen is they talk about the automation of jobs as though it's zero to a hundred. Like it, it either is or it isn't.
[00:34:09] Paul Roetzer: It's a zero or one basically pro go programming language. So it either is or isn't this job exists or it doesn't exist in the future 'cause AI's going to do it. I don't see it that way. I, there's very few knowledge work, cognitive skilled jobs that I think just disappear. What I do think happens is fewer humans are needed to do them.
[00:34:30] Paul Roetzer: So writing as a profession isn't going away. Humans will still write for a living. We'll have journalists, we'll have reporters, we'll have people, you know, content marketers, email marketers, whatever we're going to write for a living. My question is, do we just need fewer of them? Because the AI is going to make things so efficient and if you can't increase the output.
[00:34:50] Paul Roetzer: So I'm in a company where I can't just arbitrarily say, oh great, well we had 50 writers. We're were doing a hundred articles a month, we're just going to do 200 articles a month now. Well, what if the demand isn't there for the 200 articles a month? What if a hundred is peak demand, that that's all you need?
[00:35:06] Paul Roetzer: Okay. Well then I don't need 50 writers anymore. Maybe I need 20. That's my concern, is that we're looking at this zero or one, the job exists or it doesn't. So when we talk about automation, it's like fully goes away or it doesn't. And then the other thing I've struggled to see yet is projecting out.
[00:35:24] Paul Roetzer: Based on scaling laws of where this technology is going. So these AI researchers have proven to be extremely good at projecting the capabilities of these models, throw more data at it, more computing power, more parameters, and you can generally predict with relative accuracy what they're going to be able to do.
[00:35:45] Paul Roetzer: What I want to see is let's run those models and let's say what they're capable of 12 months from now. Then look at a job level and say, what does that mean to writers, accountants, lawyers, not do the job exist or not, the job will still exist. The question becomes how many people need to do that job? And that's where it's just like I'm spending.
[00:36:07] Paul Roetzer: People ask me like, what keeps you up at night? That one keeps me up at night, like trying to find a way to solve this and think about what are the jobs of the future actually looking like. So these studies to me are really helpful. Like, I like reading them and it gives me context, but I don't look at them as authoritative in terms of this is actually what's going to happen.
[00:36:24] Paul Roetzer: They're asking 1700 people who don't have a clue what's going to happen. Like no. and maybe they're a little more informed than others 'cause they're doing some of this, but they don't know. And so they're just reporting that, that these people may have that here's what they think's going to happen. So, I don't know.
[00:36:40] Paul Roetzer: I mean, I, I'm sorry, it's like a soapbox moment. I wasn't really planning on like getting into this right now, but obviously a topic we're going to be talking a lot more about, I've been thinking a lot about this postma con, like I tried to take a little time away last week and Ibasically spent every waking minute thinking about this topic because I think it's the most important thing we can work toward is what does the future of work look like?
[00:37:02] Paul Roetzer: What are the jobs of the future look like? What are. The skills and traits that remain uniquely human. Ihaven't seen a good answer to any of that. And so I'm, I'm kind of like, I'm thinking about this a lot, as you can tell. And so I think that, it's probably a topic we're going to come back to. I've been working on some things, writing a lot of ideas down, but nothing I can formalize it.
[00:37:26] Paul Roetzer: It was going to be my opening keynote for Mako this year and I just couldn't get there. Yeah. Like mentally, Icouldn't arrive at a place to say this is what I think is going to happen, but I'm hoping some point in the near future there'll be some breakthroughs and I can maybe get to that point where I can give some, some confident assessment of what I think's going to happen to hopefully help people try and get there, I guess.
[00:37:49] Paul Roetzer: Excellent. And
[00:37:49] Mike Kaput: I'm sure we will have a excited audience for that because I made that
[00:37:54] Paul Roetzer: a topic outta nowhere. Yeah.
[00:37:59] Mike Kaput: No, I think it's on everyone's mind. I mean, it's very, very tangible regardless of what level you're at. Thinking about how this is going to affect your career. Yeah. So we also have some big Google updates, in terms of some of their AI related products.
