← The Artificial Intelligence Show
The Artificial Intelligence Show
#134: DeepSeek Updates, OpenAI’s o3-mini and Deep Research, New AI Copyright Guidelines, OpenAI In Talks to Raise $40 Billion & Your AI Questions Answered
01:16:30
Loading summary
Transcript81 lines
- [00:00]
Paul Raitzer
I expect that people are going to get used to the Chinese labs putting out advancements and models, doing things more efficiently, pushing the American labs to maybe release stuff before they would normally be ready. And I think we're just going to kind of in a week or two, this will be the new world and we'll be used to it. Welcome to the Artificial Intelligence show, the podcast that helps your business grow smarter by making AI approachable and actionable. My name is Paul Raitzer. 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 Caput, as we break down all the AI news that matters and give you insights and perspectives that you can use to advance your company and your career. Join us as we accelerate AI literacy for all. Welcome to episode 134 of the Artificial Intelligence Show. I'm your host Paul Raetzer along with my co host Mike Kaput. We are recording this on Monday morning, February 3rd, 11:00am Eastern Time, approximately, I don't know, 15 hours after OpenAI dropped a Sunday night AI model on us. It was a unique end to the weekend as we'll talk plenty about the new Deep Research model, the same name as Google uses. So it was a busy Sunday night, busy Monday morning getting ready for this one. So we've got a lot to cover as always. Couple quick notes before we get started. We have dedicated artificial intelligence show social media accounts. If you're not following them. So on X or Twitter it is @AISHowpod and on YouTube it is @AI Showpod so you can find us on YouTube and Twitter and we also have, we post on TikTok and Instagram and all those things, but those are the two main dedicated accounts if you want to follow along on those channels. Also, just a quick note and thanks to everyone for the Response on the AI Literacy Project. I have my inbox, email inbox, LinkedIn inbox, and everything in between has been flooded with amazing notes and people volunteering to get involved in the literacy project. And if I haven't gotten back to you, I apologize. I'm trying to work through all of those outreach communications and we're grateful for everyone for, you know, wanting to kind of explore that. If you're not sure what I'm talking about, you can go back and listen to. Was it 133, Mike? We talked about the literacy project or 132, one of the shows last week it was 1.
- [02:32]
Mike Kaput
132.
- [02:33]
Paul Raitzer
Okay. Yeah. So or, or just go to literacyproject AI and you can learn more about the AI literacy project. And then, just as always, we're very grateful for our audience. We had a pretty remarkable week last week in terms of the, the numbers. So we did two weekly shows which is not the norm and as a result we broke 24 hour records for downloads, the 7 day record for downloads and the 30 day trend line. It actually saw a 30% jump from our previous high. So we're, like I said, always grateful for our growing audience and the people who listen regularly and, and the people that reach out to us and kind of let us know, you know, how the podcast is impacting them. So we really appreciate all that and we'll keep doing our best to keep bringing you this news every week and trying to condense it into the hour, hour, 20 minutes or so that we do each week. All right, so this week is episode is brought to us by the AI Mastery Membership Program. We've been talking a lot about this lately. Incredible response to this program as well. We've seen some huge jumps and number of members that are joining. Again, if you go Back to episode 132, I sort of explained kind of the vision and roadmap for the Imastery membership program. We made some major changes on January 24th that took the existing Mastery membership and enhanced it by bundling in our piloting and scaling. AI certification courses are now part of that membership for the same fee. And then we announced phase two which is coming in spring of 2025, just a few short months away. We are very aggressively working on this. We're going to be launching a new AI Academy platform and user experience. We're going to be dramatically expanding the courses and professional certificates and building a turnkey AI Academy solution for businesses. So we've had a ton of people outreach about the business accounts. We've had some amazing conversations just in the last week with enterprises that are looking to build AI literacy into their organizations. And that's going to be, you know, things like our AI Fundamentals course series, our Gen AI App course series, which can be like weekly reviews of products and tools and platforms, AI for industries course series, AI for departments. Those are all coming this spring. So you can go to SmarterX AI AI mastery, or just go to SmarterX AI and click on Education and you'll see AI mastery right there. And for our podcast listeners, we have a promo code of pod 100 p o d 100 and that gets you 100 off of the membership. Also the AI for writers summit. We've been talking a lot about that. This is our third one Mike, I think. Yeah, third annual summit. We had 4500 last year, I think from 90 countries. So this has been a huge event for us the first two years. I think a lot of writers and creators are really trying to understand the moment we find ourselves in and figure out what all these advanced models mean to the writing profession and copywriters within enterprises. A lot of uncertainty ahead and we're trying to kind of piece it together for everyone. This event is going to be Thursday, March 6th. It is a virtual summit from noon to 5 Eastern time. There is a free registration option so you can check that out. The agenda is now live. We're going to have My opening keynote is going to be the move 37 moment for writers and creators. I'm kind of like telling the story of AlphaGo and what we learned from that and how it's going to impact writers and creators. We have a session on AI copyright and IP which is extremely critical. We're going to talk about some updates to us copyright in today's episode got Mastering AI Prompting. Is that you Mike? Oh no, that's Andy Crestadino is doing I think Mastering AI Prompting. Mike's doing AI powered research how to trans transforming how writers discover and create which the new OpenAI model. Mike certainly great timing on that.
- [06:15]
Mike Kaput
Yeah.
- [06:16]
Paul Raitzer
And then we still have a yet to be announced keynote and then an ask us anything at the end. So AI writer summit.com again that is AI writer writer summit.com you can also find this on the marketing AI institute site. If you're bouncing around on there, just click on events. And so that again is coming up March 6th. There is a free option. Check that out. And then final reminder, the submission to speak at Macon, our 6th Annual Marketing AI Conference October 14th to the 16th in Cleveland. Registration is open for that event now, but we are also accepting speaker applications and this is like kind of on a rolling basis through February 28th. You have to get in but I would get in as quick as possible because we've been getting a lot of submissions and we're going to be kind of filling the agenda as we go. So check that out at Macon AI that's M A I C O N AI. On the homepage there is a button to submit your speaker application. So if you're interested in speaking, check that out. If you're interesting in interested in attending, you can get registered now at the lowest pricing, I think the price goes up at the end of each month, so take advantage of that. Okay, so we'll get into Deep Seek a little bit because it continued to dominate the news at least until Sunday night when OpenAI dropped a model that no one was really expecting until that day. So, Mike, kick us off with what's the latest on deepseek and the implications?
- [07:38]
Mike Kaput
Sure thing, Paul. So the fallout effects from Deep SEQ are still rippling through Silicon Valley and the wider AI ecosystem. We talked a bunch about Deep Seq last week. This is a Chinese AI lab that has created open weight models that are allegedly as powerful as OpenAI products and the products of other model providers for what they believe is a fraction of the cost of training those other closed lab products. Now, the impact of Deep Seq's latest releases, which is V3, which is a competitor to GPT4, oh, and R1, a competitor to Reasoning models, the impact of these was pretty immediate and dramatic. So as we were reporting on this last week, on Monday, when we recorded Nvidia was seeing its stock plunge. It ended up down nearly 17% in a single day, erasing roughly $600 billion in market value. This is reportedly the largest single day loss for any company in US Stock market history. The broader tech sector saw steep declines as investors question the massive AI investments being made by these companies to train AI models, when apparently Deep Seek is able to do it on the cheap. So companies like Microsoft, Meta and Google are facing a lot of investor scrutiny. OpenAI and Microsoft are also now investigating whether Deep SEQ may have used data distilled from OpenAI's systems. David Sachs, which is who's President Trump's aizar, claims that there is, quote, substantial evidence that deep seat used OpenAI's models to train its own, an allegation Deep Seat has not directly addressed. The idea here, which David Sacks is not. The only person to question is did they actually train their their models for as cheap as they say and as quickly as they did without any other help? And it sounds like there are a lot of questions if that's actually the case. Now, Deep Seq has also triggered security concerns. Hundreds of companies and government agencies are now blocking access to it over fears about data security and Chinese government access. The Irish and Italian data protection authorities have also launched investigations into how Deep SEQ handles European user data. So, Paul, first let's talk about the fallout in the markets. Like this is something a lot of listeners I would imagine will immediately be seeing. And we're feeling Nvidia is still kind of way down. Like, is Deep Seat going to Change how investors think about the economics of these big AI companies?
