# Hermes Agent: Agents that grow with you |Episode #357| Practical AI · 2026-05-20 · 48 min · https://www.youtube.com/watch?v=UTZhvPXnmwA ## Transcript Daniel Whitenack I’m Daniel Whitenack, CEO at Prediction Guard, and I’m joined, as always, by my co-host Chris Benson, a principal AI and autonomy research engineer. How are you doing, Chris? Chris Benson Hey, doing great today. How’s it going, Dan? Daniel Whitenack It’s going really well. I’ve been excited for this one ever since a conversation we had back in 2024 on the show with one of the co-founders of Nous Research. Today we’ve got another one of those co-founders with us, Jeffrey Quesnelle, co-founder and CTO at Nous Research. Welcome. Jeffrey Quesnelle Hi there. Thanks for having me on. Daniel Whitenack It’s great to catch up. We’re big admirers of a lot of the things you’re doing, and we really enjoyed the last conversation. That was in 2024, though, so it’s been a couple of years. That’s like eternity in— Chris Benson Yeah, it’s 20 years in AI time. ### Open source and Hermes Agent Daniel Whitenack Yeah, 20 years in AI time. If I remember right, at that time Nous Research, which listeners might remember as a loose community of folks on Discord and elsewhere, was starting to become more of an actual company or organization. Could you help us understand how that’s evolved over the last couple of years? Maybe share some highlights from the Nous history books? Jeffrey Quesnelle Like you said, it was two years ago, so 20 years in AI time. We started out as a loose collaboration of the homies on Discord, as I always like to say. It grew organically through Twitter and other places. People were fundamentally interested in open-source AI and making sure this technology, which we think is the transformational technology of our generation, wasn’t going to get locked behind a few closed companies. At the time, GPT-3.5 had just come out and ChatGPT had happened. Those of us who cared about open source were wondering how it was ever going to survive. It seemed like the curve of capitalization was already starting, and if you didn’t have a ton of money, you were never going to catch up. But the research space was really nascent, especially two years ago, and there was a lot of low-hanging fruit—things that could be discovered by one or two people doing independent research. We brought together some of the people who had made those discoveries. That gave us credence to the idea that David really can beat Goliath, or at least get on the same field. We brought everyone together and asked whether we could form a company around these people who were doing this for free, in their spare time, so they could devote their whole experience to it. We formed the company about two years ago and went through a path of experimentation. The North Star was always to keep this technology open and make sure as many people as possible could use it. The question was what form that would take. We looked at different parts of the AI pipeline and where we could democratize things. We started with what we’d originally done: academic research. We published several papers asking whether we could take something extremely expensive in AI, or locked behind huge capital costs, and make 1,000x improvements to bring it within reach of people like us. We have some money, but it’s a rounding error—the catering budget at some of these larger places. A lot of things in AI get better if you throw enough money at them, because everyone is scaling so fast. But if you’re a little smarter about it, there can be big efficiency gains. We spent several years doing that research and putting out fine-tunes aligned with models that respect you and aren’t moralizing. That was our original Hermes series of models, which were very popular. This was before thinking and reasoning models, before the DeepSeek-style models. We had a lot of success, and people enjoyed those models. We also spent a while looking at distributed training and whether that was a viable path. We put out interesting research about how you could train models over the internet. This was also in response to political things that were going on—and are still going on—about what would happen if people tried to outlaw even using open-source AI. All of that coalesced around the idea of needing these 1,000x hacks. One interesting 1,000x hack we tried to take on eventually turned into what is now Hermes Agent: recursive self-improvement. That’s where a model is able to train itself or do experiments on itself and automate that part of the pipeline. You can think of high-end AI research engineers—the people who really know what they’re doing—as NFL quarterbacks now. You have to pay them like NFL quarterbacks, and they can determine whether your team gets to the playoffs. They’re that important to the structure of these large, well-capitalized companies, which also have better access to this limited talent pipeline. GPUs are a piece of it, but human brains are a piece of it, too. We asked how we could hack around that. We started building an internal tool for our model research team to automate some of the research they were doing and figure out whether there was something there. It was a “let’s see if we can do this” project. That tool was internally called Hermes Agent, and we started building it about six months ago. We used it internally for several months. If