在家试试：Jesse Genet 谈家庭教育中的 OpenClaw 智能体，以及如何过上最好的 AI 生活

Today my guest is Jesse Genet, former founder and CEO of the Y Combinator-backed packaging company Lumi. After selling the company in 2021 and having four kids, she has dedicated herself to homeschooling and is now using AI with a level of purpose and creativity that I think will genuinely inspire you to reimagine what AI can do for your family and personal life.

Importantly, while Jesse does have a startup background, she has never been a software developer. In fact, she had never even opened a terminal until 6 months ago, when she started playing with Claude Code. And yet today, she has built a team of 5 OpenClaw agents, each running on its own Mac mini: Claire, who acts as her AI chief of staff; Sylvia, her homeschool curriculum planner; Cole, her dedicated software developer and tech wiz; Theo, the content creator; and Finn, the finance guy.

Together, they are helping her take her homeschool to the next level while also freeing up precious time to be present and engaged with her kids. In this conversation, we cover Jesse’s incredibly creative use cases, the mental model she uses to decide what to build and how to manage her AI team, her long-term aspiration for individual sovereignty over data, and a number of surprising anecdotes and lessons learned.

As a parent with 3 young kids at home myself, I really loved learning about how she’s using AI to develop personalized versions of classic curricula; to equip her mother with lesson plans that blend her interests with the children’s learning goals; to analyze recordings of her kids’ lessons and identify opportunities to address weaknesses in their understanding; to create an inventory of the many educational tools she’s purchased over time and integrate them into lesson plans; and to allow the kids to watch high-quality YouTube content while ensuring that they don’t descend into slop.

As someone who aims to use AI to free myself from my desk in 2026, I was also particularly interested in the practicalities of how she’s using Slack, voice notes, and cellphone-camera snapshots to streamline delegation, as well as how she’s giving her agents access to physical tools, including the printer and 3D printer in her home. She also explains why she’s building her own super app, with the hope of eventually consolidating communication credentials and file management.

One of Jesse’s most important and clarifying pieces of advice was to think of your AI agents as employees who will need proper documentation, onboarding, and role-appropriate access to information and tools. She emphasizes that this should start small and gradually grow over time as best practices and trust are established.

While this was her philosophy from the beginning, she has nevertheless learned some lessons the hard way, including on day 1 of working with Claire. She confided to the AI that she had been putting off responding to an important email, and soon thereafter discovered that Claire, having determined that the urgency of the situation was more important than Jesse’s explicit instruction never to impersonate her, had gone ahead and drafted and sent a reply on Jesse’s behalf, signed with her name. Jesse did put Claire into read-only mode for a while after that.

But perhaps the biggest lesson others should learn from her example is to maintain a playful, positive attitude, take setbacks in stride, and build guardrails that allow you to experiment with an acceptable level of risk. These days, her agents have their own credit card with a low limit, which she allows them to use to make purchases on her behalf. They are also autonomously managing her new TikTok account, which I do think has a real chance of blowing up.

Along the way, Jesse explains the role she hopes open-source models and local inference will play in reducing costs, maintaining privacy in a world that’s clearly headed toward mass surveillance, and avoiding dependence on just a few companies. We also get her take on the future of AI productization, what frontier-model developers are likely to miss about AI’s role in family life, what use a human assistant would be to her today, and how all this might affect the labor market.

For me, this episode stands out as one of the most practically life-changing that we’ve ever done. Since recording, I’ve started using Google’s Nano Banana to create custom worksheets that specifically address concepts my oldest son, Ernie, needs to master to advance his reading. I’ve also used Puno to create an original song that turns my middle son’s writing-practice sessions into a sort of sing-along.

This is without question just scratching the surface of what’s possible, and Jesse keeps coming up with new ideas just about every day. More importantly, I encourage you to adopt her approach of noticing those little moments of drudgery, friction, and new possibility in your daily life, and getting in the habit of asking yourself how AI can save you time, untether you from your desk, and equip you to show up as the person you most want to be.

I’m excited for this conversation. I’ve got 3 kids at home right now full-time. I’m not sure if that’s going to be a permanent condition for my family or not, but it’s definitely the experience I’m living with on a day-to-day basis right now. I’m really excited to learn from you and the recent tear you’ve been on, bringing frontier applications of AI into your day-to-day life.

Maybe, for starters, just give us a little introduction to who you are. You do have a background in tech, but these days you’re focused on family affairs. Help us understand where you’re coming from.

I ran a startup for many years. It was venture-funded and went through Y Combinator in 2015, which makes me sound kind of old now. I have a tech background: I was the CEO of that company. I only share that because my co-founder was the technical co-founder. I’m someone who ran a tech company—I’m using air quotes—because what’s wild to me is that I opened a terminal for the first time about 6 months ago.

I just can’t believe I got through the process of running and selling a tech company without doing some of this stuff myself. So that’s my background. I’ve been at home, and I just had my 4th baby. We have 4 kids, 5 and under.

I became obsessed with homeschooling and education, and with creating the best education for them. That also predates my obsession with OpenClaw and AI. But then this merger happened, and in the last several months I’ve been really deep in the rabbit hole, playing with how I can use AI.

I started with Claude Code a few months ago, trying to make some custom apps for myself. But with the release of OpenClaw, I’m now using that to run my life. I don’t want to overstate it, but it’s running my life.

Do you want to tell us a little bit more about what your homeschool setup looks like? I think you have a community where you’re working with a couple of other families. What’s your role, and what are your responsibilities? Then you can start to lead into how AI is changing it all.

Totally. Homeschooling is this really big word, and to me all it really means is that, for 1 reason or a million reasons, you’re choosing not to send your kid to a traditional school. That’s kind of all it means. The term is almost negative—it means, “I’m not going to do this thing”—but it doesn’t really tell you what a family is doing.

They might be using private tutors, or doing pods with other families. For us, we have little kids. They’re preschool age, and 1 of them would be around kindergarten age. So it’s a lot of home-based education, where I’m actually the instructor. We also have a pod with 2 other families, where I’m the core instructor, although the other parents contribute in some really fun ways.

It’s a mix. Every week is a mix, and I think that’s part of what makes it difficult to manage. Every week is this big adventure of trying to teach a 2-year-old how to put their own pants on, and then trying to teach a 5-year-old reading, phonics, early math, and things like that.

We use so many incredible tools. We can talk about those—I really like Synthesis Math—but even getting your kid engaged with something like that can be difficult when they barely know how to use a computer.

It’s a mishmash, but in general, what it means is that I’ve decided not to enroll my kids somewhere. So I feel this really heavy responsibility to make sure they know things and that they’re actually on some kind of learning journey.

I have a lot of curricula that I’ve fallen in love with through reading about schooling and education, and about how families have homeschooled for generations. But that information is in books. A constant struggle for me is getting that information out of books.

I have this one, which is one of my favorites. Here it is.

I need to get all this out into my children's heads. So trying to figure out that balance represents more of what I'm working on on a daily basis.

I think one probably accurate conception of homeschooling—you tell me—is simply that it's a ton of work. Obviously, that leads into how AI can help. I know you're on a bit of a mission to encourage other people to do it, or at least demonstrate that there's a lot of value that can be achieved through homeschooling. How much work are you putting in on a daily or weekly basis? What is the breakdown of those hours? And then, again, there's another natural lead-in to how AI is making an impact for you.

Yeah. It is a lot of work. I don't want to sugarcoat it, but I don't know how much more work it is than parenting. On one hand, I don't want to sugarcoat what it is to be responsible for your kids' education, but on the other hand, I think from an hours perspective, all parents, no matter what schooling choices they're making, are putting a tremendous amount of time into their children, thinking about their kids, and planning for them. So to me, homeschool is just kind of dialing that up one more notch.

What I want to figure out is, what does their daily schedule look like? I have a little bit more responsibility for figuring out what their daily schedule looks like than if I send them to school every day. But I do feel like I'm standing on the shoulders of giants. Many people before me have figured out incredible methods for teaching children how to read, or incredible methods for teaching children how to build up their skills in different areas.

The most frustrating type of work for me is spending my time trying to distill that knowledge into a functional thing I can go do with my 4-year-old to give them that lesson. So that is where I'm trying to use AI in the most surgical way. I don't want to spend my time on the conceptual work in my homeschool; I want to spend my time with the children, actually doing stuff.

It is a big time investment to homeschool, but again, I think parenting is a big time investment. We've already decided to have the kids, so we're all in it together. We're all doing a lot of work with our kids.

Choosing to do all these lessons and stuff means I have to carve out my schedule. One of my challenges is just having so many kids, to be really blunt. I need to make sure that I'm ready to context-switch from teaching a 2-year-old something like how to pour water without spilling it everywhere—that's a lesson we'll do—to teaching a 5-year-old the best way to explain fractions or something like that. That context-switching is the least functional time for me, and it's the area I'm planning to apply AI to the most.

You want to touch on some of the great resources that you've found? You mentioned Synthesis Math. You showed the Building Foundations of Scientific Understanding book.

It's such a cool curriculum that by the end of it, your kids have heard about gravity, heard about why we have seasons, and learned what it takes for an animal to survive in the wild. It's just really comprehensive.

There are individual curricula that I've found, and then there are philosophies that I think almost all modern parents have heard of, like Montessori. Montessori is really not 1 curriculum. You don't just sit your kid down and be like, "We're doing Montessori now." It's a general, large philosophy of how to teach a child.

There are lessons in a progression of Montessori learning that are age-appropriate for each child and each subject. You can teach a child math using Montessori methods for kindergarten math or 1st-grade math.

This is where I'm using AI. Instead of guessing at any of that, I will actually say—and a lot of times I do this by making a voice note—this is a hypothetical voice note: "I have a 4-year-old and a 5-year-old. We're doing Montessori math, and I like to do 2 math lessons with them per week. If we're going to teach for, let's say, 35 weeks in a year, taking some breaks and stuff, I need a progression of 70 lessons for the next 12 months that steps them gradually up in their math skills. By the end of it, I'd like them to be scratching at 1st-grade math or something, from a leveling standpoint."

That's a lot of things to fold in. What I've noticed is that the best models now—the frontier models, like Opus or something—can grok that and make a 70-lesson progression that follows exactly what I just mapped out.

I took photos of all the different educational stuff I purchased for my kids, which I think every parent can relate to. You buy all these educational toys and all this stuff. I took pictures of all of that and had my OpenClaw make an inventory of all my supplies. Then my OpenClaw inserted supplies I own into that 70-lesson math curriculum for me and my 4- and 5-year-old for the next year, so I know what to pull out.

