Here’s How to Know If You’re Getting the Most Out of AI – with Bryan McCann, CTO of You.com

30:41And there's this tension between these two types of principles. Well, I shouldn't bother automating some things that I don't know if that is even worth automating. Maybe in three months, I'm going to throw all that away. that's okay throw it away that's what i would say every time i wrote a new paper i threw everything away i started from scratch and it completely rewrote my framework because i'd learned so much that the the important thing for me to do was to get really good at building so that i could build the system the way i needed it to be not just be attached to my system obviously you can't do that with entire companies but i do encourage my team to do that too like i want a culture of building a bias towards building so much so that we're not afraid to throw things away, throw code away.

31:29We have a better idea. There's a better way to do it. Let's just do it. You talk about engineers. What's your thought on these enterprises that you now work with on kind of the vibing, the vibe coding entering in that domain so that an engineer is becoming more of a work researcher, like somebody who can use code, not because we're in this area, they know how to cope, but they can use the agents to do that. Are we going to think differently about who is a good, productive person in our organization when these things are more kind of implemented into our lives? I'll give you an example, just to think about it.

32:08Like, you know, our producer on this podcast is Emma. She used to do pre-research. Hi, Emma. How's your state of mind right now? She's state of mind. And Emma used to do pre-research on people, everybody who kind of asked to be on the podcast and then ended up vibe coding a red bit service that basically did research listen to some podcasts and basically gave like a guesstimate if this would be a good candidate or not right clearly like very kind of like brilliant way of thinking about how do i scale my own abilities for spending three hours researching how can i just do that right and not necessarily something that you would assume like a podcast producer or somebody who's doing the product producer's job, you kind of niche-chatting to think of.

32:49Nonetheless, you did. Very much for me, like an example of what I would imagine people, internal organizations that you also work in now will start to kind of think, hey, I could just code this up myself. And so curious of where your head is about Vibe coding inside organizations for non-coders. I think everybody should be doing it as much as possible, just like Emma did. They should have the same exact thought process just because they're not a coder. you know it's not part of their identity doesn't mean that if they i think the same principles hold if you're doing something three times yeah take some time to automate it see if ai can do it your time now is better spent trying to do it and and just like i said with the engineers if the i can't do it you're gonna get some really valuable intuition about where ai is right now in a way that you wouldn't otherwise get.

33:42It doesn't really matter if you're an engineer or if you're a product designer or if you're a podcast producer. Learning how to learn, learning how to create, these are essential fundamental building blocks for learning how to scale yourself. So if you think about neural networks, organizations, and people, all is very similar types of entities. where they need to learn as fast as possible. They need to have these certain building blocks and skills, very good information passing. I mean, one day or later, you know, we can go more philosophical and talk about how all this applies to, like, the individual as well, right?

34:29But anyone in an organization, any node in a neural network should be looking to do these things. It doesn't matter if you're in jail. I have to ask, and I know Jeremy's going to kill me for this now because I'm bringing you back into the dark side, right? Dude, why did you have to go there? This was such a positive conversation. I know, and we had many positive things. But we all talk about organizational design that was made probably, and you'll know this better than me, Jeremy. I think, as I understand it, it was made when the railroads kind of was introduced to the U.S., like the whole classic hit radical kind of way.

35:02Yeah, chain of command, yeah. And it does seem that increasingly we'll have to redesign an organizational structure that probably, I don't know what it would look like, but maybe it'll look more like the SEAL team, small groups that report straight to the president than the army, right? Now, I do think there is something interesting to thinking about humans also, or basically tasks of saying, hey, I have a prompt, I have something I need, and I need to basically get an output. And I can either ask an agent or I can ask a human. Is this a human or an agent question? And then some humans can take very abstract then prompts because they will then be able to pass that on.

35:38And so, Jeremy, I knew that you kind of thought it was an evil way of thinking as people as just kind of notes in a... Prompting humans does feel weird, but I go trying to ask. But I do think that there's something interesting in like if organizations are becoming more like the web with notes in an ordinary network and in many ways with pathways. than this classic heretical organizational design that we have now. So Brian is curious if I can lure you down that way of thinking. Yeah, no need to lure me. He's already down there. He's already deep in the rabbit hole. I'm there. I think we should be designing organizations like neural networks.

