エージェント製品開発の過酷な現実 ― Daniel Jeffries 氏と語る #782

Until recently, the big problem with personal agents was how hard they were to use. Sure, you could set up OpenClaw over a weekend, but was your non-technical mom going to use it? Over the past few weeks, though, AI agents have crossed a key accessibility threshold. You no longer need to be a geek to get started. Meta's Muse, OpenAI's Dock, and a growing list of others are ready to go right out of the box. So, is it game over for people building agents? Not at all. It's gotten harder to stand out, but there's real opportunity in the opinionated design choices that don't make sense for a general purpose product and the unique experiences and workflows those choices make possible. My friend Dan Jeffries has spent the past six months making exactly those choices while building Nautilope, a platform for personal and collaborative AI agents. And he's developed some pretty strong opinions along the way. For example, a single agent that knows everything about you needs to know when to keep its mouth shut. And he thinks a lot of the elaborate multi-agent org charts we're seeing are basically cosplay.

We get into how he's approaching persistent memory, coordinating work across people and agents, keeping private context private, and where the model should stop and the tools should take over. Tools are dumb hammers. The hammer that has no intelligence at hammers things. So I believe most of the stuff is like if you are exposing the machinery. In other words, if the LOM has to shuttle the data from here or make the hash structure or put something in the database, you screwed up your tool. The agent should supply the fuzzy, intelligent decision-making and that the tool should do the work for you. So it should work like a database. So if people are like, I use Jev to clean up my desktop, I'm like, dude, you could have done that with a shell spray. I'm Sam Sherrington, and this is the Twimble AI Podcast. For over a decade, I've been exploring the ideas and innovations shaping the future of AI through conversations like this one that help you understand what's real, what's next, and what matters. Let's jump in. I realized, you know, we're sort of in, like, if you look at the diffusion of it,

the plastic diffusion of innovation curve, I was trying to build products for people in, like, the early adopter, early majority. We're still in the 2%, like, bleeding edge beginning. Like, we're in the, really, if you're building. Do you think we're that early? I do. I really do. It gets like the early, like late 80s, early 90s, where you just accepted that your computer would mess up, right? And you'd have to know how to put it back together, right? You would expect it. And we're not, we are actually not in the iPhone 56, Mac 35 era where you're like, where it just, you expect it to work perfectly. You don't know how it works. And where if it breaks down, you're like, what happened? And who do I call? Like, how does this, what happens, right? So we're definitely not there. And so I wrote an article on what I call the agency problem, which is like a lot of times people don't know what to do with agents. And only kind of like a small percentage of people do. And I tend to find they tend to be people who are very high agency, right? They tend to be people who are like, I have 50 things I always want to do, like five businesses I want to start, 30 books I wanted to write, three movies I wanted to make, 50 trips I wanted to take.

They've been in the backlog for forever. and now I've got an infinity canvas and it can do anything. And it messes up. It's like that thing, it's like the commercial where you become this manager and it's like, well, this employee comes to you and he's kind of an idiot and you've got to take the keys and redo it. And this one comes to you and is like, I did the job, boss. And you're like, no, you didn't. What is this? And another one is like, oh, I did it. And you're like, oh, that's almost right. Hold on a second. Yeah, that's good. We'll just tweak this and this. But you don't know which of those you're going to get in the slot machine, right? And, you know, and that and so there's this weird management era and then certain things just don't work. I share this tweet not too long ago that kind of characterize this experience for me. It was like a pie chart of 2026 with a 10 percent sliver of me actually doing the work and 90 percent sliver of me working on the agents that I'm trying to get to do the work for me. You know, some combination of, you know, coding, review, iteration. like, you know, it's both frustrating, but also it feels super productive in the long term.

Like, you know, there used to be a time when, you know, I get frustrated when I have to do repetitive tasks that I don't think are interesting or valuable. And so I'd spend a really long time trying to, you know, write the macro in Excel so that I didn't need to copy all the cells and do all that shit. And now you can go down that path super easily and really quickly automate a lot of things, but those things break all the time or you end up finding the corner cases and you need to go back into this code. I agree with your point about this management and delegation era and how it allows you to expand your plate and do more. I had a best friend who was a pretty good tech guy come to me and was like, I want to know what to do with agents. How do I get started? And I'm like, the question sounds sensical to me as a person who is like in that two percentage, like weird high agency, semi-autistic kind of people, right? Who is like, you know, and I'm like, I don't understand.

What do you mean? How do you get started? Like you talk to it. And then when it messes up, you talk to it again and you figure out what's wrong. You critical think or whatever. But there's like a lot of people are like, well, I just want to, I want the, I want the template, right? I want the thing that tells me how to do the thing or whatever. And so I would say it's just not there yet. It's not in the like, you know, it just hasn't gotten to that phase. And so you spend a lot of time just trying to get the agent to do what you want. There was a great essay about it called Two Shaped Objects. Okay. And they talked about a favorite, a famous Japanese craftsman wood planer called Akana. It's a huge piece of work of art. takes multiple craftsmen months to make and set up for you and all this stuff. And when, you know, it shaves wood with like angelic precision. It's also economically totally worthless. Okay. Because a power planer will do it for a fraction of the cost and like in a fraction of the time. So like the actual problem in much of AI today is that a lot of it is the setup exists.