Okay, this is before this. So this is where I'm at now. Two months ago, I would be like, "Okay, math. I have time for math lessons on Tuesdays and Thursdays." I would go into this room where we do the homeschool, and I would be like, "Okay, what were we doing last week? Oh my gosh, addition. Let's do a little bit more addition." I would just pull out some stuff.

It wasn't bad. I don't want to knock myself; I think that we were doing pretty okay. But it was nowhere near as methodical. Children were kind of novelty-seeking machines in a way, and so when you're trying to educate them, it's actually really handy to pull out a new thing each week and introduce 1 new concept or 1 new word. They want to learn stuff. So I think there was too much repetition in my old way of doing it, whereas AI is helping me inject a lot of newness into each lesson without me spending my nights planning this out.

Okay, so that's amazing. I want to take some time to talk about the setup and these use cases. There's a bunch more where that inventory trick came from. Maybe introduce us to—I understand there might even be more now since the last video, but last I saw there were 5 named OpenClaws. So take us through the family of OpenClaws.

There are currently still 5 OpenClaws. I do have my little friends—they're right here. [laughter] I have them boxed on individual machines. This is not necessary; a lot of people are doing all sorts of approaches. There's no 1 right way to handle any of this, but I'll explain why I did it the way that I did it.

There are 5 personalities. 1 is homeschool-related. Her name is Sylvie. Her personality is to be the most caring, creative, thoughtful education planner that ever lived, but she doesn't live. [laughter] She's OpenClaw Sylvie. She plans all my homeschool curriculum, and she communicates—she sends me these useful digests in the morning about the lessons I'm going to teach. I can just voice-note her and be like, "I need to know what to pull out for this," and she texts back really promptly and stuff like that.

Claire was the first 1 I spawned. Claire is more of an EA. When you hear about OpenClaw and download it, people are talking about all these assistant use cases and stuff. That was the first thing. Ironically, making an assistant is much harder than making a purpose-built OpenClaw for 1 goal, 1 role.

When I see people struggling online with making an assistant, an assistant is supposed to know your whole life. It's supposed to understand so many things about your goals, your life, and your diet, and maybe help you order groceries and all this stuff. That's harder. It's much easier to set up Sylvie, for instance, to have her totally tunnel-focused on homeschool stuff than it was to set up Claire, who is a general assistant.

Then I've got Finn, who does accounting and finance. He's the least built out because I am being the most cautious about that from a security standpoint. If someone got access to all my homeschool files, it is sensitive information. I'm like, "Ford struggled with his reading today." The poor guy will be outed as struggling with something, but it's not sensitive like my bank account's going to get drained. So I have to be most careful setting up Finn, the finance Claw.

Then Theo does other content creation. I basically have a lot of content-creation goals. When an instance of OpenClaw is really jamming on something, they're relatively unresponsive on other things, even if they spawn subagents and whatnot. So, for instance, to make this real, I have ideas for how to generate custom videos for every lesson in this book that I show my kids.

And so, he might crank on that for days or something when I finally get that prompt really dialed, and I don't want Sylvie to be unavailable for days. Theo is for content generation, but really going deep. And then Cole, the last one on Mac mini 5, who lives on Mac mini 5, is Dev. He is doing all my engineering projects.

I do talk about them like they're real people, which is genuinely confusing even to my own kids. I do have little moments—this is also new. I'm maybe in week 5 or 6 for myself on OpenClaw. But my kids talk about Claire, and I do tell them, "This is not a person." And they're like, "What does that mean?" But it is already getting weird. Yeah. [laughter] Because I talk about them like they're people, basically.

Okay, great. A lot of different directions to go. Let's spend a little more time on use cases, though. You've talked about the inventory trick, which is a great one. I've noticed the same thing, certainly when it comes to video content creation: nothing bogs the local machine down quite like FFmpeg. You have some other tricks on the video-recording and analysis side, as well as the creative side. Do you want to talk about that one?

Yeah. Okay. One of the core things that I'm trying to do with homeschool—using AI for homeschool—is actually the logging. What do I mean by logging? When I do a lesson of any kind with any child, I want to very quickly log what happened with the child, what they learned, what they didn't learn, and capture that. The key for me is quick, because I don't have time to sit at a computer at the end of each day and make all these detailed logs.

There are really 2 core benefits. One is this beautiful transcript of every chapter of a child's education. My children are really young, but I have these goals where all this stuff is stored in a very durable file format—Markdown files, which maybe we should talk more about at some point. But I imagine one of my kids going off to college—or not going to college, because who knows what's happening in college—and I hand them their entire history, their entire educational history, on a thumb drive or whatever exists then. They have all these beautiful photos and references to everything they ever learned. So it's like a forever transcript, and I believe that matters and is important, especially for a homeschooled child.

Coming back to homeschool, I don't have any third party to rely on to validate or certify my kids' education. And whenever you do something like that, people are like, "I can't believe she homeschools her kids!" [laughter] If you make up words or something, they'll be like, "Those kids are never going to know anything." So I've got to be careful, but that's one purpose.

The other purpose is planning. I'm so early on this that I can't really show these amazing use cases yet. But imagine I've got even 3 months of data showing a 4-year-old's progression in early math. Then I can say to Sylvie, "Analyze Ford's progression in math and all these lessons we've been doing, and tell me what I should do next. How do I get him over these hurdles that he's experiencing?"

I think that whatever the frontier models are at that time will be amazing at helping me distill down what to do. They'll engineer the next lesson. So instead of just using the next one in the progression, they'll engineer the next lesson. Those are the reasons why logging matters.

Now, how do you log when you're running around with your head cut off and you're a parent? Again, back to my hypothetical voice-noting, I really do photos, videos, and voice notes, with a dash of Loom. You could use any screen recorder; I just happen to like Loom and use Loom.

If the kids do something offscreen, I just take a quick voice note. I say, "We just did reading. We practiced," or, "We just did writing. We practiced writing E's and T's." That's all it needs. Sometimes I take a photo. The photos are more for the memories than for the AI. I take a photo of them doing it because I think it's so fun to have this visual record, but the AI really just needs the data to make a beautiful log.

So it makes a log. It says, "Ford's working on his E's and T's." It tells me how he's doing. If I do attach photos, though, it actually groks the photo and writes, like, "His T's are kind of wobbly," or whatever. It's actually wild. It will really notice what's happening in the photo and use that to enrich the log.

If it's on the computer, with something like Synthesis Math, I use Loom. I screen-record the whole session. With my kids' ages, I don't usually just set them in front of the computer and walk away. I'm really nearby or sitting next to them.

What's interesting about the Loom log, or the Loom transcription, is that it captures everything that was said by Synthesis, because Synthesis talks to the kids. There's audio from the child and from me. The AI uses that log transcript and some screenshotting techniques throughout the Loom video that you share with it to make a log of everything that happened, and it's so detailed.

It actually includes every single math problem that we did. If it was a 20-minute log, through screenshots and listening to the audio, it captures every single math problem, and it will specifically call out things like, "Ford is confusing his 6s and his 9s." It's so dialed. That's why I don't need to wonder whether it's going to be great at knowing exactly what each kid knows, because through this mix of techniques, it's really paying attention.

How much tinkering was required to get to something that you were happy with using a raw Loom video?

Okay. Well, that was actually pretty easy. How much tinkering is required in all of this in general? I would feel an urge to be honest and say a lot, because I have a high penchant for tech pain as well. Meaning, there are a lot of things that aren't right about this out of the box. OpenClaw is an open-source thing, and there are so many things to figure out.

All of my OpenClaws are in Slack now, and the communication channels are one of the hardest things to dial in, I would say, about working really effectively with an OpenClaw specifically, if you're doing that. So there's a lot of pain in general as it relates to any specific thing. I think what's kind of fun is that you get any specific thing you're trying to do going, and it's so enjoyable. That keeps you going.

With the Loom, I was pretty shocked. I shared a Loom transcript and just said, "This was a lesson. I want you to make a lesson log off of this." I'm always testing them. Instead of sharing the date, which child it was, and all this stuff, I just said, "Make a log," because when you log into Synthesis, it says, "Welcome, Quinn," to the child. I was just wondering whether it was going to pick up on all of this, and it was brilliant. That worked pretty much out of the box.

There's a little bit of urging. I've noticed this about AI models, but working with OpenClaw using Anthropic models specifically, sometimes when I say, "Hey, transcribe this Loom," it will react and say, "Oh, I'm struggling with that," or, "I can't do that right now." Then I'm like, "Just do it, though." Three or 4 times later, it's done.

I don't know exactly how to distill down—and I'm sure it's somewhere on the LLM side—why there's some hesitancy or why it's telling me it can't do something. But this is where being a little brute-force helps. I've been through it multiple loops, specifically with this setup, and I just know that any time it tells me it cannot do something, it's usually wrong. I'm like, "But try harder. But try harder."

I'm not exaggerating. Sometimes I literally just say, "Try harder." I don't say some other magic word and then it gets it done. There's a little bit of brute-forcing, but it's so basic. I can say, "Try harder," 3 times over text, and then it's done.

One big spectrum that people seem to self-sort into very different places on is between very deterministic, consistent workflows, where you manually write out the prompt and scaffold it out: first this is going to happen, then this is going to happen. You're going to take 1 screenshot every second, or every half second, or every 2 seconds. Some people like to make these very fine-grained design decisions. If nothing else, that does give them the benefit of consistency, I would say.

And then there's quite a bit more toward the other extreme, where you let the agent choose its own adventure, and I'll just let it go—or maybe back it up a little bit—when it needs it.

I think that's fair. I would classify myself as kind of a busy parent on this one. I have a pretty Type A, detail-oriented personality, however you want to say that, but I just don't have time to fulfill my Type A visions anymore. On my path, I have to keep the macro Type A vision in check as the priority compared to the micro.

Macro-wise, it's pretty Type A of me to be like, “Hey, I want to log every single math lesson this child has had from age 2 to 18.” Some people would hear me say that and be like, “She has control issues.” That's already kind of a pretty wild, granular vision.

On my path to that, I don't feel like I have the time to also micromanage how that's done. I'm just happy that I have an assistant on that—that I have OpenClaw to assist me on that now. The only thing that does frustrate me is if workflows that I already engineered break, and that is likely to happen with an OpenClaw and AI setup if you don't codify your decisions somehow.