- [10:19]
Paul Raitzer
I don't think so. I think it was mostly an overreaction because people didn't really understand what it was or what the implications were. And it's kind of how the market works. The thing I kind of. Again, I Very, very careful. Not. I'm not providing investing advice. I say this every, every single time, but I had a lot of friends reach out to me asking what in the world is going on that day by text. And what I generally guided people is, you know, if, if, if you're worried about short term. Yeah. That you could see these drops and they may actually sustain for a little bit, but if you're thinking long term, there's no fundamental change. If anything, this creates greater demand for GPUs from Nvidia because it just proves out the ability to build intelligence more efficiently. It doesn't mean you're going to build less of it or like, require less computing power. So I just, I just really saw it as, as a significant overreaction. Not surprising. I do think that had it been an American lab that had done it, there would have been no reaction. If not, Nvidia might have actually been up that day, which is part of kind of how I assess this. So I think the, the key is it opens up these kind of unknowns and, and Wall street just doesn't like uncertainty and unknowns. And so they had to kind of let the dust settle a little bit. And I think it was pretty apparent right away that there, everything wasn't as it appeared. Like it wasn't. I think we talked about this on last week's show. Like, the training run probably could have been in the 6 million range, but that doesn't mean that's what they spent to build this thing. It's more likely they spent over a billion dollars on the infrastructure and the chips and everything that they built in to enable it. But it still makes it significant. Like Zuckerberg said, you know, he acknowledged it. Nvidia released a statement acknowledging it. Sam Altman tweeted about it saying it was an impressive model. So people took notice and it was like a note was certainly a noteworthy event in the timeline of advancements of AI. I think the, the fact that they show the chain of thought so clearly, which is still wild. Like, if you look on Twitter and you see some of the examples, I haven't personally used the model, but if you look at some of the examples where it's showing you what it's thinking as it's like answering it's wild. Like it's kind of how the human mind works that you can see it kind of bouncing around. And OpenAI has to date, as well as Google and Anthropic and all, they have all avoided showing that level of chain of thought, I believe for security purposes. I think they see that as somewhat dangerous to be be able to understand that deeply what the model is doing and thinking. So you're seeing kind of some other stuff where now these other model companies are starting to trickle out, more advanced stuff. So I don't know, we'll, we'll see what happens this week. It's going to be hard to get a read on the markets this week, obviously because of the tariff war that was started over the weekend. And you know, now the US is imposing tariffs on Mexico, although I just saw that got delayed a month now we'll see if the one in Canada goes through, it's gonna be 25% on Canada. But you know, they were having phone calls today to try and negotiate that down. So the markets are running haywire right now because of tariff wars, not because of Deep Seek and other stuff. So you can't read into like Nvidia stock this week and figure anything out. So the last note I'll make here is Andrew, um, who we've talked about many times. You know, one of the founders of the Google Brain team, creator of Coursera. He has deeplearning AI, heavily involved in, in the current and past of AI. So he tweeted the buzz over Deep Seek this week crystallized for many people a few important trends that have been happening in plain sight. One, and these are important notes. China is catching up to the US in generative AI with implications for the AI supply chain. Two, open weight models are commoditizing the foundation model layer, which creates opportunities for application builders. Now what that means there is like, if you think about OpenAI and Google, how they have kind of these more closed proprietary models, what they're saying is like these open weight models are catching up to the biggest models out there and it's going to commoditize the value of these things. And then the third was scaling up isn't the only path to AI progress. Despite the massive focus and hype around processing power, algorithmic innovations are rapidly pushing training costs down. And then the final note I'll make here is related. Sam Eltman actually did a Reddit conversation. This was on like Saturday or Friday. I remember what day this was. And Sam said OpenAI has, quote, been on the wrong side of history and need to figure out a new open source strategy. He added that not everyone at OpenAI shares this view and it's also not our highest priority. So again, the deep seek moment may have been the trigger for Sam to finally step up and say we not, we're not doing everything I think we should be doing with open source. Which is what, you know, Elon Musk's beef has been in part with Sam all these years. So yeah, just, you know, it's a continuing story. I think it's going to probably fade pretty quick. I expect that people are going to get used to the Chinese labs putting out advancements and models, doing things more efficiently, pushing the American labs to maybe release stuff before they would normally be ready. And I think we're just going to, kind of, in a week or two, this will be the new world and we'll be used to it.
- [15:45]
Mike Kaput
And just to kind of wrap this up with a bow. So it's clear there is a moment here. This is deeply important. But it sounds like also just contextually, you know, people, a lot of people were under that false impression that, oh, they did for $6 million what OpenAI does. That's just not true.
- [16:06]
Paul Raitzer
That I think it's pretty safe to say that, that those media headlines were misrepresenting what was actually going on. And I, I think you're right. Like, you know, if you look back five years from now, I would imagine if you were making a timeline of significant milestones in AI development over, you know, that five year period, the Deep Seq moment is probably on that timeline. Like that's how I kind of think about why we're talking about this for a second week in a row is I think when we look back, it'll have been a very significant moment. And I. And one of the things that may trigger that I don't think is necessarily a good thing is I think OpenAI and others are going to accelerate the release of more advanced models as a result of this, to stay ahead when they have yet to solve how to keep those models safe. And I think that's going to be a major problem as we move throughout this year.
- [16:58]
Mike Kaput
As we have said many times, it's not a good year for the AI doomers.
- [17:02]
Paul Raitzer
No.
- [17:05]
Mike Kaput
All right, so speaking of kind of effects of Deep Seek on the market, so the next big topic, we're going to talk about both OpenAI's O3 mini and deep research. So we're first going to tee up O3 a bit, Paul, and talk about that and then discuss what's going on with this pretty cool new feature called Deep Research that they have released. So first up, literally just days after all the deep seat drama OpenAI has released, Open has released O3 mini, which is a new reasoning model that is designed to excel at STEM tasks like coding, math and science, while being more efficient and cost effective than its predecessors. O3 mini can search the web and will eventually show its thinking while it goes about accomplishing tasks. Now, in a sign of perhaps democratization of access, OpenAI is actually making O3 mini available to free ChatGPT users. OpenAI is also rolling out O3 mini across its entire product line. It is available not only in just free chatgpt, but with higher limits in the paid versions as well as through the API. Apparently, enterprise customers will soon gain access as well. And technical evaluations show that this model matches or exceeds the performance of previous models on certain STEM benchmarks Marks while delivering responses 24% faster. Chat GPT plus and Pro users also get access to a model called O3 Mini High, which thinks harder and gives better answers. So Sam Altman posted on X about O3 mini, quote, a lot of people prefer this to O1 and it's just the Mini model. Now we work on the big brother. So Paul, let's first talk a little bit about O3. Like, this is, seems like an obvious response to Deep Seek in terms of moving this release forward, offering significantly more intelligence for far less cost at much greater speed. Seems like no matter how quickly these models come out, that's the trend to bet on, right? Is that we're going to see more things like O3 mini.