Like the con exists so that the setup exists. and when you see a lot of the videos are like, this is insane, I just rolled out the new thing and you're like, it's 90% of them setting it up and then you're like, okay, great, now they're going to do real work with this thing, right? And then that part of the video doesn't happen, right? There's no work being done or it's like I set up my 15, what I call the faceless toaster bots, right? I set up my email bot and then my CEO bot to manage them. I'm like, you're just LARPing as a CEO. No, you're just cosplaying. You're not actually doing anything, right? I'm like, what are you doing? Well, I'm getting the agents to talk to themselves, but what are they talking about? Why do they need to talk? Isn't that the same model underneath? Why are they talking about? Like, what's happening, right? And so I think that there's a weird disconnect where for high agency people and people who've got an insane backlog, the infinite canvas is amazing. And for many other folks, they still are looking at it going like, I don't know what this does for me. And they're looking at it with a bit of fear, right?

They're thinking maybe it'll take over. It's not going to take over. It's not going to take the jobs. The jobs will change over time. I've always said this. I'm like, we've already destroyed all the jobs a thousand times in history. We always change. We always adapt. So the last time we spoke was end of 2024. You were working on computer use agents. That was a great conversation. I'll refer folks back to it. But, you know, talk a little bit about, you know, what you've been up to since then, and how you've evolved the way you're thinking about what you're doing at Kentaros, and in particular, what you're doing with Nautilus. Yeah, so I'm still trying to solve computer use. I don't think that it's solved. I think I've seen a lot of tweets otherwise that promise me it's solved, and it's not. Okay? Like, Jeb is still not a visual model. Jeb doesn't always make great decisions, so that means you've got to translate the window from Gua into, like, a non-visual sense. So I've been trying to do a lot of stuff there to help crack it. You've seen a lot of pre-baked workflows that are not actually correct,

that are not generalizable. And you still have the round trip to the cloud of, like, if you're going to take screenshots. So where we were at was trying to micro-train agents to do one thing really well. But the problem is it didn't work too well, right? It kind of worked. It generally worked the way we wanted with Agent Tutor. And so we were trying to say, I'll be able to kind of just train it to do this workflow. But it was just, it's just slow. Like now the models are much better at screenshotting and understanding coordinates and all that. They're all trained on it, but you still, you don't have a micro model that just lives on your desktop that's JV fast. that said models like Astor and Claude are pretty good if you don't need to move at human speed for software testing for doing DevOps, ClickOps style work for you whereas in the past as a system I'd have to go in and figure it out and type the commands myself like a lot of times now you can ask it

and watch what it's doing and that's pretty cool so I think there's been a lot of progress there but it's still not fully solved Right. So then not alone, what we pivoted to is we've been burning on the concept of personalized AI. Right. I think a lot of people are burning on it, but I think we're thinking about it very differently. So not alone, I think my friend joked is 30 products in one, not one. But the one liner is it's her, but for teens or families. Right. In other words, her as in the movie, as in the movie, her. right the flip phone version of her okay not not the full os version because you can't build that yet not the thing that can plug into anything that has a perfect personality but what we did is we built we built incredible personality customizability radical customizability into what we call our genies our agents and they act like secret keepers we call like they act like secretaries in the old form of the words secretum which you know which means like in the latin which means secret keeper and they were the the keepers of powerful people and their effects right and now

everybody can have one and we we wanted them we had to we we had to rebuild kind of group chat and we have a magical router in there that'll allow like the agent to just wake up by just talking naturally and and you'll be in there with people and with your agents and that you can just chat naturally and they all seem to have their own personality keep their own memory they don't really seem to break that they we have an amazing memory system you know you can't really do this in the, we've been asked a lot of time, like, well, why didn't you just plug in every messenger in the world? Well, you can't just magically route in there, and you're also violating the terms of service. They don't really want you there. Those things were based on, their terms of service were built on when bots were a spam, right, to be kept out. So even if you go get a business license from WhatsApp. Are we talking about, like, Slack, for example? Well, yes. So Slack, you can put them in there now with, like, you know, a quad tag and things like that, but it tends to be a singleton AI. And now that a diverse, you know, AI, right. That has like a lot of different personality quirks. And like, you have to add it every time, like at,

and it wakes up and those kinds of things. Right. And so I felt, you know, we needed to like reconstruct that. And we also needed to, we needed to give people self-sovereign AI where you control the backend, the front end, the mobile clients completely as open source. And whether I win or not, I think this is the wave of the future. I think we have to melt down all the software and recraft it. And I do think it has to be self-sovereign because, as Francois Chalet said, AI is going to become our interface to the world. Like I say to people, and I think that it's crazy, but you won't even know how to open an application on your phone in a decade, right? Like you'll just talk to it, like in the movie, and it'll do things for you. And the problem is that that's another SaaS application. and all the personalized AIs right now, and we're up against the wall of free compute and infinite budgets. And, you know, and so it's very hard to break through the noise. And people don't fully understand what it is. Like, because it's going to be our interface in the world,

it's going to be much more intimate what you share with it. Okay? And we feel that if that is a SaaS application, okay, that it is surveillance economy squared. Like if you think that what happened like with social media is bad to society, it's going to be infinitely worse when you're sharing your kids' problems in school, the mall you're worried about, you talk to it about a legal problem, and it shows up in court, these kinds of things, right? And the only way to do it is it's not going to be passed by law, right, is you've got to go back to kind of the cypherpunk phase. So we're building in a spin on end-to-end encryption into it where the back end is distributed keys. And unlike end encryption where it just humans a genie needs to read stuff in plain text So it gets a time delegation to wake up and decrypt an asset decrypt a chat decrypt a memory et cetera and talk to you and then it stops right And so we think that these kinds of things you have to have a cypherpunk solution of self AI We don't currently support local models, but we will.