36:21I think you can, if you look at everything we've learned over the last five years about how to get neural networks work really well, we should apply that to organizations and a lot of those things you kind of already do but for somebody who doesn't know maybe you take one step back for somebody who doesn't know what we've learned about how to get neural networks to work really well say what we've learned and then what the implications might be for work design because i don't think that'll be obvious to some people right right so here's a very concrete example again something we kind of do but like it's worth thinking about in the context just to establish some credibility for this metaphor.

36:58So if you take a neural network, a neural network, in the simplest case, what we can think of one as having layers, and then those layers have nodes in each layer. And as you build it up, the data comes in at the bottom in this mental picture. And then at the output, we're going to have a classification of like yes or no data comes in yes or no at the output layers in between so if there are layers and you have too many layers then when you're training the neural network the signal when you make a mistake on the yes or no question it tries to back propagate down through the network right but by the time you get to the really really bottom layers there's almost no signal left That makes it very hard for those bottom layers to learn.

37:52So what is something that we do? We add residual connections or skip connections between higher level nodes down to lower layers. So they pass signal directly and they send error signal back directly. So you might have heard about, you know, meeting with your skip manager, your skip level manager. We do this. These are similar principles that have emerged independently to some extent in two different types of organizations of nodes or people, but they're nodes in a network, trying to learn as quickly as possible. So if that one makes sense to you, maybe one that would be less intuitive, but we could then transfer over is looking at transformers.

38:45so transformers have this really nice self-attention mechanism where every node in one layer if it's a self-attention layer can attend to all of the other nodes or like all the other vectors let's think about it as a sentence every sentence has a word every word gets a vector every vector gets to pay attention to all the other vectors that's what self-attention means conceptually and they get to say how similar am I to that or how useful is that for this example relative to my job right for my node in this network and keeping the information flowing very well up and down the network with those skip connections is important and keeping information passing horizontally in the network with a mechanism like self-attention and there might be other mechanisms too but self-attention is a very popular one so information of flow horizontally up and down very very important in neural networks and whenever you see a bottleneck happen it's almost always in our advantage to break that bottleneck we sometimes have tried to do that yeah to try to make like hidden latent representations that mean something and most of the time to date that's not the line of thinking that has won the line of thinking that has won has been makes everything as parallel as possible to bring it maybe a little bit more forward into the people side make sure everybody can talk to everybody if you are going to have a lot of layers then you better have a lot of skip connections but it's probably easier just to have fewer layers which we see the flattening of a lot of like new ai native companies just being then you kind of get to what's mcgill's number or whatever like there's only so many kind of call it nodes that any human can actually attend to.

40:41I think the number is typically around 150, right? So they're almost rate limited by the, yeah, Dunbar. Thank you. It's almost, who is at McGill University, which is why I thought of McGill. But you start being rate limited by whatever the human's capacity to attend to other humans and other information is there, right? For sure. But tooling can help with that, right? Just like we talked about before, I don't need all of the information from every, I don't, I think a lot of what will maybe go into that is more and more of the information passing the like the raw content of work will get passed with ai and the people will be there to focus on the people so you're right there's going to be some upper pound limit my business partner nicholas from autos he has the thesis that ai is squaring Dunbar, so it's 150 squared, so it's 22 ,500 that you'll be able to kind of oversee or have an emotional connection to with an AR I'm ensued.

41:45I have one last question that I'm keen to ask you. We had the founder of Sabia on the podcast and he talked about that they have designed a new way of getting hold of the data in their team by not necessarily doing a big data factoring kind of system with one big platform that I think a lot of big companies have spent a lot of time doing, but basically using agents to go and find the data with just wherever it was, and then make sure obviously it could be representative of you. As companies are going through this process of figuring out what data they have that unique, and then how do they make sure that the data, the information gets to the edge of their organization?

