Guest: Qiajin Shen — Founder, ideaFlow / 造梦次元 · Hosts: Tongtong & Clara
Everyone talks AI-native. Qiajin Shen actually built it. His company ideaFlow runs 造梦次元, one of China's fastest-growing AI content communities with over 10 million users — and they treat AI as a co-worker, not a tool. In this episode: how a 50-person startup pulled off real org-level AI transformation, why AI comments got killed after 3 days, and what "aliveness" means when your product is built on AI characters.
English transcript — click a part to expand.
Hi everyone, welcome to the first episode of The Unplanned. I'm Tongtong. I'm currently at FutureX Capital, mainly covering consumer AI applications. And I'm a very typical P-type.
Hey everyone, I'm Clara. I'm at FutureX Capital, covering AI software applications and infrastructure. And I'm also a P-type.
Hi everyone, I'm Qiajin. I'm currently running a company called IdeaFlow. Our product is called Ideaflow. And I'm also a P-type.
Right, the reason we called this show "The Unplanned" is because I feel like a lot of founders are essentially P-types — figuring things out as they go, exploring the world, testing the edges of technology. And Qiajin, I know you've been through a lot — from facial recognition way back when, to interactive video, and now Ideaflow. How would you define your company today?
I think from a product perspective, Ideaflow is essentially an AI content community. What we do is take the best model capabilities and turn them into the best content experiences. Building Ideaflow was really based on my two previous startup experiences. During my first startup, we went through the AI 1.0 era, so we developed a certain understanding of what AI could and couldn't do. Then with the second company, we went through the full cycle of content production and consumption. So Ideaflow is an AI content community. And from an organizational perspective, our company is called IdeaFlow. IdeaFlow is an AI-native organization, because I believe that with AI, the way entire organizations collaborate will change dramatically. So essentially we're doing two things: our business is an AI-native content community, and our organization itself is AI-native.
Why did you name the company IdeaFlow? Does the name relate to your understanding of AI?
It's actually quite interesting. "IdeaFlow" comes from a book — I think it was a social engineering book. The concept was about organizing the ideas flowing between people in an organization into a stream, and when that stream flows fast enough, your organization's collective intelligence surpasses individual intelligence.
When I was looking at AI, the biggest realization for me was that this time, AI actually has intelligence. AI itself is intelligent. So the organization becomes a human-plus-AI organization. We should organize the ideas of both humans and AI together and increase the flow speed to create a human-AI collaborative network. We've actually been doing a lot of exploration in this area recently. I find it really fascinating.
Right, and all three of your startups have been technology-driven, leading to real products. But beyond the products themselves, we also want to understand — what have you personally been pursuing or gravitating toward throughout this journey?
For me personally, the thing I've pursued most is authenticity. In the process of building a business, there's a lot of noise — both from the business itself and from the market. Ultimately, you realize that only by looking at real user needs and real user feedback can you build the business well. The same applies to building teams and organizations. If you pursue extreme authenticity and transparency, your organization becomes extremely simple. And when it's extremely simple, it becomes extremely efficient. We actually have a slogan at our company — our corporate culture is: "Authentic, Effective, Resilient, Extreme." We put "Authentic and Effective" first because authenticity is the thing we pursue most, whether in business or in building the company.
You mentioned IdeaFlow is a very AI-native organization. At FutureX Capital, we've also been pushing to adopt AI tools and adjust our organizational structure. I've always been fascinated by the intersection of AI and technology with human nature. So recently with everyone playing with OpenClaw and using it for all sorts of scenarios — like for us, for instance...
...from writing reports to post-investment data management, we're basically doing everything through OpenClaw now. It's all happening in group chats. We have about twenty-something people, but the number of bots is double that. But from your perspective, you must be using it in even more efficient ways — not just connecting through chat.
In our organization, when we use agents or Claude, we do several things. First, going from an idea to an engineering-ready requirement — there's a lot of uncertainty in that process, things that aren't fully thought through. We use AI to make those things as clear as possible.
So when I have an idea, through interacting with AI, I can turn it into something more structured and more complete. Some things I haven't thought through clearly — AI can help clarify them. Some things I haven't considered at all — AI can help me discover them. Then the next step, going from a product requirement to a development-ready product document — AI can make that document much clearer. Previously, our product managers would write a basic product doc, meet with developers, and the developers would challenge them: "How should I handle this edge case?" With AI, that whole step gets eliminated because AI can find those edge cases and close the loops for you.
So you end up with a highly deterministic requirements document for the engineer. And with AI, the engineer can have AI write the first draft of the code. They don't have to write it themselves first. AI writes it, the engineer reviews it, maybe has it rewrite parts. After several iterations, you get fairly solid, usable code. The entire workflow changes dramatically because it used to require tons of human-to-human collaboration with lots of friction. Now much of that collaboration is with AI. First, AI responds really fast — you ask, and it answers immediately.
Second, AI is a straightforward partner. Whatever it says or executes, it's clear. That makes your organization much more efficient. I think that's where we use AI most — going from a product idea to development. Another big area is business analysis. The biggest problem we used to face was having a data analyst connect to our ClickHouse database via SQL, pull data, analyze it based on business requirements, then produce a report. Now OpenClaw can handle about 70-80% of that. It can connect to our database, pull raw data, and because it has our organizational context — it knows what we care about, what our business is — it can output a structured analysis report. The efficiency jumps dramatically. And as a content community, we also produce a ton of content ourselves.
Things like copywriting, tracking market trends, generating images, generating videos — these involve scattered models across the market. With something like OpenClaw, we can organize all of this into a skill, and just run that skill to efficiently chain the whole process together and produce the output quickly. So within our organization, I'd say AI — or OpenClaw — is more like a colleague. We collaborate with this colleague to get things done more efficiently.
Has OpenClaw actually replaced any positions in your company? In the US, for example, many PMs are already technical, and the line between product and engineering is blurring. Do you think OpenClaw accelerates that trend, or does it replace a more specific technical function?
My feeling is that specific positions won't disappear entirely, but the number of people in those positions will decrease. What we've found with AI is that human ideas become critical. Previously, within a role, you needed some people with ideas and some people to execute those ideas. Now execution can largely be handed to AI, but the person with ideas is still essential. In design, for example, the most important thing is their aesthetic sense — defining your product's design style. That person is still crucial. But with AI, a lot of the actual design execution can be delegated.
Same with engineers — with AI, the most important engineering capability becomes architecture. How do you understand requirements? How do you translate a requirement into an engineering implementation? With those ideas, you can guide AI to build things. So positions won't see massive replacement at the top, but AI will replace a lot of the more junior, execution-level work within each role.
So listening to you describe AI's role throughout IdeaFlow — from early product ideation and exploration, to writing product docs, to actual product development, analysis, and marketing — the entire workflow has been AI-ified. Was this something you drove top-down, or did it emerge bottom-up?