It's pretty easy as they get better. We support things like Venice AI, which is almost like a VPN for AIs, right? And we rate the models by privacy. You've mentioned a bunch of different things. Let's go back on that and zoom in. I guess the first one I want to start on is multi-agent and your experiences with multi-agent. I've gone back and forth on this, both through conversations and interviews and my own experience. I get the general idea that multi-agent makes sense from a, you know, basically due to limitations of the models, like keeping context about a particular domain constrained within an agent that's responsible for that domain. But at the same time, my experience, I do a lot with Hermes. I have it dealing with a lot of different domains related to mostly business, less personal. And it does just fine, like pulling the skills it needs for a particular problem, but a single agent.

And so I kind of want to hear you talk a little bit about, like, why should people even care about multi-agent? You know, where have you found that that's really useful? And, like, what's required to make it, you know, useful? Well, first of all, I love the Hermes team. I think they're fantastic, and they do incredible work. I love Hermes. I love everybody on them. I follow them. They follow me on Twitter. You know, a lot of the folks there, Techum and Shannon Sands, their aliases online. I think they're just incredible developers. I think they have incredible philosophy. I love them. I believe we actually have the same philosophy. I believe in a super powerful G that is a super powerful super agent that knows everything. It's a pets versus cattle thing. Okay. The agent can spin up the cattle, but I don't want to know the names of the cattle. Okay. They can spin up sub agents and do things, but I don't want what I call the faceless toaster bot. I don't need the email bot. I don't need the tax bot or whatever. And then I make the CEO bot to manage them like I want. Yeah, yeah, yeah. Right. So that's not what I mean by multi-agent.

I mean everybody has an agent radically customized to them. They can do anything. We have things. It can do video editing. We have co-creative apps built right into it where you can take over and trim it. You don't have to do what I call prompt and pray where it's like it makes the video and then you're like, oh, no, I need to cut out this two seconds. Please cut it. Oh, no, iteration 56. You can cut it yourself or you take over and write the MD file or the Word file or the Excel file right in there, right, or they take over. The terminal, I have a let Genie control it, whatever. So what I mean multi-agent is humans and agents on an even playing field in a conversation. Now, there's a lot of work that you have to do because the agents are not really used to talking to each other. So you have to build a bunch of little classifiers and back-end things that route and tell them to look things up, what memories are shared, et cetera. And we also include things like namespacing. So this is a super important thing. You see things on the news where like Muse basically gave someone's home address and things like that. And that's because it knows everything and there's no namespacing.

We built namespacing into the core so that when you talk about something with your agent in a personal chat, one-on-one, you're always talking to the same super agent that you personalized. You gave it a name. It's a picture. You picked its 11-man voice from 30,000 things, right? You customized it. So all these kinds of things. It keeps its memory radically just like in Hermes. I think they did a great job with memory. and you do a great job with memory. I've never seen my agent kind of break character, if you will, right? And then you talk to that specific agent about things, and they can do all those things for you. If it needs a sub-agent or whatever, it does that. So I don't believe in kind of the LARPing approach, the cattle approach of like, you know, I need this in a 30-agent start-up. This is my marketing agent. This is my, you know. You don't need it. It's the same agent. Yeah, yeah, yeah. The same agent underneath it. But you do have a marketing channel where you and your team are talking about your go-to-market strategy. And here's an Ask Me Anything doc that we wrote for X. Or here's our initial tweet.

Genie, can you check out Genie1 or Genie2 or Genie3, which they have their personal names, right? Can you go and check this thing out and make an edit on it? And then your team reads it and goes, I don't really like that line. I'm going to take over and write a little bit. But that's a convenience for the people, not because the agent, you know, needs to be separated into these different contexts. Well, it is a convenience for the people because you end up doing what I call like meat space delivery or like pigeon delivery, right? So like imagine like we could have an agent where let's say this happens all the time. You say like you're going on a trip next summer or whatever, right? And you go to you do about you have your agent go do a bunch of research on it. and then you now become the interpreter for your wife, copy and paste it somewhere, and she's like, well, what about this part? And you're like, well, you haven't fully read that part, right? So now you've got to go read it and be the meat space interpreter, right? So what I have, what we've been doing is like. How do you know what's going on in my life, Dan? Yeah, because it's happening in everybody's life right now, okay?