42:27How do you think about the requirements for doing a big data science project and how much do you believe that you increasingly can just leave the data where it is and then have agents go and fetch it. So with you.com, we, I described, we do a lot of public web search and I do a lot of private data search inside companies. And there's a similar maybe analogy here where, at least with the public web, like if we had a system where everybody just, typed the things that they put into the public web and typed it into this centralized database and then chat gpt or something just like had access to that database it would obviate a lot of the need for having a bunch of like web crawlers and to some extent search itself and in that sense it would be much more efficient as a system so something that's much more like this is Amazon, right?

43:27You have like buyers and sellers just like inputting information. And you'll have services like Google Shopping, which need to know what's going on on Amazon. And they're crawling Amazon, but it means they're always behind. Because in Amazon, people are typing the information into the system. And Google has to, by definition, after that happens, crawl it, find it, discover it, and then incorporate it and then use it. So companies that centralize that data to some extent, I think it's still worth doing over time as an efficiency trade-off, but having methods to deal with the messy, more distributed, less efficient process of having data be just where it is and you go find it is going to be incredibly important for growing in this new era.

44:17I don't think you can wait to centralize, but it's still going to be an important part of the process as you, you know, if you're a massive company, eventually want to get efficiency gains. You don't always want to be running your search and your finding systems to go with the data. So I think it'll be a mix. But the fact that we haven't developed those go out and find tools to the same extent suggests that, you know, it's to date, it's a harder problem. It's much more complicated to have a system automatically do that. A lot of that knowledge is like baked in the minds of data scientists who have already left those companies.

44:58It's a mess, but it's worth doing. One thing I would love to hear if you have any thoughts on is this recruiting. The question about people. I was actually talking to Wade yesterday as I was driving to a meeting. We were catching up and he was telling me about the conversations that he has with people where he's basically saying, no, you have permission. Try it. Just do it. And there's a question of some people, we were actually talking about the challenge of hiring young people and how, as I've talked with college students and other young folks who are looking for work, the realization has occurred to me, they are looking for a job a lot of times because their whole life a teacher has told them the assignment to do.

45:43And what they need is a boss to give them their next assignment. And that's, I think that that's actually, it's a profound kind of realization that I'm still kind of grappling with, but it strikes me that it's an exceptional person who's a self-starter, who takes initiative, who can frame their own work, who can set their own objectives, et cetera. how do you think about recruiting in the age of the organization that's more like a neural net? Because I don't think everyone is, at least most people haven't been trained with the skills they need to be an effective node. It's a transition, but I think there are other proxies that often correlate well with it.

46:25I have had engineers come to me and tell me like, I just need to know what to do next. and I'm like, I'm not going to tell you. Like the job here is for you to figure that out. Like it is a different type of job. That's going to be the job here. And in this AI world, I think increasingly for more and more people, I guess I do it a little bit trial by fire, you know. I'm going to tell you the future. My job is given all the information I have, tell you what the future is going to look like one year, two years out. And then you can take that information and you can run with it you can decide what's going to be the most impactful thing you can do, the most important thing you can do with your life's energy.

47:08And as long as we're aligned on that, then you belong here on my team. And if you decide that that's no longer aligned or you feel like you're outgrowing us, then I'm going to support you for the rest of your life because you worked with me in this very crucial way and crucial time. And that's how I treat it. Like it's all of the things that we're building right now, going back to something we said before, we're going to redo it all in the future, right? We're going to be building so much faster. Everything's going to be rebuilt again, sometimes with each other, sometimes separately. And anybody who comes through my door and maybe passes those like individual, you know, tests and things like that, just to make sure they have some basics, right?

47:52They become a special node in my network. That is someone that I'm hopefully trained to be way better than me and there's way, way better things than me, you know, and I don't just mean like making multi-billion dollar companies or something like that. I mean, more like for humanity, like I want people to have impact and figuring out how to have impact is, is a skill. For me, it's this listening process. I get all this information all the time and trying to feel out where I have strong convictions. And if I don't do it, maybe nobody else will. Now, someone always will kind of try to do it. But, you know, will they do it the way that I'm going to do it?

