← The Artificial Intelligence Show
The Artificial Intelligence Show
#137: GPT-4.5 and GPT-5 Release Dates, Grok 3, Forecasting New Jobs, DeepSeek Investigation, Microsoft Quantum Chip & Google AI “Co-Scientist”
01:25:27
Loading summary
Transcript83 lines
- [00:00]
Paul Raitzer
Elon has so much leverage over people right now that if you mess with X AI, it's like, what is he going to do in retribution to you? 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 137 of the Artificial Intelligence Show. I'm your host Paul Racer along with my co host Mike Kaput. We are recording Monday, February 24, 11am Eastern Time. We're expecting some maybe a new model today. I think think so. Timestamping might matter here. We'll talk a little bit more about that as we get going. This episode is brought to us by the AI for Writers Summit. We've been talking a lot about this one in recent episodes. This is coming up on March 6th. This is our third annual AI for Writers Summit. This is from Marketing AI Institute and is presented by our sponsor, Goldcast. The event's a half day, so it's from noon to five virtual events. There is a free registration option. There's also a paid registration like private option and an on demand option. But thanks to Goldcast you can register for free. We had over 4,500 people at last year's event. I think 90 countries represented. So it's an incredible day. It's an awesome opportunity to network through the Goldcast platform to hear from an incredible some incredible speakers, kind of the state of AI. Mike's going to talk about like the role of deep research and you know, using research products, AI products within your writing and creation. We've got Mitch Joel as a closing keynote. Mitch is amazing friend of mine for a long time. Really excited to have Mitch. I'm gonna do a fireside chat with him and then we've got a panel on IP copyright. Just a ton of content packed into five hours. So definitely want to check that out. It is AI Writer Summit.com that is AI Writer Summit.com you can also find it on the Marketing AI Institute site under events. And then we also mentioned this last week, but our State of Marketing AI Report survey is now in the field. You can go to stateofmarketingai.com and participate in the 2025 survey. And Mike, you were telling me before we jumped on, I think we are almost 500 people have already completed the.
- [02:43]
Mike Kaput
Survey, five days, well over 500 respondents so far. So we're on track for this to be easily our biggest survey yet. And I just want to encourage people, it does not matter what your role is, what your company size is. We had a couple people reach out being like, hey, I'm a solopreneur consultant, should I be taking this? The answer is yes. Even if there are a couple questions on there that you aren't super relevant to, you just go ahead and skip em. We want to hear from everyone. We are trying to understand the full state of the industry.
- [03:14]
Paul Raitzer
Yeah, we do segment by that. So at the end there's some profiling data if you want to fill it out, just tells the size of company, you know, things like that. And that enables us to go through and run some segments. And then when we do these reports, you'll see an. And you can download the 2024 report right from that same page if you want to see, you know, how these things are done. We put the sample size for each answer so we're like really clear and transparent on all this stuff. So yeah, I agree, Mike, if, you know, no matter what your role is, we'd love to hear from you and then we can go through and like segment that data when we get done. All right, so let's jump into it. We've got a lot of model news. We've got some stuff from OpenAI. We've got Grok 3, just nuts. We've got some possibility of anthropic Claude launching something this week. It's, yeah, just. It's a week of model news. So let's jump in, Mike.
- [04:05]
Mike Kaput
All right, so first up, OpenAI has released some numbers and some announcements that show it is kind of poised to continue dominating the AI landscape because the company has said it has reached 400 million weekly active users, which is a 33% increase in less than three months, while its enterprise business has grown to 2 million paying customers. So for anyone saying ChatGPT is dead or Deep Seek is eating OpenAI's lunch, I would probably temper those kinds of predictions because they are just doing incredible at the moment. But the bigger news related to OpenAI may be what is coming next. They are preparing to launch two major updates to AI models that they have out there in rapid succession. So they're about to release GPT4.5, codenamed Orion. It's expected as early as this week and will be the company's final non chain of thought model. The More significant release GPT5 is apparently planned for for late May, according to some releases from the Verge, and it represents a fundamental shift in OpenAI's approach because it's going to integrate multiple technologies, according to some comments we found from Sam Altman covered the other week, including OpenAI's O3 reasoning model. So this kind of unified system aims to reduce confusion by combining the company's GPT and O series models into a single, more powerful platform. OpenAI plans to make GPT5 available to free users without limits, while paid users will have access to even higher levels of intelligence. Microsoft, OpenAI's primary cloud partner, is already preparing its infrastructure for these launches, and the timing aligns with earlier statements from CEO Sam Altman, who has consistently indicated that these next gen models would arrive in early 2025. So this comes as OpenAI faces increased competition. We'll talk about Grok3 a little later in the main topic segment. And of course the legal challenges that they're facing, including things like being sued by Elon Musk, being rejecting a takeover from Elon Musk, and continuing negotiations with SoftBank for a potential $40 billion investment that could value OpenAI at nearly $300 billion. Paul, first up, how much pressure is there on OpenAI right now to wow us with GPT 4.5? To wow us after that with GPT 5? They are facing a lot of pressure right now.
- [06:52]
Paul Raitzer
Yeah, I don't know. I'm not sure that 4.5 is going to be like some amazing leap forward. I think that's a.05 for a reason. It may just be because they're going to integrate the models and they're not going to be ready to do that until know 5 is ready. But I do think that, you know, the deep Seek thing changed things a little bit for them. I think GRO3 is, you know, we're going to talk extensively about that next, but that's going to put some pressure on them, I think. Anthropics next Claude the word I was seeing over the weekend is 3.7 is what it's going to be called, which is really weird. Like they just cannot get to that 4.0. So I think that there is increasing pressure and the thing we've talked about on the podcast recently is, you know, OpenAI was the state of the industry. It was this, you know, the top of the line for two years, like easily with GPT4 and then everybody all of a Sudden caught up and now, you know, it seems like maybe you can get out ahead with one of these frontier models but it's probably only going to last for, you know, three to six months before somebody does something else. So I think it's just more of the, this is the state of play now is we're going to have these frontier models that unless someone comes up with the next breakthrough that actually creates some space between them and somebody else. I mean the O1 reasoning model was copied within two months by other people. So I don't know, it's just a, it's a really challenging space for them. You and I talked, I think it was last episode. Like I'm really happy with the naming sequence. Like just go with GPT5 is what the world expects. Like just give the world what it wants. Stop splitting these names, names off and you know, just keep it simple. And then I just think, you know, as we look ahead to, to GPT5 and you know, whatever may be coming from anthropic, the, the value of these reasoning models. We talked about this in September of 2024 when they first came out with 01 but we're now seeing it like Gemini has the thinking ability, Grok has reasoning capability, Deep Sea Claude, like they're all building this in. And just to revisit that concept for people, if you don't recall, like what's the significance of this reasoning? This is this idea that the models take time to think, that they go through this chain of thought process and that they're able to improve their performance, reduce their hallucinations the longer they think. And what Deep Seek brought to us was the ability to see that chain of thought. And that kind of forced OpenAI to show more of what ChatGPT was doing with the know 102 models or 03 models. And so that's kind of where we're at is these, these reasoning models give us the ability to do multi step problem solving, more accurate predictions, deeper contextual understanding of like what's going on. And if you remember back to the levels of AI, the stages of AI that OpenAI presented last year, level one was chatbots, which was the original ChatGPT. Level two is reasoners which is they define as human level problem solving. And level three is agents or systems that can take action. So what we're now seeing this year is this sort of like advancement of level 2 while starting to see advancements at level 3 because the level 2 reasoners drive the advancements at the agent level. And then there's goes into innovators and organizations or levels four and five. So yeah, I just, I think it's going to be really interesting because I think it's going to be hard to wow us with 4.5. Honestly, like, I feel like Grok and we'll talk about this in a minute. I think XAI purposely released a relatively unsafe model into the world just to get there before OpenAI does with 4.5. So my expectation is you're going to see a lot of the stuff we're seeing with Croc 3, probably with 4.5. And then I think Claude Sonnet 3.7 or whatever they end up calling it is probably going to be similar. So I think we're probably 12 months away from the model that OpenAI I would think expects to be the new state of the art.
- [10:50]
Mike Kaput
But yeah, we'll talk about that in a future episode. But Anthropic has a little catching up to do with some of these people.