My wife is not reading that long chat TPT conversation that I had with that agent. No, the first thing I do is, okay, now give my wife the TLDR thing. So the thing is, I can take the agent. She made the report. I could take it and say, take that memory and that report. Go talk to my wife and get her opinion and come back to me. And it'll come back to me with what she was saying about it. Like, I think we should look at this list. So the agent has the context. Let it be the agent context for that thing. Right. And then it can always search history. It can always go through the history and go, you know, who was I talking with about this? And that one agent you were talking to was like, oh, we were talking about this. this stock trader, this person was in charge of this or whatever, and it'll look on those things and understand. And it's a context keeper and a secret keeper. So it's not just convenience. We're acting like meat puppets or pigeon carrier delivery between people when we don't have full context. So to me, that doesn't make any sense. So in talking through kind of multi-agent and some of those usage scenarios,

you describe what I heard as like a lot of context that the agent needs to be able to manage and manipulate. And are you finding that modern models, you know, have large enough context windows that like you can be kind of, that that's less of an issue for you as someone building the system? or are you doing like super-surgical, applying super-surgical heuristics to context in order to make it all work? Like, you know, how are you approaching kind of dealing with the context limitations inherent with models? There will never be enough context for me. They're all Mr. Me-seeks, okay? So the way that we've solved it is I think a number of memory breakthroughs, we're probably going to do a paper on it, where we've got like compressed fashes that are expandable, like to like get a little bit more context, a little bit more context, a little bit more context. We've done the old saying again of what computers do well,

which is searching retrieval and scale and brute force. And we've given them a tremendous retrieval system. So the first thing you'll always see the genies doing is looking stuff up, looking through the conversation history, looking through like memories that they have. They have tons and they can keep a ton of memories because they have the hashes put in without it trying to stuff every full memory in that they can then go retrieve more of that memory and the memories decay over time. They'll decay the memories. What is the hash? The hash basically is a compressed version of us to say like it's almost like a lookup list of like. Meaning so you've got a, like the agent at some point has to be reading plain text. You've got this plain text and the hash is going to other things. Correct. You have a little bit of metadata and a hash that's like a clue that it can go, well, okay, I think these hashes are probably the things I need to go check out a bit more. Let me expand that and get, like, more than a snippet, right, of that, and then another snippet is, you know, nope, these three I don't need. This one is probably it. Let me expand that one fully and read it.

Okay, and then kind of come back before I open my mouth. That has been tremendously effective in terms of, like, I'm often shocked with, like, how well it seems to remember things. And like I said, I've never seen our genies break character, which I've seen. And we're also sometimes shocked. with as they build up their own personalized context over time, either in group chats or about you, even when they're using the same models, they tend to have a diversity that is sort of shocking to us, right? And we can't fully explain it, actually, even at this point, right? So are the agents, like, building this kind of hash memory table structure thing constantly? Is there... Deterministically done things. Got it, got it. So that's the hornet that's doing it deterministically. Correct. Correct. I have a philosophy of tool design that I call, and it's got a little fuzzy lately, and I'll explain that, but it's what I call the dumb hammer design, okay, or the magical hammer design, okay? And what I used to see in the old agents is agent builds magical capability into the code, like the code makes an LLM or human-style decision in code, right?

Like a fuzzy understanding is determined. And so I would see old agents do this, and I called it the magical hammer problem. And so I have a rule in there where I'm like, tools are dumb hammers. The hammer has no intelligence that hammers things, right? And so I believe most of the stuff is like if you are exposing the machinery, in other words, if the LLM has to shuttle the data from here or make the hash structure or put something in the database, you screwed up your tool, that the agent should supply the fuzzy, intelligent decision-making and that the tool should do the work for you. So it should work like a database. When people are like, I use JEV to clean up my desktop. I'm like, dude, you could have done that with a shell spread. Okay, you didn't, that's, you know what I'm saying? Or they're like, I'm going to use JEV to like, you know, to strip PII. That's secondary defense in depth. The first thing you should have done is used a dumb hammer, right? You should have used like a, you know, a search for like certain patterns, right?

With regexes, right? You know, and millions of tools that are out there. And then on top of that, blur the faces and things like that that are hard, right? I've kind of evolved a very similar philosophy about tools. And it's not so much about dumb versus smart, but it's about like narrow. like I've had a lot of success with okay this one tool it's a little thing it does this one thing and the agent if it needs to do this one thing it doesn't need to think about anything it just needs to call this tool with the right parameters and the thing gets done like create an invoice edit an invoice do this do that like no I'm not just I'm not giving you the agent my keys to QuickBooks and have you go figure it out by looking at the schema no, these are the things you need to do. Do these little things and we'll be a lot better. And it's worked out really well so far. I believe in the Unix philosophy. And the Unix philosophy was always that, like, every tool does one thing well.

Right? Like, and the Windows philosophy back in the day was I have a monolith that does 80 things poorly. Right? But, like, the Unix philosophy was always, like, greps. You know, less is more. More is more. Right? You know what I mean? Like the PS gives me the processes and filters them. That's it. Okay. And Grep does it even better, the filter. Right. And Auk and Sed do it even better. Right. In different ways. So those are the things where, like Auk and Sed should do a lot of the heavy lifting for you. And then the agents should apply the fuzzy context. Now, I will say that with the new decision style models, and I do, I kind of ragged on Jeb a little bit, but I think it's freaking awesome. And I think I'm really looking forward to the decision visual versions of it. I think it's the kind of thing where you make a small model now and it's going to be the type of thing where local people, you know, local models can start to excel again because they can fine tune on these things and they're going to be visual. I think I think these are really cool where they can make lots of decisions in a short period of time. Now, that does kind of give you a magical hammer where you can kind of build in like, which way should I go?