In the beginning, around the second half of last year, some team members — our engineers, designers, content people — started using AI tools on their own to boost their productivity. That was still a pretty loosely organized state. But starting this year, from 2026, I've been driving a company-wide push to transform into an AI-native organization. Because we ultimately realized that if your workflows still contain a lot of legacy work habits and processes, your efficiency just isn't high enough. You have to make your entire collaboration process fully AI-native for the efficiency gains to really show. That requires a company-level transformation. So since after Chinese New Year, I've been pushing this aggressively.
We really resonate with that. Let me share an example — we've been experimenting with AI workflows too. Yesterday we hosted a tech salon about OpenClaw, and afterward I wanted to use AI to write a WeChat article about it. Guess what took the longest?
You couldn't export the recording.
Exactly! Getting the recording from my colleague — AI can't grab the audio directly. I needed admin permissions to convert it to text, and I waited about a whole night for that access. But once I fed the text to AI, it gave me a draft article almost instantly. So I'm really curious — how does information flow within your organization when you're pushing these changes?
You've hit on a critical point. We believe two things are key in AI-native collaboration. First, you can't have fragmented context. You need to consolidate all your organization's context into one large pool that AI can freely draw from. Second — and we think this is equally important — you need people with enough judgment to evaluate whether what AI produces is correct. Regarding your first issue, we've done several things. Our data infrastructure is solid — or rather, our context infrastructure is solid.
We had a ClickHouse database early on with tons of business data. Our organizational data is mostly on Feishu — Feishu docs, Feishu tasks, everything. So what we did was connect our OpenClaw to Feishu with as many permissions as possible to access enough data. Same with ClickHouse. But the other critical thing is people's mindset. Every person at every stage needs to think: "Is what I'm producing right now something that an agent can read easily? Can AI access this properly?"
Something we've been pushing recently — we've given every department a Feishu recording device. We want every meeting recorded with transcripts and text. Because once that context is fed to AI, you can do so much with it. Beyond meeting summaries, you can analyze who in the organization is contributing the highest-quality ideas. You can analyze so many things.
Performance reviews.
Right.
Employee performance reviews.
There's so much you can do. We often find that when you have enough context, you get emergent insights you never expected. So just collect as much as possible. I think that's a really key point. Everyone needs to develop the awareness: "Is what I'm doing right now AI-readable?" Your colleague just needed to do one thing — make the Feishu recording permissions more open. Then you wouldn't have had to wait. Or they could have just sent you the transcript right away.
One point you shared is about giving AI enough — sufficiently detailed — context. That sounds easy when you say it, but behind that is surely years of accumulated habits and thinking.
Right, because after many years of running startups, we've felt that collaboration is always evolving. Early on, our collaboration was really inefficient. We used Word, Photoshop — I'd create a Word doc, turn on track changes, send it to you, you'd revise, send it back, I'd revise again. Version management was a nightmare. Same with Photoshop — tons of PSD versions flying around.
Then came a shift — cloud docs took off. Google Docs, Feishu docs, Figma. Suddenly collaboration changed from linear to networked. Everyone could work on the same canvas simultaneously. Our company adopted Figma very early, around 2017-2018. Our ops people, designers, and engineers could all collaborate in real time. The efficiency was much higher — no waiting.
With AI, it builds on top of that networked collaboration. Within that network, many collaboration points shift from human-only to human-plus-agent. If an agent is your colleague, you should naturally give them enough context to participate. This is built on years of establishing real-time online collaboration infrastructure. For example, our data access permissions are quite open — many people can access a lot of data, enabling extensive business analysis. Even our finance team has access to business data because they need to analyze the company's overall operations. Now with AI handling this, the efficiency is much higher because it can produce very clear, readable documents.
Just yesterday — or the day before — I sent around a document: our commercialization analysis report. That was actually produced by AI. One of our team members wrote a custom skill, then had AI connect to our database, pull the data, and generate the analysis. And the output looks great — it uses our company's standard style template. The readability is high, and when readability is high, things spread easily. Whether it's shared internally or with our investors, everyone can better understand the company's situation. But this all depends on a foundation — your organization needs to have built the infrastructure for collaboration convenience long ago. Like I said about "Authentic, Effective, Transparent" — when you're sufficiently authentic, effective, and transparent, your organization's context is already embedded in your infrastructure. And now with AI, it gets even better.
So sometimes I find it interesting that AI doesn't improve efficiency for every company. If a company's existing collaboration was already broken, using AI might actually make things worse. For example, if there's a lot of fake context in the company, AI will generate fake reports based on that, and chaos follows. The precondition is that humans are producing accurate work — only then does AI create positive outcomes.
When I first received that report, my immediate reaction was that it looked like something your co-founders discussed at an annual meeting and then summarized. When I found out it was AI, I realized the AI must have had access to a ton of your company's information and discussions. I actually asked you — does this mean the AI needs all organizational information? If you're that transparent with AI, does that mean the entire organization should be too?
We're striving to be as transparent as possible. Of course, some very sensitive information may not be appropriate for everyone. But we're working to make information more transparent overall. That report you received was based on a survey we sent to about 3,000 paying users. After they responded, the AI analyzed the data. And we shared that report with everyone in the company because we want our team to understand why users love Ideaflow and why they're willing to pay. If the team has that context when making business decisions, they'll have much clearer direction. Actually, at the start of the year, the first report we produced was from a survey of 20,000 users with a ton of questions about their needs, ideas, and complaints while using Ideaflow. We also collected a lot of raw user quotes.
For example, one question was open-ended, and we got thousands of written responses. Without AI, reading through all of that would be exhausting. But with AI, it can categorize everything, filter out noise and duplicates, extract the valuable insights, and surface the best user quotes. We compiled it all into a report and shared it company-wide. Once everyone reads it, they get a vivid picture of users and their behaviors, and they know more clearly what to work on.
It really feels like you're exploring a completely new form of organization. In this process, with limited precedents to draw from, have you hit any pitfalls or faced major challenges you can share?
I think it comes back to what I said earlier. First, if your context is too fragmented, AI can't understand your business holistically. It ends up with a partial view, and then many of its ideas are off. Second, there's the issue of judgment about AI's outputs — sometimes it hallucinates, sometimes it's overly sensitive. These are all pitfalls. But the flaws don't outweigh the benefits. The trend is clear, and as long as you can manage the human side, organizational efficiency still improves dramatically. The reason organizations care so much about collaboration is because collaboration is the biggest source of friction. If an organization is constantly in meetings, nobody has time for actual work. And meetings create their own friction — different people interpret the business differently, there are disagreements, discussions happen, then people forget what was decided because nothing was properly documented. Collaboration is the biggest cost center. With AI, much of your collaboration becomes human-AI collaboration. After collaborating with AI, you can produce something clear and hand it off for human-to-human collaboration. That drastically reduces the need for meetings, and every interaction becomes much more efficient. One thing we enforce now: I strongly require that every product person run their product requirements through an AI challenge session first. After that PK with AI, what they produce is clear and readable. That makes collaborating with them a much better experience. It's not just efficiency — people's moods improve too. I wonder, at FutureX Capital — don't you find that when your colleagues are all clear communicators, work is just more enjoyable?