- [19:12]
Paul Raitzer
Yeah, it's going to be harder to keep up with the model releases. It's, it, I mean it's seriously getting out of control. So the, the first thing I, I note here is like, I would imagine like some other people in our audience, my initial reaction was, oh my God, the name, like, it's just. So if you go into Chat GPT right now you can choose depending on whether you're in your personal account or business account or pro account. Like now you have differences in the accounts, but in the standard paid account there's seven models to pick from from and very little guidance as to like which one is like, okay, faster responses, better responses, like, well I don't, how am I supposed to choose? I want better responses, but I guess I want the fast response, like, I don't know. And it leaves the user to actually like be testing the outputs of these different things, trying to figure out and then you can't even tell the difference.
- [20:00]
Mike Kaput
Yeah.
- [20:01]
Paul Raitzer
So, you know, I think that they're really starting to get in their own way here in terms of the average user adoption, the non developer user adoption. Because I really think developers are, and maybe like extreme power users, non developer power users. They're the only people who care about all these model choices. As I've said many times, like people just want to use ChatGPT, they don't want to have to pick between seven models. So that being said, I did my first test with O3 mini high was actually on a related topic. I said, how could OpenAI simplify model choice for non developers using Chat GPT? Right now, when a user creates a new Chat, there are seven models to choose from. Most ChatGPT users would struggle to understand the difference between these models and when to use which one. What would be a better way to handle this? And it actually had like two decent, like actually pretty obvious answers. One is smart selection so that the interface just picks it for you. You put in your prompt and it picks it. And the other is to like hide all of these other options under an advanced option. So just like, I'm an advanced one, I'm a developer, I want my choices. The average user doesn't want those things. Things also worth note that Sam has teased there's something else coming for O3 mini this week. And the Deep Research model from last night, wasn't it? He said on Twitter. So there's something else coming there. But then the really interesting thing happened last night. Just so this again, this is, we're talking on Monday morning, this is Sunday night. They were in Japan, I think is where he released this. They must, they had like an event. They were live streaming from Japan, but they released Deep Research, which, yes, if you're following along at home, is the same name as Google. But they didn't uppercase it, I'm guessing, so they didn't run into like legal trademark issues maybe. So they just like lowercase Deep Research. So when they finally name something, oh my God, that isn't like O3 mini high, they. They use somebody else's name. So Mike, I know you dove into a little bit of the background on like what Deep Research is. So why don't you give us the rundown there and then I'll offer a perspective on it.
- [22:02]
Mike Kaput
All right, so this is a new capability in Chat GPT called Deep Research. It's like an option you can select when you are actually chatting with Chat GPT. So not like A separate standalone product. But what this is is it's designed to function as an autonomous research agent that can go spend a bunch of time investigating complex questions and delivering comprehensive data answers. So Deep Research will show you exactly how it arrived at its conclusions through a detailed sidebar that tracks its research progress and cites its sources. So think of this basically like a research analyst working alongside you, gathering and synthesizing information while keeping detailed notes about their process. So you can input questions through text images or even upload documents like PDFs and spreadsheets for context. Then the system works independently to navigate through all that information, adjusting its approach based on what it finds. In the coming weeks, OpenAI plans to enhance this capabilities output with embedded images and data visualizations. Now right now the access is initially limited to Pro subscribers who get up to 100 queries per month because it's apparently very computationally intensive. Apparently though plus team and enterprise users will gain access later once there's a more efficient version that they are using for those accounts. So Paul, I guess like you know, obviously his name is infuriating at this point. It is triggering for me that like we couldn't call it anything else even we could use research still just pick another.
- [23:42]
Paul Raitzer
At least they didn't do like Deep Seeker, like I could.
- [23:45]
Mike Kaput
Yeah, oh my God, that would be a nightmare. But basically the kind of difference as far as I can tell right now is look, both this and Google's Deep Research aim to kind of do the same thing, which is go autonomously do research on topics for you. But I think Ethan Malik posted a really good kind of breakdown about how they're different. So he says OpenAI's deep research is very good. Unlike Google's version, which is a summarizer of many sources, OpenAI is more like engaging and opinionated, often almost PhD level researcher who follows the lead. So this basically seems like this is a much more fluid, autonomous kind of research partner versus something that's going and creating like a whole research brief for you. Though they are both producing briefs, it seems like this one might be more useful to more converse with in some ways, but we'll time will tell as we test it out.
- [24:41]
Paul Raitzer
Yeah, I don't know that the average user is going to like immediately be able to tell the difference. I, yeah, I did test it and I don't know that I would agree with Ethan's analysis. Now Ethan's had early access to this and done way more tests, but in my very limited test, I, I don't know, they're very Similar outputs. Like it wasn't dramatic difference between one or the other. So yeah, all right, so I have a few thoughts here. Kind of like wander on this one for a minute. My initial reaction to this, this was that the AI timeline is accelerating so our path to AGI or whatever we want to call it, the, the delta between what these models are capable of and society's understanding and preparedness grew again last night. Like I, I think this is a advancement in capabilities. It is our first interaction outside of OpenAI with the full O3 model. So the, this is not 3 mini high, this is the full O3. Now there's an O3 Pro coming also that's going to be even more powerful. But this is your only chance to currently use the full O3 model is to actually use this product. So you can kind of use the most advanced reasoning model they have. I think Perplexity is cooked. Like I've said this the last couple months that I was starting to become less bullish on Perplexity as a, a sustainable company. And I think they're done. Like they're. Because Google already is better than them at this. OpenAI is obviously going to be better and they're only going to accelerate the development of this. Everybody else is going to build the same thing. Anthropic is going to have a deep research product, GROK is going to have it. Everybody's going to have a deep research product. And once those things, you know, all have access to the web, they all become extremely valuable. Like what do you need Perplexity for? So I think if I was again on like putting odds on things, I do think perhaps complexity at some point just gets folded into somebody this year. I just don't understand what their market is going to be when everybody else can kind of do the same thing. Plus you can build these things open. So I saw somebody already like created an open source version of this on like hugging Face where they like reverse engineered how they did it. So that's interesting. The battle with Google is certainly noteworthy here. I think we are now in a race for these reasoning models and to productize those reasoning models. And maybe this is how they get away from the model confusion is you start building independent products or agents for specific things. And so deep research is that idea that like I don't even care what model they're using, does it do the analysis I need it to do? And so maybe productization is actually what eliminates the confusion with all these models. I haven't had time to think about this, but the impact on the future of search, SEO, web like changes impact on education. You cannot give research projects to students and not accept the fact that they may be using these tools to do it. Impact on the future. The work of work which I. I'll talk about has kind of go on here for a moment. Last week we talked about humanity's last exam. I think that was 133Mike when we talked about that new exam. So basically an exam was created to try and accept the fact that the current evaluations weren't hard enough for these models and we needed something insanely hard. So they went out and got over a thousand subject experts to contribute all of these questions, like the hardest questions you could create. There's a Data set of 3,000 challenging questions. And we went from, let's see. Kevin Roos tweeted, when I wrote about humanity's last exam, the leading AI model got an 8.3%. Five models now surpass that and the best gets 26.6%. That was 10 days ago in all caps. So we went from 8.3% to 26.6% percent in those 10 days. O3 mini high is 13%. Now it's text only because that's all O3 mini high does.01. It was the highest multimodal model at 9.1% and now we have deep research at 26.6% all in two weeks. So Sam actually tweeted on Saturday, the day before they released Deep Research, a screenshot of humanity's last exam, and said we're going to need a new exam soon. Now keep in mind, they're training 04 already. Like they already know what the scaling laws are for the reasoning models and they can look out ahead and say by the time we get to 05, this, this exam's done. Like we will have surpassed this too. Okay, so here's where we start getting to the interesting stuff. So Sam tweets, this was, I think last night because then he said it in his presentation at this live stream event. He said, my very approximate vibe is that Deep Research can do a single digit percentage of all economically valuable tasks in the world, which is a wild milestone. That's his, his words. So, so I thought, how can I test this model? So I upgraded to pro this morning and I was like, all right, let me, let me run my first test. So I took this idea from Sam that we are now at the point where single digit percentage, which who knows what that means? 