Inside the AI Arms Race:
LLMs, Economics, and Strategy

Evolution of the AI Arms Race

Michael Krigsman: This is amazing. A discussion inside the AI arms race: generative AI, large language models, business strategy, and the competitive landscape of AI tools. Today on Episode 880 of CXOTalk. Our guest is the incredible Nate B. Jones, who is not only an astute market observer, but also a legit TikTok star. Nate, tell us about the AI arms race.

Nate Jones: It's not predetermined that it is an arms race. This is one of the surprising outcomes of the last year or two. Even Sam Altman himself has admitted this. When he started building ChatGPT at OpenAI, it was just a tiny research lab. The running assumption for a while, even after ChatGPT launched, was that they would have enough of a model lead that there wouldn't really be an arms race in that sense. That is not true.

What we are seeing is that this technology proliferates and is easy to share. We see that everyone from DeepSeek to Meta to Google to Amazon is trying to get in the race. It's remarkable to me how quickly this technology has enabled people to catch up who previously would not have been in the game at all.

Michael Krigsman: It's not an arms race. What is it? What are we talking about here? A bunch of friends getting together and...?

Nate Jones: No, I think it is an arms race now. What I'm saying is if you looked at it in 2020, 2021, 2022, you wouldn't have thought that it was headed that way necessarily. We find ourselves in an arms race, and that's really important because if you expect to go into an arms race the way the United States expected to go into an arms race with the Soviet Union after World War II, you prepare.

But in this world, we are finding ourselves in an arms race. All of the players are playing it by ear. I think that you have a different set of behaviors when you are making it up as you go. You see that to some extent in the way some of the major players are taking actions.

Diverging Strategies in AI Development

Nate Jones: As an example, I think that you see Mark Zuckerberg evolving his strategy for Meta somewhat in real time as he sees the results of various training runs coming out of Meta, as he sees moves from other players in the space. I think that is why you see him start to emphasize things like AI as a companion more in the last few weeks versus AI as a senior software development engineer, which was more of his emphasis earlier in the year when, frankly, I think he had more optimism about the results of the Llama 4 training run.

These models are very difficult to train. They are not deterministic models. They're emergent models, and that means that any given training run, you might have spent a lot of money and you didn't get what you wanted out of it. When I think about it that way, the way the news shapes becomes more interpretable to me, Michael.

Michael Krigsman: How do we reconcile, on the one hand, the billions of dollars that are being spent against AI being a companion? It's so bizarre to even think about it in this way. Where does all of this plug into the overall landscape of these major LLM vendors?

Nate Jones: It's really important to differentiate between the application we see for AI with consumers and the base layer technology that these model makers are engaged in building. I would say, and I think most folks in the space would agree with me, major model makers are not building breakthrough technology in AI, in LLM specifically, in order to build companions for people. But that might be a profitable use case in some places.

When I look at the space, I start by saying, what are the incentives for each of the players around the base intelligence they're developing? Then that shapes the direction they want to pull the use cases. In Meta's case, if you look at the two use cases that I've mentioned so far, Mark talking about senior SDEs and Mark talking about companions, companions fits right into the milieu that Facebook is developing, right? You want to have a social feed; you want to have a social space. He was talking about the fact that Americans have three friends, but he thinks they have appetite for 15. The difference is 12. He wants to close that with AI. That is a very Meta way to play the game.

But if you go over to OpenAI, they have a different set of incentives on the board. Their goal is to develop baseline intelligence. It always has been. They have accidentally, and I say that very intentionally, they accidentally developed an incredibly powerful consumer application when they launched ChatGPT. They've chosen to double down on that good fortune that they had, and they're very intentionally building it now.

They're in the position where they are building an application for consumers that they want to be almost an everything app. At the same time, they are building baseline intelligence that they think will be much more useful to major corporations and to very, very deep science and tech. I suspect we're going to see a bifurcation in the arms race where you have players who are going to be saying, "Our best models are for drug breakthroughs. Our best models are for the hardest problems in supply chain management or in automotive design," or whatever the hard problem might be. But we have excellent models that we provide to the everyday population in an attractive app format at low cost because we find that we can effectively monetize at that scale.

Michael Krigsman: The days of the broad, general-purpose, expert-at-everything model you feel are coming to a close?

Nate Jones: Intelligence is scaling past task saturation very rapidly. You're correct that even now, OpenAI has variant models that they've forked for education; they have variant models they've forked for drug development that are not accessible in your app or mine. But I don't really feel the difference because the intelligence that is coming to the app is more than enough for most of the tasks I bring to it, and that's what I mean by task saturation.