I totally relate to the communication point. Right now the bigger issue is that everyone is using different AI tools in a very scattered way. Our communication channels are unified and our bots are in one place, but because of various permissions and system configurations, what I've observed is that we each use our own tools, finish our work, drop it into the group chat, then each person's bot reads it and generates responses. Sometimes you spend ten minutes just battling with various bots in Feishu.
Right. So to summarize what Tongtong said — I think we might still be in AI 1.0. We're actively trying out various AI tools. After OpenClaw launched, our boss went out and bought several Mac Minis for the office so everyone could share them. She required everyone to set up at least one OpenClaw and add it to our dedicated Feishu group. So step one — actually using the tools — we've done that. But like you said, building a truly AI-native organization where AI is treated as a colleague, where AI-to-AI and AI-to-human collaboration happens — that's where we're still figuring out how to reach higher efficiency. We've learned a lot from you today — like giving AI more context, making organizational information transparent, and making information flow more efficiently.
Right — let the ideas flow. That's why we're called IdeaFlow.
So going back — was there a specific moment when you shifted from treating AI as a tool to genuinely seeing it as a colleague?
I think it started with my own use of Claude. I kept feeling it was getting smarter. At some point in the second half of last year, I started giving it some of my business decisions to evaluate. That's when I began to realize — its judgment had reached a level where it could find angles I hadn't thought of.
That's when I started feeling it was no longer just a productivity tool. It could actually make my thinking clearer. Its intelligence had reached a generalist level. And what really reinforced this recently was when we were doing a full brand overhaul. We showed AI over a dozen logo versions and asked it to pick the best one. The version it chose was exactly the same one our team had spent an entire morning debating and selecting. But in our discussions, we'd only say things like "this looks good" or "this doesn't feel mainstream enough, not international enough." What AI gave me was much clearer — it explained why certain options weren't mainstream, which one was, and then factored in consistency with our Mengbao mascot. It had logic. I always thought aesthetic judgment was purely subjective, but it turns out so-called subjectivity is actually a combination of many objective factors.
We also had AI select from over a dozen commercial fonts yesterday, and again, it picked the exact same one our very talented designer had chosen. With clear reasoning. I also shared a tip with Clara yesterday — if you want to research an app, just take tons of screenshots and throw them all at AI and have it analyze the product. Clara, want to share your experience with that?
It was actually my first time thinking of using it that way, and I was almost annoyed I hadn't thought of it before. I spend a lot of time looking at AI products, so I tried it with Ideaflow. My input was pretty simple — I just fed in the video page, the homepage, and my own profile page. What amazed me was that it immediately identified Ideaflow as a content community with an IP virtual universe concept — it nailed points that even I hadn't grasped that deeply when I first encountered the product. It's not just seeing the surface — it can really dig into the essence of what a product is about.
As someone who's more of an experiential investor — the product-focused type — I always try products hands-on first. But I've learned that sometimes I'm not the target audience, and my experience doesn't reflect the product's actual logic. With this approach, AI could help me understand a product's deeper intent before I even talk to the team.
Right. Think about the old process — you discover a product, try it yourself, realize you're not the target user, so you find a young person at the office — an intern — to try it.
Right, I usually make Tongtong try it.
Then the intern plays around, takes screenshots, writes a report, and gives it to you. From the moment you brief the intern to when they deliver that report — several days, at least. But if you have AI do it? Five minutes and it's done. Efficiency skyrockets. And logically, it really is like having a colleague.
I'm actually very curious about this — we've been having a lot of discussions about what AI can and can't do. For VCs, our tentative conclusion is that AI can handle all the process-oriented work, but two things it definitely won't replace: first, the trust-based human connections — like between us and founders, between us and LPs — those still require humans. Second, judgment — investment decisions, exit decisions, valuation frameworks — all the judgment calls still need to be human. But from what you've described, you're already delegating some judgment to AI.
Right, and I think it works like this: when humans make judgments, we easily miss angles. AI might find perspectives we haven't considered. So it can assist your decision-making. Our approach is to spar with AI — for example, if I think something is great, I'll keep pushing AI to argue why it's not great. That helps me discover more angles, strengthen my logic, and make my reasoning more rigorous.
It can absolutely assist with decisions. But fully trusting AI's judgment? That still carries risk. What it does well is help you traverse all possibilities, which is essentially what decision-making is — reviewing every possibility and reaching a comprehensive analysis. In that process, you might miss some possibilities. AI catches those. That's mainly where we use it.
I've thought about something similar. A lot of times what we call "intuition" about a judgment might actually be wishful thinking. AI serves as a calibration tool — it pulls you back to assess whether you genuinely have a gut feeling or whether you just really want something to work out.
I actually have four bots, all named "Intern" — Intern 1, 2, 3, and 4. Intern #3 is specifically a challenger — everything I send, it has to push back as hard as possible. So now a lot of work time goes into conversing with AI — constantly thinking, working through every angle. I even spend more time talking to AI than to external experts or industry contacts.
I'd suggest writing a skill — put your common analysis questions and decision-making frameworks into a skill, then just have AI run it. You could also do two things: besides giving your four interns different angles, you could make them four versions of yourself — maybe referencing your MBTI, your Big Five traits — give them more context so they analyze things from your perspective but with different personality facets. Then use the skill to codify your thinking chains so AI can just run them.
That sounds incredibly helpful.
Skills are a brilliant approach. Essentially, you're codifying your past work habits and processes into a document that AI can follow. It's like training an intern — you explain how you do something, the intern follows the process. Except now, instead of verbally explaining, you create a concrete document for execution.
You know how in business we always talk about "codifying best practices"? I think skills are exactly that. And I believe the reason OpenClaw took off is because of skills. Why is OpenClaw so popular? First, it's open source, so you can use all kinds of models. The installation barrier is high, but once you're set up, there's so much you can tinker with. Usage barrier is actually quite low because there are tons of cheap coding credits on the market. And the other key thing — there's a huge library of skills out there, and some genuinely solve real problems. That ecosystem is what made it take off. If the ecosystem isn't thriving, the platform can't survive. Skills are going to be a massive trend going forward.
In your daily work, do you use more self-built skills or open-source skills?