9%, all economically valuable tasks. Keep in mind, jobs are series of tasks. So I went into deep research and I said, okay, this is the prompt I gave it when presenting OpenAI's new deep research capability. Sam Altman said, quote, my very approximate vibe is that it could do single digit percentage of all economically valuable tasks in the world, research and analyze what jobs could be most impacted in the next one to two years, including rationale and estimated impact. It went off for five minutes and started doing its thing. And again, if you haven't used one of these models, you need to experiment with them. Like watching its chain of thought is kind of wild and it does share way more than it used to share about that. It finished it and honestly, like it looked really impressive. I could see like a college student turning it in and saying, hey, I did it is awesome. But the first time you start clicking on citations, you realize it's citing everything from the March 2023 GPTs are GPTs paper and then articles about that paper. So I went through, read, read this. I was like, this is useless like this, this is 22 year old data basically. So my second prompt was, you're citing studies and articles from 2023 that do take into account all the advances in AI models since then. I'm specifically looking for an analysis of how open AI's new deep research model and then I, I put C and then I put the URL into the announcement post how this will impact jobs. So now it got way better. Now it did still pull some of those 2023 studies, but it also started pulling in more direct research. And as you watch the chain of thought, you could see it saying, oh, but I need to connect it back to the current release. Let me go revisit that current release information from OpenAI and you could like watch it going back and forth, which again is just fascinating to see. So I'm gonna, I wanna actually take a couple minutes here because I think this is really critical and I'm going to go through some of the highlights of that deep research output. So again, my prompt was all about trying to understand the impact. It starts with roles most susceptible to AI automation and augmentation. Not all knowledge jobs will be affected equally. The deep research models capabilities align closely with tasks that involve information processing, writing and pattern recognition, which means certain roles will see a greater impact. Jobs heavily centered on generating or analyzing text and data are the most susceptible, especially those with routine or formulaic components. On the other hand, roles requiring a high degree of human judgment, interpersonal skills or creativity may be less affected in the short term. Now the reason I'm reading this is because I agree with everything it wrote. So I am I'm going to now provide the perspective of someone who thinks deeply about this a lot. And this is a really good output. It and I think it's important for people if you haven't come to the stage of acceptance of where we are, I think these are really important things to listen to. So then it goes into examples. Writers, editors and content creators Professionals who produce written content. Technical writers, journalists, copywriters are highly exposed to this. Technology Researchers and analysts Knowledge workers who research and synthesize information such as market research, policy analysts, academic research assistants will find many of their tasks accelerated. Legal assistants and paralegals Legal research and document drafting involves scanning large volumes of text and producing written analyses. This means law firms might handle the same caseload with fewer support staff or focus their staff on more strategic advisory rather than rote paperwork, which is the theme we keep stressing. It's just like it's going to change things real fast. Accountants and financial analysts A large portion of work in finance involves analyzing numbers and writing reports. For instance, summarizing quarterly results, auditing financial statements or evaluating investment options. The deep research model can analyze financial data and produce well written analyses or summaries. It says while complex strategic decisions still rely on human judgment and domain expertise, the preparatory work compiling data option analysis can significantly be automated. This suggests a high level of augmentation in finance roles and potential automation of junior level tasks. Software engineers we've talked a bunch about. There's a section on evolving skill requirements for knowledge workers. As the deep research model takes over certain tasks, the skill set required for many jobs will evolve. AI literacy will become as important as basic computer literacy. Knowledge workers will need to know how to work with AI tools effectively. This includes skills like prompt engineering, interpreting AI outputs and verifying the information that provides. That's a really important paragraph that when you think about any job, any knowledge work profession, those are fundamental things. Knowing how to prompt it, knowing when to give it a better prompt because the first prompt didn't get what you wanted out of it, or you knew the output wasn't the level of expertise you needed. Interpreting those outputs, verifying the information is accurate. These are the things. So then it says critical thinking and oversight skills will be in high demand. I agree. The ability to apply domain expertise to validating AI generated results becomes a core part of the job. That is what we're doing in real time right now. This is me applying a domain expertise to assess the output of the AI the human only skills increase in importance. Creative thinking, strategic vision, interpersonal communication, emotional intelligence will differentiate employees. And an era where routine analysis or writing is done by machines. I. I'll stop there for a second. Mike, is there anything I just went through there that you would disagree with or that like.
- [35:26]
Mike Kaput
Not at all. This is perfect.
- [35:27]
Paul Raitzer
Yeah, it's pretty good synopsis. Okay, so then OpenAI says they're going to keep doing this. They're going to build more agents. This is going to get better and better and better. Then I wanted to pull one other thought in here. So I don't know if Mike, we mentioned this later on. If we do, we can kind of gloss over it, but Y Combinator. So if you aren't familiar with. Y Combinator is a startup incubator. They have, let's see, since 2005, they've funded over 3,000 companies. There are more than 60 Y Combinator companies valued at over a billion dollars. So unicorns and the combined valuation of YC alumni is over 600 billion. Sam Altman was the president of YC Y Combinator from 2014 to 2019. This is a very important organization within the technology world. They put out a call for startups. So this is last week. So if you don't think like, if you're still on the fence about whether or not we're trying to replace people, I'm going to read to you two of the areas that they have put out a call for startups. The first AI personal staff for everyone. Quote. Despite the explosion of software in the last decade, wealthy people still employ lots of human staff to provide personal services. These are things like tax accountants, personal lawyers and money managers, but also personal trainers, private tutors and even personal doctors. The list goes on. Why can only the rich afford this? Because software hasn't been able to replace these types of personalized knowledge work tasks until now. Over the next few years, we expect AI to get good enough to do most of these jobs, not just tasks, jobs. So if you are working to bring a part of this personal AI staff to every human on the planet, we'd love to hear from you. Then one other one there was like I don't know, 10 or 12, so I'm just picking two of them. The next one was vertical AI agents. What is a vertical AI agent? This is them. Quote it's software that's built on top of large language models that's been carefully tuned to be able to automate some kind of real important work. In recent batches we've had YC companies build an AI tax accountant, an AI medical biller, an AI phone support agent and an AI compliance agent and an AI quality assurance tester. The value proposition of B2B SaaS companies was to make human workers incrementally more efficient. The value prop of vertical AI agents is to automate the work entirely. Vertical AI agents that reach human level performance grow extremely quickly. Again, the reason we are spending so much time on this topic is if you're not aware or accepting of where we find ourselves, this is it. You have Y Combinator that drives the growth of startups, calling for startups to replace people. You have Sam Altman saying the current model that he knows is only going to get better in a few months is already capable of single digit tasks for all human work. This is, it's real like, and, and again like the key here is the timeline is accelerating. We're likely going to get a new model this week I think from Google. It sounds like maybe Anthropic supposedly is sitting on something as good or better than what Deep Research is doing and hasn't released it yet due to concerns. It's only going to accelerate from here and it's just really, really important that people, even if you don't want to believe it, that, that you try and step back and be realistic about what is happening because it's going to happen really fast.
- [39:06]
Mike Kaput
That is a really good kind of call to action and warning. And honestly I just keep coming back to in my own life trying to think about how do I become less emotional about this Because I get it, it's scary, it's overwhelming. But like you have to look at the reality, the writing on the wall. And also how do I go even more all in in my time on AI for X. Right. I think that's really the only sustainable strategy I'm seeing as a knowledge worker moving forward.
