"I don't want you to do my thinking, I want you to challenge my thinking, I want you to sharpen my thinking. Don't ever kiss my ass. Don't ever suck up to me. Don't ever say anything to make me feel good about myself. I'm not looking for validation, I'm looking for information."
I'm Mike Allen, co-founder of Axios, and we're gonna plunge into the ways that AI has changed Jim's mind, mood and performance. We thank Accenture for making these conversations possible. Jim VandeHei, co-founder of Axios, you are AI pilled.
Well, I'm an AI lab rat. Starting a year ago, we made the conscious decision that we were gonna be all-in on AI, both to really understand what it means for our company, what it means for our employees, but also what it means for our readers and for ourselves. And so to do that I felt like, "Okay, as the guy running the company, I have to throw myself fully in." And really I wanted to figure out like what's real and what's not real, and I'm baffled at times, but I now think I understand why a lot of people feel like, "Ah, it's overhyped, it doesn't work." It's because they don't have the time or they haven't had the interest in really putting in the work to get it to create the magic.
Zooming in on you as an AI lab rat, what do the models know about Jim?
Yeah, so I basically downloaded like probably an insane amount of information. Now, I have an insane amount of information to download. So what do I mean? Every TED Talk I've given, every public speech that I have a transcript of, every column I've written, every story I've written, every strategic memo that we've done internally, it has access to my notes function where I keep kind of my living diary and my repository of ideas. Anything that I find that I think is really interesting, I put into files. And so within both Claude and within ChatGPT, I work a lot just making sure that its memory remembers the things that are most important to me. And you can just ask it, "What do you remember about me?" and you can see, and I fine-tune that.
But then in the last six months, I've gone what I would call kind of next-level AI. And I wanna make sure like everyone understands what I'm saying, 'cause like a lot of people don't use Projects, but within Claude, if you just go into your little hamburger menu on the left, you can go into Projects. When you create a project, it is basically any task that you plan to kind of keep revisiting, something that you wanna build on. I have one for my wife's medical stuff. I have one for a book that I'm writing. I have one for CEOs. I have one for just kind of collecting all of like the views I have on different topics that come up in speeches, and then I wanna keep fine-tuning those so that, you know, if I ever need to remember like what's my latest thinking on Social Security, I can go and access it. And then when I'm working on those specific topics, I always work within that project.
Within a project, you create an instruction and then files. And again, this sounds more complicated than it is. Files is just you dump in anything relevant to that. So in my CEO one, I wrote a book on leadership, just the good stuff. So I put the transcript into a file. And so I have all kinds of files. But I also put every column on CEOs that I write, I put it in there so that every time it's thinking about how I might write or think, it can go back and look at exactly what I've written within the newsletter. And then the instructions are just you telling it, "What am I trying to accomplish within this project?" And there too, you could just voice it into Claude, or into Gemini, or into GPT, and then say, "Hey, create instructions for me that I can put into a project," and it will help you do it. And you can say, "Ask me questions to sharpen it," it will help you do that. So that creates projects. So then the next level from that has been really in the last couple of months starting to create agents. And think of agents as a project that works on its own. This is where you really get into like autonomous.
And it can work while you're sleeping.
It can work around the clock while you're sleeping. The cool thing once you get into an agent is you can put in skills, right? So in skills, I have a skills document for how I would design a product. I have a skills file for how I wanna simplify everything that I do at work and in life. I have a skills for how I like to communicate, smart brevity, with very specific principles. And so you set up that agent, you give it instructions, but it can work in the background. This sounds to most people like, "Wow, that's really complicated. Jim, you must really be a technologist." I'm not at all a technologist. The apps themselves are very good at teaching you how to create an agent. You now have a button, create an agent, and it will walk you through how to create an agent. It will walk you through creating those skills. Like you can do it at home, and not in hours, again, if you spent a half hour, I think you would have a pretty good agent pretty quickly. And you're like, "Well, I don't know how to connect it to my email." It does, it does it for you. You just say, "Connect to my email," and then you'll have to give it a permission so it's not entering something it shouldn't enter, and it does it. And, "Connect it to my Slack," okay, connect it to it. I don't need to know what an API is or all this lingo that people use in technology, it just does it in a human-fluent language.