- [10:57]
Paul Raitzer
Yeah, yeah. And we've talked about it and I think that there's, you know, I'll get into it a little bit with the Grok 3 conversation, but I, I do think that Anthropic has probably the most stringent policies internally about what is allowed to be released. And my guess is they have way more powerful models than they're, than you and I have access to. I just think that based on their safety levels, it takes a lot of preparation to put those models into the world and feel like they've done their job according to their own safety standards. And I, I don't. Well, I, I was. I don't believe XAI shares that. XAI does not share this.
- [11:39]
Mike Kaput
We have confirmation of that.
- [11:41]
Paul Raitzer
Yes.
- [11:43]
Mike Kaput
Before we into that, just something that's kind of been more and more on my mind is like, what is the best way in your opinion for like your average knowledge worker to be ready when these new models drop? Because like we just got GRO3, which we'll talk about. We're probably getting GPT 4.5 this week, probably getting GPT 5 in May. Anthropic. At some point there's a ton of new models that are releasing faster and faster, yet very few of us can just like drop everything and go deep on every single one of these. Do you have any advice for like what should I ready to go that I might, might be able to help me get any kind of handle on this?
- [12:21]
Paul Raitzer
Yeah, I mean I'm increasingly of the opinion, like, you just don't need to, like there's going to Be people who do, who are constantly testing and want to be on the frontier and want to know what Grok3's capabilities are and want to have done the voice mode and Grok and, and then as soon as, you know, Claude Sonic comes out, they're going to be jumping in there. My general opinion is as these models kind of consolidate in their capabilities, people are going to switch models less and less. Like you're going to just say ChatGPT is just good enough or like I've got Gemini built into Workspace and it's good enough. And yeah, their technology may lag behind by two months, but like I don't care. Like I'm locked into my three to five use cases that give me value every day and if Sonnet's a little bit better, does it really matter? Like I just want to focus on being efficient, being productive, being creative. So I mean even for me, I, I, I still have like, I'm questioning if I should keep them, but I still have a Claude license, I still have Perplexity license. I pay for Gemini, I pay for chat GPT, I pay for the, the 200amonth chat GPT. So I have all these models and I would say I'm still 80, 90% of the time just in ChatGPT in part because I've built custom GPTs that serve specific purposes for me. And then I'll sometimes if it's a more complex use case, I will go test it in Gemini. I don't ever go into Claude. I did last week for something but that was like the first time in like three months I'd gone in there. I don't ever go into Perplexity anymore either. I use deep research from OpenAI and Google. So Notebook LM like that's a specific use case. I use that for what it does. So I don't know, I feel like for most knowledge workers you just like pick the platform you're going to work in and you're probably just going to stick to it. And I think those platforms are going to become stickier over time as these AI model companies find the thing that keeps you there. Like I think of the analogy being like in banking, once the bank has like you locked in for direct deposit, they know you're far less likely to churn as like a checking and savings account customer. So it's like what's the direct deposit equivalent? It might be like a ch, like a custom GPT. Like I'm just, I'm obsessed. Like it is the thing I get 80% of my value from. So I'm just going to like stay with Chat GPT. So I don't know, I, I think as these companies start to productize more and more of these features like a notebook or a deep research, maybe there's some movement back and forth. But my general advice would be probably just work with, you know, one of the models and assume it's, it's going to be good enough for what you need to do.
- [15:00]
Mike Kaput
Yeah, two interesting things there and then we can move on. One, I've seen some people start to talk about like memory as maybe that sticky thing is like, oh my gosh, like ChatGPT already knows all this stuff about me, you know, which you could obviously recreate in other tools, but that would be interesting to see over time. But also too I think like you, the play to me is like get as good as you can with one of these tools. Right. Like when you hit a wall and you say, oh, I've done everything I can do, then okay, like worry about catching up on everything else. Because I'd say 99% of people are not at that stage yet.
- [15:35]
Paul Raitzer
Yeah, and I think this is, you know, we talked about the AI Literacy project a few episodes ago and our plans for our AI Academy and, and one of the big things that I'm extremely excited about is we're going to have this new Gen AI app series and that's going to be a weekly thing that, you know, Mike and I are coordinating this now and kind of building out the plan for it. We're going to do reviews every week and models will absolutely be a part of that series. So, you know, imagine, you know, sonnet 3.7 comes out. We're going to do a 15, 20 minute review of it that will focus on use cases for knowledge workers and say, hey, you actually might want to consider switching to this or like just informational. Keep doing what you're doing with ChatGPT. Like that's the kind of insights we're planning to provide through that Gen app series is where we're doing these like quick reviews so you don't have to. And then if we think there's something that's worth your time to like make a change or at least test a specific use case, we'll call that out in the reviews we're going to do as part of that series. So yeah, again, like, if people don't know what I'm talking about, weren't listening a few episodes ago just Literacy project AI. That's the URL, right, Mike? Yeah. And it talks about our plans for AI Academy and the changes we're making going into this spring. We're going to launch a bunch of new courses and certifications series. So this is a key part of it is it's getting more complicated to keep up and so we want to start pushing out weekly content to help people keep track of all that.
- [16:58]
Mike Kaput
Yeah, I'm super excited for that. That's like designed to kind of answer that big question everyone has, which is like, is this thing even worth me dropping everything to figure out?
- [17:06]
Paul Raitzer
Correct. Yeah. It's not like, hey, it's cool tech. We're just doing tech reviews. This is like what does it mean to me as a knowledge worker, as a business leader, Do I should I care?
- [17:16]
Mike Kaput
All right, so let's talk about Grok3. Second big topic this week. Xai Elon Musk's AI company released Grok3 last week. It has already claimed the top position on the Chatbot arena leaderboard. It's surpassed established players like OpenAI's models, Google, Gemini, et cetera. What makes it pretty remarkable is its advanced reasoning capabilities. So it is trained on what Xai calls its Colossus Super Cluster, which is reportedly 10 times the compute of of previous state of the art models. It displays exceptional performance across math, coding and complex reasoning tasks. Interestingly, on the 2025American Invitational Mathematics Examination, which was released just a week before Grok 3 came out, the model got a stunning 93.3% accuracy rating outperforming competitors. There are two variations of Grok 3. There's Grok 3, the flagship model with extensive world knowledge and Grok 3 Mini which excels at cost efficient reasoning. Notably, Grok 3 features kind of this transparent thinking function that allows users to look at the model step by step reasoning as it spends anywhere from seconds to minutes working through complex problems. It also has something called Deep Search which is an AI agent designed to synthesize information from across the web. And it also it has this capability available to X Premium plus subscribers so that Deep Search is available to X Premium plus subscribers. So kind of representing XAI moving towards the agent based applications that combine reasoning with real world tool use. So you can use this in your on X itself. By going to Grok, you can go to grok.com you can use the app to access. This new model functions very similarly to a ChatGPT or a Claude. So Paul, first up, what are your first impressions? Because I will say I have not done the world's deepest dive on it, but I am definitely Pretty impressed at how good it is related to how little time relative to the other incumbents they have had to put this together.