you go back to the genesis of Hermes Agent and look at the first open-source commits, it was basically all about model training. Everything we’d built into it was based on what we used internally for AI research. Once we had built it, we asked ourselves, “Why are we keeping this secret?” Our default is open source unless we can say there’s a reason not to. We released the Hermes Agent repository, and it was a floodgate of product-market fit, for lack of a better word. A lot has been said about what’s happened over the last several months with agents and harnesses, and I’m happy to talk about all of that. But it was right time, right place, and it’s affected every aspect of the company over the last several months. That’s awesome, because our North Star has always been open-source AI and bringing it to as many people as possible. We’d been figuring out the vehicle, the delivery mechanism, and in the last several months it’s crystallized for us. That’s the history of Nous over the last two years, in maybe five or ten minutes. ### Open source and geopolitics Daniel Whitenack That was awesome. You introduced a whole bunch of things we can dive into, so I’m going to pull you back for a moment. I want to go back, because I think there’s something tangential to what you were bringing us up to date on. The story of Nous can be very inspirational for people interested in getting into the space who see the big giants and the massive amount of money they’re throwing at it. At the same time, people have recently been seeing the perception of open source take a hit, with Llama coming toward the end of its life and Meta no longer championing open source among the big players. Could you lay out how you see open source working in the modern AI world? As you’ve formalized into an organization, moving from that informal network, you’ve had to make choices. You have this North Star of keeping everything open, but you also have bills to pay and paychecks to make. Chris Benson Can you talk a little bit about how you guys see it, and philosophically, how people navigate that space? If you look at Nous as a paragon to pursue and say, “Hey, I can go do that, too, to some degree,” how do you see that? What’s the role of open source, and what’s the role of being a company now? Jeffrey Quesnelle There’s the inside baseball version of open source: what’s actually happening in the industry. It has shifted several times over the last few years. We started with the golden age: Llama comes out, and one of the fabulous five companies, or whatever, says they’re going to do open source AI. It was driven largely by Zuckerberg’s personal motivations. I wouldn’t say emotional motivations, but he just decided he wanted to do it. He’s in charge of the company, so that’s what was going to happen. The fact of the matter was that this created an ecosystem of people who were able to research AI. A ton of the researchers now at different places started out by downloading Llama and things like that. So it was quite important in bootstrapping the AI researcher space we have right now. But as we saw with Llama 1, Llama 2, and Llama 3, the price tag kept going up. With Llama 4, they made some architectural decisions within the model. It seemed like a minor thing at the time; it could have gone one way or the other. But they made a few mistakes, and the model really didn’t come out that well. They had to look at themselves and say, “We just spent $200 million, $300 million, $400 million on essentially a paperweight.” That’s where the question really became: Everything’s fun and good when everything’s working, but that first failure case made people question the motivations for having open source models. ### Optimizing human interaction At one point, you have one of the largest American companies having a come-to-Jesus moment about why they’re spending money on this. Are they aligned with their business objectives? Then you have the DeepSeek situation, where a Chinese company drops a model that was trained and was, at the time, nearly state of the art. It came out of nowhere, and now it goes from being a dollars game within the Western AI space to having a geopolitical underpinning as well. Huawei was going to get here; it was forced through there. Open source was now squarely within the vectors of this geopolitical thing. It wasn’t just that there was a closed model provider out of China that was state of the art. It was that we had an open source one coming out of China that was near state of the art. Then this playbook started to get replicated throughout China as a growth hack mechanism. Within China, they realized that if you put out an open source model, you could get your company to the top and get it in the press. In America, it was realized that this wasn’t going to be profitable. As that fizzled out, it was still quite profitable in China to put money into open source models in the hope that it would elevate you to the top tier of players in the game, with the money stuff to figure out later. For various reasons, that may or may not work differently in China than it does over here. But the functional difference was that all of the open source models were now coming exclusively out of China. ### Leverage and human capability That’s more or less where we stand right now, except for what I’d call a very significant recent development: NVIDIA and Jensen Huang. At