[00:38:13] Mike Kaput: So Google's search generative experience, s g e, which is its AI powered search results, which are available now in various forms of Google search. They just got some major updates and these updates include, you can now see photos and videos in some S G E results. There have been very significant speed updates to SS G E results.
[00:38:33] Mike Kaput: Google says the time to generate these now takes 50% less time now. So from when you type in a query, to getting kind of a full paragraph of chat and search results in a conversational way is dramatically faster. And SG now shows the publish date of the links that it cites in its results. So you can now tell how recent the information is that it is providing you.
[00:38:58] Mike Kaput: Google interestingly, also took the time to reiterate that search ads will continue to appear in s g results. I found it pretty interesting to kind of see this continued evolution of S G E, and I know I, for one, have started using it quite a bit more. Also I see and click on search ads a lot less, even though they're present in sg.
[00:39:20] Mike Kaput: Do you see this like starting to potentially get into this territory of having a major impact on Google's business model and how we search?
[00:39:29] Paul Roetzer: It certainly could. Interestingly, like Idon't, Iprobably knew I had access to this, but we don't have access in our institute account, so you can't, I can't access this on my institute email and domain, but I was able to go into Gmail and I do have access in my personal Google account.
[00:39:45] Paul Roetzer: And so I started playing around with this morning, it was the first time I actually played around with it, when I saw your notes about it. So, I did, I ran a comparison search to something. I had actually been using Chad GPT to help me plan with. And, I'd pushed chat bti quite a bit in terms of, it was a business strategy thing I was working on, and so I gave the same thing to, to Google.
[00:40:05] Paul Roetzer: I don't know. I mean, it's interesting. I can see where they're trying to go with it. I could see how you could appease the publishers with it because it definitely infuses the citations right into it. I could see how you might still scroll down 'cause now you can ask follow on questions and like the prompting we were talking about with ChatGPT, it does recommend follow on questions just like they've done for years in Google with like, these are other related questions people would ask about this thing.
[00:40:31] Paul Roetzer: So yeah, I mean, it, it's an experience I could see getting used to pretty quickly and forgetting what the old experience was like. I mean, it feels pretty natural. Yeah, I don't know how good it is. Like my one single five minute use case chat, G B T was way better. But it didn't give citations.
[00:40:49] Paul Roetzer: But chat, G B T gave me the answers I was looking for, assuming they were correct. Better than this. Like, I don't, I don't want to click through all these links, but yeah, so I mean, if you haven't tried it, I, it's just labs.google.com/slash search was the link in the Google Post. We'll put up in the show notes and once you get there, if you haven't already asked to sign up on the wait list, you can, so I'm guessing at some point, Isigned up for the wait list.
[00:41:15] Paul Roetzer: I don't know if there's even still a wait list 'cause that article was from May. So just labs.google.com/search. If you haven't played around with this yet, you could probably go check it out. Again, you might have issues with your corporate account, so use your, your Gmail if you don't. So
[00:41:30] Mike Kaput: kind of at the same time, we have some reports according to Axios, that Google plans to overhaul its assistant to focus on using generative AI technologies similar to those that Power ChatGPT and its own Bard chat bot.
[00:41:45] Mike Kaput: So Google Assistant being this, voice assistant Google has had for quite a while, for consumers, developers, and Google's own employees, that this update is going to kind of change how Assistant works. Google says that some work has already begun on this revamped Google Assistant, starting with the mobile version of its product.
[00:42:06] Mike Kaput: And as a result of this, Google is reorganizing the teams that work on assistant a bit. Details are pretty thin at the moment, but Paul, this seems like it's a long time coming. Like we've talked many, many times about how it would be a game changer if Google Assistant or Apple Siri became actually highly sophisticated and useful thanks to all the developments we've had in ai, which as of today, they don't really seem to be.
[00:42:31] Mike Kaput: What are you, your thoughts on
[00:42:32] Paul Roetzer: Google's move here? Yeah. Again, it seems so obvious that this is the direction you would go and I think, you know the talk we had about personal assistance earlier in the show. It's, it, to me, the ultimate user experience is if Siri works and if it can be a true personal assistant.