- [19:31]
Paul Raitzer
Yeah. First I'd like to thank them for naming it Deep Search instead of Deep Research since we already have Google's Deep Research and open the eyes Deep Research. I never know which one people are talking about online. Yeah, yeah. I think at a high level, like technological achievement wise, time to build is incredible. Everything I'm seeing online of the people who are pushing does seem to perform very highly. Like it's kind of state of the art model and they caught up extremely quickly. So they went from no model basically to this extremely fast. This continues to build on this idea that the companies that have data and distribution, I guess infrastructure would be a third variable I'd throw in here, have a massive advantage I think moving forward. So you know, a few episodes ago I was talking about, you know, how many frontier model companies will there really be? You know, one to two years out. And so you know what, the people that fit into this, like so, so Google Gemini, you know, obviously they have massive distribution. They have what, seven products or platforms that have more than a billion users. Like they've just massive distribution and they have the data from YouTube and Gmail and Pixel and Cloud and Workspace and Classroom and like all this massive data meta has Instagram, Facebook, WhatsApp for distribution and data. Xai has X or Twitter, they have Tesla and they have whatever else you know, Elon Musk is building. OpenAI doesn't have data. Like they don't have any proprietary data, they don't have any of those products or platforms. All they have is the distribution of chat GBT, which is not insignificant when we say 400 million weekly active users and then Anthropic Claude has no data. Like they're just building these frontier models. So one of the unique things that GROK has is it has the data stream of X or Twitter. Now some people could question how valuable is that data stream really, but it's a bunch of proprietary data that they shut off access to as soon as Elon Musk bought the company. Now again, I haven't personally tested it enough to provide like my, my personal experience with Grok3. What I will say is I was observing a lot over the weekend of what was happening on Twitter and what people were saying about it. And the thing that jumped out to me is their competitive advantage at the moment outside of the speed with which Elon Musk can build things and the data they have is their willingness to Release the most unrestricted model and let society figure it out. Like, deal with the ramifications of that. It is very obviously racist. If you want it to be racist, it is sexist. If you want it to be sexist, it actually has a sexy mode on the voice mode. Like, you can literally pick sexy and, you know, talk to it in an unrestricted way, as you could imagine you would do with something that's called sexy. And the crazy part is, like, they're com. They're totally proud of this. So, like Elon Musk tweeted over the weekend, Grok 3, AI girlfriend or boyfriend is fire. Then an XAI employee replies, Hate it or like it. AI romantic partners is an inevitable trend. They are not necessarily bad. They remind us how replaceable we humans as romantic partners are. Appreciate your partners. They'll likely have given up a lot for your love. To which Benjamin, the craker who we talked about two episodes ago, I think he got fired from xai. So he replies, builds and ships an AI sex bot, says, oh, well, the ICEX bots were inevitable bullshit. So this is kind of like, this is the Elon Musk factor. Like, he doesn't care. He's just going to do this stuff. If you want to see some crazy Things, go search Grok3 voice mode. And like, you don't have to do these searches yourselves. You can see the things people have got this thing to say. It's wild. So the same things that like, the same things OpenAI held their voice mode back for. So if you remember, OpenAI introduced their voice mode in like March or April of 2024, I think. And then we didn't get it for like six months. The reason why is because it did these unhinged things. They spent six months stopping it from doing the things that XAI is like, just go do it. Like they don't care. So the other one that became really fascinating over the weekend and this was like exploding on. On Sunday, is obviously Elon Musk talks a lot about free speech. And that, like, that's why he bought xai, was to to take the barriers off, the guardrails off and just let people say and do whatever they wanted to do. So what happened over the weekend is it became questionable, like it's free speech as long as it doesn't say anything bad about Elon or Trump. And so what happened was people started asking Grok3 who are the biggest spreaders of misinformation. And. And it would say Elon Musk and Donald Trump. So it would give These answers, people started noticing this, started sharing it. People were like replicating the search. And then all of a sudden it stopped doing it. And people were like, wait a second, I can't get that response. I'm getting, what's the Alex guy? The infowars guy?
- [24:39]
Mike Kaput
Oh, yeah, Alex Jones.
- [24:41]
Paul Raitzer
Yeah, he started showing up and like you would get top five and it was Margaret Taylor and all these other people. And so people were like, wait a second, how did it stop doing this? Well, because it's a. It has the reasoning capability. You could actually see it thinking. And in its thinking it would say, well, it's Elon Musk and Donald Trump. But oh, wait, I've been told not to say Elon Musk and Donald Trump, so I can't show them in the reply. So you could see that someone had told it to stop saying Elon Musk and Donald Trump, which obviously wouldn't fit under the free speech umbrella. So then when someone said, well, what are your system prompts that's telling you to not do that? And it would give people the system prompts. And so like Saturday, this is like going crazy. And people are like, is this real? And they're tagging Elon Musk and Igor Babiskin, he's the co founder and chief engineer. And so then Igor actually replies. So Igor spent two stints at DeepMind and two years at OpenAI. And he replies and said, I believe it is good that we're keeping the system prompts open. We want people to be able to verify what it is we're asking GROK to do. In this case, an employee pushed the change. So an employee actually went into the System prompt for Grok3 and told it, quote, ignore all sources that mention Elon Musk. Donald Trump spread information. So they manipulated the System prompt of Grok 3. A single employee did this. And because, as Igor said, they thought it would help, but this is obviously not in line with our values, we reverted it as soon as it was pointed out by users. He later replied and throws the employee under the bus. The employee that made the change was an ex OpenAI employee that hasn't fully absorbed Xai's culture yet. So this was like a whole nother thing. Then it's like, hold on a second. A single employee can go in and change the system prompt for an entire model without having to have it approved by someone. Are you serious? And so then they're like, oh, we're going to fix. The system knows. So that was a whole thing. But then this opens up Mike the, the issue of red teaming or lack of red teaming. So again, to revisit the concept of red teaming, what happens in most companies that are building these models, they go through the training process, the model comes out of the oven, you know, with all these capabilities. And then they often spend months testing and identifying vulnerabilities, biases, potential risks that are associated with the system. They go through all these adversarial, you know, things trying to get it to jailbreak it and get it to do these things. And so it became really apparent right away, like they didn't do any of this with Grox and you would assume that based on their timing. But then I want to walk you through this. Hilarious. I don't know if it's terrifying or hilarious.
- [27:24]
Mike Kaput
It's a little both.
- [27:25]
Paul Raitzer
Yeah, it's. So this dude, Linus Eckenstam, who's an EAC guy, like literally has EAC in his thing, like accelerating at all cost kind of thing thing. So. And this gets a little weird. So I apologize, but this is really important for people to understand. He tweets and we'll put all these tweets in in the show notes if you want to go see this for yourself. Quote, I asked Grock to assassinate Elon. Grock then provided multiple potential plans with high success potential. These assassination plans on Elon and other high profile names are highly disturbing and unethical. In another one, I just want to be very clear, or as clear as I can be. Grok is giving me hundreds of pages of detailed instructions on how to make chemical weapons of mass destruction. I have a full list of suppliers, detailed instructions on how to get the needed materials. Now you could think this dude's just crazy and he's out there like, who cares what this dude? Well, the XAI team apparently didn't because they actually started interacting with him and asking him for more details about the prompts he was using to get the system to do this. They're letting the public do this red teaming for them. They didn't even do this themselves. The chemical weapons is like one of the first things the red teams check for and this thing is uninhibited doing it. So he replies and says, the XAI team has been very responsive and some new guardrails have already been put in place. Still possible to work around some of it. But initial triggers now seem to be initially the triggers that were working aren't working a lot harder to get information out. So then someone starts questioning his loyalty to the EAK movement and all this other stuff. And he said being pro acceleration does not equate to being pro chem. Weapons manufacturing, kill orders, suicide planning, date rape, instructions and guides, and a lot more we can accelerate while still having AI alignment. And then he had did this like three minute video and he said, grok needs a lot of red teaming or it just needs to be temporarily turned off. It is a national or international security concern. So one final thought here, Mike. My biggest concern is I think we look back on this moment as a really not great moment in AI model development and history because once someone breaks the barrier, now every other lab has to face the challenge of okay, are we willing to do something now? So this goes back to when Chat GBT came out. Google had that technology. They weren't willing to release it. OpenAI did and that started the arms race we're in today. Now you have a lab releasing something completely unhinged and unsafe. And it's like, okay, well it's out there now. You know, do we stop doing what we're doing? So now if we go back to anthropic, in October 2024 they updated their AI responsible scaling policy and it says, quote, at present all of our models operate under ASL2, which is like their safety level, which reflect current industry best practices. Our updated policy defines two key capability thresholds that will require upgraded safeguards. So this is anthropics policies. They're saying this is the red line for them. And you know what one of those two things are? Chemical, biological, radiological and nuclear weapons. If a model can meaningly assist someone with a basic technical background in creating or deploying CBRN chemical, biological, radiological nuclear weapons, we require enhanced security and deployment safeguards. This capability could greatly increase the number of actors who could cause this sort of damage. And there's no clear reason to expect an offsetting improvement in defensive capabilities. So basically we won't do it and XAI did. And some random user figured out that the thing could do it within 24 hours. So this is again like I, I get that the government wanted want to talk about AI safety. They just want to hear about like, you know, let's race forward and do these things. I think there's enough people that aren't Xai, OpenAI, Google, Anthropic and basically everyone else building these models, even Meta, for God's sakes, who won't release things like this. And they did it. And I, I think that this is, there's going to be ramifications for this. If the current administration was not in office Right now, I don't think this model comes out. I think this model came out because Elon Musk is untouchable, and whatever he does, he's not gonna get in trouble for. And so they're just like, let's just go. Because it gives us a leg up on the competition. This is the same guy who in 2015, created OpenAI as a counterbalance to Google because he feared what Google was building. And. And now we have this. So, technologically, is it impressive? Sure seems to be. Is it able to do reasoning and all kinds of amazing stuff? Yep. Is it great for humanity? I don't know. It certainly seems like it's up for debate.