his last GTC keynote, Jensen came out and said NVIDIA is going to try to carry the financial flag for Western open source models. He committed something like $20 over the next several years toward training open source models of Western provenance. ### The future of human agency I think this is an interesting development because, if you pull all the threads on who it really makes sense for to train an open source model, I think NVIDIA is a company where you could credibly make that argument. At the end of the day, they’re all running on NVIDIA. It all ends up running on NVIDIA chips, and more things need to run on NVIDIA chips. So for them, it’s aligned with their business interests. Anything where you can’t really see how it’s aligned with the business interest will always follow the money. That’ll tell you the outcome. I think NVIDIA doing that is a potential game changer, at least on the Western side, because they’re really the only people who have it. You win some, NVIDIA wins a bit, you win a bit, but guess what? The house always wins, and NVIDIA gets to be the house here. On the inside baseball side, I think there are some interesting developments right now, and I’m currently very enthusiastic, or optimistic, about where it will go in the short term. We can’t predict anything several years out. That’s the inside baseball story. On the larger question of how someone fits into the new story, whether it’s replicable, and who you are as a person in this space, I’d say there’s never been a time when a single person’s leverage could be multiplied or maximized more. AI is a human capability multiplier. I’ll give you a concrete example: Hermes Agent itself. Hermes Agent was first developed internally at Nous by a guy named Technium. You guys might know who he is; he’s on Twitter all the time. Teknium is not a developer, in the sense that he doesn’t really know code that well. Before this, he wouldn’t have been able to sit down and write a huge program from scratch. Previously, he would have been unable to make an application. But with the current AI tools, he was able to architect, write, and build an application that is now the number one open source repository on GitHub. That goes to show you that if you have the vision, the drive, and the care, this technology can make you a thousand times more capable. There’s never been a time when a few people could have a more outsized impact on the industry and the world than right now. Don’t think the opportunity has passed you by. Every industry and every vertical is ripe for someone with a laser-focused vision to use these tools to completely take over and win. The grass is still green. Daniel Whitenack I’m wondering, now that we’ve shifted into this discussion about agents and Hermes Agent, about your perspective, Jeff. It seems like we’ve seen a shift. Initially, with AI, everyone was asking, “What model are you accessing? What model are you using?” You put in an input and get an output from a model, so you need to choose a model. Now everybody’s talking about agents or agentic systems. Some of that is jargon and marketing, but there’s also a fundamental architectural change happening. You mentioned agent harnesses and how these applications can operate over a distributed system of things, have connections to tools, take action, and maintain state. We saw, for example, when we looked into things people use all the time, like Claude Code, the importance of the agent harness. That’s also a big part of the Hermes Agent Pro project. Could you help me and our listeners understand the importance of the model in these types of agentic systems? How does that shift in a world where the important part, and a lot of the potential value and IP, is in the harness, and the model is something you can swap in and out? What is the importance of the model in that world? ### Brains, bodies, and harnesses Jeffrey Quesnelle Here’s an analogy that’s not perfect, but I’m finding it instructive: the model is the brain, and the harness is your body. Think about it like that, inside yourself. The brain was something we previously couldn’t build. That’s the true magic of AI: it’s still within the model. Speaker 1 It’s all within there. That’s the zero-to-one thing that allows any of this to happen. We could build robot bodies before; they couldn’t do anything. That’s really it, right? The model is your brain, and the harness is your body. Let’s pull on this analogy a bit more. Suppose you have an IQ of 150, but you’re an invalid. You’re stuck in a wheelchair. That will limit the number of things you can affect in the world. You can try to talk to someone and tell them to do something, but there’s some limit to what you can do. Let’s imagine that most of the work you want to do needs to somehow touch the world itself. You’d be limited in that respect. Meanwhile, if you had someone with an IQ of 95 but an athlete’s body, there are more things they’re able to do just by virtue of being able to manipulate the world more. Both sides are important, and I think they exist on a sliding scale. A better harness with a worse model can beat a better model with a worse harness. They’re both symbiotic and work with each other. The harness itself is the thing that ultimately touches what we humans care about: the world we live in, this thing we call reality around us. The model itself has very little ability to touch