[00:42:49] Paul Roetzer: So not only can it give me answers, but it can go take actions on my behalf like this. I ironically was listening to the big technology podcast this morning, and Matt Wood, who's a VP of product at Amazon Web Services, was talking about how they're doing the same thing with Alexa. That they're, you know, infusing these language model capabilities into Alexa to try and improve Alexa's capabilities.
[00:43:11] Paul Roetzer: So I think you're just going to, you're going to see, you know, Siri, Google Assistant and Alexa racing to, to win at this in the next six to 12 months. Like, I think this is a really key thing to watch for. And what I would do is watch the major, developer conferences, from Google and Amazon and Apple and, I would think that's where you would hear major announcements about these capabilities is, and I will have to check the calendar and see when those are coming up.
[00:43:36] Paul Roetzer: I think Google actually has their next conferences. August 29th to the 31st in San Francisco. I was like, I was thinking about going out there because Igotta guess they're going to announce a bunch of stuff at the end of August.
[00:43:50] Mike Kaput: Interestingly, YouTube is also experimenting with some new AI features. According to the Verge, they're experimenting with the use of AI to auto generate video summaries.
[00:44:00] Mike Kaput: So these summaries are only starting to appear next to a limited number of English language videos, and they will only be viewable by a limited number of users. These AI generated summaries will appear on YouTube's watch and search pages, and they're basically intended to give a brief overview of a videos contents without kind of replacing the existing description written by a human.
[00:44:22] Mike Kaput: This will just be another additional summary of what's in the video. Now, Paul, this seems like a pretty useful, straightforward feature that we might soon see in YouTube. I mean, can you kind of talk a bit about how valuable summarization is as an AI use case, especially
[00:44:38] Paul Roetzer: for video? I mean, you and I have like dreamt of this use case for three years.
[00:44:43] Paul Roetzer: Like this is one we talk about all the time. Summarization is so valuable in so many ways from summarizing long form PDFs to transcripts, to, you know, videos, all of these things. And I think it's interesting because in the spring, I remember you and I were like testing different tools who were claiming to be able to do summarization of videos.
[00:45:02] Paul Roetzer: Right? And really they were just pulling the transcripts of the a p I and summarizing 'em. The summarization sucked. Well then we got GPT-4 and Claude two and some of these other innovations on language model side where they got really good at summarization. But, you know, so I think this is, it's going to be a widely used capability.
[00:45:20] Paul Roetzer: It's going to be one of those things where you're going to realize like, how did, how did we not remember YouTube even before that had summarization. I could see it just being such a natural value add. It's also, to me a great example of, I. How hard it is to build moats. So like, let's say you were a startup or had an idea for a startup to summarize, videos on YouTube.
[00:45:41] Paul Roetzer: Like that was your idea for a product and you didn't understand the context of how motivated these big companies would be to build these in as features you might've raised ahead and built, you know, raise some money and, you know, trying to build a, a video summarization tool. Well, it game over like as soon as YouTube introduces that feature.
[00:46:00] Paul Roetzer: And that's, it's always that debate as like a startup entrepreneur is like, am I just building a feature? Of a larger platform. And if you are, there's a chance you're just going to get run over at some point. And so I think that's the kind of thing you see happen here. And so Iwould assume that, you know, with Descrip and YouTube and some of the Vimeo, like some of their big, you know, video platforms, they have to have native summarization capabilities.
[00:46:24] Paul Roetzer: Big. Yep. Like their teams, their product teams are probably racing to build this kind of thing right now, right into their platforms. So, yeah, I think this is, this will be a hugely valuable feature for a lot of people, especially when, you know, it's like, you know, the human level accuracy of the summarization.
[00:46:41] Paul Roetzer: We're not there yet in a lot of cases. My experience has been, some of, some of these times the summaries aren't just, just that. Good. And like you and I have talked about the, depending on your use of the summary, you're probably still going to watch the video yourself. Right. Like, I think I was saying when, when I am, when you and I are synthesizing information, like the McKinsey report, I could just take that, that article and put it into Claude two and Prop and say, summarize this for me.