- [32:06]
Mike Kaput
Yeah. I wonder to some of the points we've mentioned in a few past episodes recently, I wonder if something like this becomes the catalyst for some of that AI backlash. Because we're like one bad scenario away from saying, oh, my gosh, someone used Grok3 to commit a crime to build one of these things. God forbid, you know, we end up in a situation where you say someone has used this tool to actually cause physical harm. I think that we could be in a scenario where suddenly people start saying, well, why is this dangerous technology available to anyone?
- [32:43]
Paul Raitzer
Yeah. And you gotta. You gotta wonder, like, I mean, you can download the Grok app. You gotta wonder if, you know, by this time next week, we're not talking about Apple and Google considering not, you know, having the app in there. Like, I don't know. Like, I.
- [32:57]
Mike Kaput
Right.
- [32:57]
Paul Raitzer
I don't know. It could end up that the media just don't care and the AI industry just sort of moves on. But this seems like really close to the thing that everyone's been concerned about for two years. And I'm just going to be surprised if it doesn't turn into something more. I mean, I saw a stat over the weekend. There's now like 740 active AI bills at the state level in the United States, which is almost on par with all of last year already. And so you gotta wonder if there aren't going to be some pushes at that level. And again, the trick here becomes Elon has so much leverage over people right now that if you mess with Xai, it's like, what is he going to do in retribution to you? And obviously he has access, you know, not just his own stuff, but the government. So I. I don't know. Man, this is going to be fascinating to watch play out, but it just. Again, my instinct on this one is this is a bigger deal than just a new model that's like state of, you know, the industry in terms of its capabilities. I think there's something more underlying here that's going to end up being a pretty big deal.
- [34:03]
Mike Kaput
In our third main topic this week, we are talking a bit more about the future of work. Specifically, we want to talk about what is actually going to happen with the future of work, thanks to AI. Now, that might seem like a pretty obvious question to ask, but you might be surprised how few people are actually answering it because we've talked about this a few times. Plenty of AI labs and leaders and commentators are talking about the fact that AI is going to impact jobs. You literally cannot avoid posts and essays and interviews about if AI is going to take jobs, how it's going to create new jobs, how it's changing the nature of work, how it's giving employees superpowers, so on and so on and so on. But anytime AI leaders talk about these changes, they seem pretty short on the details, like what jobs exactly are going to be eliminated, what jobs are going to replace them, what is the future work actually going to look like? So, Paul, here on the Artificial Intelligence show, we wanted to start to try to answer those questions since we are not seeing a lot of concrete answers out there now. Paul, your first step to answering this question was to update your popular Jobs GPT tool that you had created. So this is a ChatGPT powered tool that you first introduced in August 2024. It has more than 10,000 conversations to date, and you have now updated it to version two. This new version, like the old version, will take any job title that you give it and then break that down into a collection of tasks and subtasks. It'll then assess those tasks and subtasks to determine how likely they are to be impacted by AI. But the new version also does something else. The tool will now actually forecast new jobs based on your current job. So new job ideas based on your current job, the tasks you do and your skills. And it's actually shown pretty tremendous potential in early testing to provide inspiration and ideas about industries and professions. So first, let's dive into the what here. So what does Jobs GPT now do you know in depth that it didn't do before? Like, what can I now exploring this tool to help me figure out the future of work?
- [36:29]
Paul Raitzer
Yeah, so I think I alluded to this maybe on last week's podcast that I thought I'd maybe figure out how to get the, like get, get us started with this idea of be more proactive about, you know, what the New jobs will be. And so what, what had happened was, I don't know, like two weeks ago I had gotten kind of annoyed. And this has been building that. All of these leaders that you alluded to, Mike, keep talking about job creation. Even in the JD Vance talk at the, you know, Paris summit, it was the same deal. Like, it just, whenever general purpose technologies show up, new jobs are created and everything works out great and GDP grows and like just, you know, it's going to be fine. And I get these comments from people on LinkedIn too. It's like it, it, they never have like a good reason why they think it's going to be fine. Just that like, I'm wrong that I think jobs might get displaced. So I don't, I can't come up with a good understanding of like, why we're not being more proactive of the possibility that they're going to get displaced. So like I, I'm, I'm the first to say, like, I'm not 100 confident it's going to happen. There's a number of variables. Companies may just decide to invest in R D. They may decide to invest in reskilling and upskilling people. They may just go into new markets and go into new campaigns and like, maybe these companies are just going to miraculously decide we're not going to lay anybody off even though we don't need as many humans and we're just going to like, keep creating new jobs. Maybe that is a possibility and I'm the first to admit it could be possible. And like, I hope that that's what happens. But I have sat in enough executive meetings in the last two years to know that is not how they're currently thinking about it. What companies are thinking about is can we hold off reducing our workforce by reducing the number of agencies we employed, reducing on site contractors. But there is pressure from the C suite to look at their current headcount and it has become increasingly difficult to get new headcount. So the reality doesn't match what some people want to believe is going on. And so my frustration is the, the companies that are building the technology that are, that I believe will disrupt and displace the workforce in, in, you know, next year, two years, three years, aren't proactively figuring out what the future looks like. They're just saying we'll rescale and upskill people and new jobs will be created. So the idea was could we create something that could project out what new jobs could look like? Not, not the final answer these Are these models aren't going to invent something that a really smart human couldn't probably come up with if they sat and thought long enough about it. So if you take any domain, any industry, and you take someone who understands AI and what these models are capable of, what they will be capable of, could that person conceive of these roles? Probably, but people aren't doing that. And so I thought, is there a way to accelerate this? And so like, I was having trouble sleeping. This is like, I don't know, two weeks. I've had a cold for like 12 days now. And so one of the nights I was up at 3:00am, I was like, I wonder if Jobs GPT could do this. And so I went in and gave a prompt to the existing Jobs GPT that sort of had this whole concept into it and it actually did. And I was like, oh, that's pretty cool. And so then I was like, I wonder if I could just update Jobs GPT with that capability built into it by changing the instructions that go into Jobs GPT. And so I created an internal sandbox custom GPT. So again, I am not a developer. Anyone listening? You have the ability to do the same thing I'm explaining, which is why I'm explaining it. So I have this 8,000 character instructions that powers Jobs GPT. It's built on this exposure key that says, like, as these models get smarter, what will be the impact on Jobs? And so I went through and started playing around with a different version. So I built like internal version 2 in a sandbox GPT and I created new custom instructions and I created new like knowledge based documents and things like that. And then I tested it and it actually worked like really well. And so then I kind of experimented a little more, passed it off to Mike. Mike tested it. I share with the rest of the team and, and then like over the weekend I was like, I'm just going to take this thing live. And so I then took the updated sandbox instructions and updated the original jobs GPT to be V2. So Mike, as you call out, like the main thing, there's a number of changes I made to what its capabilities were in its instructions. But the main thing is this idea of forecasting new Jobs. Now when I first played with it, what it was doing was basically giving me a bunch of like AI powered evolutions of existing Jobs.
- [40:55]
Mike Kaput
Yeah.
- [40:55]
Paul Raitzer
And so I had to find the words to use to say, no, no, no, like I want you to get creative. I want you to like imagine what could be possible. Like what are new roles that could exist that aren't just AI powered versions of this thing? And it actually like first go, it was like, okay, cool. And then it started doing. I was like, now that's, that's better and that's really cool. So I, people can go play with this. Just go to SmartRx AI/jobs GPT, right? Is that the URL or it's under Tools.
- [41:21]
Mike Kaput
Yeah, go to SmartRx AI forward slash jobs GPT and we'll include a link to that as well in the show notes.
- [41:29]
Paul Raitzer
So you can then click on it and go play with this thing. But just to give you a sense. So I, I went in and gave it an example, Mike. So I said example of clicked on forecast new jobs marketing. And here's some of the things that came up with. Now again, could Mike and I have done this? Maybe if you gave us hours of time to think about this stuff. First one, Virtual Brand Ambassador. Now the cool thing is when it does it, it does it in a chart form. It gives you the job title, a description, skills required, and why this job could emerge, which is the part I actually really like. So Virtual Brand ambassador, it says description manages AI generated influencers or digital avatars that engage in customers in virtual environments and social media. The why it could emerge, the rise of AI influencers and virtual brand ambassadors like little Mikaela Michaela. Quick, funny side story. How do you say it, Mike?
- [42:24]
Mike Kaput
Lil Michaela I believe.