the world. All it can do is output tokens. We need to allow those tokens to become instantiated in the world around us. We ourselves are instantiated in a world where we have persistence, where we have causality, where time moves forward in a direction. The harness allows the model to exist within the reality we all care about. As far as ownership—where the value accrues and those sorts of questions—I don’t want to get ahead of my whole story and make some kind of Silicon Valley pronouncement about where all the money goes. I’d phrase it differently: we’re getting to the point where people will want to pay for outcomes. When I hire an employee, I’m paying them to do a job, and ultimately I want that job done. Otherwise, I wouldn’t have hired them. The prime motivating factor for me is getting some material, economically valuable outcome, or something I view as economically valuable. ### Hermes Agent architecture I think that’s the similar model we’ll eventually get from this whole AI system: these systems will be paid to do certain things, and they’ll be valued insofar as they get the job done. Who owns the harness, who owns the model—all of that will get worked out in the end. But if you pull back and ask what the prime motivating factor is, it’s that I want to pay for something to get done in this world. It would be cheaper, faster, or more cost-effective for me to do it this way than any other way. That’s water going downhill. Follow that path, and it’ll take you to the promised land. Speaker 2 I love that analogy about the brain and the body, the harness and the model. I think that will be really helpful for people. I’m wondering, in terms of Hermes Agent specifically, could you walk us through some of the agentic or automation work you were seeing out there as you started to develop Hermes Agent, and some of the key opinionated choices you made about what you wanted Hermes Agent to be, uniquely, among all these other things out there? Speaker 1 The key idea was that Hermes Agent ought to get better the more you use it. That was the motivating factor. Maybe that’s different from how people used to design software. We’ve always been trying to break out of the idea that the paradigm is the same. We have to act like the paradigm is different. How did people used to design software? They’d say, “I think I know a problem. I think I can design the best solution for that problem.” They’d whiteboard it all out and build all these systems. We said, “We don’t have to be like that anymore. The AIs are smart enough to see what we do and figure out what they should be doing.” The idea is to get out of the model’s way. Give the model the ability to do what we want it to do, and lean on it as much as possible. One of our opinionated decisions was to create a very limited number of built-in, hard-coded features. We gave it the minimum ability to touch the world, run code jobs, browse the web, and things like that. Everything after that is an emergent property that we encourage through the prompts. The 2 main ones are the memory system and the skill system. The skill system is extremely interesting to me because the first time it happens to you, it feels like magic. You use the agent to solve a problem, it does something for you the way you like it, and it notices without you telling it that it’s learned something important. It creates a skill that says how to do a specific thing. The next time you do something similar, it uses what it figured out before. Here’s an example. Our team was going to an offsite, and we were flying in. I wanted to see if I could get Hermes Agent to make a restaurant reservation for me during the flight. There are a bunch of things on those websites to keep automated systems from doing that, like bot captchas. It was going around, saying, “Okay, it’s not working. I’m getting blocked.” Eventually it found the API backend they use for bookings, and it said, “Oh, I see all the things here.” Then it said, “Okay, I found the data. I booked— which one do you want? Book it.” This took about 30 or 45 minutes of actual agent runtime. At the end, it said, “Oh, wow. I learned all these things,” and you see a skill created: “Las Vegas restaurant booking.” The next day, it knew how to do it. I came back and said, “Dinner was great last night. Can you book this other one?” Boom, instantly it knew exactly how to do it. It took those learnings and applied them. Rather than designing all these different skills, we let the model do it. It’s encouraged through self-reflection. Internally, in our prompts, we have a couple of people who I’d consider to be the world’s best LLM prompt engineers. They’re LLM whisperers. They talk to them all day and know the right words and phrasing. Karan is number one in the world, I’d say, in my opinion. We put prompts in to nudge the model to reflect: “When you see things being used all the time, or if it feels like some important achievement has been met that would be useful to you in the future, make note of it and put it in your skill system.” That’s all we really do. As I said, we get out of the model’s way and let it use its judgment about what it thinks is important. We encourage it to do that. We give it guidelines for how to think about how to solve the problem, not how to solve the problem itself. Then we let the model go. That was our first opinionated decision: let the model drive