[00:47:08] Paul Roetzer: Or we could do the same thing with Bing, or you could do the same thing with GPT-4 and it'll summarize it. Your understanding of the article is very different when you read it yourself. Yeah. And so it's like, you know, when high in high school, I still remember trying to like pass tests, reading the cliff notes instead of reading the actual book.
[00:47:24] Paul Roetzer: And then you get questions asked and you're like, well, I wasn't in the cliff notes. And so like you have a deeper understanding, but I think there's a lot of use cases for summarization where I don't need the deep understanding of this topic. I just need the surface level of what was talked about.
[00:47:38] Paul Roetzer: Now I'm going to move on because I'm not getting quizzed on it, or I'm not having to talk about a podcast. I just want to consume more knowledge. And so I could see absolutely like using this kind of thing where humans expand our capability to consume information, but the stuff we really have to go deep on, you're still going to have to put in the work and learn it.
[00:47:57] Paul Roetzer: That's going to be a hard thing to teach the next generation though. Because they're all going to get raised with summarization at their fingertips. And like we have to ingrain in them. You have to still go deep on stuff when you really want to understand the topic. Absolutely.
[00:48:12] Mike Kaput: So we also have a few important updates from Meta.
[00:48:15] Mike Kaput: There are some reports right now that Meta could actually be launching AI powered personas in its services, including on Facebook and Instagram as soon as next month. And these personas basically provide features like search or recommendations and basically do this with different personalities like you're engaging with an AI bot or agent.
[00:48:35] Mike Kaput: Some of the personalities cited, seem pretty diverse. There's one that kind of offers travel recommendations in the style of a surfer. And another that speak click Abraham Lincoln. So, These built-in chatbots essentially could be a way to boost engagement with services like Facebook and Instagram, especially as they face competition from sites like TikTok.
[00:48:59] Mike Kaput: They also could show off more meta's AI capabilities. We've talked about a bunch of, you know, AI research updates that are really impressive from meta. These could be showcasing that in actual commercial use cases. Now, Paul, this is not new for social media platforms to be incorporating AI agents into their products.
[00:49:18] Mike Kaput: I mean, Snapchat notably rolled out something like this a little earlier this year. I mean, is it possible we're going to see a sizable portion of our interactions on social media start happening via AI agents or
[00:49:31] Paul Roetzer: bots? I assume so I honestly, I don't get the persona thing. Like if I want travel recommendations, I don't really want to take the time to find like, the surfer version of it.
[00:49:42] Paul Roetzer: Right. Or like the Abraham Lincoln version. It's like I did all my travel recommendations, like gimme my two choices of, you know, the, the voice and I'll pick one. But II don't know, I'm not sure what their play is around, around that. But the idea of these trained bots that have the capabilities, like again, it's almost like the vertical knowledge idea.
[00:50:02] Paul Roetzer: Yeah. Where they're trained on these specific things. Iabsolutely see that. And I think, again, what happens in the future is you may have a single personal assistant that becomes your go-to, and maybe it's like a SIR or Alexa or Google assistant, or maybe it's hyper writer, whatever you're going to have like a go-to.
[00:50:17] Paul Roetzer: And then that agent will go interact with all these other agents. So it'll go find specialized knowledge and specialized assistance and it'll interact and. You may only have to have the single interface versus you having 15 personal assistants, one for travel, one for this, one for that. And I think that's what a lot of the VC firms are trying to figure out right now is, do we bet on people are going to have one or two personal agents and those agents will go communicate with all the other AI agents, or are there going to be just hundreds of vertical agents built?
[00:50:47] Paul Roetzer: And as consumers we're just going to interact with individual agents for everything from, you know, travel and leisure to entertainment, sports, whatever. Yeah. And that's, again, nobody has a definitive answer to that. So I would just pay, you know, pay attention to that space because it's going to affect businesses.
[00:51:02] Paul Roetzer: And whether you build your own personal agents for your company or you know, people are just going to interact through their agents with you.
[00:51:10] Mike Kaput: So meta also just announced details of something called Audio Craft, which is a new family of generative AI models that they built for generating high quality, realistic audio and music from text.