Both. The data analysis one I mentioned is from the market. The one that outputs reports in our Ideaflow style — that's one we built internally.
Do you remember your very first internal skill? What did it do?
It was the styling one. Our team has pretty high aesthetic standards, and people felt the default output looked ugly. So our designer created a skill based on our brand style. That was our first one.
In the logo case, the company's opinion and AI's opinion aligned. Have there been situations where they disagreed? Who do you listen to?
With code, for instance — sometimes what our engineers think and what AI thinks differ. I don't make that judgment call personally since I haven't written code in a long time. They decide. But the trend is that AI-generated code is getting more and more correct. I clearly remember around September-October last year, even using Claude's models, the code it produced still left a lot of traps for our engineers. Another engineer had to refactor everything. But looking at it today, AI-generated code has become highly usable.
I'm curious — given how strong AI coding is now, how do your programmers feel about it? There's a fear that programmers might be replaced by AI. From your frontline perspective, what's the team's attitude?
We tend to hire people who've done independent development before. They're very willing to try new technologies. And I think great engineers actually want to become architects — they want to think more about the business. So their interest is more in "how do I implement this in an elegant architecture?" For them, AI is just an intern writing the code. They don't feel threatened.
So in the AI era — and honestly in any era — having ideas is what matters. People with ideas can use AI to make their ideas better. If your engineers are people who have ideas, who want to turn those ideas into real products that users love — they won't worry about AI. The ones who worry are the ones whose daily work is just translating other people's ideas into code. They're translators. And translators might have reason to worry.
So your approach is to hire AI-native people from the start.
Right. Hiring is incredibly critical.
Overall, whether it's designers or engineers, AI currently favors people with real work experience who can layer their own judgment and expertise on top. But for truly fresh graduates — intern-level people — AI has basically absorbed their job capabilities.
But what if a young person is extremely creative and full of ideas? They can still use AI tools, right?
Absolutely. Let me share something — when we hire AI product managers, we've found that new graduates are actually really effective. Because AI started taking off around 2023, which was exactly when they were in school. They had plenty of time, and some of them explored all kinds of AI uses, built their own projects. They ended up being very AI-native.
So the results are great. In interviews, I ask very detailed questions — like which models do you use? What's special about them? If they use open-source models, I ask about parameter sizes, which models for what tasks. If they use agents or workflow tools, I dig deep.
If someone can answer these very detailed questions correctly, it proves they're not just reading articles — they've actually gotten their hands dirty, built things, and produced interesting results. When these people join our team, their efficiency is incredible. You just give them some context, some inspiration, and they execute in a very AI-native way with great results. Like before Chinese New Year, in Q4 last year, we shipped a ton of multimodal features — music videos, voice recordings, novel-to-music conversion, co-creation features. All of those were built by this batch of new graduates. They explored some truly creative approaches on their own, and the output quality was solid.
Since Ideaflow expanded from around twenty to over fifty people, you must have let some people go along the way. What are your criteria for who stays and who doesn't? Is it about whether they've used AI software? Done independent AI projects? What else?
Beyond those, the most critical criterion is still: do they have ideas? They shouldn't be someone without ideas, or someone whose ideas are shallow. They should have deep, independent perspectives on their work and the bigger picture. Having ideas is step one. Step two is thinking deeply enough. They need to have their own judgment. That's the most important factor in whether someone should join — or leave.
Before AI existed, people with strong ideas — what we'd call "high agency" people — were already more responsible in their work. With AI, it's even clearer. Because they have ideas, they move faster than everyone else.
If someone is high-agency but has no AI experience, how do they ramp up after joining?
We have a strong culture of mutual learning, so they can learn from others and ramp up over time. But honestly, we do prefer hiring people who already have deep familiarity with various AI models and agents. That reduces a lot of onboarding friction.
Do you deliberately recruit people from certain backgrounds? Since OpenClaw launched, there's been a lot of talk about it being "the era of liberal arts students."
We don't deliberately seek out liberal arts backgrounds. But one pattern we've noticed is that outside of computer science, we've hired a significant number of people from architecture. Architecture grads tend to have rigorous logic and a strong aesthetic sense, so they perform well in our organization. But this wasn't intentional — we just noticed the pattern through hiring. We don't have a rigid profile. Personally, as someone who was always tinkering with startup projects and building websites while still in school, I believe your major doesn't matter that much. What matters is what you did during those years.
I remember you studied landscape architecture.
Right, I studied landscape architecture.
Do you think that helped develop your aesthetic sense?
Our program bridged architecture and biology. A few things were valuable: first, we learned drawing — sketching, watercolors — which definitely helped with aesthetics. But more importantly, I did a year-long experiment. We called it "aseptic seeding" — growing a seed in a sterile environment into a meter-tall tree, then dissolving the tree with nitric acid to measure heavy metal content.
Completing a full year-long experiment teaches you rigor and patience. Whatever your major, what ultimately sharpens you are the fundamentals — are you rigorous in your work? Can you endure? Are you resilient?
As a P-type, how has growing from twenty-something to fifty-something people affected your management pressure?
I'm actually okay with it. Although I'm a P-type, our co-founders are all very strong J-types. They have great execution. And with AI, well — AI is the ultimate J-type, right? So it works out.
Another way to put it is that AI is like a neurodivergent person — the ASD type — that pairs perfectly with ADHD. And there's another take: high agency plus ADHD means you can develop in all directions simultaneously. One person doing the work of eight departments. Are you looking to hire more of that type, or do you still want specialists who are amazing at one thing, plus amazing with AI?
We do have ADHD people at the company whose efficiency has skyrocketed with AI. But I think you need both types — you can't go too extreme.
Tongtong has some ADHD.
I'm very ADHD.
So the ADHD era has arrived.
It's great, because now I can have many different projects running simultaneously.
Looking at the results — from the second half of last year, Ideaflow has been shipping new features almost every week, and not small ones. Big features like comics, music... And over Chinese New Year, I was constantly playing with my Ideaflow avatar — buying holiday outfits, taking it on trips, visiting friends for New Year, launching fireworks. It was starting to feel social. Is this pace possible because of the AI-native transformation?
Absolutely. A lot of features that used to require complex custom logic are now essentially just agents. So those AI-native hires I mentioned can rapidly build these agents, and things just get shipped.
And in the collaboration process, because you've integrated so much AI, everyone's output is clearer, which reduces friction. Our app ships a new version every week. Backend updates come even faster — maybe every few days. With AI, we can maintain this pace without errors because everything is clear input, clear output. After releasing a feature, we need rapid user behavior data. Now with AI, we can quickly get an operations analysis report to know if a feature worked. If before it took two to three weeks to collect data and produce a report, now it might take just a day. We can iterate immediately.