it. The second was the memory system. We have a hierarchical memory system that creates notes about you and how it seems you want to be treated. There’s also a layer of previous sessions it can go through, and a multilayer memory system as well. The model’s own discretion mainly determines when to activate those pieces. What that means is that when GPT-4—or the next GPT, the next Opus, or whatever—comes out, Hermes Agent just gets better. It has better judgment in using those things, and the new model drops in and gets better. Chris Benson That sounds really cool. As you’re talking about this, people listening to the podcast may be wondering how to position it. One thing that has hit people is that a lot of harnesses have come out over the last few months. Some are proprietary, some are open source, and they have different capabilities. Daniel Whitenack There’s just a lot of overlap. How would you position Hermes for both users and the architectural context, like cloud versus local? There are a lot of different options. As you’re working on this and trying to get it out there, what use cases and specific audiences do you see it being really well suited for compared to the competition people may be looking at? Could you talk a little bit about its market fit? ### Enterprise and multi-tenancy Jeffrey Quesnelle Its primary market fit first is the power user who wants to use AI to its full capability. That was the first segment we aimed at. Like I said, why use it? It’s the agent that gets better as you use it more. That’s the best tagline we can come up with. If you want something that just works and gets better the more it works, Hermes Agent is the agent for you. It’s open source native, in the sense that we keep everything open. You can run it locally using local models if you want; that’s a first-class option for us. Or you can have us host it for you. Wherever you fit on that line, that’s where we are. We feel pretty good about where the story is right now and the broad product-market fit. We’re currently working on expanding what it means to have an agent like Hermes Agent in a workplace or collaborative setting. I think we’re pretty happy with the one-agent, one-person model as it is now. We’re exploring multi-tenancy: how do lots of these things work together? We have a couple of high-level partners we’re working with right now, but we don’t have anything to announce yet. That’s where our thinking is going. We’re dogfooding this ourselves 100%. Hermes Agent is written 99.99% by Hermes Agent, and we also have several Hermes Agents deployed within our organization doing real work, to help us figure out the enterprise, or at least organizational, story. It’s really powerful. I’ll give you an example. We have one agent whose job is to connect via MCP to all our back-end infrastructure and servers. It can access the databases, but it’s read-only, so it can see everything that’s happening. We first used it as a debugging tool. If someone came to us and said, “This isn’t working on the website,” our systems on the back end were kind of spaghetti: databases, Vercel, and all these different providers that get it all going. It was a bit of a mess, with all these different services. We started asking, “Can you figure out what these problems are? We’re seeing 500 errors in this area.” Our engineers who know the systems would talk to the agent, and it would say, “I can’t find this. I’ll go look here.” Over the course of about a month, it built up a huge repertoire of skills for doing what we need within the company. Now our support person, who works in our Discord or handles email, can go to that agent, ask about a specific customer account, and it can do what we need, provide reports, and find root causes. That wasn’t something we had to code ourselves. The skills emerged as the engineers used it. The difference is that an engineer had to use it once, show it how, and work through it once. The agent then passively got better and better. Now the whole organization can use what that one engineer taught it. We’re trying to work out that story, too, and we think it’s quite powerful and interesting. That’s where we’re taking it from here. Daniel Whitenack I have an interesting question. At least, I hope anything I ask on a podcast is interesting, but this is particularly interesting to me. Chris and I have talked about this on the show a lot: the way you should automate or mechanize things is often different from the corresponding human process. You shouldn’t assume a one-to-one mapping. Since you’re finding that the agent develops those skills, have you learned from how they develop how to do things better as humans within the company? On the flip side, what are some creative ways you’ve seen agents solve things that might be totally different from how a human would approach the problem, but work because you’ve fully bought into the agentic process? People are gradually learning this through agentic coding tools. There are some sacred cows they have to get rid of, right? It doesn’t matter if a function is duplicated in a couple of places. That’s fine for the agentic system; I don’t need to think about it. Have cases like that come up for you? Jeffrey Quesnelle As a guiding principle for what to automate and when, think of agents as humans with infinite patience. They have infinite patience but very low creativity. The question