Thank you, we're gonna take a quick break for a message from our sponsor, Accenture. We thank Accenture for making these conversations possible.
I'm very excited about our reinvention story today, which is literally 160 years in the making. The invention was trying to bring a company into the trends of the future.
So because you speak a lot, write a lot, your JimGPT is unusually rich, you're able to put in like all your transcripts, all your videos, all your writings. But if I'm a normal person, a normal writer, what could my corpus, my canon include that I'd want to inject to create a JimmyGPT?
Well, if you have anything that you've written that you're proud of, if you've had anything that's ever been taped that you're proud of, I would plug that in. But I think you're saying, "Well, what if I don't have any of that?" Then what I would do is I would go for a walk and I would hit the voice-
They're very good on voice. And I would just start talking, start talking about how you think, how you write, how you work. And I would start to talk in great detail about how you wish you thought, how you wish you worked. And then I would ask it to engage you in conversation, "I want you to start to ask me questions that will help sharpen a way for you as the AI to better understand how I think, what I do, what I aspire to do, and I want you to help me build out kind of an instruction manual for not just you, this specific AI, but all AI, to be more useful to me." And it will go back and forth, and it will fine-tune it, and it will write the instructions. You think, "Oh my God, it must be so hard to make instructions for an agent." Not at all, you could voice it into the machine, it will put it into a much more like AI-friendly format and AI-friendly language for you, but then you can start to get real results. And then once you do that, you can't stop there. Then once you start interfacing with it, it'll still screw things up, it'll still not be quite right. Just stop and say, "That was dumb, do it this way. Oh, that was great, more of that." You have to have that loop. And then once you create that loop, it's like you and I, right? If we are in conversation, let's say we didn't know each other and didn't hang out all the time, little by little, you would start to, "Oh, Jim is a little cranky here. Jim's really smart there. Jim's arrogant there. Oh, Mike, he always flatters me."
Right, like, we would figure that out. And that's all AI's doing, and it's just doing it at a superhuman speed.
Last thing on this, because talking to the phone is the most useful thing of all. What's an example of when you say how you think, what would be an example for you or someone you love?
Well, you know, this video we're doing is based on a column I wrote called "Confessions of an AI Lab Rat." The origin of that, I'm about to do a climb, and so I've been rucking, and so I had a lot of time. And I was walking and I was like, "You know what? I'm gonna do what I've been saying." I started interfacing with it. I said, "Listen, I'm gonna write 'Confessions of an AI Lab Rat.' You know my obsessions about AI, I wanna make sure I'm not missing anything. And so I'm gonna just free associate on the things that I think I've learned, and I want you to ingest that, I want you to remember it, because I'm walking and I'm gonna need to type this up, but I want you to ask me questions." And so we went through 10 different things, and it would ask me questions. It took like an hour, like it was back and forth, very specific questions. "How do you do that? When do you do this? Did you really get that type of result?" But it ended up like helping me think more broadly and more deeply about a topic. And that's a really important thing, 'cause there's a legitimate fear that you just outsource your thinking to AI. And if I could say anything that sticks with people on this video, it would be, "Don't do that." That's what you did with the algorithms. You just sat there and you passively consumed content. And a lot of that content was just crap. With AI, it's easy to just like ask it a question, get the answer, send it to your boss. That is kind of slop or it's like mediocrity. Don't do that. Like really train the AI to say like, "I don't want you to do my thinking, I want you to challenge my thinking, I want you to sharpen my thinking, I want you to deepen my thinking. Ask skeptical questions, speak to me clinically like a doctor. Don't ever kiss my ass. Don't ever suck up to me. Don't ever say anything to make me feel good about myself. I'm not looking for validation, I'm looking for information."
You estimate you use AI one to two hours a day?