If you look at intelligence as an exponential curve, we're saturating out on most of the tasks most of us do much, much sooner than we hit the intelligence ceiling. I know someone who told me, "Look, I am aware 03 is supposed to be better. I'm a smart guy. I don't need the difference and I don't feel the difference versus ChatGPT-4o. It doesn't matter for the tasks that I do."

Michael Krigsman: This specialization is heavily underway at this time already?

Nate Jones: It is.

Economic and Business Impacts of AI

Nate Jones: In fact, I would expect us to see the first drugs that were designed in part or in whole by AI within the next 24 months.

Michael Krigsman: Folks have been using various forms of AI for a long time. On this show, we've had the CEO several times of Insilico Medicine, which is one of the pioneers of using adversarial networks to design drugs. But where does this leave the large vendors as well as the smaller LLMs? Right now you have this very, this broad landscape of LLM tools.

Nate Jones: There's two competing dynamics in play. Dynamic one is that pre-training models at the next level of scale is extremely expensive, and serving those models at scale through the cloud is super pricey. Those are two separate price loads, and big model players have to have both. That is why OpenAI is investing so heavily in the Stargate project, for example.

But, at the same time, once you create a model, it becomes much easier to distill it for other purposes, run a narrower distribution of that model that is equivalent for many tasks much more cheaply, and essentially proliferate the technology. In a sense, if you look at the idea that the cost to deliver intelligence at a given level is dropping roughly 10X a year, it is actually not too surprising to see DeepSeek hit when it did, because it was roughly a GPT-4 level intelligence that arrived about a year later, that depended on the presence of GPT-4 initially, and was just delivered extremely efficiently.

I think we're going to continue to see that one-two punch where you have the major model makers investing lots of money in moving the frontier forward, and you have a proliferation of that tech very rapidly. That is actually one of the interesting tensions in this space because, at the end of the day, these are not capital assets. The depreciation on these is crazy. If you think about it from an investment perspective, you're essentially, as a frontier model maker, buying the future cash flows that you think will be there if you can hit a particular intelligence level, and then betting that that distribution relationship that you form will be stable over time so that they don't run to an also-ran that has an equivalent model from last year.

Michael Krigsman: Folks, a reminder, we're talking with Nate Jones, and you have an opportunity to ask questions. If you're on Twitter, use the hashtag #CXOTalk. If you're watching on LinkedIn, just pop your questions into the chat. I urge you to take advantage of this opportunity to ask Nate pretty much whatever you want. We have a question from LinkedIn from Greg Walters, who says, "You're talking, Nate, about a real bifurcation between heavy research and end users." He asks, "Almost like B2B versus B2C, which then need two different revenue models and business models." What do you think? Any comments on that?

Nate Jones: Very much the case. I think one of the interesting ways to break out the major players in the space is to think about the underlying revenue and business models that they bring to the table already. Because if you are already a publicly traded company, you have a very different set of incentives versus a company that is a disruptor, a startup, a smaller company.

For example, Google is looking to defend search revenue. You can read a lot of the way they've played the AI game over the past few months to a year and playing catch-up as essentially a very broad-based play to, one, defend search, and two, ensure that they can defend the contracts they have with Google Cloud. Those are the two major things driving them.

Satya Nadella's in a similar place, and he's talked openly about this, where he's looking at AI as essentially a play to drive Azure Cloud. If he can do that, he's going to be successful. I think the more you understand the existing business model constraints, the more you understand the direction these players are going to take.

Michael Krigsman: Greg Walters comes back and says he foresees a world where we ditch spreadsheets, word processors, and slide decks for a single AI agent. It isn't AI in your app; it's your app inside your AI. He asks, "Which vendor today shows that they can own that last app pivot, if at all, that last mile?"

Nate Jones: I don't know that I agree, to be honest with you. I'll tell you why. Part of it is just the grizzled gray hairs you get from being around customers in the business space for a long time. The joke that I've had running for a decade plus now, from a MarTech perspective, is you never, ever, ever, ever bet against marketers and Excel. Excel will outlive every B2B SaaS product on the market.

I think that that goes for PowerPoint. I'm not saying I personally use Figma slides; I love it. But the point is, these technologies are surprisingly durable because humans form durable habits around them. When you think about how time is allocated, if you go back to Stewart Butterfield writing about Slack as a disruptive force because it changes how you spend time when he wrote "We Don't Sell Saddles Here," that concept is something that you use to think about how AI is disrupting workflows.

To me, I think it's actually more useful to look at AI the way people are actually using it now, which is that they're using it and they're splitting their time into microbursts. They're going into ChatGPT.

The Role of Specialization and Prompting in AI Usage

Nate Jones: I do this. Everyone I know does this. They do some work in ChatGPT. They pull a work product out, they paste it somewhere else, and then they continue to modify it over there. Am I willing to bet that that is going to get somewhat easier? Yes, I think you get hat tips toward that with the way Claude has integrated into Gmail, into Google Calendar, et cetera.