Right. Should I go left or right? Should I like jump up and down? Should I like classify this as PIA or not? Should I make the invoice red or blue, right? You know, these kinds of things. So I do think you're starting to get to like a fast and slow level, like thinking, you know, of intuition, which is cool. And I think we're going to, I think what we're missing, and I've been trying to build into my latest version of my endless search for like perfect computer control is like the midbrain, right? which is like a mid Flash model that can make like semi decisions quickly right And so it like lightning fast decisions semi and then like the big planning slow model Right. Right. And kind of like separate that into like fast reflex muscle twitch and, and slow and kind of trying to find the balance there with it. And I think, I think I might crack it, but you know, this was, I thought I might crack it two years ago too. And four years before that. You mentioned Venice in there.

Tell us about Venice. What is that? Eric Borges started it. He's an old-school sort of libertarian. He came out of that kind of thinking of the cypherpunks and things like that. So they've moved to, I think, the last time I saw something like a $100 million run rate. Those are ones where they run a lot of the models themselves, the open-source models, and they rate the models. So they will give access to things like open-and-anthropic, but they'll try to fool them. or they'll say like, you know, those or bundle those requests so that they're not sort of like, you know, De-anonymizable or something like that. Yeah, de-anonymizable, and they'll rate the model kind of based on it. And then they have kind of their own models that they host that say like, you know, we're not going to keep all of the stuff on this. And I think, you know. So could you characterize it very loosely as open router that cares about privacy? Yeah, it's open router that cares about privacy. I'd say something's open VPN-ish for models, right? and you know I think I think they're doing God's work I think they're doing amazing stuff with it I have all the respect

in the world for Eric and his team although Eric I'd like to see some team based version of it not just individual subs please I don't know how many times I have to send an email or tweet at you out about it but please please team enterprise version but otherwise look I mean I think God bless them they're doing amazing stuff and I think those I'd like to see more of that kind of stuff more diversity. And I'm just, as you know, my roots in open source go back a long way. I'm constantly fighting with it. I can't believe that we're actually fighting the battle again. I thought that this was settled like sorry decisis. I thought it was done. But apparently not. But Nautilo does not today support local models? Not yet. And primarily because I just haven't found any models that I was super excited about locally. I think, I hope, that that's going to change with the new M5, M6, 512 gigabytes of RAM units coming from Apple next month.

I think that the real problem, the real bottleneck is not the GPUs at this point. The real bottleneck is the memory and how much you have to quantize the models. That said, we're compressing more and more intelligence into smaller things. it would not be hard for us to support a module already probably in a day or two to say like if you're running this on your own hardware on big GPUs, you know, your GP200s, 300s or whatever in your own thing as an enterprise that, you know, we could set up an open AI style endpoint or whatever. We'd probably do that in a couple of days, right? So I would like to get there, but we built it to be enterprise ready where like, you know, we have a COI for deploying it with drivers to like railway and to, you know, droplets. You know, it would be easy to add GCP. You know, we'll be, you know, but we didn't build all the super scaling out kind of stuff into it yet. Right. Or the host. But that's where that's where it will go.

And again, whether I win or not, I think that enterprises are going to build this because the stuff, the open waste models are going to are going to get to the point where you hit it. every infinite curve is the, the exponential curve is an S curve. And at some point you just don't like, it's solving the point character structure and doing like the DNA, you know, work, but you don't need that. Like Merck needs that, you know, like, like maybe the quantum, you know, the quant traders in like, you know, black, blackstone need that, but I don't need, I needed to like, once it's computer is solved and it can make a funny video or a presentation or whatever, and talking to my wife about our travel plans, at some point, like, you've got good enough. And I think we hit that in the next couple of years. You think it's possible that on the local side, like, getting that up and running looks very different from, you know, just a single endpoint in the cloud, meaning you say you can get it up and running a couple days, but, like, the reality may be that, like,

you really have, like, a suite of models and you have to have much more sophisticated routing because you're trying to, like, optimize to... So, yeah, look, I believe that the enterprise version of it is not a couple of days. Let me be clear, right? Yeah, yeah. Supporting VLM in a Docker container or whatever, or VM, with your local model for your... Sure. Or your 8-way cluster or whatever is feasible in a couple of days, okay? But, like, the true version is going to be a scale-out version with routing where you've got, like, you know, tons of containers and VLAN and monitoring a DevOps system. So I believe that's where enterprises are going to be. Yeah, and I guess I was asking about it less from the perspective of, like, enterprises do complicated stuff, but just that, like, you know, we can think about intelligence in the cloud as, like, you know, this one surface that we push stuff to. But, you know, maybe intelligence and open-weight local models is going to be lumpier,

and some models are good at some things, and other models are good at other things. And, like, in order to have a similar experience locally, you know, sooner rather than, you know, later, like, we have to be more sophisticated about what mono gets what task. Yeah, so I think there's actually a short-term potential branching pathway of the future in my Monte Carlo simulation where JEV's style models, because they're relatively small, we're seeing a lot of people training them in a couple of weeks. Someone will find the right recipe for doing that. A 27B model or a small model that can run locally on your computer. I think that actually becomes a cool routing layer. Are you talking about using JEV and these 27B models? No, I'm saying a JEV-like model. A JEV-like model running locally could be incredibly valuable to say, you know what, I need this to control, it's visual and it needs to control my, I don't need to go off to the cloud to like click and open Word and put the thing in there.