[00:51:23] Mike Kaput: So this is a single code base that works for music, sound, and more audio craft is currently being released publicly, for research purposes with meta stating, quote, responsible innovation can't happen in isolation. Opening up our research and resulting models helps ensure that everyone has equal access.
[00:51:42] Mike Kaput: Paul, could you maybe put this into context for us, because it seems like there's. Two really important things going on here. First, you know, kind of next phase of generative AI is upon us as we get really powerful audio generative AI tools that can synthesize voices and music. And at the same time, we talked about this a couple podcast episodes ago, meta is continuing this commitment to releasing really powerful AI to the public with few strings attached.
[00:52:10] Paul Roetzer: Yeah. So when we talk about generative ai, again, we'll go back to kind of summarize this. So what we're often talking about categorically is text, which, you know, language models is the core there. Image, generation image recognition, so images, video, audio, and code. Those are kind of the five we generally speak about when we're talking about ative ai.
[00:52:31] Paul Roetzer: Now, there are other categories you could throw, synthetic data, virtual beings, like, there's all these other things you could create, but for us, many of the use cases that marketing and business leaders are going to care about are going to fit within those five categories. 2022 was certainly the year of image.
[00:52:50] Paul Roetzer: Like I think we had dolly two, we had mid journey, we had stable diffusion, we had kind of this explosion of capabilities and access to image generation tools. And then obviously the end of last year was the explosion of the text, the language models and, and language generation capabilities that carried into this year.
[00:53:09] Paul Roetzer: Images continue to accelerate. So your kind of exponential growth curve in terms of capabilities of image and text. Then we had this year with gen one and gen two from runway. So you started seeing video starting seeing announcements from meta around video. So video started kind of, it's it's climb and now audio absolutely is going to see its second half of 2023 into 2024.
[00:53:32] Paul Roetzer: And all along code's been the underlying one. Just as marketers and business people, we don't pay attention to the code one, but anyone doing development work has been paying attention to the code capabilities for a while. So I think audio, we're going to see a lot of innovation in the next six months and certainly Meta's willingness to put this stuff out for better or for worse, changes things very quickly.
[00:53:54] Paul Roetzer: When you put this kind of power in people's hands, and like you said, very few strings attached, that not only do you have the ability to use it, but you can commercialize it, it affects a lot of stuff. And, you know, audio's a space. I haven't really sat and thought about a lot, you know, we've thought mostly about text.
[00:54:11] Paul Roetzer: We've thought you know, quite a bit about image. But video and audio are the ones that we just haven't explored a ton 'cause the use cases haven't been as prevalent yet, but I think they're definitely going to be, and it's going to be a space to pay attention to. Or just songs on demand, you know, music on demand, whatever you want to create, just to be able to create it is a pretty wild concept.
[00:54:33] Paul Roetzer: I think that's how we did our opening video this year, I think. Yeah, Gracie, I think so. Used AI to create the music for it. So, yeah. Again, stuff two years ago, the idea that you could just have stock photos at your fingertips. Like, think about that for anything you want to generate, audio, video, co, anything, just like prompt it and you, you got it.
[00:54:55] Paul Roetzer: It's a wild thing that we're almost taking for granted exists. When two years ago it would've been insane to think this stuff would exist.
[00:55:05] Mike Kaput: So the partnership on AI, or p AI for short is an influential nonprofit that's supported by some big AI players like Google Meta, Amazon, Microsoft, apple, and others.
[00:55:16] Mike Kaput: And it's starting to seek public comment on a document. It's calling, its AI Procurement and use guidebook for newsrooms. So p a I says this guidebook provides a roadmap for newsrooms navigating the difficult questions accompanying AI procurement. Like should your newsroom use AI tools for what, which tools should you choose?
[00:55:37] Mike Kaput: How can you adopt AI tools that support business but are also responsible and ethical? Now, they're actually going to take the feedback people give them, and incorporate it into the final guidebook, which is going to be published according to them later in 2023. Right now it's in a Google Doc format that we're linking to, an article about it in the show notes, and you can definitely, access.