- [39:34]
Paul Raitzer
Yeah. And I, I think it's a great point Mike, but, but this like it's hard to emotionally disconnect ourselves from what's happening. Like I was, someone was asking me over the weekend about like tariffs and like why would we do this? Like our allies, like what is going on? And it's hard to not get emotional about it. But I said like just factually like the chaos and the pain is the point. Like the strategy they're employing is to create chaos and pain and so they can leverage it to get what they want in the end. And that is not a right or left like opinion. It is just that's what's being done and so you do, you have to kind of like try and be able to step back. And the same thing applies in the AI world. Like, what is going on? Why are they doing it? And like, what does it actually mean? And to disconnect from the fact that, wait a second, I'm a writer. Wait a second. We do analysis every week on the podcast. Like, this is our world. My wife is an artist. Like, we live in this world where this impacts us personally. But you do have to step back and try and remove that and be objective about what's actually going on and what does it really mean.
- [40:37]
Mike Kaput
All right, our third big topic this week, the U.S. copyright Office has issued a landmark report that provides updated guidance on how copyright law applies to AI generated works. This report is titled Copyright and Artificial Intelligence Part two Copyrightability. And it comes after their extensive consultation with over 10,000 commenters from all 50 states and 67 countries. So they did an executive summary of their core findings. This is a 50 plus page report. I'm just going to quickly touch on the main points here. So first, existing copyright law is adequate to handle AI generated works. They say that no legislative changes are needed to move forward here. Right now, the use of AI tools to assist rather than stand in for human creativity does not affect the availability of copyright protection for the output. Copyright protects the original expression in a work created by a human author, even if the work also includes AI generated material. Copyright does not extend to purely AI generated material or material where there is insufficient human control over the expressive elements. Whether human contributions to AI generated outputs are sufficient to constitute authorship must be analyzed on a case by case basis based on the functioning of current generally available technology. They say that prompts do not alone provide what they call sufficient control. And they say that human authors are entitled to copyright in their works of authorship that are perceptible in AI generated outputs, as well as the creative selection, coordination, or arrangement of material in the outputs or creative modifications of the outputs. So while the report offers quite a bit of clarity actually on some key questions people have, it also acknowledges that standards may need to evolve as AI advances. So they're going to plan on continuing monitoring the developments going on in this space and providing ongoing guidance. So, Paul, I guess, like with the caveat, as always, we are not lawyers. You should check with your lawyers before taking any path forward here. But it seems like this is at least some of the guidance we've been waiting for for a while. Like, it kind of seems like decently big news that AI generated outputs can, in certain Circumstances get some type of protection. Like, is that how you're initially reading this?
- [42:58]
Paul Raitzer
Yeah, I mean, so you, you know, you and I are in the midst of a pretty large content strategy that this affects. Like, and we'll. Mike and I'll talk more about this in the future, but we're always assessing how AI can be used in outputs that then impacts our ability to hold a copyright to that output. And so we were just on a call with our IP attorneys 10 days ago on this exact topic, and then this came out and the immediate email was, get this to the IP attorneys and ask if this changes anything based on what we just discussed. So I'll read two quick paragraphs from the release from the Office as the Office confirms that the use of AI to assist in the process of creation or inclusion of AI generated material is a large. A larger human generated work in a larger human generated work does not bar copyrightability. Now, that's way more clear, I believe, than what they previously stated. So. Right. The previous thing we would always say is like, hey, if I generate AI generates it, you can't copyright it. Now it's moving more towards like, well, as long as you remix it enough, as long as you have enough human involvement, you can actually copyright the stuff, even the stuff that came from the AI, as long as you've made changes to it, basically. So then there's a quote. It says, after considering the extensive public comments and the current state of technological development, our conclusions turn on the centrality of human creativity to copyright, says Shira Perlmutter, Register of Copyrights and director of the U.S. copyrights Office. Where that creativity is expressed through the use of AI systems, it continues to enjoy protection, extending protection to material whose expressive elements are determined by a machine. I don't know this means. However, rather than further the constitutional goals of coverage, that's a word salad to say that, like, they've kind of moved a little bit on this. So here's my overall take. Don't go changing your generative AI policies on your own without input from your attorneys. Go talk to your IP attorney, share the information with them in case they haven't seen it, but be proactive here because your team is doing things every day with generative AI that isn't considering these things. And you do need to kind of quickly assess how this evolves. I would then take the information from your IP attorneys, adjust policies internally as needed, as well as any policies you may have with outside contractors, agencies, freelancers, things like that, and train them how to do it. So if this now creates some level of freedom to use AI more, make sure that they're trained how to properly deal with the outputs of the AI so that your copyright is protected. It's not enough just to say, okay, you can now use generative AI. We can still get a copyright. That is not what the copyright office is saying. They're saying you have to have human elements within it. So come to an agreement what that means within your company with the help of your IP attorneys and make those adjustments. In a quick related note, the Authors Guild, one of the largest associations of writers in America, is launching a new project to certify books that have been written by human rather than machine. The new human authored quote, unquote certification will help authors distinguish their work and let readers know what they're. This is what I think I said this last year, like this was inevitable that we're going to have human authored or whatever you call it, human certified, songs, books, articles, everything. So I'm not surprised at all by this, Mike, but I think we'll see a lot more of this, like human created stamp. I know with my Exec AI newsletter I do on Sundays, I put at the bottom, this is 100 written by me, right? Yeah. Maybe I need to get like a, a Paul stamp on it or something. I don't know.
- [46:38]
Mike Kaput
That would be great. Yeah, no, this is cool stuff to see and I certainly understand the motivation. I'm a little curious how they're gonna actually do this because, like, I don't know. If you figure it out, Authors Guild, go talk to every higher education institution in America as well because it seems like a huge lift to even figure this out.
- [46:55]
Paul Raitzer
Yeah, I don't know if you have to like turn in your word docs and like show the provenance of it all. I don't know.
- [47:02]
Mike Kaput
All right, let's dive into some rapid fire topics this week. So just right after Deep Seek has rattled the tech markets and investors are questioning, you know, do we need to be spending this much money on AI models. OpenAI is preparing to raise what could be the largest private funding round in Silicon Valley history. They are apparently in talks to secure up to $40 billion in new funding that would value them around $300 billion. SoftBank is leading the charge, looking to invest between 15 and 25 billion. The valuation represents a pretty significant leap from OpenAI's previous $157 billion valuation just a short time ago in October of 2024. Now, OpenAI says that they plan to use the funds or it's reported that they plan to use the funds partly to fulfill their $18 billion commitment to Stargate, their recently announced joint venture with SoftBank and Oracle to build AI data centers across the U.S. they also need capital to fund operations. They reportedly lost about $5 billion last year on $3.7 billion of revenue. Interestingly, this would make OpenAI the second most valuable private company in the world, trailing only SpaceX. And it also deepens, appears to be deepening the relationship between Sam Altman and SoftBank CEO Masayoshi San, who appears to be making OpenAI his primary vehicle for betting on the AI industry. So Paul, people were already turning their head at $157 billion valuation. Now we may almost double that, raise an additional 40 billion. How much is enough here? Is there a danger this Stargate thing blows up? Like what, what's going on here?
- [48:53]
Paul Raitzer
I mean, there was the rumor last year that Sam was seeking trillions. And I don't, I don't, I think there was something to those rumors. I don't know that it was 7 trillion like reported, but I do think that when they look out, you know, over the next 10 years, they expect to spend trillions on building out the capabilities and this intelligence. They're big numbers. I, I did a little quick research just to see. I did not use deep research for this. This was a traditional Google search believe the most valuable companies in the world, not just private, but publicly traded. So just for context, at 300 billion, how big is that? IBM's market cap is 234 billion. So it's bigger than IBM. Samsung's 237 billion, T Mobile's 266 and Coca Cola is 273.
- [49:38]
Mike Kaput
My God.