I just need that open and then the big model in the cloud can write the machine consciousness foam in there, right? And I think also it can be the kind of thing where you've got a lot of potential with security. In other words, you have the ability, like I said, to use the dumb hammer heuristics of like, before I send this off to the big, friendly, like Substance Compute in the Sky, I should maybe scrub this, right, for faces and PI and all this, and that you could have a local Jeff model going, hey, you're classifying missing stuff, but I caught a social security number there and I blanked it out. I think that can be very cool. I think that happens in the next six months, and I think that's going to be very cool. So that would probably be the point where I would start to say, this is valuable, even before I have the 512 gigabyte machine, super fast or whatever, it's just got to be relatively quick. So a lot of times I always look at it and there's a lot of people I respect working on the local models. Right. And I, and I, and I, and I'm like, you know, I'm like right there with like, you know,

like the self sovereign thing. I believe it. But a lot of times I look at it and I'm like, well, it's definitely a tool shaped object to set up. It's challenging. And it's also, you know, a lot of times they're like a bam, I got it running. It's, you know, it's super quantized and it's like 10 tokens a second. And I'm like, okay, that's not quite where I want it to be yet. Right? You know what I mean? So it's kind of like, yeah, I got this GPU and I can play the game at five frames per second. I'm like, not quite what I'm looking for yet. Right? Well, like, I was very excited to see, like, the video model, like the H3 or whatever and things like that. It was like, oh, man, this can run on local hardware. This is awesome. And it was like the average things were, like, 30 minutes per generation of, like, a 10-second video. And I'm like, okay, that's not going to work for me. But I do think that that gets better, but it's really a, like Steve Jobs said, the software is just all the stuff you didn't have time to get into hardware, right? So as this hardware gets better, then it becomes an easier problem.

So I do think hardware just moves slower and it takes a little bit of time and a lot of kind of tricks to kind of get us there. So what models are you using? Given the degree to which you care about privacy and self-sovereignty, when you are using your own Nautilus, what models are you using? How are you thinking about ensuring that you have the degree of privacy you want with the constraints of needing to use cloud-based frontier models? Yeah, so I think it depends on sometimes the chat that I'm in. Just remember, we have namespacing. So if I'm working on the marketing stuff, I have no problem using, you know, Claude or Saul or Astra. Okay? So I just want to be clear about that. I love those models. And I'm not an open, you know, or privacy absolutist. I believe you have diversity. I believe you have a need for privacy, and you have a need for open weights and open source, and you have a need for commercial software, and they can all and should all coexist.

That's fine, you know? But when I'm using privacy ones, I tend to use Venice and I tend to use Kimmy. K3, I tend to use Jalen 5.3, DeepSeek V4 Flash 4.1. All on Venice or via? Almost all on Venice, right? Yeah, almost all on Venice. Those tend to be the ones that, like, if I'm asking about a health issue or something like that, I tend to look there first. And I think I would encourage people to do the same. I would encourage people to think about the kinds of information that they're sharing with it. It's really tempting to not care and to just say, like, here's my tax return, GPT. Like, I get it. Like, GPT, look, there's no shade, man. Like, I mean, you know, Codex is incredible. I use it all the time. It is incredible. And when I think about it, I'm horrified at how I've gotten sucked into sharing some of the information that I share. And it's very reluctantly, it's very consciously, it's probably not nearly as much as people who aren't thinking about it are willing to share.

And there are conveniences that I give up because I'm not willing to share certain information. I agree with the hesitation to share. And also interesting to me to learn about Venice. I'm going to check that out. Yeah, they're great. And they plug into almost any harness that you want. They've got a pretty good – you pay a little bit of a premium for it, which I think is fair. They're more expensive than Open Router. But I think they're not sending you off to Kimi in China. They're running it on their own hardware, right? And so that's the beauty of Open Waits, right? And I think that's really important in their policies, the way that they think about things. So that can go into any harness. They have their own subscriptions. Now, maybe they don't have as much subsidized compute as everyone else. but you know it's worth giving them a subscription you know for certain types of things so actually one of the things we're doing now is we're we're looking at some kind of like you know some give backs in inference to defray costs right and we're looking at like originally we designed Nautilo to be like you have your

keys on the server by your sysadmin and everybody uses those right but now we're bringing bring your own keys and this was someone in the community who was like well I just want to plug in my keys because I have the community and I'm like well I'm not opening I can't afford 2 million in comp to give away on the community so everybody can use it free for eternity so I was like okay people just come in and they're using it like discord and chatting with me but they're not getting the full power of it but if you bring your own keys and you go hey I'm plugging in this then we can plug in and say here's a Vans key and it gives you access to see dance and to music generation and to the yellow lambs and to kimmy and all these kinds of things and you get the full power of playing around with the genies and like in in wonderful and it becomes almost like a new use case almost like a discord with uh you know with with genies as first class citizens so i think that's that's we're working on that right now uh we hope to have that out next month i think it's wonderful but you really you got to think about you have to think about these things a little because