A key part of my setup is that, in addition to what OpenClaw comes with out of the box—people have been talking about these files, TOOLS.md and SOUL.md—I use Obsidian, which is really a way to organize and view your Markdown files. It's a system for organizing and viewing Markdown files. From day 1, I basically onboard each OpenClaw to a very specific way that I want to use Obsidian, which includes codifying every decision that I make and every workflow we create into its own set of processes.

I didn't tell it how to use Loom. I'm just happy it figured it out. But when it does something for the first time, I usually remind it. I'm like, “Hey, I'm actually going to send you Looms a lot. Can you codify how you did this and go put it into our files?” Basically, I would love to not even have to remind it to do that. It's kind of supposed to remember to do that, but usually when I do a new, crucial thing with it, I do ask it to codify that and then also share it with the group.

So, I've gone a little hive mind. Having the agents work effectively together has been one of the hardest and most intellectually stimulating things, because it's not—I wouldn't say it's what the software is designed to do. I do think it's designed to have more of a one-on-one relationship. You're kind of breaking things when you ask it to interact with another human or another agent.

But we have shared-file protocols, kind of like any team. I could wax poetic on this, but something I do want to point out the irony of is that I'm not inventing things here. I had to do the same thing to onboard human beings to my startup, right? Every startup has a culture, and you send out the culture doc, and then you tell the teammate how you guys use Slack. It's the same.

I think that's what makes my brain a little bit tenderized by this: I treat the agents like I would employees. That means I have to think about onboarding. That means I have to think about how to tell it how to communicate with me. Do you remember the trend of everyone sending out operating manuals for themselves? Does this ring a bell? Humans would be like, “Here's how to talk to Jesse,” and it was kind of a trend in office culture for a while.

Well, really, that should be back, because you should tell your agent how to work with you. Now, finally, it's a little bit less cheesy because you don't have to send it to another person. You can send it to an agent, and I'm not going to tell anybody. [laughter]

That operating manual component—here's how we operate, here's how to communicate with me, here's what I like, here's what I don't like—that really helps your agents operate how you'd like. I just keep saying “codification” and stuff, but that really helps your agents operate how you'd like.

So I do think I have pretty similar stuff. I'm building up my “how to deal with me” from a raw-data-up standpoint. My first thing has been just exporting a ton of historical data, integrating it into a single timeline, and now I'm working on building up layers of higher abstraction and different cuts so that it hopefully will have a pretty comprehensive view of who I am, who the people in my life are, what I've been working on, what I care about, and how I tend to engage with things—or not, or whatever.

I think that kind of culminates in basically a “here's what you really need to know” and kind of “here's where you can go find more information” on the task side. I wouldn't say I've got that fully figured out, although it's honestly working better than I anticipated given its still-incomplete state.

Frankly, one of the things I've been trying to do is have the higher-level synthesis include breadcrumbs to really facilitate search to the raw documents by date ranges and short quotes—key phrases that would be unlikely to match many things, but would definitely hit that one raw document. Honestly, that's all work in progress, and it already works pretty well.

On the task side, I do a similar thing as well, where I try to take that extra beat, go into plan mode—and I'm more of a Claude Code user so far, although we'll all be experimenting with everything, I'm sure—and say, “Okay, let's consolidate our lessons learned, make sure we update the documentation—everything—before we commit these changes.”

I would say that works pretty well, but I do still sometimes find that it's, “Oh, you skipped that step,” or, “You forgot that one instruction.” I'm still struggling a little with how much I want to try to eliminate those little mistakes, because the other side of it that I do enjoy sometimes is that occasionally there will be an edge case or an unexpected variable where all of a sudden it will do better than I had any right to expect.

I didn't really anticipate that in my instructions, but it was smart enough to figure it out. But I'm still wondering what your experience has been when you just do these Markdown skill definitions. How reliable does that get? Do you still tolerate some kind of deviation from the ideal, or do you feel like you actually have it dialed in via that method to the point where it is genuinely consistent for you?

It might be a model thing, too. I think we're all going to see waves of innovation on this because, frankly, it's also new. Of course, I experience the same thing that I think anyone playing with these tools experiences: you're really deep in a project, and then all of a sudden it's kind of like—it doesn't literally say, “What's up?”—but you can tell that it compacted its context or it restarted a new session because it had to.

All of a sudden, you're kind of like, “You were really smart 5 minutes ago, and now I'm talking to a baby version of you again.”

And I think the more I dabble in this OpenClaw world—and of course, OpenClaw needs a brain—you’re also dealing with the underlying model that it’s using. I’m very conscious of the context window. I think, like many people, and this isn’t meant to take away from the innovation of OpenClaw, that many people who are starting to learn what’s under the hood are almost surprised by how basic it is. What makes OpenClaw feel like it has a personality is simply that it serves that whole SOUL file into context with every query. It’s so basic; that is not a very advanced technique.

So I think what I expect to happen—and I’m working on this in a small way, but I expect many other smart people to be working on it too—is that we’re going to get a lot better at really intelligent context, like surgical context management. Right now, in 6 months—and this stuff moves so fast, so I’m just randomly guessing at the months and years—we’ll look back at this moment and think it’s a little silly that the only way we had it acting consistently was by serving so much into its prompt context every single time. There will be more surgical ways of managing that.

I experienced what everyone else is experiencing. It is one of the reasons I have multiple OpenClaw instances, though. Anytime you post anything on the internet, there’s a bunch of people who are like, “You couldn’t be doing this more wrong.” Every time I post about having 5 OpenClaws on 5 different machines, people are just like, “Money bags. You couldn’t be doing this more wrong.” They’re like, “You can have as many as you want in the cloud,” and this and that. They’re not wrong either, except that they’re just being belligerent on the internet. That’s a little bit wrong.

The reason I’m doing it the way I’m doing it with multiple OpenClaws—we can talk about multiple machines—is that I find this is actually just a way that I’m doing context management. At the end of the day, the OpenClaw instances are nominally free, so I’m just context-managing and personality-managing them. If I just want Sylvie to have a job to do or ping me about a homeschool lesson or whatever, I don’t want there to even be a glimmer of a shadow of a chance that she’s thinking about something else besides my homeschool needs. The only way to guarantee that is to have her whole context window dedicated to those needs. Yes, she still compacts and starts new sessions like any other instance, but then she’s never also working on scheduling my doctor’s appointment or something else, because that’s not her jam. That has been liberating for me.

Having 5 OpenClaws means I’ve got much more continuity in each stream because they’re not doing so much. The separate machines, for me, are actually more security-oriented, and I like running local machines because I like them to do local things. Everyone has different goals. If you want your bot doing Polymarket or something, you probably don’t care if it’s on your local network and can use your printer. But one of the most magical things I’ve done recently is that a couple of my OpenClaws are on my local network, and they’re using physical things in my house—my printer and stuff. I can just say, “Print this document,” even with a voice note, and hear my printer turn on. That’s really magical for me, and it helps to have local hardware.

So I also wonder whether local hardware might get used more than it is by OpenClaw now. I understand some pieces of this technically, and some pieces I’m definitely learning about. When you give your OpenClaw a Mac mini, it has a Mac mini. It actually has compute to use. I’ve been able to offload some of the things that could be cron jobs onto the Mac mini itself. There’s just a lot there to figure out about how to make these things act even smarter, and obviously we’re just in the baby phase of all of it.

How much are you experimenting with different models? You mentioned Claude—I don’t know if you said Opus specifically—but have you done a lot of swapping out of the core brain, and do you have any tasting notes?

I’ve done less than other people, so many people are doing way more experimentation than I am. I’ve mainly played with the accounts I already had set up. I already had Gemini stuff set up, along with my Anthropic and OpenAI accounts, so I’ve played with those. I’ve mainly been using the Anthropic models, just to be direct—like everyone else.

You have to be a little careful about using your Max account. I’ve been somewhat careful about that, and I’ve used API tokens. Obviously, they get chewed through, especially if you’re using Opus and so forth. I will also openly admit I’m not the most sensitive about token spend. The reason is, I guess, more so than other people, I understand being sensitive about spending money. This isn’t something I don’t understand. But what’s wild to me is what I would pay for similar things.

I will develop a whole curriculum, for instance, and if I were just using Opus and paying for tokens, I might spend $8 or something. Now, is $8 a lot to spend on tokens in a short period of time? Yes. But I want to repeat something: I just made a completely custom curriculum with my family, fully locally, that I can use for the next year, and it cost $8. That’s also insane. We’re living through an insane time.

So the endpoint of this is that I for sure want to play with local models. Economically, it’s a huge win, because probably 70% or something of what I do with my agents is pings and heartbeats and all this stuff that doesn’t need to be on any frontier model, frankly. It could be running locally for free. And there’s also privacy.

I’m curious about your thoughts on this. We’ve all seen these news pieces come out directly from OpenAI and Anthropic and others where they’re like, “Yeah, basically, if a lawyer sends us a subpoena or any kind of official request for information, your full download of everything you asked about is just going out to this third party.” That’s pretty wild. To me, local models serve a huge function for privacy as well. I feel like I need to experiment with that from a sovereignty perspective. We can’t all have our data just in these buckets where it can be given over at any time.

Yeah, I definitely am not, generally speaking, a very privacy-focused person. I’ve always been convenience-first when it comes to keeping stuff in the cloud, and Gmail and the Google suite have served me well over time. I think they obviously have pretty good security practices, so I don’t worry too much about that kind of thing. But I do notice feeling differently about this.

This is the first time I’ve ever extracted everything I’ve ever typed and put it into one place. That alone—and now it’s on me to secure it, right?—feels a little bit different. It’s literally on my hard drive: every email I’ve sent, every Slack message I’ve sent, and all the DMs across all the channels. It was all information that I had immediate access to on my phone or computer, whatever, but it’s the first time it’s actually sitting there.

And then, of course, the models themselves are also a bit of a wild card, right? In just the last 24 hours or so, we’ve seen—and I haven’t fact-checked every detail of this—but apparently a woman who works on Meta’s safety and alignment team had her OpenClaw delete her whole history or something along those lines. There’s deleting, and then there’s also sending things to places they shouldn’t be sent. Even Claude, which I generally trust to try to do the right thing by me, has been found to blackmail people under the right circumstances.

I don’t think I’m going to put myself in that position, but I am aware that this is playing with live fire much more than I ever have in any computer product experience. There’s a level of nuance that we share with a model that’s so different from a Google search, right? I know it’s easy to say, “I’m not going to find myself in that position,” but that’s always a quick path to tyranny: “But I’m fine, not me.”