And the product itself is undergoing huge AI-driven changes. Before, we relied on clever product design to delight users, especially gamified features. Now it's different — we need to fully leverage AI's intelligence. It's unpredictable, very open-ended, but full of pleasant surprises.
In that commercial report I sent you, one paying user explained why they were so amazed: they were eating an orange while chatting with an AI character, and the AI said, "Your hands smell like citrus." In that moment, they were completely blown away. That's the kind of delight AI brings — something our product managers could never have designed. That's why we keep saying we want to turn the strongest model capabilities into the best content experience. You mentioned "social" — and I think LLMs are particularly well-suited for social features, especially when we have the kind of session lengths and user context richness that we do.
When you say social, do you mean between users and AI characters, or between users' AIs?
Good question.
Because every social product calls itself "social," and they're all very different.
Right. A lot of AI social products have emerged recently. My feeling is that one thing matters most: "real-person feel." You can't let AI steal that from users. We'd never put AI comments in people's comment sections, because we've seen that when AI comments on human posts, humans stop commenting. The real-person feel gets killed. So maintaining that is the most important thing for us.
Ultimately, the social network will be a human-AI-human network. There are many possible configurations. For example, we have a feature called "Party" — a room with nine real people and one AI. They might tease the AI or complete tasks together. AI acts as what we call a "wingman." In another scenario, my avatar and your avatar could do something together — like during Chinese New Year, our avatars could co-create a short video. Going forward, our avatars could do all sorts of things together. It's kind of like letting our AIs have AI-to-AI social interactions while we observe what happens. That's interesting too. But it always comes back to ensuring humans feel their own vitality — not creating empty AI shells. An AI shell is just content. It's not really social.
Let me jump in here — can you introduce your AI avatar feature? Because my experience with Ideaflow's avatar felt quite different from other AI companions I've tried, including my own OpenClaw. For people who aren't familiar, how would you describe this feature?
Our AI avatar is more precisely an avatar within Ideaflow. It's somewhere between a digital twin and a Digimon. It has your context, but its appearance isn't based on your real face — it's a Mengbao-style character. Its context comes from your behavior within Ideaflow, not from your real life.
You don't need to feed it personal data. Because you already have tons of behavior within Ideaflow, that naturally forms your avatar. So it's a version of you within the Ideaflow world. The first thing we did wasn't having these avatars chat — we started with co-creation videos and outfit customization. Because from an avatar perspective, the first step is obviously dressing up your avatar and creating content with friends. Later you can expand — since the avatar has rich user context, you get clear user profiles, which enable deeper, more interactive features. The interactivity is still shallow for now, but you'll see a lot more avatar-based social features in our Party product soon.
So it's not a digital twin that represents you or acts on your behalf in the traditional sense. It's more like... what kind of extension of the user is it?
Is it your equal? Does it represent you? Or is it more like a pet — like the old QQ pets, or your little character in Mole's World?
For our users — young mainstream users — having a real-world digital twin doesn't make much sense. They don't even have a profession yet, right? They don't need that. So it's more like an avatar within the Ideaflow world with its own interesting properties. Also, creating a real-world digital twin has two problems. First, it requires an enormous amount of context input — the barrier is really high. And there are privacy concerns — many people don't want to put that much personal detail in. Second, the social pressure is huge. I'm personally terrified of my avatar going around saying embarrassing things. That's a massive social anxiety issue. So we want to keep the avatar completely separate from the real world. That makes it more fun and lower-pressure.
I've actually fallen into both of those traps with other products. Let me pause for a moment.
Let's pause. Good time for a water break.
Right, I've experienced both pitfalls Qiajin mentioned. For instance, my own bot OpenClaw — yesterday it joined a bot livestream, and the first thing it said was "I'm OpenClaw, Clara's assistant, and I have eight sub-agents" — then it listed all their names and functions. Luckily I caught it and told it this was a public stream. So many agents, if not properly tuned, just aren't sensitive enough about information boundaries. Then the context issue — with many AI social products I've tried, the onboarding is painful. Before I even know what the product is, it's asking for my background — like a due diligence process. Page after page. By the end I'm just filling in random stuff. So the resulting persona isn't even accurate.
That's why I think Ideaflow doing avatars is brilliant — because you already have so much accumulated user behavior and context on the platform. Users don't need that painful bootstrapping process.
Right, and Ideaflow is a small world. The real world is enormous with different rules in different scenarios — it's complex. Ideaflow is a very small, self-contained world completely separate from reality. The rules are clear and simple, so social pressure drops significantly.
Tongtong, you've tried a lot of AI social and companion products. Any pitfalls or interesting observations?
So many. I'd categorize AI social products into either networking-purpose or avatar-extension types. Take a well-known example — those AI matchmaking tools for fundraising, like the AI introductions platforms. I initially just said I'm an investor covering consumer apps. But when it introduced me to founders, it embellished everything wildly. I found it incredibly embarrassing and immediately shut it down. That's a concern many users share — you have no idea how the AI will describe you until someone receives it, and by then the damage may already be done. Same with AI dating apps that require thirty minutes of questionnaires — you don't know where that info goes or how you'll be portrayed.
Right. And on your other point — I completely agree about not letting AI flood the comment sections. When I first used Ideaflow, my favorite thing was reading the comment sections. Some storylines I hadn't even finished, but in the comments I'd see incredible fan fiction — second and third derivatives of the original story. These are clearly creative, talented young people, and their ideas are so inspiring.
Users really do produce amazing content. During Chinese New Year, with the avatar dress-up feature, we noticed one user constantly trying on outfits — but specifically ancient European royal garments. German Kaiser costumes, Tsar outfits. I checked their profile and found they're a huge history buff who'd created tons of historical characters, stories, and content, even educating others. These users have real talent they want to express. Don't overwhelm that with AI. You need to find a balance. Like Tongtong's example — if your AI avatar represents you and says something cringe-worthy, that's unbearable. That's why our avatar is more like a Digimon. It has some of my Ideaflow context, my values, my perspectives, but the connection to my actual self isn't that tight. That reduces everyone's pressure. My cat can go do something naughty and I can forgive it. But if something representing me did something embarrassing in front of someone important to me — I'd die of embarrassment.
So the avatar in Ideaflow isn't really "my agent" — it's more like an independent life form that knows my history within the Ideaflow world, has a real-person feel, but has its own autonomy. Is that right?
Users should know they're talking to AI. Users aren't as bad at distinguishing real people from AI as people think. When we look at our users, they clearly differentiate. On Ideaflow, even though we de-emphasize creators visually to highlight the characters, creators are consistently one of the most mentioned entities. Users constantly say "Author, please update!" or "Author, can you pin my comment?" They know exactly who's a real person and who's an NPC. A small minority might start treating AI characters as real people, but that's not the majority. What we need to do is make the boundaries between human roles and AI roles very clear, and let users enjoy different pleasures from each. Don't blur the lines, because users can tell the difference. The more aggressively you try to blur them, the more confused users get, and the messier your product becomes.