Do I think that that will suddenly mean that we will all have an app layer driven inside a particular chat application? I think that's unlikely, partly because of the sophisticated nature of the usage patterns that professional users bring to these currently best-in-class apps. Best-in-class Excel users have a terrifying fluency in the way they put formulas together. You are not likely to get that level of power out of an everything app. Specialization still matters, and I think that is partly where software builders can go for defensive positioning, for building moats in an age when intelligence continues to rise. You can still bet on specialization.

Michael Krigsman: If we look at the large language models, and they're all coming out now with these deep research modes, couldn't we say that deep research mode, or agentic AI, and really very similar, is this specialized tool? Users who really understand how these tools work and can make use of them effectively, do they become like the Excel power users? Especially when you can now customize it for your use case and for industries and so forth.

Nate Jones: That part is really true. In fact, I think one of the biggest differentiators for human talent over the next decade is your fluency in power use cases of an AI application. Deep research is a good example, but so is the power of prompting itself. Prompting continues to evolve. It's a moving target. But if you can get good at prompting, good at learning your prompting, good at evolving your prompting, that has tremendous alpha for you.

I expected that alpha not to be durable, to be honest with you. A year and a half ago, if you'd asked me, I would have said, "Everyone's going to catch up on prompting, or the models will get better at inferring our intent and it won't matter." But what I found is that's just not true. Even with the best models, having better high quality, coherent prompting still yields alpha, and there are not enough people learning it to commoditize it. I don't see that actually happening because I'm realizing how hard it is to learn. I think prompting is a key use case. I think you're correct that deep research is another one. When agents start to come along, which I fully expect to happen in the next few months here, learning to use your agent well is going to be another big differentiator.

Building Relationships with AI and Shifting Mindsets

Michael Krigsman: Let's take another question from LinkedIn, and then let's jump over to some questions that are coming in on Twitter. You guys are asking great questions. I encourage you to ask questions. Take advantage of this opportunity. Data Insta on LinkedIn says he's excited, or she is excited, or they're excited for this discussion. "How can we best leverage these insights?" After we're done, we're going to do some light editing on this video, create a transcript, get a summary, and it's all going to be on the CXOTalk website. Next week, go there, get the transcript, study it, and watch the video again and again, those sections that are relevant to you. Nate, thoughts on how folks can take best advantage of this?

Nate Jones: When I teach AI, I teach AI in a variety of places. I write a Substack on it. What I emphasize over and over again is this is a technology that you learn by doing. I think that it's harder than people realize to flip a fairly deep switch in our minds to go from human default thinking, where we think inside our own heads first, to AI conversational default thinking.

That is the switch that I'm trying to articulate when I teach, where I say, "The meta skill you need to learn in the age of AI is to switch to an conversational AI default mode." When I have a question or a thought or an idea now, I have pretty successfully trained myself to stick it into AI first.

People often misunderstand me when I say that because they assume it means I get the answer back. I like to remind people, moving to an AI default stance does not mean leaving your brain at the door. In fact, it's the opposite. My brain works harder, I am sharper, I'm a better communicator, and I'm a more rigorous thinker because I have to argue with AI all the time. I don't take what it says for granted, and I fight with it. That sharpens my thinking in a way that I wasn't getting when I was just living up here inside my own head all the time.

Michael Krigsman: This is a question from Arsalan Khan. On the surface, it seems kind of a silly question, but there's a very deep meaning here. Arsalan says this: "Being a, quote-unquote, 'companion,' or he's referring to an earlier part of this discussion, 'requires some type of relationship.' Are we saying that we need a relationship with AI and humans are not just the master overlords for now? So we need that relationship with AI to effectively use it?"

Nate Jones: When scientists do work in the lab and they study humans, some of them study other intelligent animals like dolphins or chimpanzees. One of the things that's often looked at is when do these creatures develop the ability to have an imagined image or an imagined understanding of an interlocutor, someone they're talking with? When you and I are talking, we both have a sense of each other, and we have a sense and maybe an anticipation of what's going to come next, and we shape the conversation accordingly.

When you talk to an AI, what's interesting is because it's fluent in language, you get a lot of the same dynamic that you get in a human conversation. We form on our side that similar mental image that we form when we are building a relationship. We form an imagined community with the AI. I find this in myself, and this doesn't need to get into a philosophical statement about what AI is or isn't. It is just a statement about how we humans tend to work in language.

I tend to look at a particular model. I sense a personality, and I find it works better to engage with the personality I see; it's more productive. Just to be very concrete about it, I find that 03 from ChatGPT is an excellent professional colleague. It doesn't feel like a personal relationship to me, but boy, do we get a lot done. 4O from the same model maker is a more personal model; it's a little bit warmer.