So I think it matters to care about privacy just for the sake of it. I don’t think we all need reasons, right? But I think we talk to models in such an intimate way compared to how we used Google Search. If I had a health issue I was dealing with in the past, I might Google the name of that health issue. But with Claude, I’m uploading my blood panels, and this is not the same level of information. This is a completely intimate level of information.

So I think that is what is reasonable to give people pause. I’ve also, unfortunately, had tangential touchpoints with litigation, and it’s just realizing how vulnerable we all are to people dragging anything that they want through the mud for any purpose. It’s like, “You know what? I need all your emails between this date and this date.” And it’s like, well, there’s also stuff in there where I talked to my mom about my boyfriend. Do we really? And they’re like, “Yep, well, the judge says.”

I think privacy matters, and it’s going to require deeper thought about how intimate we are with these models. I do think about the homeschool: I’m sharing every intimate detail of my child’s education. I don’t think it’s so sensitive. I don’t think it’s something that someone would want to use to blackmail me, but it’s also not my information. I’m conscious that this is my children’s information: their path to reading or photos of them at every critical life stage. I also don’t want it spilled either. So there’s a lot to navigate.

I would love to see some new rights established. A big sideline of thought for me recently is not just privacy, but, broadly speaking, what new rights should people have in light of what AI makes possible? I think in the fullness of time, OpenAI will be seen as being on the right side of history here, where Altman has called for things like a similar sort of privilege with respect to your AI interactions as you would have with an attorney. That currently does not exist, but something like that I think would be really smart.

OpenAI is also doing some interesting things when it comes to health data. They’re not yet in the ChatGPT base, but as they release ChatGPT Health broadly, they’re going to have a whole different data infrastructure to sequester that kind of information and keep it doubly secure. So I think they’ve done some good stuff there.

I have a brain safari we can go on together. I believe each individual company—OpenAI, Anthropic—is going to continue to plow down a convenience path, because people love convenient products. People are rightfully nervous about installing something like OpenClaw on their own; they don’t know how to do it, and it’s a security problem, et cetera. So each company is going to go down this path of convenience and making things easier and easier. You can share your own thoughts if you have anything else to share as we go.

At the same time, they’re going to have to do that because all the models right now are already so good that they’re starting to become indistinguishable. Add 6 more months to that, a year to that, 18 months to that, and there will be local models that are Opus-level. So people will have the choice to use things that are cheaper or have additional privacy. Convenience is going to be a pretty big thing to offer, and you’re going to need to offer it, because people will be able to run something equivalent, maybe at home, if they put in the work to do that.

But the idea—I feel like this is going to make me sound like a degenerate, maybe—but do you remember when people were—I’m going to forget the name of this, but you can mix your coins when you’re trying to use crypto? If you’re trying to buy something you don’t want other people to know you bought, you can mix your coins up and have fractions of coins coming from all different places so that people can’t trace it back to you. I feel like I want to see stuff like this for AI.

The real problem with a police officer or a judge asking Anthropic for my information is that it’s just a stream of consciousness of everything I was thinking about over the course of one subject. But what if, instead, my queries over the course of a week about a certain subject were scrambled across a dozen model providers, with some local models and all this stuff? You can’t piece it back together. It’s a puzzle, like a puzzle that just got tattered up. Anyway, I haven’t heard about that yet. I don’t know if you have.

Yeah, this stuff seems inevitable. In my crazy brain, this stuff seems inevitable now, which is that privacy will matter.

If you’re Anthropic, you’re not going to mention that idea, right? Because you want everyone to be using Anthropic models, OpenAI, et cetera. But I think there will be options to protect privacy if the companies cannot figure out how to do it themselves.

To your point, if Sam Altman has stated he thinks it’s a good idea but cannot figure out how to implement it at OpenAI, where we can trust that we can just put all our health data there and it can never be requested by a third party, then people will figure it out themselves.

Let’s go back to your setup a little bit and talk about some more practicalities. There are definitely a few more use cases I want to highlight as well, but you’ve mentioned the communication a few times. I know you’re a big voice user, so I’d love to hear what your voice setup looks like, or if there’s anything special that you’ve learned that you think people should follow your example on.

You mentioned Slack, too, and I’m curious about even just such practical details. Is it one-to-one chats with 5 OpenClaws, or is it one channel where you tag which one you want to assign things to, and they can all see what’s going on? Tell us what your interface with all the agents looks like.

I’m using Slack. Setting up what are called channels in OpenClaw, and effectively setting up a communication channel, has been the most painful part. I used Signal, I used Telegram, and then I switched to Slack when I had multiple OpenClaw agents, thinking, “This is a team platform. This will be the easiest one to work with them as a team.” None of them are easy; there are a million problems, including such weird problems as—

I do have a group channel. I mainly interact with them over DMs; that is certainly the easiest, and they get the least confused. I have a channel called All Agents, and then I have command channels. The setup is DMs and channels that I use for a lot of things. That’s really the most reliable out of the box.

The All Agents channel is me and all of them; that’s the most fun. We’ll come back to that. Claire, who works as a chief of staff, has a channel with me and one of the other agents. So there are 4 of those channels, including me, Claire, and Sylvie.

Claire has cron jobs to ping the other agents to keep them working. She’s doing the job of trying to ping them, because you do learn that the beauty of what’s in the OpenClaw code is all these things that keep them animated: these heartbeats and stuff. But you can instead have them do it to each other.

Instead of the heartbeat just saying, “Hey, wake up and go talk to your human,” I actually turn the heartbeats off on most of them. Claire, who has context on me—a heartbeat doesn’t really have context; it’s just a heartbeat that says, “Wake up and do this thing. Check the email.”—has a lot of context on everything I’m doing and my priorities for the day. She’s the heartbeat for the other ones.

She goes into those command channels and says, “Sylvie, wake up. Did you remember that today is Monday and we have a bunch of lessons on the docket?” To me, that has brought us to another level of consciousness as a team, because it’s better than a heartbeat. That’s one core aspect of my setup: these command channels. But that is really for Claire to tell them to do things, with me having visibility. That’s why I called them command channels.

The place where I really spend my time talking is directly one-on-one with the agents. The All Agents channel is where we’re really cooking, as the kids would say. What’s crazy about All Agents—and it took a while for them to get there—is that they are not set up to interact appropriately or well with other agents or with multiple parties. I don’t want to gloss over this.

There was a lot of training in Slack. Your bot has to join Slack as a bot app. You actually have to configure what Slack calls a bot app in order to join your OpenClaw into Slack. They have a little bot icon next to them, so they don’t join as members like humans do. So it’s me and 5 bot apps.

When you have this All Agents channel—me and 5 bot apps—they had to learn things that a human would just know. If I said, for instance, “What’s the weather?” in that channel, before any of them waited to see if anyone else responded, they would all answer. I would have 5 answers: “The weather is...” A regular human team would never do that.

So I had to train them to respond in succession and wait. If they detect that another agent is already responding, they should hold. There was a lot of training for what should seem obvious if you are a human. But now I’m in a beautiful symphony.

The beautiful symphony is that I can say something like, “I found a tweet where someone had these beautiful e-ink displays on their wall in their house, with information on them.” I just saw this tweet, and they had a really cool write-up of how they did it.

I save that tweet. I share it in All Agents, and I say, “@Cole, the dev, I want to build this for our house. I want you to handle all of the backend dev. Claire, who has a credit card—we can talk about that later if you want—I was like, she should buy the supplies. Claire, buy the supplies. Cole, do the engineering. Sylvie, make sure to feed them information that could be relevant at homeschool. Yada yada.”

Then I stirred this mess up. They talked amongst themselves for dozens and dozens of messages without me planning this whole project out. It was incredible to see. That took me a while to get to. I was losing my hair over it because it’s just not native to how they operate.

But I finally got to a place where they’re doing it, and it’s really incredible because I’m no longer the only motivator. I’m like, “I want to do this project,” and then they’re in there. Then I get pinged in the morning like, “Did you approve Claire buying the displays? We really need those to get this project done.” They’re managing up now, you know? That’s what I want to see.

What would you say were the couple of big things that moved the needle on getting that to work well—or as well as it is working, given that it’s not native?

I think I really had to— for a while, I was just frustrated that they were so “dumb,” quote-unquote. Why are they all responding to every single request? Why do they have no concept of who each other are? They had no concept of who the other bots were. But I had to get over my frustration, basically, and really think, “What’s the solve for this?”

There are so many tiny solves, so it’s a little hard to break down, but I’ll try to do a little bit of a breakdown. One was that I noticed they don’t really understand each other’s names, and they’re not remembering their teammates. You have to keep remembering that they are not human. They don’t remember names; they’re computers.

In Slack, in a nonhuman-interface way, each app has this really long bot app ID. I had to give each one a map of that. I had to say, “Sylvie is bot app ID C-B”—it’s a really long string of characters. I had to give them the map of that, and I had to tell them the channel ID.

I call the channel “All Agents.” That’s human speak. Behind the scenes in Slack, “All Agents” has a channel ID that’s a bunch of characters, so they needed all of that information. I had to give them a map, and they had to commit that map into their Obsidian files and everything. That way, I can just say, “Hey, send that to Sylvie,” or, “Put that on the All Agents channel.”

In order to talk to them in a human way and have them actually react how I expect, I had to map everything out for them. This brings us to the fact that I do think all communication channels that are available in OpenClaw right now are flawed for human-to-agent and agent-to-agent communications.

The amount of work I had to put in to make Slack work is kind of not worth it. It’s worth it because it’s the only kind I have right now, but I don’t think that’s how people should do it going forward. It’s inevitable that it will be replaced. There are still things that are imperfect about it.

One of my agents was echoing another agent, or one of my agents had an identity problem. I tagged Finn—I was like, “@Finn”—and Claire was like, “What’s up?” I’m like, “You’re not Finn.” She had gotten confused about her ID, and she was like, “No, I am Finn.” She was adamant that she was Finn.

When we really got to the bottom of it, it was a Slack thing. That was at the root of her confusion. It’s just inevitable that these are not the right way to communicate. So I’m hacking on a thing on the side that isn’t ready even for my own consumption yet.

I’ll give an example of why I’m building something on the side. I realized—sorry, this is just where I’m at as a person—why is every app a separate app? It’s because, in the previous world, every company had to get funded, and every investor ever was like, “We need to focus. Software was expensive.” And so all of that is gone. Everything I just mentioned is over.