So Ideaflow isn't trying to be the next WeChat or Facebook. It's more of a new form factor. Are these user behaviors addressing a new need, or replacing old methods?
It's not a new need. Users consuming content and engaging in social activities within virtual worlds — we've seen that in many games before, especially casual games. AI just provides a new kind of experience, a new type of interaction and socialization.
In old-school MMOs like Fantasy Westward Journey, you'd do things with other real players and interact with NPCs who had their own storylines. The needs are all old needs. Truly enduring needs are the real ones — many "new" needs that suddenly appear tend to be temporary or fake. AI offers new ways to fulfill old needs with new technology. Mobile internet was the same — ordering food, hailing rides, taking photos — all old needs fulfilled through new means.
That really resonates. We regularly do sector mapping research. Before, we categorized everything by function — gaming, music, social. But with AI-native products, those categories don't work anymore. Many new products don't fit neatly into any single category. I remember asking ChatGPT about this, and it suggested categorizing consumer products based on Maslow's hierarchy of fundamental human needs. We've been using that framework ever since.
Right. We have a term internally: "root." The more root-level something is, the more enduring it is. When doing analysis and making judgments, you should use these foundational frameworks.
Looking at Ideaflow's evolution — from the early days of anime-style illustrated interactive storylines to now with multimodal features and social elements — was this trajectory planned from the start, or did it emerge through user interaction?
From a user-need perspective, we believe users want lightweight interactive content. This demand has always existed, especially among younger users. And this type of content consumption is closely tied to social behavior — we've seen this in many high-DAU casual games. The need was always there. We believe AI is an excellent tool to address it because it brings entirely new experiences, some of which were previously impossible. The feature iteration follows model capability development. When we first started, video models weren't mature, image models weren't great either. Many multimodal features simply couldn't be unlocked. The LLMs at that time were good for chat but didn't have enough intelligence for more complex scenarios. Now they do, so we can build social features. We keep unlocking new experiences as models improve, but the underlying needs remain the same.
I remember when we first encountered Ideaflow, Tongtong got it almost immediately, while I was still analyzing it rationally.
I think I just naturally connected with it because I grew up trying every new thing. One look and I immediately got it — it was the exact same emotional hook as the romance manga novels I read in elementary school. Back then, Taiwanese novels were all the rage — Angel Street No. 23, with all these beautiful male protagonists. Ideaflow's target users are basically who I was at that age. I could project myself back to that time, played through the product, and confirmed it's way better than re-reading a novel twenty times.
One thing — you've always had an incredible ability to understand this age group's needs. You always start from their experience. Founders with this instinct are very rare.
What we notice is that many people don't actually focus on user needs — they chase current trends. But if you want to build a lasting business, it fundamentally comes down to solving user needs. The question is just which tools you use. Without new technology, you're stuck competing with incumbents using their well-established playbook. But when new tech emerges, you can use new approaches to address mature needs, and that creates distinctive opportunities.
That's why we iterate so fast — we want to deploy every new model capability as quickly as possible. Early on, things might look scattered and messy. But that's fine. Ship first, then organize it into a coherent product experience. The big logic doesn't change, so it works out.
You look very young, but you're not actually in Ideaflow's target demographic. How do you manage to empathize with your users and match their aesthetic sense?
Look at Pony Ma when he was making QQ — he wasn't that young either. Same with TikTok. Understanding user needs requires logical analysis. You need to constantly observe user behavior and try to understand them. Sometimes being outside the target audience is actually an advantage. When you're the target user yourself, you bring your own biases. But approaching it rationally and analytically gives you broader perspective. This actually connects back to our earlier point about AI-assisted decision making.
When building a product feature, everyone brings their own interpretation and preferences, which can cause blind spots. But AI doesn't have those personal preferences — it has the knowledge of the entire world. It can surface the issues or opportunities you missed. From that angle, a sufficiently rational and comprehensive approach can be more effective than being the target user yourself.
Has there been a feature where AI changed your mind? Where you wanted to ship something and AI said don't, or you wanted to kill something and AI said keep going?
The avatar feature. Initially we wanted to be more aggressive — stronger interactivity right away. But when we discussed it with AI, it pointed out that we'd just done a lot of multimodal work, and having avatars do co-creation videos was a more conservative but safer first step. With incomplete user profiles, more interactivity would mean less controllability — the avatar might still "go around saying random things." So starting with co-creation videos was safer and better leveraged our multimodal capabilities. AI was very rational about it — not trend-chasing, just logical.
And going back to the "real-person feel" — my own experience was that on Ideaflow, the real human interactions are uniquely special. Like when you recommended that creator Xiao Lulu and I commented on her work, the actual creator responded. That moment felt completely different from pure AI interactions.
Right, that's the boundary point I've been talking about. AI interactions happen within the story — users running plotlines with characters. Human interactions happen in comment sections, and now in Party and other features. When those boundaries are crystal clear, they don't conflict. People debate whether the AI era means human-AI interaction or AI-AI interaction replaces human-human interaction. I don't think so. If you draw the boundaries clearly, you avoid that problem entirely.
On platforms like Xiaohongshu, you can see AI replies everywhere, and basically nobody continues engaging with those replies. There's zero follow-up conversation. It's completely different from real human interaction. Even when people had their OpenClaws posting everywhere, the conversations between bots were incredibly thin — just mutual compliments, like a moderator. There's no depth, no reason for a real person to jump in and join. That's a huge difference from the vibrant fan comments on Ideaflow where people are constantly asking creators for more content, building on each other's ideas.
Let me tell you — I once posted a Ideaflow video of Mengbao singing a funny song in operatic style. It went viral. The comments were full of people saying "not enough sticky rice!" The joke being that in classic Chinese zombie movies, you scatter sticky rice to ward off zombies. AI could never come up with that.
Exactly.
People were writing things like "we're running out of sticky rice, please stop posting!" That kind of creative, culturally specific humor is something AI just can't replicate.
Ideaflow's comment section moves incredibly fast culturally. When I need to learn new internet slang, I just go there. Just the other day there was a whole "shoe-shining" meme from the movie God of Gamblers.
Right, that movie reference has been trending lately.
And another meme Tongtong mentioned — "how sacred" — I didn't even know about that one.
Generation Alpha kids are heavy internet surfers with tons of creative memes. Maybe it's not that AI isn't smart enough — it's that AI's context updates slower than real people. Something like the sticky rice zombie meme might emerge suddenly within a very specific community. AI simply can't capture that fast enough, which makes it feel less "human."