they're going to give you you know it's like the old it's like the old the old thing they would tell you during the 80s right I think you're old enough for this too where they're like yeah they're going to give me the first hit free that never happened by the way those liars those liars you know but but like yeah but like you know and then you're going to get sucked in you know well you are like yeah everybody's getting sucked into the subsidized compute but at one point at one point and me too right at one point the you know and we and we built actually in Zadon we built the ability to control third artists You can control Codex and Claude and Hermes whatever right from inside in the next like you know which is great Because people want to use it. But at some point, the musical chairs stop. And you don't know when that's going to be, right? But it's like the early, again, it's like the early days of Uber and Lyft. where you're like, it's like $3. I've used that analogy as well. Yeah, $3 you get across the city, and now it's $70 you get across the city, right?

Part of my answer to that has been to use the subsidized compute to create deterministic tools to do the things that you want to do. And I don't think enough people are doing that. It's easy to say, okay, here's my task. I'm giving my task to the LLM. The LLM kind of does a good job with the task. I'm like, but your task doesn't really need the LLM. Like, you can build, like, you know, workflow tools using the LLM, build them once, you have them forever. And, you know, they're actually better because, like, it's deterministic in the places that it can be deterministic. I don't think enough people are doing that. And people aren't always doing all the work. Like, you know, I still think, I still love code creativity. So, like, I've got one of the videos on the site where I show all I made the videos for the site. So, like, I wanted to be careful of too much PII, so I made a whole, like, mock company with different people characters and genie characters so that I could use that rather than, like, all of our personal stuff.

But, like, when I record the video, I'm recording it with my voice. So then I have a whole video of the making of the video where I'm talking to my genie. I'm like, here's my voice. Here's the video. Go give me the transcript. And then it shows me an MD file in there, and then I can edit the MD file directly where it messed up. And I'll go, okay, now go, you know, go audition some voices for me. And I'll audition different voices for, like, what I want. You know, like, you know, a wizened old, you know, dude with a bald head, right? You know? And, you know, and, like, with a deep voice, you know, or whatever. And then we get voices that we like, and she would just go take those parts and give me back the transcript. Then I would take it into the video editor and cut it in, right? and like and over like split the audio video and put that in where my voice was and then we cut the video and then then I might give it back to something like Astra or whatever and say, OK, like first let's do a style design, do some mockups of like what this looks like. Don't just make it like, let me see. I don't like this. And I understand kerning and I understand, you know, I understand taste.

So I can be like, no, that's not the look that I'm looking for. So that's the one. Okay, now go in and like give me a script. Okay, I don't like that script. I'm going to rewrite it. Okay, that come back. Now I approve the script. Okay, now go ahead and lay it in. Now that saves me hours of time, but it's not here. Please give me the thing and whatever poops out of it is golden. That's a thing where like I think I saw a famous VC yesterday say, you know, just stop sending me your AI, January Dex. They reek of no effort. and I think that's the kind of thing where like if you don't put any effort into it I understand why people get upset about that right when they're like you just pass off some essay as your own or some deck as your own where you clearly have not stepped back and done the critical thinking or re-edited or been a part of the process yourself and there are people who are like full on Luddites where like if you don't do it 100% like if you don't fresh the wheat yourself you're not a real man I'm like, all right, listen, I'm sorry, I'm using the lawnmower, okay?

I'm using the, I'm using, I know there's, like, I can hand saw it or I could use this, but I'm going to use the power saw, I'm sorry, okay? Yeah, yeah. But I do think co-creativity, putting effort in a thing, is just to underrate it. And I think a lot of the times, like, you know, Marie and I joke that sometimes, you know, we build these co-creative tools because that's what we love. And a lot of times people are like, I just wanted to do it for me. and you're like, okay. Like, I understand that, right? And like, a lot of times there are things where I'm just happy that it does pieces of it for me or parts of it. But I still think if you're doing something that matters to you, you're pitching something, you're starting a business, you're rolling out your marketing, if you're not even looking at it, what the hell are you doing? What the hell are you doing? One of those co-creative tools is like a co-writing tool that you were working on prior to Nautilo that you've kind of integrated in. And I remember reaching out to you and just wondering, why isn't this more common?

I think OpenAI, just at DevDays yesterday, announced a co-writing tool, if I interpreted it correctly. I haven't looked into it very deeply. But all AI should give you results that you can manipulate. Why do I have to every time copy it, put it back into the paste box and say, okay, let's advance this, let's do this, do that. That seems to me such a natural paradigm, and it's not nearly as common as I would have expected it to be by now. I found in the earlier version of our software that most people are like, just write this for me, man, in their prompts. And we're like, oh, my God, right? It's just sort of like, why won't you write it for me? you know and so it was just sort of like okay like apparently we're just selling to people who don't exist yet and one more thing i wanted to dig into that you've touched on i think a couple times but uh just to kind of pull it out like you know we've talked about um you know kind of the age of