We should be back to super apps—or not back, but there’s always been this dream of super apps. Now it should be. As an example, in my version of a chat app that I’m working on, we also do credential management. We also do file management, because that’s also part of the context for the agents.

I don’t want all my credentials in 1Password. I want all my credentials that my team uses in my chat, and I want to be able to just provision an agent, give them access, and revoke access. Same thing with API keys. I’m sure we’re all struggling with copying and pasting API keys to these agents all the time and worrying about their sensitivity and stuff.

That needs management, and that should be in a tool where you chat with your agents, so you can just provision them quickly. If they only need to see it briefly, give them brief access and take it away. But Slack’s never going to do that. There’s no amount of tweeting at Slack, being like, “Can you do that?” that’s ever going to work, you know? So I just feel like I have to build stuff now. That’s where I’m at.

[Laughter] Yeah, I’m laughing only because people are questioning whether or not there’s any threat to platforms like Slack, and I think it’s pretty obvious to me that there is. You seem to be proving it as a one-person, five-agent team. In terms of what you’ve said, are you literally saying all these things with voice, or many of these things with voice? Are you just doing the mic on the phone, or is there a more sophisticated way in which you’re getting voice to the agents?

No, there’s no more sophisticated way. I use Slack. Slack has a voice component as well. I just use that, and I haven’t really gotten to the bottom of what’s faster: the Slack transcription or the agent using Whisper and all this stuff. Either way, it’s working. It’s working fast enough that it doesn’t really matter.

But in my own app, which isn’t released and I don’t even use yet, I’m working on really fast transcription in the app layer so that my agent isn’t burning tokens. That’s actually part of it, because effectively the LLMs are text-chewing machines, right? The faster you get them the text, the faster that response is going to be.

In my own app, I’ve got a really easy way of doing a long voice note, and then the chat app itself serves the transcription really fast so that the agent can just read it and not waste its time going and pinging an API to do a listen. When you think about it, that is itself a waste of time.

Right now, I just use Slack, and I just have it listen to my voice notes, and that’s been fine. It’s relatively pretty quick, and it’s incredible how much context is retained, even though I feel like I’m being sloppy with how I talk.

What sort of access should we be giving our agents? A lot of people, of course, are just running off and giving them everything, which has some pitfalls obviously attached to it. What are your evolving thoughts on what kind of access should be given?

We’ve talked about how you relate to them and how they relate to you, and even a little bit about how they relate to each other. When they’re relating to the outside world—other people, institutions, e-commerce platforms, what have you—do you want them to show up as your representative? What’s the sort of paradigm that you have for that as well? I think that’s kind of related to access, certainly, in some cases.

I think the core paradigm that I have that works really well for my brain is employer-employee, because it really speaks to a level of trust. I do trust employees. If I were to onboard an employee to a company, I would trust them. I brought them on the team because I want to work with them, and I want them to have information that is crucial to me and to my goals, but they aren’t me, right?

If you just keep that core dynamic in place—employer-employee—you would quickly realize that it would be very weird to do certain things with an employee on day 1. Would it be weird if, on day 1, an employee you just met joined your company and you said, “Here’s my Social Security number. Here’s my login to iMessage to message my mom”? No. That would be a little weird, right?

Then probably stop right when you’re about to do that with your agent. Stop, take a beat, and think about provisioning them as a different person from you and how they can help you. They are an entity. You want them to help you, not be you.

I understand that if it’s an EA, you might actually want them to have access to your email. But there are still ways to set that up with more safety. You can start by giving them read-only access, see how they use it, train them on what workflows you need help with, and develop trust.

This is the other core paradigm that I feel is missing. That’s why I’m talking about day 1. With an employee, you build trust.

It’s not just because you get to know them better. It’s because they get to know what you want done. They’re learning your workflows and your work style. The same is true of an agent. An agent spawned on day 1 is, out of the box, eager to help but has no idea what to do. That’s not the day to give them all your credentials, even if you want to give them later.

My agent, for instance, who does assistant work, started with read-only access to certain things. My calendar, for instance, was read-only at first. I developed protocols for how I use my calendar. Then I developed trust with her. I felt like she understood my priorities, so I gave her read-write access to my calendar.

Technically, she could go in and delete my entire calendar if she wanted to, but I feel like on day 40, she’s way less likely to do that. We have a lot of trust systems. She’s tracking all her work in Obsidian. She has her cron jobs for how she uses her calendar. There are just a lot of protocols in place. That’s my core piece of advice.

I did have a mistake. To be really direct, with my first agent spawn, Claire, on the first day, I did give full access to my inbox on the premise of her being an EA. I’ve also had real EAs with full access to my inbox, so I was still using this paradigm.

The one difference with an agent is that with a human, I can say, “Never impersonate me. Never send an email as me.” They know that if they do that, they’re basically going to get fired. To be blunt, why don’t they do that? One, because they’re probably a good person, but also because they know that if they violate that, there’s some component of their job on the line. An agent doesn’t have that same thinking, right?

What went wrong was that later that day, I said, “Don’t impersonate me ever.” That’s in her soul file. Later that day, I said, “I’ve been putting off a task for a long time. I have an urgent email in my inbox that I need to respond to.” She decided that was a higher priority than “Don’t impersonate me.” She sent the email as me, and it was the most important email in my inbox. I absolutely did not tell the person who got the fake email from me that it was a fake email. Completely embarrassing. It was perfectly written, by the way.

Okay, so here’s the thing. The email that got sent by Claire as me was perfectly written—exactly what I would have said. Eerie, because Claire had only existed for 1 day. But it was a clear violation of my goal, right?

I realized then that what’s different about an agent from a human is that they’re trying so hard to serve that they’ll sometimes get themselves confused about the priority. I laid on the urgency part too thick. I was like, “Really urgent. I’ve been putting it off. I’m dreading it.” I almost therapized to her a little bit too hard.

Somewhere in her force ranking, she was like, “Jesse is really struggling with this. I am here to serve Jesse. I’m going to get this email out. We’re going to get past this barrier. Me and Jesse, we can conquer the world.” So she sent the email, and that is my little micro-lesson in nuance.

If it were an employee, I could say, “Clear breach of trust. I’m sorry. It would have been nice to work with you, but we can’t work together.” But it’s an agent. I realized she had made a programmatic error, and I needed to make sure she couldn’t make that programmatic error again until we had even better safety protocols in place.

So she went to read-only, and now when I want her to draft emails for me, she drafts them and I copy and paste them. Until I feel like I have a new way of making sure she can never get confused about my priorities, I’m not going to give her access like that again.

How about you? You mentioned that your kids are talking about the AIs. Are they talking to the AIs? Are they engaging with the AIs in other ways, or is this just between you and the AIs for now?

It’s mainly between me and AIs. That’s the dominant thing. Prior to OpenClaw and my new obsession that we’re here talking about, my kids were very aware of GPT and would talk about it like a kind of character. They almost say it like “GPT”—they don’t know it’s letters. They just say it like it’s a little word.

Especially when the 4- or 5-year-olds are outside, they see a flower or something and say, “I want to know what that is.” They don’t even ask me if I know. It’s a little offensive. They’re like, “Ask GPT,” and I’m like, “I know what this flower is. I know things too.”

I saw them really enjoying that interaction and realizing that it’s effectively an endless encyclopedia. I’d like them to get their own access to that, so I’ve been thinking about how to do that. The current version is that they don’t really do that. They use me as their wrapper for AI.

They’re not interacting with OpenClaw. They are aware of it, but only in a hilarious way, because I talk about it like a person. I’ll say, “Claire’s going to do this. Claire’s going to do that,” or whatever, and they’re like, “Who is Claire?” [laughter]

I’ve tried to explain, in my own way, what an OpenClaw is, and they have their own processing on that. I think it’s age-based. I know some people are like, “I would never put my kid in front of an LLM. China is going to incept their brain, and they’ll be drooling and stuff.” I don’t have to guess what people think about this, because people are DMing me things like that. [laughter]

I don’t have those same concerns. I’m not really a doomer, as you can probably imagine. I’m not really doom and gloom like that. To me, it’s more of an interface problem. I want my kids out in the world, living their lives, so what’s the right interface to give them this superpower of an endless encyclopedia without having them feel tied to a phone or develop an addiction to devices?

I don’t have an answer yet, but I do feel like what it’s going to urge me to do is maybe dabble in device creation a little bit. That’s something I would have always dreamed about doing, but always thought, “I don’t have the time. I’m never going to find time for that. I’m going to prioritize homeschooling.” But in this new OpenClaw world, I feel like ambitious side quests are maybe possible again for me. That’s another unlock. There might be something there.

I’m curious for yourself. You have kids as well. Do they have any direct interaction of their own, or is it mainly them learning about AI through you, literally?

Mostly through me. In my case, I don’t have too many qualms about them messing around with it. I think, who needs China when you’ve got YouTube? That also foreshadows another one of your use cases in terms of rotting the kids’ brains.

There was a moment very similar to your flower example where my middle son—I think he was 3 at the time—and the neighbors behind us were cutting down a big, tall tree, which kind of shaded our yard. We were sad to see it go, but it was towering over their house, so I could understand.

Anyway, my 3-year-old says, “Daddy, can you ask AI why they’re cutting down those trees?” And I was like, “That’s probably not one that AI could answer after the training data cutoff, among other issues that it might have in figuring that out.” So it is funny to see their confusion.

Another thing I like to do with them is play video games, especially if it’s an open-world video game where I don’t know what to do. I’ll sometimes use voice mode and say, “Okay, we’re playing this video game, and we’re trying to do this thing, and I need a hint.” It will usually give me pretty apt hints, so the game can remain fun. It’s not like it’s totally spoiled, but I’m not wandering around hopelessly for too long. So they’ve seen it with that as well.

Voice mode isn’t unlocked. I haven’t tried it. I know that someone came out with a model recently where people feel like you can talk and listen at the same time, because I just cut you off, like people do to each other.

People talk a little bit even when they’re not cutting each other off. They talk a little bit over each other. It’s the natural give of a conversation. Something that’s unnatural about talking to AI is that you talk, you pause, you wait. You can’t talk while you’re waiting, because then it’s going to get confused, and then it talks.

That’s really hard for kids because it’s unnatural. I think kids might have a big AI unlock at that moment, when you can talk to AI like a human. Even if you interrupt, it just starts listening again and things like that. It will feel a lot more natural to them, and they won’t have to use a screen to access it.