[00:56:02] Mike Kaput: The document through that. Now, Paul, this seems kind of like the latest in what's looking to be where we're headed, which is newsrooms are increasingly having to deal with the impact and effects of ai. Do we fully anticipate kind of fully assisted AI assisted newsrooms like anytime soon? It seems like they're adopting more of the technology.
[00:56:24] Paul Roetzer: Yeah. I mean they have to. It's going to be infused, but Ithink the interesting thing here is just replace newsroom with marketing department. Go back to those questions you just said. Roadmap for marketing departments navigating the difficult questions for AI procurement. Should your department use AI tools for what?
[00:56:42] Paul Roetzer: Which AI tools should you choose? How can you adopt AI tools that support business but are responsible and ethical? It's the exact same problems. So I would say for marketers and business leaders follow what's happening here. 'cause partnership on AI is a major player supported by all the big tech companies.
[00:56:57] Paul Roetzer: And so whatever they're doing for newsrooms, you are probably going to be able to, to use to inform the generative AI policies for your marketing department and your organization. And so that's why I think when we put this one in our sandbox, like I wanted to, I'm sure we have journalists that are listening and maybe some, you know, publishers and things like that.
[00:57:16] Paul Roetzer: But more so for me, we tout all the time. You have to have generative I policies that guide the use. You have to provide points of view on this stuff. You have to te tell your employees what they're allowed to do, not allowed to do, which tools are permitted. So I would follow this as almost like a template in some ways, or a starting point for what your Generat AI policies could look like or how to evolve your generative AI policies moving forward.
[00:57:39] Paul Roetzer: Yeah, and
[00:57:39] Mike Kaput: if you look in the Google Doc, I mean, just the way they've got it organized right now, you know, five principles of AI adoption for newsrooms. These are, from what I can see very clearly, you could sub in marketing for news and get some value out of these. There is a extensive step-by-step guide to AI adoption.
[00:57:59] Mike Kaput: It looks like some of these steps are directly, would be directly relevant to marketers as well. So that's a really, really useful template and good advice.
[00:58:06] Paul Roetzer: Yeah, and like you said, we'll put it in the show notes 'cause it is, it's just a Google doc and it's got a ton of really helpful stuff to look at.
[00:58:13] Mike Kaput: All right, so last but not least, this is breaking news as of Monday, I believe. Zoom just released a blog post that has it explaining changes to its terms of service because. There were some terms of service changes in March that people have just discovered could be somewhat problematic. And so in March, zoom made changes to its terms of service that appeared to give the company very sweeping control over and ownership of user data in order to use that data for AI training.
[00:58:47] Mike Kaput: Now there are a couple sections that were updated that people took major issue with. I believe it first started going a bit viral on Hacker News. People realized that Zoom had changed these terms to be very, very sweeping. Now, one of the, part of the terms in particular that had users kind of crying Foul said that customers quote, agree to Grant and hereby Grant Zoom a perpetual worldwide non-exclusive royalty free sub-license able and transferable license to use their data for things like product service.
[00:59:20] Mike Kaput: Product and service development, machine learning and artificial intelligence, that is according to some reporting from the Verge. Now this blog post released on Monday comes from Zoom and Zoom Basically in it is clarifying, saying we do not train our models without customer consent. And that quote, our customers continue to own and control their content.
[00:59:41] Mike Kaput: Paul, what did you make of the initial changes to the terms of service and then zoom's rebuttal to the controversy
[00:59:49] Paul Roetzer: here? Like we said earlier, I mean the terms of, of service, terms of use for these tools, it's something most marketers are just so used to just clicking through. I mean, even our consumer life, it's like, yeah, whatever.
[00:59:59] Paul Roetzer: I accept it. And they can make little changes that have a pretty big impact. So in their case, I thought it was interesting, their blog post that I love when it says in section 10.4, our intention was to make sure we provided, our intention just always sounds like, yeah, we kind of screwed up here, but here's what we actually wanted to say.
[01:00:23] Paul Roetzer: It seems like what they're saying makes sense. the, the use case. I think the fear was, oh, so when we record our calls, you're recording everything. So we're having confidential conversations that that data belongs to you and you can do whatever you want with it. Use it to train models. That's a viable concern.