48:38Am I uniquely qualified, positioned, burdened to contribute in that way? I think it's beautiful. Well said. And everybody in my company should be doing that every day. And if they're not, then we should fix it. And being the timekeeper, I, Brian, very much appreciate this conversation. Like, it's wonderful to have somebody who thought so much about all these things. This was super fun. Yeah, very appreciate it. Thank you so much. I'm so inspired. I hope we might be able to get you on another day. That'd be great. Round two. Thank you, guys. Best along with you to come. All right, a lot of fun, too.

49:15See you around. Thank you, sir. Cool. Take care. Cheers. Adios. Bye. Bye. Wow. You know what, Henrik? That was a really fun one because you and I had some disagreements about where we wanted to take the conversation, didn't we? Yes, I know. And I was asking Chachibuti the other day how we can improve, and it suggested that we should disagree more. Oh, did it? There you go. Wow. And did it suggest that you suggest that to me, or did it suggest that you keep that knowledge to yourself? It suggested that we should have ideas or points that we were prepared to talk about in the podcast where we knew that we came down to different sides.

49:49Oh, that's cool. Okay, so where do we come down to different sides in this episode, Henrik? Well, I think the reason why I really enjoy Brian and how he's thinking about it is because I am… I mean, so thoughtful. What a thoughtful dude. Wow. Really, and just clearly, like, incredibly smart. I am quite fascinated about this idea that the future of work, the future of organizational design is going to mirror neural networks, which she talked about. And I have for a while been, and I'll give you a little bit of background. And so I think a lot of companies are going through these kind of three phases.

50:26Phase one is how do you upgrade the capabilities of your staff? Two, how do you create agents or genetic workflows to be more efficient? And then I think the third one is kind of like, what's next? And so a lot of companies seem, people that are really on the cutting edge seem to be kind of passing between two and three. And they are hitting your Parkinson law where they go like, I don't understand. We've implemented 40, 50 agents in our organization, but we're not really seeing the efficiency gain. My thesis is that you basically have now people that can do jobs that belong in other teams. and because that they are not uh happening at the same time you won't be able to reduce staff in one department because they can from it so you basically have this weird kind of uh jigsaw or mismatch of skill and then what people do is of course they just fill their time up and so they've got 30 percent more time and then they get busy doing another and so i do speculate that in the next year or two we will see more organizations that will need to ask themselves this fundamental question, which is how do we design ourselves for an organization that is increasingly agentic?

51:37When we see that the agentic workflows, for example, the one is the case I mentioned with people that are applying with AI and then their job applications reviewed with AI, it basically shows that the way that the workflow was designed was for humans in a world of scarcity. And now that you have agents in a world of abundance, then you need to reinvent a new workflow. And so getting inspiration of what that might look like, how you might be more efficient and get an exponential kind of upside to your AI work, I find to be fascinating and don't hear that many people talk about. And so when Brian kind of got to speak about this, I was like, you know, you really have my attention now.

52:22And so, but I think what you might have reacted to is that there is this kind of like build-in assumption there that you are giving tasks and sometimes you will give it to an agent and it will be as good as a human. And then there is, I think, the unfortunate consequence that sometimes there will be people that won't want to upgrade, that won't be the people that will understand how to put on the Ironman suit. And at that point, they'll just lose their relevancy in the organization because the agent will be able to perform that specific task more efficient than they do. And you can't upscale them to be kind of the managers of many agents or tasks.

53:03And so you being such a kind, thoughtful person, I think sometimes come to the table. I'll say the statement, but I mean as a question that people are upgradable. And so it is just about finding the way of teaching them the skills. And I think I might come from a more cynical kind of perspective where if you are a startup and you don't have the resources to do that, there's a time, as Brian was saying, where you just have to say, well, your way of working or your expectation of how we work in this organization is just not how it's going to be. So I love you and leave you. Love me. Leave me. I, um, I was really, uh, inspired by his, I love the phrase, keep the GPUs full, um, as a kind of personal mantra and then as a team mantra.

53:58And I think that's actually, you know, right now the bandwidth limitation is mostly, you know, the reason most organizations can't grow more. If you say what's got a stranglehold on your business, folks would say, well, if we had more smart people, we could do more, right. But we can't hire them. We got to grow at the right rate. But intelligence is no longer the rate living factor and no one can reckon with that. And because why you, you have to think in terms of, we got to keep the GPUs full. That's how you max out the intelligence of the organization, but individuals and most forget teams, you know, very few individuals are thinking in terms of what can I pass?