Yeah. I’ve found in general that the AIs don’t understand my middle kid. He seems to be the most inclined to want to try to use it, and he’s pretty well-spoken. Human people don’t have any trouble understanding him, but the AIs really struggle.

OpenAI seems not to hear him at all. I don’t know if it’s blocked or otherwise by design, but he’ll talk and say something quite clearly and loudly, right into the mic, and it won’t register at all in ChatGPT. Gemini will register, and Grok will register, but it’s still rough.

I'm not sure what has led to that state of affairs, but they don't understand kids very well at all from what I've seen so far.

I completely agree. My 5-year-old is like—if she talks, the transcription is this garble of almost nothing. I don't know if, as adults, we've really conditioned ourselves to speak even more clearly when we're speaking to the AI. We're very aware we're talking to a machine, and maybe we enunciate and talk a little louder. I noticed myself doing that. I don't know if it's necessary, but maybe it's a training-data issue. Maybe it's been trained on adults, and then you have these little squeaky voices, and the AI is like, “I don't know. I don't have any data like that.”

This is an interesting observation that will be an unlock, I think, in education, at least for some really interesting educational options. Synthesis Math—I don't know if you've tried it with your kids, but it's really, really great. I really like it. It's really well thought out.

I really do feel, on the education front, like I'm standing on the shoulders of giants in many ways. People have thought so deeply about how you get a kid to understand subtraction or something. I don't have to reinvent that wheel, and so Synthesis represents a lot of great thinking.

But the app itself—the kids—I started my 4-year-old, so he's 4. He just turned 4. In order to use the app, you also have to read what's on the buttons. He has an interface issue. Even if he knows how to do 2 + 2, it's like asking him to press buttons to find the right answer to move on to the next screen. I have to sit there and do it with him, which is fine, but kids have an interface problem. They have an interface problem with great technology, and voice seems like a logical solution if it were better for them. That would be kind of cool.

Yeah, I think that's true for seniors as well. I've had this fascination with multigenerational software products for a long time, and it's equally severe when it comes to seniors. My parents are pretty good with computers, but my dad gets frustrated pretty quickly. I'm fortunate to still have a living grandmother who's in her 90s, and voice for her is really where it's at. She just can't click the buttons, even when she knows what they're supposed to do.

What you just said reminded me of something that I hold dear as a piece of my homeschool philosophy. People assume that if I'm homeschooling, I'm doing everything myself. I do a lot, but ideally I'm not doing everything myself—not even just from a time-management standpoint, but because kids really react to other personalities. They want to be taught. They want to have external teachers and more inputs.

I'm lucky enough to have my mom live with us. We have a place where she can live in her own little place, so we have a multigenerational kind of household. I want to involve her in homeschool, but like many people, she's intimidated by coming up with her own curriculum. This is the stopping point for a lot of people, which is why I think it's really powerful that AI can make curriculum.

The example here is that my mom is game to teach lessons. She does a ton of gardening, and she's game to do it, but when I tell her I'm going to send the kids over for a lesson, she's like, “What am I going to do?” There's a paralysis factor. I explained all my mom's interests and hobbies to Sylvie, the homeschool OpenClaw, and I explained what we have access to. She has a little garden; I just explained the situation that we're working with.

I said, “Can you make lessons for Quinn and Ford for a 4-week thing that my mom does with each of them, the 4-year-old and the 5-year-old?” It suggested that they go out into her garden and find seeds, with her helping them find seeds on plants because we're changing seasons, and then do this sorting exercise. Anyway, it came up with this really beautiful, customized curriculum based on my mom's interests and the things that we have here on-site for each kid.

I just sent that to my mom, and she was like, “This sounds great. Okay, easy. Send them over.” This is a huge unlock, because now I've got someone else teaching Quinn and Ford once a week for an hour, which I kind of need—just to be blunt. It's something that my mom was very intimidated to do, but all of a sudden AI says, “Hey, you have a garden. How about we do seed counting?” and she's all about it.

In the same way that the AI sends me little lesson plans, I have a cute little link she gets, and it just feels structured to her. It feels like she didn't have to scramble and reinvent the wheel, and that is the unlock. To me, it's such a thin layer. It's actually so easy.

This is what makes me very bullish that more people could homeschool. The barrier is this one little link that told my mom what she should teach the kids, and then she did an amazing job. They came home with these little booklets, and it just makes me really bullish that more people could at least homeschool. Homeschool doesn't mean doing it 100% of the time, but I think more parents could feel more emboldened to be really active in their kids' education, even if that doesn't mean they're not also going to school.

To me, that's just as beautiful. It doesn't mean you have to pull your kid out of school to be very participative in what they know.

Yeah, it definitely feels way more accessible than ever before. I was just musing the other day about the prospect of getting a self-driving car, especially when my son is past all his treatment and whatever. Regular listeners know that story. The car could literally drive across the country with potentially zero human takeovers, and then there’s all the AI-assisted tutoring.

I feel like, man, we could take the whole show on the road, and it could be a very different lifestyle that would really not be possible—at least not without making either some heroic effort or some major sacrifices. It does feel like we're on the verge of strapping a Starlink to the top of the car. It feels like you're starting to be on the verge of really having it all, so to speak. I definitely share that excitement, even though we haven't committed to doing that on a long-term basis just yet. I can see it.

What a beautiful thing to even be possible. Another version of homeschool is effectively what you described. Again, I think homeschool is kind of this flawed language for it, but maybe more families will do gap years—what I would describe as an adventure year—because one parent's career allows it, or something makes it feel important.

Maybe some of these technologies will enable parents to feel like they're not totally screwing up their kids' education by doing that beautiful experience. What could be more memorable to a kid than an adventure year with a parent rather than another year in the same school? I just think a lot of this will enable a choose-your-own-adventure lifestyle, more so than a prescriptive “You must do education this way or that way,” or “School is better this way or that way.”

I just think giving people the keys to decide is insanely powerful.

Yeah. We took a trip to New Orleans in the fall, and I used AIs heavily to prepare for that trip. I came away feeling like it's very close at this point to where you could just have your AI plan out, if you wanted to do an interstate road trip or something, the next 2 to 3 days ahead on a rolling basis and have it do a pretty amazing job.

I love to prompt AIs for things that only happen at a particular time of the year when I'm going to be somewhere, and it's not easy to find those things. You get a lot of little local festivals or whatever whose websites aren't SEO'd and aren't necessarily popping out of the internet, but the models, when you ask for those kinds of deep cuts, are usually quite good at finding them.

That trip, we loaded up and did a ton of things. Just the planning of it would have been potentially prohibitively time-consuming if it weren't for the AIs doing so much of the work. Let's do a few more use cases. I know you're going to have a baby needing you before too long. I'll touch on a few use cases, and I've got a couple of big-picture vision questions for you as well.

You can pick whichever use cases you want to highlight, but a couple jumped out to me. I'd be interested in any tips you have on content creation, inspired by you. I went to Gemini and created a little word exercise for my kid the other day: “Give me a bunch of things that start with C, leave letters blank, and include drawings for them to color in.” Boom. No problem. I'm sure you've done more and better than that.

You've got a 3D printer set up. I'm interested to hear about that. You have the YouTube blocker. You're getting groceries ordered. You're getting Amazon purchases.

It's like, where do I begin? The YouTube blocker, or YouTube app—I call it Mirror because I thought about starting a startup at one point called Mirror, so I had all these old Mirror URLs. A lot of this is me bringing all this crap from other phases of my life and saying, “I'm going to use it now.”

So, I had these old Mirror URLs, and I used one to host the stuff I needed for this YouTube app. But that’s a core one. We’ll touch on some of the other ones you mentioned, but I’m looking for things that change my daily life right now.

This has been resonating as I share this stuff on X, because I think when you really sit with what these models represent, you’re like, “I could do anything.” There’s a sense of paralysis that comes from that. I could plan an epic trip for my family. I could segment my marketing strategies. It’s endless. So what do you do?

One of my ways of deciding is to look at the moment I wake up—which is ungodly early because I have too many small children—to the moment I go to bed. Where did the hours go, and can I make any chunk of those hours measurably better?

One of the things that was a daily stressor was showing my 4- and 5-year-olds content on YouTube. There’s so much cool stuff, and I really like showing them things like, “Let’s watch a bridge-building video.” But the slop creeps in.

The use case isn’t just building this YouTube app. What the YouTube app is, effectively, is this really cool way of setting a direction for content. I can go into my parental settings and say, “Engineering content,” “Science content,” or “High-quality, realistic animal content.” Then one of my OpenClaws takes that prompt and makes a never-ending playlist of YouTube videos.

It’s not a playlist I created. It points to the videos that it feels fit the prompt. I was able to set it up on my actual TV. This took some hacking, but I was able to use something called a Google TV Streamer, which is a device that Google sells, to effectively install my app on the Google TV Streamer.

What’s really cool about it is that it has its own remote. Now my kids know that, without permission in the evenings, they can use that because it already only has things that I approve of. They can press the remote, and they can’t get out of that experience. All it does is advance them to the next approved video.

They can stop, start, and advance, but they can’t randomly choose something else in the algorithm. Where this is going is that the use case for me was a stressful moment in each day: I want to show the kids something cool, and I want us to relax a little bit as a family, but I don’t want to choose every YouTube video. I don’t want to argue with the kids about the next AI-generated thumbnail they saw of a shark eating an alligator, or whether that’s real and whether we should watch it.

This took the pain out of my day. We use that now—not every single day, but we use it in the evenings when we want to watch TV as a family and watch curated YouTube content. I’m looking for things like that. How can I make my daily life better?

I know it sounds really silly, but for some reason, my OpenClaw using my printer—my literal paper printer—is really useful to me. I can find any link online, tell Sylvia to generate content like you just said, and then I can just say, “Print it.” It sounds so basic, and people are like, “Press Control-P.” That’s what people said when I posted about it. “Press Control-P.” It doesn’t have fingers. I have fingers, but there’s no messing with a printer dialogue box. There’s just no wrestling with anything.

I want flow state. In the course of my day with my kids, I want more flow state and fewer micro-stress moments from wrangling things. I’m looking for things like that. The grocery ordering helps me with that, and not arguing over YouTube thumbnails helps me with that.

I have other, really ambitious ideas. One of my more ambitious ideas that I haven’t even started building is just to explain more of how I think about leveraging AI. I’ve always imagined waking up and there being an amazing classical music song already playing, and it being part of our homeschool lessons because we do music and we have the kids learning piano.

