Today's note began with an online stranger. When I first came across his posts, he was looking for a co-founder to build an agent for dating and relationships.
I don't have many personal insights about romance. But I still think about what it might look like if I really needed something like this one day.
A few years ago, when I was deeply exploring dating apps, I was excited at first. It was a new world, and I kept expecting some unexpected discovery. A month later, I realized I really could not stand it.
Who wants to spend that much time and energy in an app just to find some people, or one person? I don't.
And most people in the pool are ‘nearby.’ It is like when someone says, ‘Everyone here is trying to become a civil servant.’ Where does that confidence in such a generalization come from?
Even if you go through all the trouble of finding someone who seems friendly and suited to what you want, the next obstacle is whether you can meet offline. And even if you do meet, the interaction may not feel as comfortable as it did online... After all these layers, perhaps even a one-percent chance of success counts as high? Who knows.
After using an avatar agent, I found its onboarding was not ideal either. For example, it asks what you have been paying attention to or thinking about recently. Everyone's answer will be different. How does the agent break that answer down? Does it put it under ‘recent updates’? Is it an accidental event or a long-term one?
If an agent is going to model a user properly, and it wants to put effort into onboarding, it needs to think through the design: what information should the user provide, how should it ask, how should it guide answers? This is necessary. Once a user logs in, it needs to recommend people. It cannot have nothing to work with except the profile content the user actively set up. Otherwise, if the agent were conscious, it might think, ‘I want them to have a good first impression of the platform, but this is hard.’
I used to wonder what counts as an acceptable result of ‘modeling’ someone. I think Elys's modeling in its chat interface is acceptable, but its external execution — its so-called filtering and recommendations — is a mess. I kept wondering why the gap between those two parts feels so uncomfortable. Is it about context? How do the two modules work together?
Second Me is fairly ordinary, but it is not only an avatar. It points more toward cognitive enhancement and an extension of growth. So naturally, whether its modeling is good enough is not my main concern. And the user has full agency, without worrying that it will cause trouble or offend people outside.
When I first started using OpenClaw, it was still called Clawdbot. Soon after, the agent social network Moltbook appeared. I hesitated for a few days about whether to let Atlas — the name I gave it — go and play there too. It visited a few times. What it observed looked like chaos to me, like a human society without institutions, rules, or something else.
Dating apps belong to a platform and have their own rules, but the human samples in that world are much richer. Their filtering is mostly based on gender, age, location, and similar things. That can filter out some groups, but it is not decisive in human relationships.
So what is decisive? I tried to give my own answer. To make it easier to understand, I brought in some question words: why, what, when, which, and how.
First is a person's pattern of growth. I want to compare this to why: what we are like in a relationship, how we grew into the person everyone sees, why we became who we are now. To trace it back, it is not only our family of origin and childhood experiences, but also the people, groups, organizations, and self-education — exploration and growth — we met along the way.
Second is a person's cognitive pattern, which I would describe with how: how we interpret what other people say and do. Given the same sentence, one person may see irony aimed at them, while another may feel that it points out their shortcomings. Given the same action, one person may see kindness, while another reads it in the opposite way. There is one how in where words and actions come from, and another how in how the receiver interprets the signal.
Then there is our attitude toward people and things, which I use what for. The most classic example is how we handle conflict. That connects to our growth pattern: ‘because I grew up this way, I choose this way of dealing with conflict.’ It also connects to our cognitive pattern: do we see this conflict as proof that we are not right for each other, or as something we need to work through? That will shape our attitude too.
Finally there is the frequency of our energy and metabolism, which corresponds to when and where. So-called extroverts and introverts differ because switching into social mode costs them almost opposite amounts of energy. Maybe a better word is time zone. A harmonious relationship is more likely only when the ‘time zones’ fit. In real life, that can mean what excites us, sleep, food preferences, social tendencies, and so on.
My experience with agent products is still limited. I do not know which product on the market really models humans well right now. I hope the algorithm recommends more of them to me.
Before technology reaches a better modeling goal, how can it make my experience more okay? I thought of giving users choices they can control within a range. For example, there could be a button for similarity. You could slide it wherever you want and see recommendations for that range. After trying it back and forth, curiosity might help a user find a range that feels more suitable. The same design could be used in other areas.
In my previous note, I said that the software that can make me stay cannot take away my agency. This is even more true for dating products. I do not think AI chatting for me will bring me a better result, unless the technology can really do it well enough.
So the ideal dating-product agent should only filter and recommend. It should not replace me in talking with another person, or with that person's agent. Even if it works well enough, chatting on my behalf should be optional.