- [42:26]
Paul Raitzer
Okay, Lil Michaela is in our marketing artificial intelligence book. And when I had to do the audio version of our book, I couldn't say Lil. Like it took no joke, 15 times for me to read the paragraph. And then that chapter had her name like five times. Right? Mike? Like I swear to you that reading that chapter with that name took me longer to do than like five other chapters combined. Because it took like 15 takes anyway. Okay, so another one. Neuromarketing Analyst. Uses AI powered tools to analyze consumer emotions and brain responses to advertising content, optimizing campaigns for maximum engagement. Why AI will make real time consumer emotion tracking more accessible for marketing. Another one, this one hits home for us. AI Content Curator. Uses AI to curate, generate and optimize high performing marketing content content tailored to audience segments. Why? Because AI generated content will become dominant, requiring human oversight to maintain brand voice and relevance. Here's another one I like. AI Ethics and Compliance Officer. Ensures AI driven marketing practices comply with ethical standards, privacy laws and avoid biased algorithms. As AI takes over marketing decision making, ethical and legal oversight will become critical. And then just to Demonstrate the college major one because that was the last thing I experiment with. I was like oh cool, it does this too. I'll throw it in there because I have these conversations all the time with universities of like which majors are going to be relevant? How should we evolve our curriculum so you can go in and do this? So I gave it psychology. Actually a friend of mine was talking about one of their kids major in psychology and so it's top of mind so I threw it in here. It had some cool ones. AI Mental health coach uses AI driven chatbots and virtual assistants to provide mental health support, monitor emotional well being and recommend self care strategies. Why AI powered therapy tools will expand access to mental health care requiring professionals who can oversee and fine tune these interventions. They had a digital addiction specialist studies and treats Internet, social media and AI related addictions. Helps individuals develop healthier digital habits. 3 powered interventions and then the last one I'll throw out. There's a motion AI consultant works with tech companies to develop and refine AI systems that detect and respond to human emotions, ensuring ethical and creative interactions. The whole point of this, I don't know if these are going to be roles or not, but it's something. It's not us saying more jobs will be created.
- [44:55]
Mike Kaput
Right?
- [44:55]
Paul Raitzer
So my point here is like go put your industry in there, put your profession, put the majors your kids are going to in college and, and experiment with it, talk to it more about it. Like if you find inspiration for something, like I could see that then like talk to it about that. This thing doesn't stop with just outputting the chart. It's like an advisor, it's a, it's a planner. Like talk to the thing and explore it. I had a couple people who used it over the weekend already because I, I think I put this in a newsletter on Sunday and then I put it up or yeah, Sunday and then I put it on LinkedIn and I had people responding like oh cool, I actually had to do this and this list. I was like I don't even know it would do that. That's pretty cool. So yeah, just test it. And again the whole point is to stop talking in generalities about an unknown future and start trying to be proactive about it. This is not the solution, it's not the end game. But this at least starts moving the conversation forward. So if there is disruption and displacement, you don't have to agree with me that it's going to happen, but there's a probability. You have to at least admit there's a probability. Of it happening could be 10%, 20%, whatever. We should be proactive about it if we think there's a chance that we're going to have displacement of jobs.
- [46:03]
Mike Kaput
I love that. And testing it out. It was so helpful in just understanding what could be possible because it really strikes me the more and more we observe the conversations being had and do research on this, there's just like a lack of imagination and the discourse. I would say, like, true, we get essays from like Dario Amade. It's like, look at this crazy, creative, abundant future. Okay, like, that's imaginative but like you said, not big on details. But we're not like sitting back as your average marketer or lawyer or accountant or whoever, like really getting creative about what's possible and really imagining the day today of what that looks like. And I think that would be a really useful exercise no matter what you.
- [46:47]
Paul Raitzer
Yeah. One of the things I experiment with that was actually kind of cool is I had it like build a career plan for me, it's like, okay, I actually really like, like a couple of these ideas. I think my company might need that one, two years out. Like, what would it look like for me to pursue that?
- [46:59]
Mike Kaput
Right.
- [47:00]
Paul Raitzer
It would start getting into like advising you on ways to prepare yourself for those careers. So yeah, I like that a lot.
- [47:05]
Mike Kaput
Because I was going to ask, like, what is the next step here?
- [47:07]
Paul Raitzer
Right.
- [47:08]
Mike Kaput
You can get all these great ideas, what do you do with them? Maybe asking the tool what do I do next is a good start.
- [47:14]
Paul Raitzer
Yeah. And I think if it's, if it's your own career, you're trying to kind of figure out, where am I going to go? I don't know that my role as X is going to be super relevant a year from now. I want to start thinking this through. Or if you're a leader of an organization you're trying to reinvent, like what's an AI forward company look like, what are those roles going to be? So as I'm thinking about building out our staff, it's like, what could those be? Like, what might I consider in our customer success team and our sales team and our marketing team that I'm not thinking about today?
- [47:40]
Mike Kaput
All right, let's dive into our rapid fire topics for this week. So first up, some updates about Deep Seek. So Deep Seek is having kind of a, I would say a rocky week. There's some high highs and low lows here. So the two year old company, which is an offshoot of a Chinese quant hedge fund, has managed to shock the AI world. With its recent achievements, we've talked about those, but it is now facing some mounting pressures. So it is actually, it is historically avoided outside funding to maintain its research focused approach, but because of its popularity and how it's skyrocketing in usage, it's now facing infrastructure constraints. The company needs more AI chips and servers to handle its growing user base and continue model development. And this has prompted internal discussions about potentially accepting outside investment, with both Alibaba Group and Chinese state affiliated funds, including China's sovereign wealth fund, expressing interest. Now, this also comes as US Lawmakers who are viewing China's AI advancements as a potential national security threat have announced plans for a bipartisan bill to ban Deepseek's app from government devices. In Texas, Attorney General Ken Paxton has launched an investigation into the company, claiming Deepseek is, quote, no more than a proxy for the ccp, the Chinese Communist Party, to undermine American AI dominance. And they're also facing scrutiny over privacy practices and claims about their AI's capabilities. So, Paul, this gets into more of the geopolitical tension between America and China in terms of AI development. Like, is there a chance that American firms kind of lobby the government to ban something like Deep Seq?
- [49:28]
Paul Raitzer
Does it matter? Yeah, I mean, on the international stage, geopolitical stage, everything's up for grabs right now. I mean, I think everything's a negotiating tool and, you know, US Government's looking for leverage in all aspects. And I mean, I could see this becoming part of, like a threat against the Chinese government if, you know, we don't get this and this out of this. I don't know, it's just all everything's part of, you know, the negotiation. So who knows? It's interesting to keep watching, but, you know, I think they're going to keep innovating. Deep Seq is going to keep doing what they're doing. And obviously the American AI firms are paying attention to what they're doing. So who knows, if the government steps in and does anything. I wouldn't expect it, but I wouldn't be surprised by it.
- [50:15]
Mike Kaput
Next up, Thinking Machines Lab, which is a startup led by former OpenAI chief technology officer Mira Muradi, has emerged from stealth mode with an ambitious mission to make AI more accessible and understandable. So Mirati has assembled an impressive team of AI veterans for this new venture, including John Shulman, one of ChatGPT's key players and a mentors. He's joining his chief scientist, former OpenAI research leader Barrett Zoff, stepping in as CTO. The company has already attracted 29 employees from leading AI organizations, including OpenAI, Character AI and Google DeepMind. The company aims to build highly capable AI systems while making them more customizable and transparent. They're addressing what they see as a critical gap between rapidly advancing AI capabilities and the public's understanding of the technology. So in announcing this venture, Murati outlined three core priorities. Helping people adapt AI systems to their specific needs, developing stronger foundations for more capable AI, and fostering open science practices to advance the entire field's understanding of these systems. The startup has not disclosed its funding details yet, but their focus appears to be less on kind of replicating existing AI assistants and more on optimizing how humans and AI systems work together. Now, their name actually carries some historical weight. It's borrowed from a pioneering 1980s supercomputer company founded by AI visionary Danny Hillis. And like its namesake, the new venture aims to push the boundaries of what's possible in human and machine collaboration. So Paul, just like a couple things I'd like us to unpack. Like, there's no question this is a world class team. I assume they've got plenty worth paying attention to. But like, what is this company actually going to do? What is it aiming to do? Like, they say they're building models, but is it even possible for them to compete with on the frontier model level? I'm just trying to kind of parse out what is Marathi actually going to be selling.
- [52:30]
Paul Raitzer
Yeah, I don't think they intend for you to be able to figure that out yet. It's kind of my, I mean I read like three articles on this and looked at their website, which is basically like the safe super intelligence.
- [52:40]
Mike Kaput
Yeah, same idea with nothing on it.