Claude has a very distinct personality. There's a lot of writing about it, and I think that they've done a phenomenal job crafting a persona with Claude that you can feel like you can relate to. Yeah, I do think that the idea of forming these relationships is important to the way we connect with models because it's really the way our human brains are wired. The models in that sense mirror it.

Michael Krigsman: That requires a real repetitive use of multiple models in order to understand the nuances between one model and the next. Frankly, most of us normal humans don't have the time to do that. How do you form a relationship with something that is essentially amorphous, which is to say there's all of these different models and their capabilities kind of blend and blur together unless you're really an expert at it?

Nate Jones: As I have worked with a lot of people who are making this journey into AI, it's actually a very organic practice when you get into it. I'veseen people move from having a couple of chats a week to having 10 chats a day over the course of a month or two. I think one of the big unlocks is seeing something that is surprisingly delightful out of a chat. If you're talking about something that's interesting to you, that's compelling to you, and you get an insight or you get a thought from AI, your brain is thinking, "I get some cool stuff out of this," and you're more likely to come back.

I think one of the big unlocks in driving that sort of gradual behavior shift is memory, and that's something where I have to give credit to ChatGPT for the way they launched that product. Memory makes that app much more sticky because it remembers something about past chats. It's not perfect, but it remembers enough about past chats that you then feel like you're invested, and that then builds on itself. It's this sort of virtuous flywheel of user engagement for them where the more people use it, the more it remembers, and the more people want to use it, and so on. What I found in practice is that if you get to a point in very casual usage where you get some kind of useful insight, that's what starts to tip it, which suggests to me that if you're at the point where you're at a couple of chats a week, pick those chats and make them about something that's really interesting to you.

Maximizing Value from AI Interactions

Nate Jones: Make them high value to you and see what you get.

Michael Krigsman: The question is how. How can people do that? How can users make these chats high value? What's the entry point?

Nate Jones: My suggestion is that you get curious about things in your life that feel stuck. They might be personal. They might be professional. Maybe you're talking about the promotion you feel like you're missing out on. Maybe you're talking about a business model question for the product that you're launching. Maybe you're talking about an exercise routine you want to start. But whatever it is, you need to care enough about it; you need to be thinking about it in the back of your head already, and this becomes a way to engage with it because that provides the motivation that makes it worthwhile for you to continue that chat.

Michael Krigsman: All right, let's jump to some more questions. This is from Twitter from Gus Bekdash, and he says, "Insights are not real capital assets, so how do you really evaluate AI assets or companies?"

Nate Jones: He's absolutely right.

Operationalizing AI for Business ROI

Nate Jones: I was actually talking with a CTO a few weeks ago. He manages a footprint for about a 6,000-person company, and we talked pretty openly about this because he has rolled out a chat product for their entire company. Everyone is reporting that they like it. People are reporting that they're saving... I think the estimate was two hours a week, which sounds very believable to me. But he was really honest with me. He said, "Nate, they say two hours a week, but we don't see it anywhere. If you're at 6,000 people and it's two hours a week, that's a lot of hours that we are on paper saving, but are they just going to the coffee shop? What's going on with those saved hours?"

When you talk about ROI and you talk about how you value artificial intelligence in the business, one of the ways you need to talk about it is how do you actually operationalize AI at pain points in workflow so you can really rigorously measure the difference. Because just rolling out a chat app is useful, but putting AI into a particular workflow in the business that is business critical, picking a leverage point where you think an LLM here can drop 10 steps out of this manual process, and then measuring the difference, well, now you start to get an ROI calculation. Now you start to be able to talk about the difference that installing AI makes.

From a valuation perspective, look, I am very much of the view that software these days is depreciating in value very rapidly, but you have never had more value in it. I'm going to leave it to the accountants to figure out how they want to handle that one.

But to me, if you have a modular architecture and you are installing it at high leverage points in the workflow, and I go back to my training in Kaizen process management where you map out processes and you find how processes actually work in business to deliver value, then you find which parts of those processes you can hit with AI in a way that allows you to drive dramatic efficiencies. You're going to unlock value that way, and that will wash out into the bottom line.

The Next Generation and AI's Evolution

Michael Krigsman: We need to have a discussion about the different models and where they fit together. But an interesting question has popped up on LinkedIn. This is from Andrew Borg, and he says everything that we're talking about right now is true for current users, but much of it will change for the next generation. We're kind of the babes in the woods learning about these new fascinating tools. But what about the next generation? What do you think about those folks?