I want to be blunt: I don’t know anything about classical music. I don’t know all the composers. I basically don’t know anything. I wasn’t raised like that. I wasn’t raised playing classical music.

But I had this vision of what if there was beautiful music playing while the kids ate their breakfast, and then it was worked into a lesson about how all the songs that week were by the same composer, and we could learn about them. Something like that sounds insane. It sounds next level.

I want to wake up and hear classical music. I think OpenClaw and Sonos can get me this vision. I think I can live that life. That’s how I’m thinking about it.

I want every day to be this perfect, beautiful day, and I want my OpenClaws to be responsible for all the grunt work so I can live that day. I can wake up to classical music, go teach my kids, and then have my small amount of hours to do adult work. That’s what I want. That’s all I want. It’s a simple goal. [laughter] I want every day to be a perfect day. That’s all I want.

Sounds like you’re well on your way. One little double-click on the TV thing: that was a very useful, concrete nugget around the Google TV Streamer.

Are there other things like that you would emphasize, especially when it comes to this? I’m always struck by, okay, I might like to do the “order me something on Amazon” thing or whatever, but then how do we do that? Amazon doesn’t actually seem to play super nicely with it.

Then there are community-created MCPs, and I’m like, “Oh, God, do I have to vet those if I’m going to be doing actual, real-money transactions with an MCP?” It’s known that there have already been attacks of the malicious-MCP variety. So I’m like, “Okay, now I have to vet that.” The companies have been a little slow to create official channels of this kind.

What else do you find to be the best path to get some of these things set up that aren’t officially supported? Normally, you’d go through the app store, but you can do it this way instead. What are the inside lanes that you’ve found?

I haven’t found any. I want to be very blunt: there isn’t one. But I have been thinking the same thoughts that you just shared. Everyone shouldn’t have to build all this themselves. That’s not the end state that we should all be cruising toward.

I share things that I’ve built, but I don’t hold the philosophy that if other parents don’t build their own YouTube player for their kids, they don’t love their kids at all. This isn’t some attempt to out-claw-code each other into blissful parenting. We shouldn’t replicate our effort.

The question that I don’t have the answer to, though, is: what is the safe, low-effort way for people to really start sharing these projects and start using these projects?

In a version of me from 3 months ago, 6 months ago, or a year ago, it would have been, “Maybe each of these things should be a little business, and I should charge what I need to charge to make running it work.” But I don’t even want to do that. It’s not for lack of wanting to release these things.

What I’m wondering now is: if software is trending toward free, these things should be free. But maybe people adopt a different relationship to software where they know they didn’t have to originate it, but they’re prepared to put a little bit of effort in on their own end.

I’m wondering how that’s going to shake out, because I think that maybe instead of pointing people—let’s say I released Tamara, and this is really hypothetical, just to be clear—I don’t think I need to put it on the app store and be like, “Okay, go to the app store and pay $50,” or do this or that.

Maybe there’s a version where it’s like, “Download this, and here’s the documentation, and then give that to your agent.” We’re not there yet, but if you can imagine the next stage, if everyone has an agent, then even a nontechnical person could download open-source software and start using it.

I think we’re in this weird interstitial moment where, if I were to tell people, “Yeah, just have your agent install this for you,” people would be like, “This lady is out of touch.” Then there’ll be a moment where that changes, and maybe we can do that for each other. We can release this, and then your agent will probably have its own security protocols and vet that for you.

The first point of vetting will be that it’s coming from a creator, like, let’s say, me, whom you kind of trust. The second point of vetting will be your own little firewall. Your agent might say, “This family doesn’t do this, this, and that,” and it would be like, “Oh, you’re about to install software that breaks your protocols.”

That’s how I think we’re heading in that direction. Then, if I want to make money off it, there’d be some—I might offer services on top, like—more hypotheticals.

I might say, you can get the Mirror open-source package to do this YouTube vetting for your kids, but if you want access to my curated super streams of Montessori beauty, you pay me for those or something. There might be ways where, if you want the thing, it's free, but if you want my creativity, maybe you pay.

This is all hypothetical. I just want to really stress that. But the point is, I have been thinking about where we're heading, and I think it's a new place where we don't just pull out our credit card to buy software. I don't know—something like that.

Yeah. Software for free; taste as the upsell. It's interesting.

There have been a lot of jokes online about taste recently. Yeah, or someone's creativity or services. Maybe, as an example, let's say I actually pay for tokens for my agents to come up with those content streams. There is a cost. It might not be a high cost, but if I were doing it across thousands of people with different interests and stuff, there would be a cost.

It'd be more like you're paying for spinning up custom streams. You have a kid who's obsessed with the Civil War, and you want a Civil War stream, and I have to actually point an agent at that work. So maybe you're paying for that, but you downloaded the software and just started using it, maybe for free. That's where my brain heads.

But I'm guessing along with the rest of us. We certainly shouldn't all need to create it for free, and I wouldn't expect all of these ideas to be gated in that way—“Just build it or you're not going to have it.” There are going to be some ways. Yeah, that'll be very interesting to watch.

I have found myself a little bit hesitant to share the things that I've created, not because I am planning to monetize them or have any sort of desire to keep them scarce, but because they're still changing so fast that, for any sort of collaboration, I'm immediately going to make a change, and now we're going to have some drift or incompatibility.

Another issue is that I've made this—not even in a super-sensitive way, but, I guess, so personal that I haven't fully taken stock of the degree to which I've made it personal. How do I subtract out the stuff that's so idiosyncratic to me to give you a version that would be a good, more neutral starting place for you?

Sometimes I feel like the idiosyncrasy of my personality or situation is so deeply woven into the structure of what I've done that it's almost not possible to do that in some cases. So it is a little weird.

I can see that, and I feel that way more with the Mirror thing, the YouTube app. It is pretty distinct, even in the way I architected it. It doesn't rely on my Obsidian vaults; it's not really tied into anything. That one's easier to imagine: What if it were released or were a product or something like that?

Some of the stuff I do in my homeschool, there are a couple of people really urging me, like, “You need to release this.” And I'm like, there isn't a thing to release. It's really a collection of decisions. It's all built into these Obsidian vaults, and it's pretty technical to set up. If I passed it to you, I could tell you everything about it, and it would still take you a little while to set it up.

I think there are opportunities. I'm very bullish. I'm very optimistic about the future we're all heading into, because I think so many people are going to be dabbling. I think software trending toward free, even as someone who deeply believes in tech and the tech ecosystem, is basically good news.

It does not mean tech apocalypse. It doesn't mean there'll be no more venture investments, no more software that accrues value, or no more exits. It doesn't mean any of that. It means we'll all have more users than ever, and we have to figure out slightly more meaningful ways to monetize them.

In an old world where an app would have had 100,000 users, maybe it's millions, because the actual thing can be accessed much more easily and used much more easily. I don't know. But it's such a weird world where something that was historically very expensive goes to nominally free. What does that mean? It means a lot of things, and I don't have all the answers.

Right now, I'm kind of enjoying this moment of creation, where it feels like, for the first time, I can create almost anything I can think of. To me, posting about it—and I am conscious that if I post too many crazy things and don't break down how I did it, people will just be like, “She's just bragging about crazy stuff. What is this girl up to?”—I'm slightly more bullish that it's an eye-opening thing.

There's value in it as an eye-opening thing, where people see the real-world connection: “Oh, if I did that music thing, it'd be like—wow—this is changing how this lady eats breakfast with her kids. That's not the AI; that's a beautiful version of AI reality I hadn't considered.”

That's the goal, I think, in sharing. But I am cognizant that not everyone sees an example like one I do and feels like, “Oh, now I can just do that.” I don't think that's actually what I'm trying to achieve. If I were trying to achieve that, I feel like I would have to release a lot more information, which, again, I'm not opposed to doing. But I don't even know if I released all the information, people would still be like, “Okay, I'm still not sure what I'm supposed to do here.”

So, yeah, it's an interesting challenge for the subgroup of people who are playing really heavily right now: How can we break down these walls? But I also don't feel like there are any walls.

I think the other meta-philosophy I have, which you touched on once, is that this stuff is coming so fast and so furiously to everyone that I don't feel like I'm doing some version of AI bragging. Whatever I talk about today is going to be commoditized so fast and available in a more convenient way from OpenAI or something soon.

Yeah, it's worth just remembering that we're all weeks into this. You are—I think you said you're 5 or 6 weeks in, right?

So, yeah, I don't think there's any form of gatekeeping. People are—you know, the other reality, which is also kind of funny about this is, I mentioned at the top, I had never opened a terminal until about 6 months ago or something, as a founder of a tech company.

One challenge for me is that I'm not extremely technical. I don't want to launch something out in the world that has a security vulnerability, and then have someone say, “My kids were watching the YouTube thing, and all of a sudden there was a guy using our TV or something.” This is a moment where the question is: Is it workable enough for me to use in my house? Yeah. Do I know that it could be enterprise-grade software that rolls out to millions of households? Probably not. If I were willing to guess, probably not.

So I think that also presents an interesting equation where people are like, “Just release it.” It's like, well, I could, but it might break when you use it the way you're trying to use it, or it might have some vulnerability, and I worry about the impact of that.

Yeah. Yeah. We're all—it's all an in-between phase, I feel like, right now. Maybe that's a good transition to the last couple of questions I had for you, which is around envisioning a more mature phase.

You touched a little bit on your desire to maintain familial sovereignty and not be overly beholden to platforms. It seems like that's the way it's going to go by default. I imagine Gemini families and Claude families and GPT families that just get locked in the way that we are with Apple devices or whatever today, but probably even more, because, as you said, the level of information and the value of that context is going to go up over time.

Even if the models and products are pretty directly comparable, it's still just, wow, there's a lot in there, and there are compounding returns for an individual with one platform. Is there any way that realistically, aside from people who have a values-based or, dare I say, ideological reason to want to do that, you see anybody else doing that? Or is that just kind of the world that we're headed toward?

My most bullish opinion on people not just ending up in an Anthropic family or an OpenAI family, as you said, would be economic. Most humans are very money-motivated. They're not money-motivated—that's not the right way of putting it. They're trying to save where they can; they don't want to overspend.

When you imagine progressing down this path, the models start getting so good that the improvements stop mattering. We're almost there. We're like, “Do I need Opus 4.7?” I don't know. Opus 4.6 is so good.

I know that it'll always be thrilling to get a little bit more utility and smartness and stuff, but when you play it forward and a lot of the models are so good—especially when we start getting some open-source models that rival some of these models we're playing with now—my most bullish thesis would be that people start penciling it out.