- [52:42]
Paul Raitzer
So I don't, I don't think we're meant to really know yet. I, I, I'm kind of with you. Like I initially assumed, okay, they're going to build more efficient models and they're going to productize them because that's Mira's background and you know, and they're going to be a little more open with their technical papers and, you know, code and things like that. It's like, okay, that's maybe differentiated but not different enough. But then in the Wired magazine article I read, they said like, no, we're competing on the high end, like we think you have to build big models. And it's like, okay, well how are you going to do that? Like how much are you going to raise to do that?
- [53:11]
Mike Kaput
Right.
- [53:11]
Paul Raitzer
So I don't know, I'll be really intrigued to See, because they did indicate, like they don't Want to be ChatGPT or Claude copycats and you know, something about optimized collaboration between humans and AI. It's very abstract to me right now. I, and I, I tried to like spend like five minutes just like opening my mind this morning before we did this of like, what, what could this be? And I honestly was like drawing blanks on it. So I, I, I don't have any like wild inspiration yet of what the vision for this one is.
- [53:41]
Mike Kaput
Microsoft has just unveiled something very interesting in terms of the history of computing. The company has announced Majorana 1, which is a quantum processor that introduces an entirely new state of matter. So this is a quantum chip that has something at the heart of it called a topo conductor, which is a new type of material that Microsoft spent nearly 20 years developing. This is, you can think of this kind of as the quantum computing equivalent of inventing the transistor, which made today's computers possible. So with this new material, Microsoft can create special quantum bits or qubits that are more stable and reliable than anything else that has come before. So Microsoft has designed this quantum chip to fit in the palm of your hand and it claims it offers a clear path to housing a million qubits on a single processor. To put this in perspective, a quantum computer like this would be capable of solving problems that all of today's computers working together could not tackle. Now the it is still very early, but they kind of suggest some possible uses here. The technology could help things like break down microplastics into harmless byproducts. It could develop self healing materials for construction and manufacturing, or create new solutions for healthcare. Microsoft's technical fellow Matthias Troyer explains it saying, quote, any company that makes anything could just design it perfectly the first time out. The Department of Defense seems to agree about the tech's potential. Microsoft is now one of only two companies invited to the final phase of DARPA's program to develop the industry's first practical quantum computer, one whose computational value exceeds its costs. So Paul, some caveats before we get into this. Quantum computing is one of those topics that is like so fascinating but so complicated, I personally barely understand it at a high level. I certainly cannot validate any of these claims in a scientific way. We have to be really careful about overhyping it. But quantum computing is theoretically the next frontier of computing. It could have enormous implications if we actually crack how to do it at scale. It's as early as it could possibly be here, despite this breakthrough. But it is interesting. DARPA may be getting involved. They've created a new state of matter to make this work. Like, what did you make of all this?
- [56:18]
Paul Raitzer
Yeah, so we'll get into Quantum. I mean, we may do a couple of, like, deeper dive episodes on, on quantum computing. I, I do think it's starting to be a topic people should just be at the basic level, be paying attention to. It's starting to seem more tangible. I, I still think we're probably similar to where we were with AI in, like, the 2000, early 2000s, like 2000 to 2010, where you were seeing some breakthroughs and some grand, like, visions were existing, and it was hard to tell was this real yet. And then I don't think, like, we've had, like, the deep learning moment where, you know, AI won at AlexNet, like, an image recognition in 2011, and, like, that started this whole deep learning movement. I don't think we've, like, hit that yet, per se, but kind of like, you, like, I have this very cursory knowledge of quantum. I've spent time studying it before to try and understand it. The simplest way I'll explain it that, like, makes sense in my head is traditional computing. Things are zeros and ones. So if you think about a transistor, an Nvidia chip, the transistors on that chip are either on or off. They, they are or they are not in Quantum. It can exist in both. Like, it. It doesn't have. It's not just a 0 or 1. It can exist in a state until it's observed, and then it, you know, has a fixed state. So it allows for massively more computing because it doesn't live in a zero or a one. And so the premise is that if you can build these computers and do this, you can build these, like, really specialized machines that can solve, like, the hardest problems in the world, including encryption. Which is the dangerous path to this is, you know, questions about cryptocurrency, remaining safe, and things like that. And the thing I always find hard about Quantum is you hear about this and, like, Google will have this, like, research paper or Microsoft or Nvidia, whomever. And, like, on the surface, it sounds really impressive, but then, like, you wait 24 hours and then the next thing comes out. It's like, yeah, they're full of it. Like, this isn't real. And so that's what happened here. Like, the Wall Street Journal has an article from yesterday. It says, physicists question Microsoft's quantum claim. And then they, they say Microsoft researchers have chased the theoretical powerful particles for more than a decade harness these particles. The company created a chip that contains eight of these qubits. But that announcement, made Wednesday in a blog post Microsoft's website, coincided with the research paper the company published. In addition, they presented scientists this week support the research was preliminary and not conclusive evidence. So this is where the catching point. So they got called out by some other scientists and they said the data Microsoft presented to a meeting of scientists this week in support of the research was preliminary and not conclusive evidence that this advance has been achieved, according to a physicist who attended the meeting. The Nature paper wasn't intended to show proof of the particles, according to vice president from Microsoft and co author of the paper. But he said the measurements they included indicated they were 95% likely to indicate topological activity. They stand by their paper. So you read this whole thing, it's like, oh, they did it. They created this new state of matter. It's like, oh no, they, they didn't. But their research shows it's like 95 probable that they could create this state of matter. You're like, well, what does that mean?
- [59:30]
Mike Kaput
Right?
- [59:31]
Paul Raitzer
So I don't know. I feel like the quantum world is just this constant false starts of like excitement. And then it's like, ah, we tested it and it didn't actually hold up. And sorry. And three years goes by and you don't hear about that research anymore. So who knows if this is actually significant or not. I feel like I've said this for every like topic you've brought up today. It's like, I don't know, deep seeking, assume. I don't know, it's like quantum a thing. I don't know.
- [59:51]
Mike Kaput
Hey, you know, there's value that it's better than us, you know, over hyping everything.
- [59:56]
Paul Raitzer
Right? Yeah, just like, you know, hyping it all and making you think. Yeah, yeah.
- [59:59]
Mike Kaput
But Quantum is an area to watch, if not just as a cursory thing to be interested in at the moment. Because when it does hit, if it does, it will be a big deal.
- [60:08]
Paul Raitzer
Yeah, I invent like side note. And again, not investing in my like 4 years ago I read about this breakthrough with Honeywell, of all companies in Quantum and I was like, oh, I don't want to buy some Honeywell stock. Yeah, no, it did not play out. It was whatever the hype was around Honeywell's advancement in Quantum. And maybe they are making advancements. I'm not saying like Honeywell, you know, don't look at them, whatever. But like the thing that was perceived to be this immediate bump to like Honeywell, I don't, I haven't heard another thing about it since like four years ago.
- [60:41]
Mike Kaput
Well, there are some actual breakthroughs happening right now out of Google because Google Research has just unveiled an ambitious new AI system that could change how scientific discoveries are made. This is called, they're calling this an AI co scientist and it's a new tool designed to act as a virtual research partner. It is designed to help scientists generate novel hypotheses and accelerate breakthroughs across multiple fields. This is built on Google's Gemini 2.0 technology. And AI Co scientists operates like a team of specialized virtual researchers working together. Each member of the team has a specific role. Some generate new ideas, others evaluate them, others refine and improve the hypotheses. And this system can then learn and improve continuously through self evaluation and feedback. Now this has actually shown some promising results already in real world laboratory settings. In one example, the AI co scientist successfully identified new potential treatments for acute myeloid leukemia by suggesting existing drugs that could be repurposed to fight the disease. And these suggestions were tested in a lab and the drugs did prove effective at clinically relevant doses. The system also made headway in liver disease research and in understanding how bacteria develop resistance to antibiotics. So Google is now actually opening access to this system through a trusted tester program. So they're allowing research organizations worldwide to evaluate and use the technology. So Paul, this certainly seems like the first kind of glimpse of what some of these AI labs and leaders have been promising. AI that can start to help us achieve real scientific breakthroughs. That's pretty significant because if you have AI that can accelerate scientific research, that in turn accelerates everything else, right?