is, where would I need that? If there’s a place in your workflow that doesn’t require creativity but requires infinite patience, that’s a great place to put an agent. With our log system, theoretically, a human could read all the logs and step through them, but no one has the patience to do that. No one has the long-context memory to see it all at once. Those are the sorts of areas to think about, especially if you have to do it all the time. What you don’t want to do is make the one-to-one mapping you mentioned: “We need a CEO agent, a CFO agent.” That’s the wrong way to think about it. You need to think operationally: what is the thing actually doing? A human might be able to do it, but would they really want to? Those workloads are the best ones to put through an agent. We start with anything that could theoretically be done by a person who isn’t a genius or super creative, but that doesn’t make sense for a person to do. As for what we’re learning about solving things differently, I’ll put it this way: I started coding when I was 7 or 8 years old. It was the thing that made me different growing up. I could do this cool thing, and it opened every pathway in my life. All my success has been rooted in the idea that I was a really good coder at a specific point in time. I was born in the late ’80s. I got to be there right as the internet started coming on, when software was starting to eat the world for real. I was young at the time, and all those things lined up. I created a whole bunch of things. Until about 6 months ago, I didn’t really use AI at all. I wrote all my code myself and thought I was still better than all the AIs. I thought they couldn’t do what I do. Then, in the last 3 to 6 months, through using Hermes Agent, I had to admit that time has come and gone. Rather than being disappointed about that, I’d say it means I get to run the clock back and be early again, like when I was young and software was eating the world. That’s freeing in some way, if you’re willing to let it be an opportunity. To get back to your question about learning to do things differently, I’d say anything aesthetic is based in aesthetics. The AIs have no discernment for that. They don’t care if something is pretty. It’s not even that they don’t care; what I’m finding is that they really don’t know. They don’t have human taste. They’re incredibly smart in the way we’d consider a genius person to be smart: someone we’d try to select for Stanford, someone who can ace SATs, solve math problems, and write code. We needed to win the genetic lottery to do that. They can do it right out of the box. Jeffrey Quesnelle But what they can’t do, and what almost every human can do, is discern—to have an innate sense of aesthetic. I think the reason is simply that our human sense of aesthetic is grounded in the lived experience we all share as common men and women on this planet. AIs are so divorced from the experience you and I went through: being born, living in time, growing, and even our biology, shaped by the evolutionary pressures that created us in the first place. Our aesthetic is quite unique. There could be a million different ideas. Animals have different ideas of what’s aesthetically important in their pack structure, and there are a million versions of that. There’s only one math, really, and there’s only one science. It’s interesting that the universal computational things are what the AIs can take in. The things that are very unique to us don’t feel unique because of the fish-in-water argument: we’re so immersed in them. It’s only when we’re faced with something that’s finally another example—when there was only us to begin with, and now there’s another example—that we look at it and go, “Oh, why is this so different?” You can think about it as slop, or whatever you want to call it. Getting back to what you said, they will solve things in an aesthetically unpleasing way, and you just have to learn to deal with that. At least right now, you shouldn’t rely on them for things whose success depends on being aesthetically pleasing to a human, whatever that might be. Chris Benson Yeah, I want to comment on something you said a moment ago and then draw out the thread you’ve been following. I think what you said connects to something I say a lot: the models regress to the mean. They may appear creative, but they don’t do truly novel things. They have to have something they’ve been trained on that they can associate with. If I’m doing something truly novel, there’s nothing written that they could have been trained on beforehand, so they’re usually not very useful in that way. Having said that, that still leaves 99.9999% of the usage out there. One thing I’ve noticed, not only in our conversation today but in previous conversations we’ve had on the show and privately, is that we are retraining ourselves. With the capabilities of a Hermes agent and what harnesses can do, and some of the things you guys have done to make Hermes so amazing at learning skills, we tend to go to the common use case of coding. Daniel pointed out that maybe you should keep just one version of a function. We all grew up hearing, “Don’t do that; keep everything really lean.” But that’s really for us humans, and you pointed out that we have so much more capability now, though it may not be aesthetically pleasing and may require a different way of doing things. As people use