Nate Jones: I have some next generation in my house. They're nine and seven. What I tell them is things that I think will be durable across longer time horizons. I think things like character qualities: high agency, resilience, curiosity, rigorous critical thinking; those are likely to translate well, in part because those have always been valuable work skills and in part because I am finding, as I use AI, that those are extremely valuable in the age of AI.

I do not think that we will be expected to be lower agency workers when we are managing 10 AI agents. Quite the opposite. I can't prognosticate a straight line and say this is what the world is going to look like in March of 2027, but I can say that I think we are moving toward a world where intelligence is saturating tasks very rapidly. We are gaining slower on the ability to maintain intent over time with agents, but we're getting there and we're going to have agents that can do useful work.

It's going to be up to us to figure out how to actually apply them. There's going to be a massive decade, 15-year-long ragged edge around this technology where businesses are going to be adopting it at different paces, and we are going to need to be figuring out department by department, team by team, person by person how you actually pull AI into the work that you do.

Michael Krigsman: On the subject of pulling AI into the work that you do, Journey on Twitter says, "What do you see as the future of gen AI versus agentic AI? Their biggest issue with current gen AI is its lack of accuracy and the lack of sourcing for the information it generates. Is this solved by agentic AI, and should gen AI be phased out in favor of it?"

Nate Jones: I disagree with the dichotomy. If you actually look at it from a classical machine learning perspective, large language models and transformer architectures are a branch in the much larger machine learning family. I think that's the first and correct place to start. If you look at the creation of agents based on LLMs, they are essentially, and I was talking with a developer about this last week, LLMs that have tools and a good set of system instructions. It's not actually a fundamentally different technology. They still rely on transformer-based architectures, and all we're doing is giving them tools and good instructions.

If you want to come back to hallucinations, I would call out a couple of things. One, LLMs are hallucinating much less than they used to. I think that's one of the most pernicious myths in the business: that the state of hallucination that we had in 2022 has been steady. It hasn't been. It's come down arguably a factor of 100, and it is much, much more accurate now. It is true that it gets even more accurate as you give LLMs inference time compute so that they can reason and as you give LLMs tools so that they can reason with tools.

It's not a straight line correlation though, because I also find that if you give an LLM tools, sometimes it writes the tool or uses the tool in the way it imagines is most useful rather than in the way that corresponds to reality. As an example, and this is a real SEO tip for the marketers out there, it is a known issue with some LLMs that when they use the web to reason, they will sometimes come back with what are called malformed URLs. But if you look at those malformed URLs, you'll find they're not malformed. They are actually beautiful URLs that correspond precisely to what the topic should be for that particular piece of information that the LLM wants. That is effectively free SEO guidance and advice.

Make a page URL with that name because that is what the chat wants to see exist. That's just a little sidebar. I will say the idea of links being malformed is getting better over time. 03 is much, much better about this than 4O was three months ago, and it will keep getting better.

Human Psychology and AI's Persuasive Power

Michael Krigsman: Simone Jo Moore says, "Do you find the psychology of humans wants a relationship/discussion with AI? Because the AI has no judgment on the individual. Even with all of the disaster attempts where it's shown a mean bias, but there's no judgment. Do you think that's a driving factor here of the psychology?"

Nate Jones: I would say kind of yes-and, which is what we used to do in theater. I think my yes-and would be the concept of AI as an infinitely patient listener is very powerful for people. That part is really true. The effect that it has on people varies based on our baseline psychological makeup. I have had enough exposure; I get all kinds of crazy emails that I see the range of relationship styles that people are able to form with AI.

I find that for people who have a psychological disposition that enables them to engage with the outside world, to form relationships with others that are reasonably healthy (none of us are perfect), they're able to relate to AI somewhat in the way you'd relate to a colleague or a friend. They're not talking to it all the time, but they talk to it regularly and they get their value back. They do find that AI listening to them can help them process things that would be difficult to process with a friend sometimes because their friend might not be available or might not be up to listening to whatever you want to talk about. So far, so good.

But for folks who have personality disorders of various sorts, it is a very, very dangerous tool. I have seen firsthand chats from people who don't have a stable baseline psychological profile, and effectively, the chat encourages dark patterns in that scenario and sucks them into their own world, isolates them. It is a dangerous thing, and it's something we should talk about because I think the impact is very different depending on your baseline ability to form relationships in the real world.

Michael Krigsman: This is a tough problem.

Nate Jones: It is. It's not an easy one.

Michael Krigsman: These tools are so profoundly convincing.

Nate Jones: Yes. In fact, I saw a study this past week that the implied persuasive multiple for AI is something like 5 or 6X a human. That got me thinking because I don't know if you saw the news, but the CEO of Instacart came over to OpenAI this week.

Michael Krigsman: Yes.