- [49:39]
Paul Raitzer
So bigger than all those. And then right on the heels of Salesforce at 319 billion and SAP at 334 billion. So 300 billion is no joke. It is like top 35 companies in the world in terms of value. So yeah, and it just keeps jumping. I, I would not be surprised at all if they aren't a trillion dollar company, you know, by the end of 2026, if not sooner. Did you see anything in those notes, Mike, about any updates on their move to change the structure of the company that could free them up to ipo.
- [50:12]
Mike Kaput
It's possible I missed something, but I was actually surprised at like the lack of, of talk around that when I scanned it.
- [50:18]
Paul Raitzer
I didn't see anybody mentioned. I was like, that's a pretty important part of the story.
- [50:22]
Mike Kaput
Especially because I think their last funding round.
- [50:24]
Paul Raitzer
Right.
- [50:24]
Mike Kaput
Also had conditions around how the money was used based on that conversion. So I, I'm not sure.
- [50:30]
Paul Raitzer
Okay, well, we'll look into a little bit more. I mean we'll ask Deep Research. But yeah, I do think like, I assume they are moving forward with those plans to restructure the company so that they can eventually ipo. That would be the most logical path.
- [50:45]
Mike Kaput
Yeah. And it's interesting too, just as a final note here, you know, I think it is easy to say like, okay, this is a crazy valuation. You're charging 20 to 200 bucks a month for these licenses for ChatGPT or whatever. But in the context of what we just talked about with Sam's comments around Deep Research, you have to start thinking of like the tam, the total addressable market is not other software licenses.
- [51:08]
Paul Raitzer
It's jobs, it's intelligence, the serve, it's.
- [51:11]
Mike Kaput
The services software concept we've talked about a few times. Like when you start thinking, oh, not what's, what's the market for accounting software, what's the market for accountants? That's a very different question. In terms of the numbers involved.
- [51:23]
Paul Raitzer
Yes. And what, and they won't say that out loud. Like they're not going to tell you they have a deck that has the total addressable market of all knowledge work. But that is basically what we're talking about is what is the value in the US of a hundred million knowledge workers and what they do and their contribution to GDP and yeah.
- [51:44]
Mike Kaput
All right, one other story this week about kind of the Deep Seek fallout. So Mark Zuckerberg actually just laid out a pretty ambitious and urgent vision for Meta's AI future and addressed some of the turbulence caused by deep sea in a company all hands meeting. He told employees to quote, buckle up for what he called a quote intense year ahead. He thinks in 2025 it'll be the year that a quote highly intelligent and personalized digital assistant reaches a billion users. He wants Meta to be the company that gets there first. A key part of this vision is llama4, which is their next generation open AI model. Now in a social media post he revealed that Llama 4 will be natively multimodal, which he calls omnimodal, with built in capabilities for autonomous agents. The smaller Llama 4 Mini has already completed pre training. He also predicted that this will be the year when companies can build AI engineering agents. He's talked about this a couple times, saying that we're going to have AI that can be at the level of a good mid level engineer. And he called this potentially One of the more important innovations in history. Now this roadmap obviously comes as Meta is navigating the turbulence caused by Deep Seek. He was pretty noticeably positive when asked about Deep Seat during the all hands and said the company's quote, novel infrastructure optimization advancements could actually benefit Meta since they were publicly published. So, Paul, like, when I read this and hear this from Zuckerberg, like, is Meta's competitive advantage here really that competitive? Like, I get that it matters that they're one of, if not the main kind of open model provider among the big US labs, but it also seems like deepsea kind of takes direct aim at that.
- [53:31]
Paul Raitzer
Yeah, I think that Zuckerberg's putting on a good face publicly and internally. But the reality is Deep Seek did what he was trying to do. Like they, they disrupted things with an open weight model, which is what he's been trying to do with Llama, and they got way more love for theirs than I think Meta's got for what they've done. So he's a really competitive guy. I can't imagine he was too excited about that. And I know it created a lot of headaches internally. And we talked last week about the, you know, these rooms they created to focus on what did it do, and they're already using it in their training and everything. I did have to laugh though. I was trying to scan and see if I couldn't find. Oh, here it is, I think. So the, the article about this, like, you know, tough year ahead. It was like he, I think he let off the meeting, was saying about how annoyed he is that everything he says at Meta leaks and like he's frustrated with everything. And so the headline was like in, in a leaked article or recording, Zuckerberg says, how angry is that? Everything leaks in the company.
- [54:31]
Mike Kaput
Yeah.
- [54:32]
Paul Raitzer
So, yeah, I guess that's the nature of having a high profile tech company.
- [54:37]
Mike Kaput
Our next topic this week is AI expert Ethan Mollick, who we talk about all the time, just published a really fantastic guide to AI models. It's titled which AI to Use now, an updated opinionated Guide, basically, which he.
- [54:50]
Paul Raitzer
Now needs to update because this was on 26th and there's been like four models since then.
- [54:54]
Mike Kaput
Exactly, right, yeah. That's the danger of having published these takes. Right. So the reason, though this is important is because this really does start to help. At least people answer a big question, which is not always obvious, which is, what the heck should you use? Which model should you pick for your particular use case? So he kind of outlines, look, there's three clear front runners, Anthropics, Claude, Google, Gemini and OpenAI's ChatGPT. Each of these bring something unique to the table. Chat GPT currently leads the pack with Live mode, where you can like have a conversation with the AI while it is seeing what you're seeing in real time. Google has similar capabilities that they demonstrated with Gemini, but ChatGPT is the only one offering that feature to all paying customers. Then obviously the companies are having more and more reasoning models, so AI that can think about a problem before answering. The most capable reasoning models, he says, currently come from OpenAI, though obviously deep Seek is possibly changing that, offering competitive models too. Google has their own thinking models as well. Web access is another key differentiator. ChatGPT, Gemini, several others can access the Internet for current information while Claude cannot. And then some of these, if not Most of these AIs can process images remarkably well. So video analysis is still rare. But for documents as well, Gemini stands out with the ability to process up to 2 million words at once, which is way more than any of the others. So Paul, this advice I thought was useful to note to people just because it is a challenge everyone struggles with. Like it does map to a lot of things we've seen what we advise people to do. I mean, we talked about this before. I think it's very hard to avoid having at least a ChatGPT plus account these days. I mean, for the ROI you get, I'd even go as far as to say paid tools for the other two are a huge benefit. But that seems like generally the advice we're offering to people is like you want to focus on those three while you're experimenting with everything else.
- [57:00]
Paul Raitzer
Yeah, definitely. You know, I. And that's, yeah, that's what we always say is at least just get the ChatGPT account and figure out how to use it. But I think so much like I haven't made a note here as you were talking, like, I would love to see what the usage rates are on Live Mode and Voice. I don't know that there have been as disruptive as maybe we would have expected. I know personally I used the Live mode where you can show it what you're looking at and ask a question. I used it five times in the first two days. I don't think I've used it since it's not quite Apple Vision Pro level of not being used, but it's probably in a similar category where I just don't use it that much. And then voice, I used Voice a lot, but I normally Use it when I'm in my car. And some reason it always like, drops the connection and it drives me nuts. And so I stopped even using the voice in my car because it kept, like, restarting my conversations. Yeah. So anyway, like, I don't know, like, I think it's great Ethan did this summary. I just find myself starting to think more broadly now about these, like, adoption rates. And it even leads to this, like, the reasoning model and what we're willing to pay. So I'm now paying the 200amonth for the pro license rate. I think you do the same thing. Yeah. And the way I justify is like, I'm fairly confident I will get $2,400 in value out of this, like one or two examples or use cases or projects will pay for itself for the year. So I don't mind if I pay the 200 bucks and then forget to use it for 30 days. Like, it's okay, I'm going to get the value. But I know a lot of people probably aren't in that same boat of like, willing to just, you know, spend the 200 bucks. But I do wonder if, like, even these reasoning models, as amazing as they are and passing humanity's last exam, or like, you know, keep rising the charts, the reasoning models are harder to figure out how to use than the chat interfaces. So, like, you and I had a packathon internally to try and figure out how to use reasoning models when Oh1 first came out.