a tool like Hermes Agent, how should they think about opportunities to do things better in ways they’ve never thought about before? Let’s stay away from coding for a moment and look at the rest of the world of work activities. How should they look for those opportunities when it’s no longer human-centric, but agent and model-centric? How do you optimize the use of a tool like Hermes Agent if you’re the human trying to retrain yourself? Jeffrey Quesnelle Yeah. Number one would be: don’t tell it how to do something. Describe the outcomes and the conditions for what you’re trying to get to. Like I said earlier about getting out of the way of the model, I think people, when they first start using these things, want to keep control. They want to say, “Do this one step, do this one step, do this one step.” The models are only going to get better and better at long-horizon planning. Ultimately, I think you should be goal-oriented and evaluation-oriented: what would be the conditions under which you’d think the goal was met? If you can describe those succinctly, that’s what the agent can work toward. Let it work toward fulfilling those success criteria. If you leave those success criteria unstated, let me give you an example: you’d say, “Write something,” and then you’d say, “Well, this is slop. This is not X, but Y. Not X, but Y. All the paragraphs are the same length.” But you told the AI to write something, and it did write something. It doesn’t have the judgment that it shouldn’t sound like slop, because it thinks slop is the best thing. It’s the mean average of every word that’s ever been written, and it thinks it’s great. Your unspoken assumptions or evaluation of the outcome, if left unspoken, will remain persistent. That’s why, as I said, we have these LLM whisperers who can articulate all of their internal suppositions about what they want in a succinct way. For humans, that’s its own skill set. Knowing how to actively communicate everything you want, think, and say is not something we normally do. We have the shared history of human common understanding. I don’t have to say it to you because you and I feel the same way about most things. It’s so ingrained that I don’t know how to put it all into words. Instead of thinking, “Explain it to me like I’m five,” explain it to me like I’m an alien, because that’s really what the AI is: an alien that never grew up on Earth. You have to say what you want much more explicitly. If you can say what you want explicitly and write out your evaluation criteria—everything that goes into your own judgment—it can find its way toward that goal. Chris Benson As we draw things to a close, I’m excited by the possibilities moving forward. As you say, it’s a time when we can look at a lot with fresh eyes and experience some of that newness, even though we’ve been working in technology for a long time. We always like to ask our guests at the end: what’s on your mind as you step into the next few months or year of your life? What’s at the top of your mind, whether it’s something exciting, a challenge, or something about the ecosystem? We’d love to hear it. Jeffrey Quesnelle I have the business answer, which is making Nous Research continue to grow through our enterprise offerings that we’re working on. I think about that a lot. It takes a lot of my time. But something a little more overarching would be thinking about the place of people in work in the years to come, and not only from an economic standpoint, but from a humanity standpoint. What will using these tools do to us? Society is already moving toward a doom-scrolling, TikTok-y kind of tic. This could be one more domino in that effect, because now we can check out of thinking as well. If your job is just telling the agent, “Do the thing, do the thing, do the thing,” over and over, what does that say about your growth as a person? At Nous, we call ourselves human-centric AI. AI should make you better today than yesterday, and better tomorrow than today. I think we’re on that path right now, but it also requires responsible usage by the people using it. I wonder how someone could get trapped into developing no critical thinking for themselves. If you take this 10 or 20 years from now, what happens to someone who only ever grew up with this as the paradigm? Human innovation, thought, and critical thinking might be things they don’t even see. The answer is the computer knows, and that’s the only answer there ever could have been. It only takes one generation: whatever you grow up with is what you think is normal and right. I had a baby 5 days ago, and I think about what it means for her to grow up and live in that. That’s what’s on my mind right now. But ultimately, you can take that one of two ways. You can be pessimistic and be worried about it, or you can have agency. You can have agency and be part of the world and try to be part of the solution. So that’s why Nous Research exists. That’s what I’m trying to do. Like I said, better today than yesterday, better tomorrow than today. Speaker 1 Awesome. Great way to end it, and also, congratulations. Thanks for joining us on the show. We do hope to have you back. Maybe someone from Nous Research will be coming back on the show in less than 2 years. We’ll try to make that happen next time. Thanks for joining us. We’ll talk to you soon, Jeff. Really appreciate it. Jeffrey Quesnelle Thanks. Bye.