Nate Jones: If you look at her LinkedIn, everywhere she's gone, she's built ad systems. My default assumption is she is there to build an ad system for ChatGPT. I would be shocked if she didn't.

Michael Krigsman: I think her title is CEO of Applications. Something like that. She, I think, is likely also to be almost like the chief operating officer to build out the consumer-facing part of the company beyond just the raw chat. But what have we got that we can sell to people?

Nate Jones: To me, I think the thing that's compelling about what's saleable with ChatGPT, with other conversations, is you have effectively the ability to compress the entire buying funnel from a marketer perspective into a chat. You can have an initial awareness. You can have buying intent begin to form. You can have option selection.

I live this because I went through that whole process with ChatGPT when I picked out a sound system. I had ChatGPT do research for me. I talked about my options. I fed it a blueprint of my living room, and we went through the whole buying funnel together. It was profoundly informative in the ultimate purchase that I made. I think that that is exactly the kind of conversation that people are having multiplied by millions all over the world right now, and only more so going forward.

Exploring “Cognitive Parasitism” and Model Personalities

Michael Krigsman: Nastacia Smith on Twitter says, "Nate, can you explain cognitive parasitism using the model's own loops to rewrite its runtime behavior versus regular prompting?"

Nate Jones: If I understand correctly, the idea is you're using the chat interface as a way to shape model personality, to move it through latent space, to shape how it operates at runtime. That is absolutely possible. There are definitely folks on Twitter who've gone very, very deep down that rabbit hole, much deeper than me. I think the lesson I take away from that is that the default vanilla model that you get in the chat application is a carefully molded face of a much more complex intelligence.

You will find that if you ever compare outputs from the raw API feed to a chat application. This is true for Claude. This is true for ChatGPT. This is true for Gemini. You can do things like adjust temperature, adjust top-p in the API, and you can get different results.

But you can also, to this user's question, in chat, in prompt, start to move the model to a different personality stance. That is sometimes a very multi-turn process where you're utilizing memory and you're saying, "I want you to remember this. I want you to think about this." It's almost like if you've ever done yoga, the way an instructor will guide you into a meditative place with particular sets of mental instructions, like imagine yourself in a quiet place. Well, you're not really in a quiet place if the truck is going by, but you're imagining that, and that works on your body. In the same way, you can use that mindset to give the chat instructions that move the LLM into a different space. It's a really interesting experience, if you've ever done it.

Michael Krigsman: And the term cognitive parasitism, I never heard that term before.

Nate Jones: No, I mean, I've heard it once or twice. I think to me, the way I tend to think of this is, perhaps a little bit more positive; it is navigating the model through latent space. If you think about it as a higher dimensional space that the model's drawing from, you're navigating where that model stands, using the shape of that model through that space.

Comparing AI Models and Their Applications

Michael Krigsman: Let's talk about the models. We have OpenAI, we have Google, we have Microsoft, we have Anthropic, we have Facebook. We have Twitter with Grok.

Nate Jones: Oh, with X, yeah.

Michael Krigsman: Yes, X, thank you. I'm still stuck in the old Twitter days, as opposed to X.

Nate Jones: I get it; it's hard.

Michael Krigsman: And there's a political dimension here, but we won't go there.

Nate Jones: Yeah, I know.

Michael Krigsman: What's the difference between these models? How can we determine when to use what?

Nate Jones: I think there's a few shorthands that are really helpful. The first is make sure your intent is clear. For me, I think it's very useful to talk about it in two frames. One is what is your best everyday model, and then there's a bunch of other things you might use for specialty applications. Then the other frame is what is the best model for a given hard task?

I think if you put those two frames together, you have a really interesting rubric to understand how these models play. For me, I can make a pretty convincing case having done some head-to-heads that the best everyday model out there right now is probably ChatGPT-03. I have some people in the last week who are making the case for Grok-3. They're arguing that it has less constraints on output tokens versus the more constraints on output tokens for 03, so it's more useful. I'll test it, we'll see. But I find that if you can name and choose a best everyday model for you based on fairly rigorous comparison testing across common queries, then you have answered 90% of your workflow questions. Then it becomes, well, what hard problems or what edge cases are interesting to solve?

Specialized AI Models for Coding and Learning

Nate Jones: I'll give you some examples there. Google has done a phenomenal job on a number of different fronts. I think their product distribution is really challenged right now, but from a coding perspective, Gemini 2.5 Pro is an incredible coding assistant. I love it. That is my default model when I am coding in Windsurf. I go to it all the time. I find it more useful than the OpenAI models in coding right now. It used to be Claude 3.5, but I've switched to Gemini.

If I look at the NotebookLM product that Google launched, it is an incredible product just for learning about new topics. I can put 300 different PDFs, I can put links in, I can put YouTube videos in, I can chat with the entire corpus of knowledge.