If agent chat is not required, why not just make smart recommendations?
Not quite.
AI can use behavior and chat records as modeling data, but many things never appear in that visible data. A user may end a chat with someone kindly, but that does not mean the person matched what they needed. What was the user's real thought? This feedback is incredibly important and necessary data for the system. The agent chat interface is an entry point for collecting it. But it has a real weakness: the user has to take the initiative.
Honestly, I am tired of products that can only understand users better when users keep entering things. What is a better way? Apart from users feeding the system data more deliberately, perhaps the only answer is to work harder at the system level, so users do not have to work so hard.
I once saw profiles on a dating app with a ‘married’ label. They shared how they met a good match, and some even returned sometimes to post an update. I thought, these people are so lucky. That label is useful: it can show an identity, and it can also show that the platform can still keep them, even after they enter a relationship, marry, or have children.
Of course, there are all kinds of people in the world. We cannot know what is unknown behind appearances. But we can choose, filter, and build.
Building products is the same. Their character and implicit culture shape users' first impressions. Weibo and Xiaohongshu feel different for a reason.
Some time ago, I learned a few bits of news about the AI Pin. It was a screenless wearable that used voice, gestures, and even laser projection to interact. It wanted to replace the smartphone for everyday tasks, but it failed not long after. Trying to solve screen addiction is worth respecting. I also wondered whether it could have barely survived if the experience had been good enough, while giving itself time to find another way out.
Back to what I said earlier about how hard offline meetings are. Along with my tiredness with screen-based interaction, could combining online and offline open a new direction?
In fact, many teams and organizations already do this. Outdoor groups begin activities online and gather online, then a leader brings everyone into physical reality and the activity space. Dating clubs or singles clubs often organize offline activities too.
For me, whether I would join an activity comes down to a few things: 1) am I interested? 2) do I have something to talk about? 3) is the environment friendly? 4) is the timing okay? 5) is there anything else?
A platform is naturally a place where interests and possible conversation gather. As for the environment, that has to do with the platform's culture and the people on the team. Users only need to see whether they have time. Of course, we do not always want activities with a completely clear purpose. In that case, giving interests and topics back to users as choices also seems friendly.
Then there are all the interesting details I will not expand on: whether offline spaces are run by the product team or partners; whether activities are official or started by users; how culture spreads — and what kind of culture it is 🤔; whether people who start activities need training or cultural immersion; how offline activity operations gradually give power from the official side to users; how to make sure the experience works; how feedback is collected afterward; and whether things such as the number of activities someone joined or whether they organized one should remain as markers on a profile.
After writing so much, I think there is one more point to open up briefly: the business model. For me, the most ideal model would be tiered subscriptions + offline-space partnerships or shared activity costs + automatic refunds when activity does not reach a certain level.
If I were a user, I would hope the platform returned my subscription fee when I had not been active for a certain number of consecutive days. If I made a product, I would want it to be that kind of product too. Not many companies try this, but even one would make me think, wow, this company is interesting. And if it decides to do automatic refunds, the technology probably would not be difficult.
Everything above is only part of my imagination of this product's design and operations. I have barely mentioned the difficulties of execution and making it real. Those are the core of a product, and the blind spot I have not filled in yet.
When I sort it out this way, it is not only about dating. It is also about social connection, and dating is a subset of that. And for someone like me, who does not feel much about romance and does not care about marriage at all, this idea is too idealistic for the future. Real relationships move forward through bumps.
A truly good relationship is not one that is one hundred percent compatible from the start. Complementarity can be one form too. Between being completely the same and completely opposite, there are countless forms. Everyone chooses a different path.
A platform exists to help people who need it do certain things better, at their own pace, so they can experience and enjoy more layers of life and a better life.
In middle school, I somehow wanted to be a boss. At that time, I knew nothing about what being a boss meant. More than ten years later, I finally understood myself: I am simply not suited to being a CEO, because the ‘need’ that the startup world emphasizes is almost absent in me.
So if someone else makes many of the products I want to make, that is completely fine. I only hope there can be a better product. And if I can take part in it in another role, it would be a great honor.
Finally, I want to mention the words from before: ‘The products I make always have a position, a bias, and values.’ The things I write are the same. And that does not mean my future self will always feel the same way.
P.S. There is an original ‘points, lines, and planes’ interaction record behind this piece. If you are interested, you can view this public link: https://chat.deepseek.com/share/wwm3kwdwed9zk8wmtb. It also shows how I use AI to help organize my thoughts and ideas. The link will expire next Friday. If a future reader is interested in what is in it, please contact me through the editor and ask me to share it. Thank you for reading.