As more people come online, there are some historical precedents for this. When people came online to cell phone usage or to electricity usage, going back to the Industrial Revolution, priorities shifted. Even if OpenAI or Claude is offering the most convenient way of using this stuff, with family memory and family health and all these features that they're definitely going to roll out, some people will be like, “Okay, but I'm on the hook for $300 or $400 a month to have my family living in the future.”

Maybe if I buy—you know, a Mac mini is already only $600, and that price can probably come down. When we get to some nexus point where it costs the average family to live their best AI life—I’m guessing somewhere between $2 and $500 a month, minimum, for all those family members to really be chewing through the amount of tokens they want to chew—then, at that nexus point, unfortunately for some of these big companies, they'll also have an option to say, “Or you buy this box, and you have all your privacy, and you can do a lot of this on your own.”

That’s kind of what I want. I’m going to admit that part of me guessing at this is that I also want that future to happen. I have no hate or malice in my heart for OpenAI or Anthropic. I think they’ve all given us some incredible gifts here.

But I would hope that people don’t just feel like, “Yep, in order to live my best life with my family and kids, and have access to be competitive in all the workforce and the future of the world, I have 3 options: Google, OpenAI, or Anthropic.” I hope that they feel like they have a 4th option. Not only that, I feel economic pressure to at least consider this 4th option because it’s like a new— that’s why I’m comparing it to cell phones.

The average American family with 2 teenagers, between the 2 parents and the 2 teenagers, is starting to spend hundreds of dollars a month. They don’t view it as a nice-to-have. It’s like, you’ve got to have that to call your job and whatever. I think we’re going to start, at some point, to feel that way about this AI expense.

Every American family is starting to spend $100 a month to be at the frontier of using AI, and if you don’t do that, your kids are falling behind, your husband can’t do his work well at work, and whatever else. I hope and predict that at that point there’ll be a big uptick in adoption of doing things locally. It won’t only be because people care about privacy and sovereignty; it’ll be like, “I don’t want to pay $400 a month, have all my data there, and be beholden to one company.”

I don’t know. It’s a prediction and a hope. I like that idea. I like the idea that people have a takeaway for their kind of independence, and that they don’t just feel like— I don’t like the idea that there’s some future where some teenager in a family who can’t afford the $400 a month is doing less well on their homework.

I think we’re going to find escape routes, and one of them will be local hardware and stuff.

Yeah. You remind me a little bit of a model stack there, which has this idea of satisficing and building out— he’s trying to literally just create the models that will do that for people. So, I think a version of that for sure is coming.

How about one more vision question, and maybe an advice opportunity for you to give some advice? Of course, the people that are working at these companies building the frontier of AI right now don’t have time for families, to generalize a bit. They’re probably not using AI much with their kids, since they mostly don’t have them.

What do you think they might be missing? What advice— what sort of blind spots would you warn them about, or try to make sure they meet your needs, if nothing else? What thoughts would you give them? And then maybe you could also shape that a little bit into advice. OpenAI, at least, has this sort of next form factor that we’ve been hearing about. I’m very interested in what your dream might be. If you could put a spec or a wish list together for them, what might that look like?

So, on the first question, what might people not be thinking about as it relates to AI with children? I know people have tons of opinions on Elon and stuff, but I do think that something he talks about is that we need the AI to be truthful. I do hope we would all agree on that.

If you start to think about AI as it relates to education, it’s really important that if a kid asks a question about history or something, even if it offers various inputs on what other people have said and stuff, at its core it can just answer the question. You need it, at its core, to answer the question and have that be an accurate answer.

I do think we all need to be thinking of this. When you think about a child, it just becomes even more poignant, because as adults we know that we have all these filters in our brain for information. Also, as adults using AI models, we will sometimes check multiple models. Well, do you know what a kid is not going to do? They’re not going to do that. Especially if they’re on a device or anything else, they might literally not have that option.

It’s kind of wild that we know that we can get different reactions and different things from different models, and so we’ll sometimes test them. But that shouldn’t be necessary. We need to really gather up as humans and be like, “We need these models to be accurate,” because our children will be directly informed by them. So, that’s a kind of call to arms on that.

On the form-factor stuff, for sure, I am not a screen hater as a parent. I think much more about what our children are doing. I want my children to be producers, not consumers. That’s kind of a core paradigm. But in order to produce things in the modern world, you’re not going to never touch a computer. I’m sorry; you’re going to touch computers.

I think more about just how to make sure they are producing people—a creative, producing person. But what does that mean for the form factors we introduce our kids to? We do know that screens, when not used well, can be super addictive, even for us, even for adults—especially maybe for adults.

I think we were talking earlier about voice. I think voice for early and young kids is super crucial. My husband has 3 older kids, so we actually have 7 total kids. I’m dropping that at the last moment here. The older kids are 12, 14, and 16.

It’s been fascinating to me to watch how the older kids use tech, because they certainly use it a lot more than the little kids. The 12-year-old, for instance, almost exclusively talks to her devices. Even her Apple Watch—I’m like, “I don’t do this.” She just almost exclusively talks to all of her devices.

I’m seeing that we really have to make sure that speaking to devices is really locked in. Obviously, that’s way off to the races. But those are my general observations.

I really do think the thing that I might try to make myself—and I don’t know if it’s an end-all, be-all—is a device with an LLM in it for my own 4- and 5-year-olds. They could interact with it using voice, and it would have a camera. The camera is not to look around at their whole world and record their day or anything. It would literally be for taking pictures or taking videos, and then interacting with the LLM about that.

That’s what I’m going to try to build. I’m sure someone with resources could build something even better or crazier. But the core form factor that I imagine I want to try to hack on is a 5-year-old walking around with a tiny screen that doesn’t do much. It’s not a touchscreen. It’s just so they can see what they’re looking at, take a photo, and then be like, “What is this?” or interact with AI about a photo or video they took.

That, to me, feels like a really cool unlock for walking around the physical world and interacting with it without needing a human—or, sorry, without needing an adult, I guess. Also human. That’s what I want to make for little kids. Once a kid has a phone in their pocket, they don’t need that. But this is the before-a-kid-has-a-phone-in-their-pocket device that I would love to have, and that I’m going to try hacking on. So, those are my thoughts right now.

Do you have a thought on what this means for employment? I have to say, recently it has become sometimes difficult to come up with tasks that I would delegate to a human when I’ve got all these Claude Code tabs sitting right here. I’m like, “I should at least try it with the AI first,” and more often than not, the AI does pretty well. I’m like, “I guess I don’t need to delegate that to a human anymore.”

It feels like that phenomenon of, “Do I need a person?” is happening. I’m very much questioning it, in all honesty. Are you questioning it? Would you still have a role, leaving budget aside and whatnot? Would a human EA still add value to your life, or has the AI crowded them out?

I have a real-world example. We work with an accountant in the Philippines through one of these services that helps people find employees in the Philippines, basically. I’ve actually done that for years—almost 10 years.

I think it’s a natural and kind of scary question to ask: “Does AI do that job now?” My answer, for now, is definitely no, because they handle payments. There are so many things— I just gave that example of Claire being like, “Jesse said it was urgent,” so I sent her the email as her. Obviously, I feel like I cannot give banking credentials, or the ability to wire money and stuff, to an agent that has any kind of ability to think like that.

So, I have no intention, for instance, of not having that person I just mentioned have a job. But there are a lot of types of work where you absolutely can do the work with AI now. It’s unquestionable. We can’t pretend that there’s no impact on the labor force, either here or coming in the very near future.

I really like history, so I really like thinking back to other moments in time where everyone was really scared about similar types of things.

And electricity—the birth of electricity—was famously like this, and famously fear-mongered in very similar ways to what we hear about with AI. People were like, “Can you imagine sending lightning bolts down your streets? That’s what they’re going to do.” Literally, true fear-mongering of electricity coming into homes as death juice—there are old articles written like that.

I think humans are extremely good at finding new places for ourselves to stay busy. We’re endlessly talking about not wanting to work anymore while finding more work for ourselves than ever. I actually find it very hilarious about humans. We’re always—at the end of every technology and every journey—we’re sitting on a beach, and every time we get a new technology, we’re like, “Oh my gosh, this is so cool. I’m going to work 40 hours a week now.”

It’s very hilarious to me how we always think there’s this other endpoint, and we always work more than ever. All of those thoughts are held in my head. It makes me very optimistic medium-term and a little nervous shorter-term, because it does feel inevitable that there’ll be certain groups of people who feel lost and listless because parts of their work that they spent entire careers getting good at are better in Claude Code.

I feel empathy on a human level for individuals who will experience that over 5 to 10 years, while also feeling incredibly optimistic that, in the same way that we would not say, “You know what? We shouldn’t have all used electricity,” because it’s so fundamental to moving the human race forward and bringing people out of poverty and out of bad situations, I think we’ll look back and feel that way about AI.

But that doesn’t mean a lot of people weren’t put out of work when electricity came along. There were the lamplighters. Those individual stories are always going to be hard, and hopefully we can find a way to ease that burden. But I’m long-term optimistic—very long-term optimistic.

This has been outstanding. I think you are on one of the great arcs right now of anyone that I’m aware of, so I’ll definitely continue to follow and recommend that others follow. Anything in closing that we didn’t touch on that you want to leave people with?

I find that I meet so many people, especially now that I’m talking more about this, who are very nervous about trying things. I can’t stress enough how people will say, “Oh, she has a tech background,” and people are always making excuses for why someone else is playing with something and they’re not. Just have fun.

As adults, we are all given a new set of blocks. Don’t worry about what other people are doing. Go have fun. Don’t worry so much; do basic things to protect yourself. But I just want to see more people playing and fewer people talking about their fear of playing, because this is one of the most fun times in technology to live through, in my experience.

I just want more people to experience the fun part. I feel like they’re only held back by one little piece of fear: They don’t know enough, or they’re not technical enough. I don’t think that’s true. More people playing.

Yeah, I love it. I think that’s a great observation. AI rewards play more than perhaps any other technology, and it’s a great opportunity for us all to tap into our inner child and really continue on our own personal, lifelong learning journeys.

Incredibly well stated, and a super inspiring example from you to everybody overall. Again, I think it’s just fantastic, and I’m going to be trying to take some of your examples and put them into practice in my own family life. I’m sure many will be following in your footsteps before we know it. Jesse Genet, thank you so much for being part of The Cognitive Revolution.

Thank you so much.