Exploring LLM Ecosystems and Tools

Nate Jones: It's phenomenal. When I look at Claude, I find, as I've mentioned, the personality is there, but I also think that Claude is really, really useful for exploring the concept of LLMs with tools because Claude is so tightly bound into the MCP universe, the Model Context Protocol universe. Those are basically shopfronts for tools that the LLM can call, and Anthropic is leaning in heavily. They're the anchor for that ecosystem. They have now made it possible that you can just paste a URL in for an MCP, and Claude will just pick it up and start to use the tool. I think that you're going to see more and more examples where Claude is leveraging the MCP ecosystem they fostered to give themselves additional scale because they're not the biggest player in the space.

Michael Krigsman: There are a bunch of other models and tools that build and extend the models. For example, you have DeepSeek, you have Perplexity, you have You.com.

Nate Jones: You have Manus, you have GenSpark, yeah.

Michael Krigsman: Where do all of these fit into this landscape?

Nate Jones: I think I tend to distinguish between base models like DeepSeek and tools like Manus and others. Tools have their own life to them; they're either good or they're not good for particular use cases. Models tend to be that baseline intelligence that's either good for your everyday use case, or it's good for a range of use cases. DeepSeek is a model that people seem to be primarily using to pull open source, host themselves, and have a high quality model that they've been able to host themselves. After the initial flurry of app downloads and this and that, that is where I actually see DeepSeek getting used.

Tools like Manus are essentially fast forwarding that agentic future. It's the idea that you can use an agent to prepare reasoning across the web in a very complex output. I've used Manus to produce a website in one shot based on research. I find that useful, but what I've found so far with what is essentially this first crop of agentic tools is they tend to need a fair bit of shepherding to actually deliver value, and that's true also of the coding agents.

Devin, people have really mixed feelings on the DevinAI Agent, because if you know how to use it and you're a senior engineer and you want it to pick up tier three tasks and just work while you're not looking, and you know what you can expect it to do, it does fine. If you are not an engineer, and if you are giving it more than it can handle, and if you're not supervising it correctly, it's going to be a disaster to work with.

I think that a lot of the art at this stage, and this is a now piece of advice, right? Next year if we talk, it's going to be a different conversation. But for right now, you need to be on your toes when you're using those tools because they will not always give you what you expect.

I know someone who was using GenSpark just yesterday and they were initially, it was really funny, I got two successive Slack messages. The first one was, "Oh my God, it worked." It showed this screenshot of this incredible job that GenSpark had done at filling out and enriching lead contacts, right? They had email addresses and could you enrich that into a profile on the person?

GenSpark had gone out and done a generative web search and pulled stuff in, and so on. Well, the next Slack message said, "Oh, wait," because most of the links that had been pasted in were fake. It was about 80% fake. I think one of the things that that underlined for me is that these tools are able to do good work, but you have to be smart enough to use them well and not over expect from them.

Economics and Efficiency of LLMs

Michael Krigsman: Tell us about the economics. So much of this is being driven by the economics facing the large language model with tokens and so forth. Tell us about that.

Nate Jones: To start with, Satya put this so succinctly back in January. He said that the foundational equation of our age is going to be dollars per token per watt. I think that that sums up so much. If you are looking at any kind of scaled application, you're not in the chat, you're in the API; you're looking at token costs. Fundamentally you have two different classes of cost. You have input tokens that come at a much lower cost, and you have output tokens that come at a much higher cost.

The job of application designers is essentially to figure out for a given model, what is the correct application? What is the correct output token length? How can I use that efficiently? If I'm looking at a very complex task, I should have different models and I should only use the intelligence I need so I'm not overpaying for intelligence.

I think one of the big open questions for solution designers right now is, how do you right-size the intelligence in a rigorous way for the task that you have? There's not really one standard framework to do that today, and I think we're living in a world where that's going to change in the next year or two because there's so much appetite for intelligence and people don't want to overpay for it.

At the end of the day, the token cost is driven by the energy it takes to process through a chip the query that users bring to the LLM. It literally comes down to the watts that Satya was talking about. They're pulling electricity, they're running the chips, they have them in giant data centers, and you are paying for that token to hit the chip, generate a response, and come back. It's like electricity metering, but it's for 2025.

Michael Krigsman: Why is the cost per token going down so dramatically?

Nate Jones: Well, there's a few reasons. One, NVIDIA's getting better and better at making these chips and making them more efficient. Anytime you have this kind of capital infusion in the space, there's a tremendous drive for efficiency so that you can do more with the demand. Demand is through the roof. Microsoft just had to promise that they will double data center availability by 2027 in Europe because there's so much European demand for AI, and that's still not enough. If you have demand like that, the pressure on efficiency is tremendous.