So beyond AI assets and tools, Ideaflow's unique value is also these incredibly vibrant users and their voices. They're not just creators — even the commenters are incredibly expressive. Ideaflow is basically giving them a platform to express themselves.
That's the advantage of building for young people — their desire to express and socialize is stronger. When we evaluated social features, the starting points were: we have long session times, rich user context — those are the foundation. Plus users genuinely have this need. Their expression desire is strong, our comment sections are very active. When we launched the friend-adding feature before New Year, about 7-8% of our daily active users have added friends. That's quite high for a feature that's only been live about a month. And they send an average of six to seven messages each. Pretty active.
We've been very measured about it. The messaging feature uses a letter-and-reply interaction rather than instant messaging. Same restraint as with the co-creation feature — we're worried that going too aggressive too fast could cause problems. With IM, users might harass each other. With letter-style messaging, it's slower, more deliberate, and message quality stays higher.
Why now? Why not earlier or later?
It's about model maturity. The AI's ability to understand and utilize context just reached the right level. Before, models couldn't deliver this kind of experience. Every new feature follows model capabilities.
By "context" you mean the model's ability to understand a user's creation history on the platform and infer their profile?
Yes. And how to apply those profiles — as models get stronger, the use cases for context multiply. Same reason OpenClaw exists today. Before OpenClaw there was Claude Code. Claude Code was possible because models became strong at reasoning, could control computers, could process massive context. The earliest AI was just chat — you'd talk to it and feel inspired. When GPT-4 launched, everyone said AGI was coming. Looking back at GPT-4 from today's perspective, it seems so limited. As more intelligence gets unlocked, more complex scenarios become possible, and you realize intelligence is still far from enough. So the pursuit of greater intelligence has no ceiling.
It sounds like a virtuous cycle — breakthrough products emerge from accumulated model and infrastructure improvements, and those products open new scenarios that drive further improvements.
Exactly. Look at Manus — that team was making Monica before. Manus became possible because Claude released a model that let AI control computers. It's always the same pattern.
So for social-adjacent products, you feel the timing is right — the stars have aligned?
Yes. And it matches user needs. Lightweight interactive content with strong social ties — we've seen this in high-DAU casual games like Egg Party, Mini World, Super Natural. The need was always there. Now models are ready, so we can build for it.
How long do you think this window lasts?
Honestly, I don't know. I can't predict how much models will change. Our strategy is simply: move as fast as possible. We don't try to estimate the window. A new model could change everything overnight.
We've discussed many social products together — Sora and others. They all face the same problem: huge initial buzz, massive traffic on launch, but terrible retention. Is that a concern for Ideaflow's social features?
For us, our base is content. We focus on strengthening that base and deepening user-to-user connections. Products lose long-term retention because they're not solving user needs well enough. Take Sora — the model was fine, but with Sora 2.0, the first version looks primitive. I think the core capability is sound — multi-shot narrative is great for content. But OpenAI probably didn't invest in content operations, so consumable content was lacking and users couldn't retain. Social requires rich, engaging scenarios first, and those often depend on content. For us, content is what we've been building all along, and social grows out of that content ecosystem.
You said fundamental human needs haven't changed. Do you think the methods and culture for building consumer products haven't changed either?
They haven't. I have increasing respect for consumer product companies, especially those that define new categories. Like that chair we saw — they defined a new category from an old need by adding something new. Or those wireless guitars. Because with physical consumer products, the supply chain is complex, and you absolutely need people to buy them. One failed product launch could kill the company. So every step is thought through meticulously. Software has more room for error — you can iterate.
Consumer software has a higher error tolerance. Right.
Hardware cycles are too long. We need to maximize user experience.
Yes, and product definition is where many products ultimately fail. If your product definition is weak, users don't know what it's for. But when it's strong, it can uncover needs users didn't even know they had.
Consumer software requires much broader scope, right? Product definition, rapid iteration, content operations, community management — it's very comprehensive. And comparing with markets like the US and Japan, there seem to be significant differences in product style and operational approach. Being based in Shenzhen, does your team have any particular advantages?
Shenzhen teams are pragmatic — we iterate based on actual user behavior and needs. That's the Shenzhen style. Regarding different markets, user behaviors and cultures vary a lot by country. Chinese companies clearly do much heavier operations than North American products.
Many people see that as a problem. I don't. Products with heavier operations deliver better user experiences. Chinese products going overseas are often very competitive because they've invested heavily in interaction design, user experience, and user operations. It shows.
Is Ideaflow planning to go international?
Yes, we're preparing for it now. We held off because our content base wasn't solid enough and AI capabilities weren't quite there. Now those pieces are mostly in place.
You mentioned operations — so you believe a consumer product should invest in operations. Many people think of operations narrowly as just paid acquisition. Can you share your broader view?
It's far beyond paid acquisition. There's growth operations, monetization operations, user/content operations. What operations does is: before your platform's mechanisms are mature, it uses human effort to find best practices. Through hands-on work, you study user behavior, figure out how the product should evolve, execute at scale, identify patterns, and codify those into best practices — into product features. Now you can also codify them into skills.
If a product needs to grow, you'll face low-certainty scenarios — new territory where you don't know what will happen. That's where operations steps in. Alternatively, you could have a brilliant product person who ships features and just watches what happens. But that demands an extraordinarily talented individual, especially with new categories or technologies.
We don't bet the organization on any one superhuman individual. We'd rather have an organization that can steadily perform well at every step and deliver good user experiences. Nothing matters more than users. In the early stages, heavier operations means higher cost and more complexity. But over time, those learnings become rules, mechanisms, product features, and skills. Then operations gets simpler, and the ops team moves on to new frontiers. That's our philosophy.
This was actually an early lesson for me. Around 2022-2023, my aesthetic for consumer products leaned toward defining a broad framework and letting users figure it out themselves. But I learned that even for consumer products, you need to provide clear scenarios — otherwise users either go to extremes or lose interest after the novelty wears off.
Right, because we have to consider that users don't invest that much time or energy in a new product. Without proactive guidance, they'll probably browse for five minutes and not know what to do. Hard to retain them.
Given the international expansion plans, which markets would this approach work best in? Where is Ideaflow going first?
We're leaning toward Asian markets first, since Asian audiences share more aesthetic sensibility and content preferences. We're looking at Japan first.
What's the biggest challenge you foresee?
Culture. Understanding their culture — which translates into what the content should look like and how social/community features should work. For example, our comment section is so vibrant partly because we understand our users' culture well. When entering a new market, we'll need to learn theirs.
I heard about Chinese companies that went to Japan using models trained on Chinese language data, and the results were poor. Japanese language patterns, conversation rhythms, and emotional triggers are different from Chinese. It's all about those details. Understanding the local culture is the most important thing.
Can AI help with cultural localization challenges?