54:39What's a full night's worth of work that I can get to the point that I can pass off by the time I leave work today. Because I value the nighttime cycle of experimentation so much, I wouldn't dare go to sleep without giving an agent something to do, right? And so to me, when he talked about scaling himself up, and then he said, quote, every minute of my time is better spent scaling my team up. To me, maybe it's a humanistic thought, but set aside whether people can be scaled up. the fact that he's now observing the best use of his own time as a leader is enabling his team to get scaled up to me i what i it was not a humanist thing it's more just a practical how do you manage in a world where you're trying to give everybody the mindset keep the gpus full well one thing that he was mentioning most people want to clock off and they want to be done one thing that he was mentioning also which i and i don't do this normally that's your your trade is to write down like a census and he said figuring out how to have impact is a new skill and i do think that that is so true and i think you know and i guess as an entrepreneur i get excited about because that is like an innate entrepreneurial capability like what do i have to do next what is the next task that i throw myself out how can i have an impact i'm not going to move that forward and the way that i look at entrepreneurship is often through these capabilities which are not do you code do you design or do your product manage?

56:02It is more, do you have agitation? So do you kind of propel stuff forward? Do you have gravity? Can you tell a story in a way that people understand it? All models understand it. And do you have resourcefulness? Can you do a lot with a little? And those three things are often the trademarks that I look at when I invest in entrepreneurship, for example. I think in the same way, and maybe I'm just seeing it because I believe more entrepreneurship is good for the world. But I do think that these deeply entrepreneurial skills are what is required in the future organization because everybody will be able to do everything.

56:36And so what they have to do is to figure out how do I, as an individual, have impact. Uniquely at impact. Yeah. I mean, he said everybody in my company should do that every single day. Everybody should do that every single day. I really think that that is such an important kind of lesson for people who wants to have a real thriving career. And I think for organizations that want to figure out who should we identify to really promote. And then I'm not an AI researcher, so I'm just so fascinated when you start to take these somewhat abstract models and then you apply them as a philosophy. Right, to org design.

57:18Yeah. Yeah. That part of the conversation definitely reminded me, I think it was with the team from Applied Intuition, Kasser Yunus, talking about this idea of guilds, right? And what if organization of the future is actually guilds of, you know, small groups of people executing experiments, as you say, with SEAL team kind of accountability. What did he call it? Not skip level. That's kind of the managerial term. But what was the neural network term? Is that redundancy? I can't remember exactly what it was. This is where his agent, he talked about the next level of agents are these predictive agents that basically jump in.

57:55So in his will, an agent would jump in and give you the right term. Well, I mean, that, by the way, a great context aware, you know, feature. And I love that vision of providing immediate value. I think it's super cool. It's a super cool vision of the future. That would be to give me my next line. I mean, the ultimate obviously is help me know exactly what the right thing to say is. But that's not a foul way, right? You could imagine that, you know, I have granola running on these calls, right? And so I could probably just ask it now. But from there and it coming up with feeders and saying, hey, you should remember to think about this or ask about this or, you know, it's time to stop interrupting or whatever feedback could be, right?

58:38It doesn't seem to be way into the future. going back to the management or leadership thread for a second one of the things he said that i really uh enjoyed and wrote down was when he's working with his engineers he said step one is try to get ai to do the thing step two is if it still can't try again and step three if it still can't commission a research project and come tell me because the stuff you can't figure out how to get AI to do, that's the stuff I want to know about. I thought that was a really, really cool way to think about. At no point am I saying, don't use AI. In fact, what I'm saying is your job is to figure out how to use AI.

59:19And the only time I want to know is if you can't figure out how AI could help with this. That's a, that's like such a total paradigm shift. I agree. Do you have anything else that we should add? No, I mean, I would say this. Thanks so much for listening. If you enjoyed this episode, hit like, hit subscribe, share the teaser or the full episode with a friend who needs to know they've got to level themselves up and they've got to think about what impact they want to have on the world. Bye-bye. Bye-bye.