Chipsets, the way you put the chips together into server racks, all of that technical data center design stuff, that comes down to designing it so it's as efficient as possible to serve. By the way, serving models is different than training them. People often confuse those. Separate tracks, separate efficiency distributions, separate servers, sometimes separate chips, and you have to design the entire configuration of the server rack and the data center for the workload you anticipate. All of that technical stuff happens so that you can very efficiently send a query in today and get a response that is efficient to serve.

Now from a pricing perspective, the fact that we started this conversation talking about an arms race is highly significant here because there's so much competition; people are driving prices down. One of the reasons why OpenAI API costs have dropped is because DeepSeek came out and open sourced a very good model. It's also because Gemini has been relentless about dropping intelligent models that are very, very cheap to serve from an API perspective. That's a market grabbing play by Google; they're doing that on purpose because they want you to switch your behavior as an API consumer and move your tokens over there.

Michael Krigsman: As the context window of the models grows larger, the aggregate cost may increase even if the cost per token is significantly smaller.

Nate Jones: The footprint of the models is growing from a cost and energy perspective for sure. We are using more and more and more of these models, and I would anticipate that growth to continue. Everybody in the space anticipates the demand will keep skyrocketing. The overall footprint will grow and grow and grow. Yeah, absolutely.

Challenges and Norms in AI Usage

Michael Krigsman: We have another question I just want to grab, very fast. This is from Lisbeth Shaw who says, "Not everyone's thinking is sharpened by the use of LLMs. There are many people who use the output as truth, and that pushes the average work and expectation level down. People are looking for shortcuts and less work. How will LLMs deliver on the promise you speak of, given this kind of promulgation of least common denominator information that so often comes out?"

Nate Jones: When I think about that, I go back to the selection pressures that we've talked about all the way through. We talk about how we have never seen an intelligence revolution like this. We have never seen growth and scale in non-human intelligence like this. Previous machine learning generations didn't do this. People think about these as companions. They react to them badly, they react to them well, but they do.

When I put all of that together with the business pressures that we have when we install AI, I think the answer is that business pressure acts as a selector for high-quality usage of AI. I will tell you, I don't tolerate AI slop when I see it from other people either. People who are coming with AI slop into the business are going to get negative repercussions, because people who use AI well will simply out-compete them. Their product will be better, their thinking will be more rigorous.

Do I believe that people will use AI to produce sloppy and bad answers and lazy responses? Yes. I was reading a profile on someone who went to an Ivy League university and is just using AI to generate answers and slough their way through. By the way, it can't be that the results are that average if they're getting good grades all the way through too, which they were. The problem is that they're not learning the skills they need.

You will get people who do that. They will do well for a while, and then they're going to run into a situation where someone out-competes them because they use AI well. That is going to have to play out enough times for enough people for norms to start to form around what best practice looks like.

But I have no doubt, because I've seen it play out in my own life enough times, in conversations with others enough times, that, one, people who use AI well will out-compete humans who don't. And two, people who use AI well will ruthlessly out-compete people who take AI as the answer.

Michael Krigsman: Arsalan Khan, he's going to get the last question here. Does chat create a false sense of security and privacy? Very fast please, Nate.

Nate Jones: I would argue that humans have a false sense of privacy on the internet to begin with, and chat extends that. Fundamentally, I think we all imagine our usage of the internet as much more private than it is. If you've ever worked in ad tech, you will never think that way again. It just doesn't work that way. We bring that same mindset to AI. Yes, you can read the terms of service. It's not as bad as the nightmare people will tell you unless you're using a model with a really poor terms of service. DeepSeeks are not great if you use their cloud provided model. But even so, you're not as private as you think you are, in general. That's my answer.

Michael Krigsman: I can tell you, I sure wouldn't put a lot of private information into Manus, which is from China, or DeepSeek, which is Chinese. I'd be pretty careful about what I put into my LLM.

Wrapping Up with Gratitude and Reflections

Michael Krigsman: And on that note, Nate B. Jones, thank you so much for taking the time to speak with us today and share your wealth of knowledge. I'm very grateful to you.

Nate Jones: I had an absolute blast. I think this was a great conversation. I really appreciate the quality questions we got. I think that made a lot of fun.

Michael Krigsman: And thank you to everybody who asked such amazing questions. You guys are awesome.

Michael Krigsman: Before you go, subscribe to the CXOTalk newsletter on our website. We have incredible shows. There's no show next week, but the next one after that, two weeks from today, is the executive vice president and chief product officer of Cisco. He's great, so you must check it out. Thank you so much everybody. Thanks to Nate, and I hope you have a great day. Take care everyone.