- [58:46]
Mike Kaput
Right.
- [58:46]
Paul Raitzer
And like, what are the prompts that would use? What would be the use case for this? Like, what's something really hard we would be trying to solve where we could actually test this model to where it's actually better than just regular chatgpt. So I do think that again, these, all these companies, Google Open, Anthropic, they're going to continue to have this issue of, like, people just aren't really sure what to do with these things. And as they get more intelligent, it actually becomes harder to figure out what to do with these tools. So I don't know, it's, it's a good reason why, you know, articles like Ethan's are helpful. But I, I do think that in an enterprise, like everything about business adoption, you really just have to be very proactive with helping people figure out how to use these in their specific career, whereas otherwise they just have no idea what to do with them.
- [59:28]
Mike Kaput
Yeah, it just strikes me the more I talk about this with people or publish about it, like, I, I think that this is way bigger of an issue than A lot of people perhaps very close to this realize like the moment you start talking about this, people are just like raising their hand, being like, yeah, I have the same problem. I've had to start building GPTs and prompts to like help me with this issue, which is fun, but also what are we doing here? Like the average person's not going to.
- [59:52]
Paul Raitzer
That you're the 1%. Like the rest of the people are just like, yeah, I don't know what to do.
- [59:56]
Mike Kaput
They're just going to not do it. Yeah, yeah.
- [59:58]
Paul Raitzer
And that's why I think like they OpenAI may be really bullish because they're seeing these like crazy adoption rates of the pro license and people are like paying the 200amonth and they're getting a ton of revenue and they're probably seeing like high usage. That's like the very, very early adopter. That's the innovator stage. Those are like less than 1% of users. I would imagine they're going to plateau real fast on people who actually know what to do with these models and how to use them in their, you know, daily lives or careers.
- [60:24]
Mike Kaput
So kind of related to this venture firm, Andreessen Horowitz actually just released a really interesting market map kind of detailing where the AI voice agent space stands today. So in this A16Z partner Olivia Moore writes, quote, Voice AI is now nearly at human standards allowing tech to replace labor on the phone. This has huge implications for businesses who can answer or make calls 247 at low cost. So they go through all these interesting kind of startups and use cases. It's well worth diving into. In B2B companies are developing specialized voice agents for different industries. There's like home service companies like Rosie AI and Revin are creating voice agents to handle customer service and scheduling. In the restaurant industry, companies like Slang AI and Loman are developing voice systems to manage reservations and orders. Some companies are building voice AI for research to conduct and analyze voice based research. Interestingly, there's a couple use cases and companies in the legal sector like caseflood and legal 27 creating voice enabled legal assistance and some voice AI for banking and financial transactions with companies like Salient and Domu. There are also a couple cool examples they cite in the B2C space. So there's a particular focus on education technology with companies like Speak AI and Practica English are developing voice AI tools to help adults learn languages while startups like Synthesis School and Buddy are creating voice enabled educational experiences for children. So Paul, while we're kind of seeing. Yeah, shoot. Like, advanced voice mode is maybe not reaching its full potential. It sounds like maybe with more of the. At the app layer. There are some interesting things going on here, it looks like.
- [62:10]
Paul Raitzer
Yeah. And I don't know if you have this one in the end, the funding and product updates, but remind me, let's.
- [62:15]
Mike Kaput
Talk about it now.
- [62:16]
Paul Raitzer
Yeah, yeah. So Google's got Ask for Me, which just came out in their search labs. I don't have access to this. I tried to get access, but I don't. I don't know it wasn't already in there. I thought, again, this is the complexity of, like, I thought I already had access to these things and you have to join another wait list. But anyway, in Google search labs, I think you have to go into your personal account. I don't think you can do this through a workspace account. There's a new tool called Ask for Me, and so a Verge article says Google is trying out a new tool that lets AI call businesses to ask questions for you. The feature called Ask for Me collects information about the pricing and availability of a service, but it's only available for nail salons and auto shops right now. So, yeah, this whole, like, movement of trying to find the market fit for these voice technologies, that's going to be a big deal. Now, again, like, this is different from you and I using the voice in ChatGPT. This is like building products around, around these kinds of things.
- [63:07]
Mike Kaput
And the fun dark side of this, I've already gotten noticeably AI spam calls that aren't just robocalls, like, so that's starting too.
- [63:18]
Paul Raitzer
Lovely.
- [63:20]
Mike Kaput
So in this week, we wanted to actually kick off a bit of a new segment that we're hoping to do kind of consistently, which tentatively calling listener questions. So we get a ton of questions each and every week about AI, both through the podcast and through other stuff we're doing, like webinars and classes. So we wanted to start kind of answering some of those questions to the best of our ability on the pod. So if you have a question for us, like, just reach out to Paul or myself. You know, LinkedIn's a good place to do that. Or go to marketingai institute.com click contact us. We'd love to hear from you. We cannot guarantee we're going to get to every question, but there's.
- [63:59]
Paul Raitzer
We definitely won't.
- [64:01]
Mike Kaput
I can guarantee we won't. There is a chance we'll be able to answer yours live if you send it in. So, Paul, I'm just going to throw this week's question at you. And this question says, I do not understand the difference between a custom GPT and an agent, or when to use one over the other. Can you clarify this?
- [64:21]
Paul Raitzer
Yeah. So this is a really good question. I think these are a lot of times people are like, too shy to like, ask these what seem like obvious answers. And then you realize like, oh, you don't understand it because it's actually super confusing. And everybody has different definitions of these things. So an AI agent is basically like, it's become kind of convoluted because everybody's sort of defining different things as agents. But it's an AI system that can take actions to achieve a goal. So in some ways, the reason this one's kind of confusing is that a custom GPT can be a form of an agent, like a very simple agent. So it can go do things. It's using the language model, you know, that it's powered by to actually complete a couple of tasks and deliver something for you, in this case, like an output output. It's different from a computer use agent that like, takes over your browser and fills out forms and, and things like that. Or from the deep research agent we just Talked about from OpenAI and from Google that builds its own, like, plan, goes and executes that plan. So you're not setting the rules for it, you're just saying, hey, I want you to help me with this thing. And it's capable of building a list of tasks, executing those tasks, and creating a deliverable that achieves your goal of doing an analysis of something. So there's different types of agents, different levels of human involvement in how those agents work. You know, the data they use, what they're integrated into, what the process it follows is. But that's where the confusion comes in is, is that like, technically a custom GPT can certainly be categorized as a. A basic form of an agent is kind of a way to think about it.
- [65:54]
Mike Kaput
Yeah, I actually ran into this issue last week. I was talking with a friend of mine who wanted some help with his executive team, thinking about how to use AI agents in his particular domain. And he was showing me some documentation around one of the systems they use. And he's like, yeah, looks like you build AI agents here. Like, this seems really complicated and a big thing because I'm hearing so much about them. You go through the directions and it's very clear. It's just a wrapper over custom GPT. It's literally the same options, which is not a problem. But I was like, man, this is actually going to be a lot easier for you to use this piece of this, the software, than you might have originally thought, because everyone's rebranding this stuff.
- [66:29]
Paul Raitzer
Yeah. And I think this, I don't know if this is overly technical, but the only real difference from like a traditional automation or what we used to call like a bot or an automation, you know, people would use like zapier and set up these zaps to do things.
- [66:41]
Mike Kaput
Yeah.