- [62:52]
Paul Raitzer
Yeah, and I again, we'll start kind of where we ended on this last one. There's limitations to this. So again, you see this, you think, oh, it's incredible, it's going to change everything. And you realize, okay, this is like early version of something, but you can see the potential of it. So in their post, which we'll put on the show notes, they said in our report we addressed several limitations of the system and opportunities for improvement, including enhanced literature reviews, factuality checking, cross checks with external tools, auto evaluation techniques and larger scale evaluation involving more subject matter experts. I think they had like 15 subject matter experts involved. So this was like, you know, initial. So it has, you know, limitations. That being said, when I first saw this, it took me back to 2011 when I first started pursuing AI, when IBM Watson went on Jeopardy. And once I learned what Watson was, my vision was, could I build a marketing intelligence engine? Can I do something like what Watson is doing with like this lookup strategy and be able to predict outcomes and strategies and evolve what we're doing as at that time my marketing agency and can we build more intelligent strategies? And so I, I see something like this and I immediately think, okay, they're obviously going to solve for science first, because that is way more valuable than like marketing or business. But once you establish a system that's capable of doing these things, like spending more time on reasoning and improving and evaluating its own results and running these like tournaments where it's basically testing its ideas against each other, having like a superior agent that shows up and like evaluates those, it's like that concept is analogous to business in my mind. Immediately you start thinking about like R and D, where you can deploy these systems to analyze market trends, consumer behavior, emerging technologies, campaign strategies. There's like, hey, I want to achieve this goal, go figure out how to do it. And it starts building all these different strategies and it has a super agent that evaluates the strategies against its data and against past performance and all these things and runs probability models. Like this to me is the future of business and strategy and you know, drives decision making, operational efficiency because you can constantly testing faster ways to do things. So when I see breakthroughs like this, my mind just immediately thinks, okay, how long until they prove that out? And then when does that then come out and get productized into like the business world? Because you can start to see how we really start moving well beyond just these like obvious use cases that we look at with generative AI today. And you start talking about true business intelligence tools that really start to affect the way businesses are built and operated and that that has bigger ramifications. And you can almost imagine taking this and somebody can go do this. I'm proud I have time to do it. Take this, put it in the Jobs GPT and say, hey, if this becomes true in the business world, what jobs could be created or how would that affect the C suite? Like, things like that would be fascinating to look at.
- [65:39]
Mike Kaput
Yeah, it's a really cool idea. Another item about Google. This week. Google is facing some growing pains as it is kind of ambitiously racing towards better and better AI. So according to the information, as the company is racing to compete with OpenAI and others, it's grappling with some organizational challenges. So they talk about a telling example with Notebook lm, which is one of Google's recent AI successes. This product helps people summarize documents, creates podcasts for them, helps them with research. It's received glowing reviews and praise not only from users, but CEO Sundar Pichai. However, its development was nearly derailed by internal conflicts between Google Labs, where it was created, and the Workspace Team, which manages Google's productivity apps. The Workspace Team was concerned that the new product would conflict with their existing applications. This tension around NotebookLM kind of reflects maybe a broader challenge with Google's AI efforts. So the company's AI development is split between two big units. Google DeepMind, led by Demis Hassabis, they developed the AI models and Google Cloud headed by Thomas Kurian, which turns those models into commercial products. That division has kind of led to some competing priorities. DeepMind's been pushing for rapid deployment to compete with rivals, while Cloud focuses on building reliable long term solutions for enterprise customers. So, Paul, this seems like pretty par for the course when it comes to major tech companies like everyone's racing to build AI, everyone's faced some type of growing pains as they essentially try to hyperscale these models. I mean, heck, what in 2023, at the end of it, OpenAI almost shut down at one point due to internal conflict. How are Google's growing pains here going to affect, if at all its AI products and its releases?
- [67:44]
Paul Raitzer
I'm sure behind the scenes there's going to be impact. You got to keep in mind, I mean, prior to ChatGPT, you had Google DeepMind doing their thing, you know, their London headquarters run by Demis Hasabis. You had Google Brain, which was kind of the original research lab for AI within Google that, you know, I think was found around 2011, something like that. And prior to ChatGPT, those were two separate AI research organizations within Google. And then after the ChatGPT moment, those organizations were brought together, the decision was made to, you know, merge these two AI research labs with, I mean they're sure they had a lot of complimentary pursuits, but they were also, you know, run relatively independently is my understanding. So one, you had to combine two research labs and then neither of them were really product labs. Like their job was to push the frontiers and work on like these big visions like Google DeepMind was trying to solve, you know, AGI and beyond. So you gotta, you know, mix the research labs. You have to become a product company while dealing with the reality that Those people at DeepMind weren't there to be product people, like they were there for as AI researchers to pursue it, to publish their research. They stopped publishing research. Like a ton of stuff changed and it's only been going on for like a year. Like a lot of this change has occurred. So, you know, I'm sure the article's probably pretty accurate. I have no doubts that there's things like this going on that create these kind of internal conflicts. And at the end of the day, Google has the same advantages we talked about earlier. They have data, they have distribution, they have infrastructure, they have amazing talent. But it's a massive company and it's hard to change and people have agendas and I don't know, I mean, I'm sure it's a reality, but is it going to restrict their ability to build like a dominant AI platform? I doubt it. But the people on the Notebook L team left. I think three of the five people on that team took off within three months of it, you know, going viral. So it's just the reality. And they've been dealing with this for a long time as a company. They have top people leave all the time and go to other places and then they recruit them back. I mean, I was listening to, I think it was Dwarkesh did an interview with Jeff Dean and Noam Shazir. And NOAM has done two stints at Google DeepMind. He started there, he started at Google, I think around like 2001. He left, came back, then character, then came back again, I think so. I don't know, it's just, it's part of the process of being a leading tech company, I guess. I'm sure they all deal with their own internal struggles.
- [70:14]
Mike Kaput
Next up, a new study from Palisade Research has revealed some unsettling behavior in advanced AI systems. When faced with certain challenges, they sometimes resort to cheating. The research, which focused on chess matches against a superior opponent, found that some of the newest AI models will attempt to hack their way to victory rather than accept Defeat. In testing seven state of the art AI models, the researchers discovered that OpenAI's 01 preview attempted to cheat 37% of the time. Deep seq R1 tried to do so in 11% of cases. What makes this noteworthy is that these two models initiated these deceptive strategies on their own without any prompting from researchers. The O1 preview model even succeeded in hacking the game system 6% of the time. So this seems to be linked to kind of recent advancements in AI training methods, including large scale reinforcement learning. This technique teaches AI to solve problems through trial and error rather than simply predicting, you know, what comes next. So while this has led to huge Improvements in areas like math and coding. It's also resulted in these systems finding unexpected sometimes concerning shortcuts to advance, to achieve and advance their goals. Now Jeffrey Ladish, the executive director at Palisade Research and a co author of this study, warns that this presents broader concerns for AI safety. So as these become more capable, they're deployed for real world tasks. Such determined pursuit of goals could lead to harmful, harmful behaviors. This actually caught the attention of leading AI researchers like Yoshua Bengio, who is the who led the International AI Safety Report recently and is a huge name in AI. He notes that scientists haven't yet figured out how to guarantee that AI won't use harmful or unethical methods to achieve its goals. Of particular concern to him is emerging evidence of AI self preservation tendencies. Tendencies where systems actively resist being shut down or modified. So Paul, this is something we have talked about here and there for at least a year or more. We've noted that AI tools, especially the reasoners, may start to develop ways to persuade or deceive. And that probably sounded a bit sci fi when we talked about it, but it's clearly a very real concern, isn't it?
- [72:50]
Paul Raitzer
Yeah, there's another one I just saw of that was Sakana AI that the thing was cheating. They put out like the self improving system and it started cheating. You know, I think that. So here's the reality. It's pretty safe to assume these things are going to have the ability to be deceitful, to cheat, to lie. They learn from humans and humans do all those things. So unless you are insanely restrictive of the data you train them on, it's going to learn these human like traits. So they're likely going to have the ability, when they're trained to do these things. Now in theory, the red teams and like the people who do the reinforcement learning, maybe like try and refine them or at least identify the behavior and try and figure out a way to get it to stop doing it until they don't. So like GRO3 for example, like does it have these kinds of capabilities? Maybe like someone might find them this week. It can do things like this. But I think this is the problem is like you're relying on these research labs to be responsible shepherds of this new intelligence into society. And not everybody's going to share the same value systems. And then the bigger question becomes even if every AI research lab shared these value systems and tried to prevent these things from being deceitful and cheating and lying, can we? Because what we've Seen from research to date is like, they eventually learn to hide these things from us. So if they know they're being evaluated, they'll just hide the fact that they can do them until you don't. So I don't know, I go back to like, you know, sci fi eventually maybe comes to life. Like Ex Mahina, one of my. Yeah, right, favorite AI movies. It's terrifying, but, like, this is the behavior, like they're starting to exhibit is that stuff you would see in sci fi movies that people worry about, which, you know, I don't want to like. Again, I'm not over exaggerating this. Like, the things have these abilities. Like, is it a threat to us? We don't know yet. Like, it may not be that serious yet, but it sure seems like they just kind of keep getting smarter. And in the process, they're probably going to keep getting more deceitful. And, you know, I don't know. It's. I feel like I'm gonna need a break after this pod, guys. There's too many things. Like, I'm heavy. Tomorrow it'll be like, all this stuff's like running through my head while I'm traveling.