I think it depends on the day, and it depends on the subject. Like, you know, like we run a company, I have a job, like other people do. And my job is to be CEO, and we write a lot, we speak a lot, we talk to staff, we make strategic decisions. But on most of those, I'm deeply involved in AI, and the company is. So I probably spend one to two hours actually inside using AI.
Right, that sounds right.
Probably spend another half of my time, at least half the time talking about it.
You know, we can go deeper into this, but like most of my leadership team, you know, my head of finance, my head of operations, my head of technology, revenue, most of them have AI agents that they've created, right? So we did that as a team. A lot of them did it individually. We have meetings constantly about how to rank and file Axions, that's what we call Axios employees, use it. And so yeah, I mean, I probably spend way, way, way more time thinking about it.
Yeah, I created one of those agents. How have you found them to be useful?
I mean, if you can get it right, and that is an if for a lot of people, it's working, doing your work on your own when you're not there. So I'll give you a good example of this. So I write a weekly newsletter to CEOs, it's called Axios C-Suite. So if you're an executive, you should subscribe, you have to be a member of the C-suite or C-suite equivalent to get it. But in that, I have created an agent, what we call kind of my CEO scout, that all day, every day, it's off looking for anything that would be of interest to CEOs. And it's set up so that it identifies the topics, it then finds adjacent information, it cites and double-checks all the sources. It gives me direct links and direct dates to all the sources. It puts it into smart brevity, it develops that and attracts that. And then I have a feedback loop between me and it, saying, "That's good, that's terrible. More of this, less of this." And it just helps like spur ideas. Now, it doesn't write the newsletter, and I wouldn't want it to. It's a pretty good writer. I'd say it's a better writer than most people. And again, there's like academics who argue with me. I just don't think most people are very good writers. I think it actually could be trained to be an excellent writer, but I think I'm a better writer than it. Maybe I'm too arrogant on that. But I do, and I have a very distinct voice and cadence.
I get your scout, and I think that we thought that we saw everything, but it's amazing what it finds.
Yeah, and when we say we saw everything, like, Mike, you probably read more words, more stories-
than any person in the world.
And I feel like I consume a silly amount of information. It always is finding reports and stories, especially stuff overseas or in research papers and data things.
And they're really important, it really matters. And it's like, "How did I miss this?"
How do you miss it? I mean, there's so much information. That's what AI's really good at, it's find information, synthesize information, cite information. But even there, like, there's just a next level to it. Like even like people are like, "Oh, it gets stuff wrong." Well, you could set up an agent or a project to basically, you could tell it exactly how to fact-check something, you could tell it how to find the root data, and then you could tell it to take it out of the agent, put it into a different AI and have that AI fact-check it. That's not perfect, but I guarantee you that's better than 99.99999% of humans.
All right, reality check. You say you found it's more confusing, clunky and chaotic than you would hope.
For most people, yes, I have. I thought by now, because it's so good, that, like, you'd have wider adoption, you'd have deeper adoption, and that it would really start to work itself more into work in particular. And what I found is, like for maybe 1% to 2% of people, it does create almost superhuman capacity. I've seen it at our company with a lot of people, I've seen it with myself. But for a lot of people, because they don't have the time or haven't had the interest, once you introduce it into your workflow, they do find it kind of clunky and they do find that it just creates chaos because they haven't mastered AI. And there's limitations to AI once you get into the workplace, and limitations obviously on the personal side, but the workplace, once you're trying to automate something, even if a technology's really good, you still have to decide who has access to it, you still have to decide what data it has access to, you have to deal within security protocols. If you're a regulated industry, you have to deal with the regulated environment. If you're a unionized shop, you have to deal with your union. So there's a whole bunch of obstacles that it hits, which is why I think you're seeing much slower adoption inside of companies. I think on true serum, most companies are not seeing massive, massive returns yet. I think they will, I think they will. I see the companies basically spotting what we're all seeing are the obstacles in real time and then fixing them. I just think this will take a period of time, whether it's 18 months, 24 months, to really fix itself and then work its way into your life stream and then our work stream.