[01:00:39] Paul Roetzer: It doesn't seem like that's what is happening. It appears as though they, they're, protecting your data and that's all yours. So I would check it out. I think it goes back to the law of uneven AI distribution. And knowing who you're working with and whether you trust those companies, because it's going to get very complicated as to what data you're providing to people and how they're using it.
[01:01:03] Paul Roetzer: And I will tell you from. Having talked with some larger software companies, in the last six months, there are some decisions that have to be made by companies that have lots and lots of data that was never intended to train models. But certainly could train models. And there are conversations being had at these companies about whether it's legal and or ethical to potentially use anonymized data that previously wasn't intended to train models.
[01:01:36] Paul Roetzer: To train models because there's lots and lots of money to be made. You could certainly say there's lots of value to be created for your customers if you use that data to do things like what Zoom saying they're doing. And there isn't a software company right now that isn't probably grappling with. What are we willing to do with the data?
[01:01:56] Paul Roetzer: What are we allowed to do with the data? And the ethical and legal aren't always the same, so there's going to be a lot of things that are legally allowed to do. At least right now in the United States in particular, the laws haven't caught up that saying, you can't do this. And so I think there's, over the next couple years before the laws do start to catch up, there's going to be a lot of ethical decisions that have to be made within organizations about whether they do or do not use the data in ways that they may be legally allowed to, but probably isn't in the best interest of the consumer.
[01:02:29] Paul Roetzer: So yeah, check terms of use, get a good lawyer, make sure the lawyers are looking at the terms of use. Make sure that, you know, it's such a free for all right Now with individuals going and getting AI tools and testing them and using 'em. And this goes back to like having policies around what tools you're allowed to use.
[01:02:45] Paul Roetzer: Nothing. Stopping an employee from going and getting a personal assistant that all of a sudden has access to all of everything they do on their screen and. It's recording everything. If you're on rewind, like it's a, it's the wild west. We've said it a million times. Like anything goes right now. And so until you get these policies in place and you have some guardrails to protect your employees, protect your team, you just don't know.
[01:03:09] Paul Roetzer: So yeah, get a good attorney. Iknow as a marketer coming up, I always hated having to talk to the attorneys. Like I hated having to send stuff to the attorneys. It cost a lot of money. It took a lot of time. It always seemed to be, you know, roadblocks to just moving quick. And time and time again, I have learned that it's time and money well spent.
[01:03:28] Paul Roetzer: And you know, I think sometimes marketers and entrepreneurs, Risk is great, but I think you have to learn to balance the risk moving forward in areas you don't understand, like data, terms of use around this technology. A lot of us are still learning how to maneuver in this space, and I think it's wise to lean on the people who know what they're doing when it comes to areas you're not comfortable with.
[01:03:52] Paul Roetzer: And as an entrepreneur, Ideal with that Every day Stuff comes up. It's like, this isn't my area of expertise. I have the time to make it my area of expertise. I'm going to pay a human, I'm going to use Chee to help me, but I'm going to pay a human who does this for a living to figure this out and give me confidence.
[01:04:07] Paul Roetzer: I'm making the right decisions. So again, hope for the future that we're going to need humans no matter what. I continue to rely on them every day.
[01:04:18] Mike Kaput: Awesome. Well that's all we got for this week in ai. Paul, thanks again for your time and insight, breaking down what's going on for our audience. I know we all
[01:04:25] Paul Roetzer: appreciate it.
[01:04:26] Paul Roetzer: Yeah, thank you everyone. We will be back again next week. And again, the make on on demand thing, if you missed it and you want to grab those, just make on.ai. And those are available for you now. Alright, we'll talk to you next week. Thanks, Mike. Thanks.
[01:04:40] Paul Roetzer: Thanks for listening to the Marketing AI Show. If you like what you heard, you can subscribe on your favorite podcast app, and if you're ready to continue your learning, head over to www.marketingaiinstitute.com. Be sure to subscribe to our weekly newsletter, check out our free monthly webinars, and explore dozens of online courses and professional certifications.
[01:05:02] Paul Roetzer: Until next time, stay curious and explore AI.