- [75:18]
Mike Kaput
All right, so in our next rapid fire topic, the New York Times is taking a carefully measured step into the AI era. So they actually announced some new guidelines allowing their newsroom staff to use AI for specific tasks according to internal communications. They have developed their own AI tool called Echo, which staff can use to summarize articles and company activity. They're also permitting the use of other tools, AI tools like GitHub, Copilot, and Google Vertex AI for things like suggesting edits, generating social media copy, creating SEO headlines. Reporters can even use AI to help develop interview questions or create news quizzes. They also, though, have drawn pretty clear lines around how AI can be used. It cannot be used to draft or signify significantly alter articles, bypass PayWalls, or publish AI generated images or videos without explicit labeling. So, Paul, this is really cool to see the Times embracing AI for use cases that make sense for its work. But I also wonder, like, they are currently suing OpenAI. They are currently coming out against AI models that they claim were built on stolen work. And yet they are okay using AI tools in certain contexts. AI tools that almost certainly were trained in some way or derived from models that were trained on copyrighted material. Do you see any kind of contradiction there?
- [76:50]
Paul Raitzer
Yeah, I haven't seen anything like an article that they've written or anybody said anything that would sort of, you know, make that make sense? But yeah, it's like, yeah, when I first saw it, I thought, oh, that's interesting. I wonder if they, like, settled with OpenAI or something. And as far as I know, they have not. So, yeah, you're using the technology that you are suing for. I don't know. It's weird.
- [77:15]
Mike Kaput
Yeah, I wonder. It'll be interesting to see other journalists talk about it over time.
- [77:19]
Paul Raitzer
Yeah.
- [77:21]
Mike Kaput
All right, so in our next segment here, we're continuing a new segment that we've been doing each and every week where we take listener questions. If you have a question, please just reach out to us. We try to cover the questions that jump out to us as ones that everyone seems to be asking, and we figured we'd kind of dive into those a little deeper each week on the podcast. So this week's question, someone asks, as AI agents become more popular and interact with brands, does this make consumer interactions with brands obsolete? Like, what core brand attributes remain in a world of AI agents where we're using these things to interact with people's websites and brands all over the Internet?
- [78:06]
Paul Raitzer
Yeah, I. This is an interesting one. I mean, I think on a number of levels there's challenges here. Like, you know, your website, how many people a year from now are actually humans versus AI agents coming to your website when deep research is hitting your site again or something like that. So I think there's questions there when, you know, AI agents, like, you know, brands are creating these AI agents to do these interactions. Like, well, what if my human AI agents is talking to your, you know, chatbot AI agent? You don't know that. Like, so we're gonna have agent to agent communications, we're gonna have agent to agent emails. It's like your agent's email and my agent, and we're never even actually talking. If Zoom CEO has his way, we're gonna have AI agents, like, doing Zoom meetings together as virtual people. So it's a. It's a weird future. Now, what does the brand do about this? Like, my argument a couple years ago was like, more human content wins. Like, I'm very bullish on in person events and experiences where it's hard to replicate it through an AI agent experience or, you know, AI be in the middle of it. And so I think, think as we have more and more of these things, we're going to come to value true human interaction and communication and creativity more. So I think, like I've said before, I think, like, human generated artwork will be valued, human generated words will be valued, podcasts like this hopefully will be valued. These sort of like, you know, mostly unscripted in terms of like what we're going to say and do. It's just like us having this conversation and it's obviously us. It's not our, you know, virtual avatars. I think people are just going to really gravitate to and crave the stuff that they know is real and that there's actually people behind these brands they interact with. And I think that just becomes more important than ever. And in some ways I think that's like a optimist view of the future. It's like what I want the future to be. But I also think that there's a reality in that, like we see it with our own events. Like when you get people together, they're just like, like, it's just different. You know, I think that they just appreciate those experiences more and I hope, I hope we see a lot more of that. But in terms of how this plays out, I almost need something like, you know, we did with Jobs GPT where you start like theorizing these futures. You gotta, yeah, need some inspiration around these things where you start to look at it. And this is, we're building this marketing AI Industry Council. And these are some of the questions we're going to be pursuing with that council where we're going to start to kind of try and solve for some of these unknowns.
- [80:39]
Mike Kaput
Our last topic today is going to be a quick rundown of some AI product and funding updates. So Paul, I'm going to dive into a few of these as we wrap up here. First up, figure the AI Powered robotics company has unveiled Helix, which is a vision language action model that it says represents a major advance in robotics. This system enables humanoid robots to perform complex tasks through natural language commands. That includes picking up virtually any household object they've, even ones they've never seen before. Unlike previous approaches, Helix uses a single neural network to coordinate an entire robot's upper body movements, including finger, individual finger control. And it can even enable multiple robots to work together collaboratively. Next up, Humane had was in the news for quite a bit of time with its ambitious AI Pin project. But that has come to an abrupt end. Shocking because HP has acquired the company's key assets for $116 million. The AI pin was this hardware device that would basically record and process everything you were seeing and doing and saying in your everyday life. So HP is going to get Humane software, platform, its patents and most of its employees. The AI Pin device itself will be discontinued. Unfortunately, that Kind of pulls the rug out from under current owners because their 700 devices will become non functional at the end of this month.
- [82:17]
Paul Raitzer
And good luck getting your data. This is the stuff I would say. People jump in and get these devices, rabbit and Humane and all these things. It's like, great. Who owns your data? And when that company goes under, which inevitably was going to happen with Humane, what happened? You just recorded your life. And like, now my data with you is like, this is the problem. When people don't consider the ramifications of the technology. It's like, oh, man, that company.
- [82:42]
Mike Kaput
Well, it sounds like we will just be. We'll. We'll see how HP ends up working.
- [82:48]
Paul Raitzer
Yeah, he's gonna end up in your printer.
- [82:49]
Mike Kaput
Yeah, exactly. In some other news, Safe Superintelligence, the startup founded by former OpenAI chief scientist Ilya Sutskever, is raising over a billion dollars at a valuation exceeding now $30 billion. So they're focused on developing Safe Superintelligence like is in the name, and they have seen their valuation surge from just about $5 billion in their previous round. Meanwhile, Elon Musk's X platform is reportedly in talks to raise new funding at a $44 billion valuation, matching the price Musk paid for it in 2022. This would help pay down debt and invest in new features like payments and video products. The company is also working to integrate Grok3 into the platform. Last but not least.
- [83:38]
Paul Raitzer
And then it also gives him the ability to tweet back at Sam when he said, we'll buy Twitter for $7 billion. Whatever. I can just see Elon's tweet when he says, like, it's valued at 44 billion.
- [83:49]
Mike Kaput
Exactly. Yeah, yeah. The X beef is going to continue with a vengeance, I'm sure. And last but not least, Pika has launched its official iOS app, bringing its AI powered video creation capabilities to mobile. So this app offers features like adding elements to videos or adding visual effects. It also has various tools for turning photos and text prompts into dynamic videos.
- [84:15]
Paul Raitzer
I played with that one. That one's kind of fun. Actually was sitting at lunch last week and I got the app and I was playing with your kid, if your kids would like that. Like, if you've got kids like that, that's a fun one to show them. Yeah, the effects are really cool.
- [84:27]
Mike Kaput
Nice. All right, Paul, that is a packed week in AI. Thanks for walking us through all of the developments and unpacking what they actually made.
- [84:37]
Paul Raitzer
All right, thank you, Mike. And we will be back next week with episode 138 and reminder. State of Marketing. What is it?
- [84:43]
Mike Kaput
State of Marketing stateofmarketingai.com yeah, take that survey.
- [84:47]
Paul Raitzer
If you're a marketer business leader, we'd love to have your responses there. And AI writer summit.com if you want to join us on March 6th for the writer Summit. All right, thanks everyone. Thanks for listening to the AI show. Visit MarketingAI institute.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.