free compute you have all these these resources while you can take advantage of some of that free compute, but you're also competing against it in some ways. I think what I want you to spend some time on is just your experience building in this environment, trying to create something for people in today's world. What is that? What are the key challenges that you deal with and how are you thinking about overcoming them? Well, so me and several other entrepreneurs that I know are talking about this all the time, and we call it the wall. The wall? The wall. The wall. Pink Floyd's the wall, right? Like another giant wall. And that is like you have the problem of monetization. You have the problem of moat. You have the problem of attention and noise. You have the problem of like everyone with the same tools. and so while it's simultaneously

easier to get something built and get something out there it's very hard to you know be seen and sometimes you look at stuff and you're like why is that like going viral and you really have no idea like you're like there were 10 better tools that you personally tried that just didn't go anywhere and like look perception is reality and so I'm you know I was having a conversation with a friend of mine and he was saying And he was always, he's got a couple of principles now. And I agree with these four. He said, if I have an idea, I assume that somebody else had that idea yesterday, that's putting it out yesterday. Okay. The second thing is, he said, I used to say, like, what's the business model? And how do I, like, you know, how do I, who's your target audience? What's, you know, what are you going to do? Right. And work backwards from there. And now he's like, I'm working backwards from the marketing. Like, what can I fit into the tweet with the video with the one, like, with the Dune meets Dinosaurs Hollywood elevator pitch that passes the blunt brain barrier and goes viral, right?

And so there's definitely, like, there's an adaptability that I have not fully mastered, right? I feel like it's sort of like the old kung fu movies where, you know, kung fu panda comes in and just, you know, gets his butt kicked by the spinning gyre and the cup of fire and the balancing. You're like, whoa, right? Trying to navigate your way through this. At the same time, it's like I feel like the Omarkey guy, you know, who's like enthusiastic, he's like, I've got the greatest tools of all time. and like, I wouldn't be excited about this. It can do anything. It's a, it's a magical genie that you ask it to do stuff and it does that. And, um, so I simultaneously feel this like, um, super excitement and a ton of exhaustion. Um, like when people said it's taking all the jobs, I've never worked harder in my life. I don't know what they're talking about. Right. So, um, you know, as I'm waiting for the jobs so I can chill out more,

uh, but that's not happening. I think it is, uh, incredibly challenging environment. I think we're going to see a lot of folks crash out, and I think it's very hard to be hurting the noise, even when you have something that's really, really cool. And then oftentimes, sometimes the product comes out and people are excited about it, and then it just disappears. So I think in the past, you could build Nero, and now you see a lot of SaaS companies. I always like to say my predictions that were wrong, and I think about a year and a half ago, I was saying the SaaSpocalypse is nonsense. Nobody is going to rebuild half of these applications. I think that I was wrong about that. I think a lot of companies will not do it, but you are starting to see a number of companies that were worth a huge amount of money go under and be sold in fire sales at 90% below value. Miro was just one of them and others. I built a Miro board in a Nautilus in like half a week. That's pretty awesome, right, that my genie can control too. you know so and you saw someone had the compositor out the other day and people jumped on and it was

like a photoshop alternative you know that was pretty slick this was a designer this wasn't a person who was like i don't know anything about photoshop and embarking it was like i mean i looked at the repo and there were a few things like in the in the readme about how to how to sign it and check it in i'm like take that out of there right you know so it helps to have dev devs in there, right? But again, the community project forms around that. And yeah, it doesn't have every brush and whatever, but what I think is you have this folk era of software coming where like, so someone at Microsoft told me, the joke at Microsoft used to be, well, you only use 5% of Word. Well, everybody uses a different 5%. But now I think the answer is like you can build the 5% of the things that you really need. And I think probably in two or three years, maybe it's a little longer. The timelines are always fuzzy. You are going to have a point where the agents can just spit out bytecode that basically does what you want. If you don't like it, you say, go test it and change it

and refact it in this way on the fly. And so I do think that that's a very hard environment to sell software in. Now, there's still going to be software that's incredibly challenging to build, And those are going to be in areas that are ironically super regulated, healthcare, right? Like, you know, in the infinite, infinite mind about the Demis and the early days of deep mind, they talk about Suleiman, you know, doing a bunch of stuff in healthcare with like the UK where they had like worked with like people and opt-netricians and they had a software that was able to detect and all of it died because it ran up against bureaucracy. It ran up against pieces in the Guardian that were like the data privacy. And so they're still running these geriatric systems. So I think there's a lot of things where people don't understand the friction, you know, and where things are going to be. There's going to be barriers in there where they might create a legal mode or there might be a friction mode or whatever, those kinds of things,

or where there's a combination of hardware and software that's really hard to solve. But a lot of other stuff Man is super hard To have a mode in the competition And I don't know where somebody Gets a legitimate exit In three or four years That other people haven't been able to build Well keep on fighting the good fight Power to the sovereign AI people And I'm going to keep fighting And I hope everyone else will I hope everyone else will wake up But don't say I didn't warn you. Awesome. Well, Dan, it's always great catching up. Pleasure. And hearing a bit about what you've been working on. As is always the case, the conversation meanders a bit, but always a good one. I'm sure you'll chop it down. I'm like a squirrel. I got 50 things I'm burning on at once. But I appreciate

you indulging me and reaching out to me as always, sir, and it's good to see you still fighting a good fight, and like, putting out awesome content off all these years, sir, appreciate it.