- [66:42]
Paul Raitzer
In a traditional automation, the human wrote all the rules and the AI just did the 10 things that the human told it to do. It wasn't thinking, it wasn't reasoning, it wasn't like creating anything, you know, on its own. It was just following a set of rules or tasks. Now there's a level of autonomy, a level, not full autonomy, there's a level of autonomy where the AI actually determines some of its actions. It does things outside of what the human told it. So when you give it a prompt, you're not saying do these 10 things in the prompt. It's sort of like taking some ownership of trying to figure out what it's supposed to do and how to do it. And so now within this, these steps is some level of thinking, reasoning, you know, true automation and creativity from the AI, and that's what makes them different than traditional automation.
- [67:31]
Mike Kaput
So this week we're also going to highlight a couple of quick, practical use cases for AI that we're finding useful or interesting or worth discussing at the moment, which is also something we're going to try to aim to start doing, perhaps in a request recurring fashion. So I'm going to kick off one Paul, that I was exploring this week. Just quickly describe it and I also know you had been doing something with ChatGPT's tasks that I actually wanted to double click on a little more with you. So this past week I actually created a pretty extensive prompt and we'll link to a post that describes it that turns AI into your personal writing critic. So it doesn't just check grammar, doesn't just look for errors, it analyzes seven different areas of your writing, like clarity, logical flow, how engaging it is, precision, persuasiveness, tone, and writing mechanics. Also provides a bunch of actionable scores and specific improvements. So recently I used it for a number of very long form content pieces, think literally thousands of words long, and got really good success with it. I was trying to kind of fill a gap of like I needed something more than just, you're out of the box, here's some tone, here's some style, here's some clarity I needed to be like, hey, does this actually make sense? And found it pretty useful. So you can always kind of be thinking, I think about whether it's writing or anything else, what are the limitations you're running to into in your own tools and then create custom prompts to address those. So that's one and then two. Paul, you had posted this week about some experiments that you were doing with ChatGPT's tasks function. And like can you maybe share with us like what you were working on, what was going on around with that?
- [69:07]
Paul Raitzer
Yeah. So when tasks first came out a few weeks ago, I set up three of them. One I just thought was like to test its real time nature, I put summarize Cleveland Cavs games for me and daily at 11pm like send me a summary of what's going on. Those have gotten better. Like sometimes they'll send me a summary of a game from three days ago. But you know, interesting because I was actually trying to test like impact on sports journalism, that kind of thing. Like where maybe I don't even need to go to ESPN anymore. I can just get real time updates. I did one for mentions of my name. So anybody in the audience who remembers Google News alerts, back when we were running our agency, Mike there was we had Google News alerts set up for every client, every executive at the client companies, the board members like you just set up all these alerts. And so I thought let's see what ChatGPT can do with that. So I get a daily email with any mentions of my name basically online. But then the interesting one I wanted to experiment with was sendai new something memory. So again, the way we curate the information for the podcast each week is probably 99% my Twitter feed. So I have notifications from about 150 different brands and people. And if anything about AI is talked about throughout the day, I will see it within that feed. And I grab those links and I read them and I put them into Zoom. That's kind of how we curate what we talk about. So I thought, well, let me see if like maybe I'm missing some stuff. Let me test the new AI news summary. So this one's fine. Like honestly, it sends me like five or 10 things. I didn't go in and say, like here's examples. I didn't like build out how it should really function. I just want to see how it functions. Kind of out of the box. So relatively unimpressive. But then what day was this? This was February 2nd. So this was Sunday. So I get an alert from Chat GPT that I have news and the headline was OpenAI GPT 5 beta released. And I was like, oh well, that's news. So I click on the alert. Now keep in mind, this is from OpenAI in my inbox. So I click on it and it takes me there. And in my news summary it says OpenAI launches GPT5 beta for enterprise testing. It has started beta testing GPT5 with select enterprise customers, focusing on improved contextual understanding and multimodal capabilities. I was like, wow, I can't believe I missed that. So I click on the Read more and it takes me to an OpenAI news page that pops up with a 404 error, but in the URL is GPT5. So it's possible that OpenAI's own task thing preempted their announcement of GPT5 and maybe it's actually coming this week. Well, we will see. But yeah, so either their tasks thing doesn't work super well and does hallucinate, or they actually had a news page that announced GPT5 and it wasn't public but their task tool found it.
- [71:52]
Mike Kaput
That's wild. Yeah, I'll be honest, I need to put out a call to listeners to send me tips on using tasks because I'm sure this is just me needing to spend more time on it, but man, I've had some really like lackluster out of the gate. Like I can't get it to work in ways that are actually useful for me.
- [72:10]
Paul Raitzer
Super buggy.
- [72:11]
Mike Kaput
Yeah, that's so interesting. It'll be really funny to see if it did actually preemptive when this news comes out.
- [72:18]
Paul Raitzer
Hilarious.
- [72:21]
Mike Kaput
All right, to wrap up this week's episode, we have a few AI funding and product updates. I'm just going to run through these rapid fire. First up, Eleven Labs, an AI audio generation platform, has secured $180 million in Series C funding, which values them at 3.3 billion. They are a major player in synthetic voice technology. Their tools are used across media, gaming and tech. Their technology already powers voice features for prominent names like ESPN, Chess.com and the Atlantic. Next up, Google has rolled out Gemini 2.0 flash formally across web and mobile applications. This is the latest version of their AI model that promises faster responses and improve performance across key benchmarks, particularly for everyday tasks like brainstorming, learning and writing. As someone that's been playing around with this a little bit, I would validate that and say you should definitely try out the new model now. Alongside that, Google has also upgraded its image generation capabilities with the latest version of Imagen 3. This new iteration promises to deliver richer details and textures with improved accuracy in following user instructions for creative projects. Now, to use this, you just ask Gemini to create you an image of whatever you want and it will default to image N3 and generate your image. I will say, like, just very initial tests of this recently, like this blows Dalli out of the water for me. It's really cool. I would go recommend you test it out.
- [73:55]
Paul Raitzer
Does anybody use Dall E still? It's so bad.
- [73:58]
Mike Kaput
Yeah, it's gotten to the point where it's. I mean, we're so spoiled. But it does look dated bots at this point.
- [74:04]
Paul Raitzer
And they all look the same. Yeah.
- [74:07]
Mike Kaput
So the last update we've got this week is Meta has announced an update to its AI Assistant. The company's chatbot will now be able to access and use information from your Facebook and Instagram accounts to provide more personalized responses. There's two kind of new capabilities here. First, Meta AI can now remember details from conversations across Facebook, messenger and WhatsApp. So you can explicitly tell its AI to remember certain preferences or things you like, and it will factor these into future interactions. The more substantial change is that Meta AI will now also automatically tap into the user's broader social media activity so the assistant can access information like home locations from Facebook profiles or recently viewed Instagram videos to shape its recommendations. Paul, it has been a wild week in AI. I think we're probably getting more news this week that's going to be big and relevant, but we'll see. Appreciate you breaking it all down for us.
- [75:08]
Paul Raitzer
Yes, there's more model news coming. I think we'll have multiple launches in February for sure. Maybe multiple launches this week. We will do our best to keep up with all of them for you. I put this gif on Twitter earlier this week of the cats heads moving back and forth, back and forth like as like model releases be like. And it's just like, I think that day like Gemini 2.0 flash had come out and then O3 mini came out and then it was just like, oh my gosh, it's crazy. Well, good luck everyone. Don't get caught up in the madness. Focus on use cases that actually matter to your job and just like stick to those. Just keep nailing those and stack those. But do not get overwhelmed by the fire hose of AI model news like Mike and I have to do every Monday morning. And we appreciate you being with us and we will talk to you again next week. 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 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. Until next time, stay curious and explore AI.