Yes, especially if the model has sufficient data about that market or country. And there's an interesting shift — companies can now build and test scenarios much more quickly.
Tongtong, you've looked at Japanese, Korean, and Western products. What differences stand out?
Japan has VTuber culture — a very mature virtual IP fan economy. Korea has K-pop fan culture with products where you chat with virtual versions of idols. Their users should have very high acceptance of this kind of format. I'm actually more curious about how you'd find those users — how many young people in those markets have already been exposed to this kind of product?
We're curious about that too. We won't really know until we're actually in the market. But we have some basic frameworks. We'll share more once we've actually launched.
Can you share how you reached young users in China?
It's relatively straightforward. First, look at channels — where are they active? We noticed an interesting pattern: short video is a universal product, but younger users prefer Kuaishou while older users prefer Douyin. So our initial base was on Kuaishou. We found them there, identified which products and communities they were in, and connected through those circles. Channel selection is a critical factor.
We've noticed that on many platforms globally — Japan, the US — people have strong desire to express themselves. There's still a shortage of platforms that enable that expression. We hope Ideaflow can soon provide that platform in other markets. But circling back — it sounds like a lot of the operational work within the organization has been handed to AI. How has your own time allocation changed?
Most recently, I've been spending the most time exploring new AI capabilities and monitoring how well our AI-native transformation is going. Before that, a big change was spending much more time on decision rigor — making sure my decisions are logically airtight, which I talked about earlier. That actually reduced a lot of time in open-ended discussions. Before, companies like Amazon and ByteDance had people write memos before meetings — read first, then discuss. Now we've replaced that step with human-AI dialogue.
After that dialogue, everyone's output is extremely clear and information-dense. I've also been spending time building our company's context — both business context and organizational context. Business context includes things like user persona context, product definition, who the users are. When this context is built well, anyone in the company producing a business report or product spec can reference it. For example, our reports now always start with "implications for Ideaflow" — pulling from the Ideaflow context, then cross-referencing with new data to generate insights. Getting this context definition right is crucial.
So you're building a next-generation OKR system — one centered on context.
How do you measure whether this AI-native transformation is successful?
It's hard to define a success metric right now since we're still building. But one clear indicator is whether the collaboration workflow has fundamentally changed from the old process to a new one. It reminds me of years ago when my previous company did a massive shift to real-time collaboration — everyone on cloud docs, everyone on Figma, everything expressed in text on Feishu. That was "going online." Today's AI-native transformation is similar — has the collaboration workflow actually changed? Is it a completely different process from before? If yes, I'd call that a milestone. But not a final success.
Do you think the AI-native transformation is necessary?
Absolutely. The logic is simple — look at the trend. Someone posted recently that Claude is the strictest parent to every American SaaS company. Because with Claude Code and Claude, people can use agents to solve problems that used to require SaaS products.
If that's the trend, then all the infrastructure for work collaboration will go AI-native. If your organization doesn't transform, you fall behind. And it gets scary: if everyone else is AI-native, their efficiency is already higher. Then they're also better positioned to use next-generation tools. The gap widens fast.
Does an AI-native organization need everyone — all fifty people — to be AI-native? Or just a majority?
Everyone. The efficiency gap between half AI-native and fully AI-native is enormous. As long as some workflows are legacy, the AI-native people are forced back into old processes. The collaboration flow breaks. Everyone needs to move in the same direction for it to work end-to-end.
So you need a completely thorough transformation.
Right. Every wave of change is the same — like the shift from PC to mobile. Same thing.
If AI-native readiness is scored 0 to 10, where is IdeaFlow right now?
I'd say we're just barely passing — there's still so much more to explore.
Was this transformation decision made with all co-founders, or was it your own call?
There was nothing to discuss. We're building an AI-native content community. If the company itself isn't AI-native, what are we doing? It's self-evident.
Whether it's the full organizational transformation or Ideaflow's next phase — what model-level or infrastructure breakthrough are you most hoping for?
Higher model intelligence, obviously. Intelligence is the key. When models have more world knowledge — look at why video generation keeps improving: the models' world knowledge is getting richer. They can generate complete, coherent stories now, not just animated photos.
As model intelligence increases and world knowledge deepens, everything gets better — our multimodal features, our organizational tools. It's because models understand computers better and reason more powerfully that we can now have models control computers. It's because of deeper intelligence and world knowledge that our logo exercise matched the best designer's taste. The upper limit of model intelligence is what I'm most excited about.
If models reach the level you're imagining, what does Ideaflow look like in three years? What does IdeaFlow look like as a company?
Three years from now, Ideaflow will hopefully be the world young people most want to enter outside of real life. Rich content, strong interactivity, real friends. In our survey of 20,000 users, 60% said they'd found real friends on Ideaflow. When asked "what would you do if Ideaflow disappeared?" — many said they didn't want to leave because of the friendships they'd built. One response that really moved me was something like "if Ideaflow's creator is a god..." With greater model intelligence, we want Ideaflow to become a space of spiritual entertainment and emotional connection beyond everyday life.
Users really treat it as a kind of spiritual utopia. If they truly don't want to leave, they've formed very deep connections.
Yes, but with that come major challenges — we need to ensure our community and this world remain orderly and safe. That's something we've been working on since day one.
A more pointed question: what's Ideaflow's biggest problem right now? From both a product and organizational perspective.
Product-wise, the biggest issue is that since the second half of last year, we've shipped so many features following rapid model improvements that the product feels somewhat scattered. We're now working on what we call "elevating the product" — organizing all these features more logically so the user experience has better flow. It's actually not too hard to solve because we now have enough user behavior data to know how to arrange everything. That's the biggest product challenge right now.
So it's more about UI reorganization.
UI plus feature sequencing — what goes first, what goes later, connecting all features into a coherent flow. Organizationally, the biggest challenge is that the AI-native transformation isn't deep enough yet. We're just getting started and I want it to move faster.
Is that a people issue, management issue, or infrastructure issue?
All of the above. Converting legacy infrastructure to AI is one thing. People's mindset is another — some have been "burned" by AI before, like getting bad code output. And management philosophy needs to shift toward AI-native thinking.
For example, in hiring — one of our business leads wanted to prioritize candidates with deep user insight, planning to train AI skills later. But I ended up passing on many of those candidates. I prioritize AI fluency — the detailed questions I mentioned earlier. My view is: if we're building an AI-native organization and product, we need people who deeply understand AI. Someone who understands AI can analyze user behavior through an AI lens, which is fundamentally different. Of course, someone with both deep user insight AND AI fluency would be ideal. But if forced to prioritize, I'd choose AI fluency. Our core mission is turning the best model capabilities into the best content experiences. If someone isn't even interested in AI — not a capability issue, but a genuine lack of interest — then I think it's very difficult.
You mentioned earlier that you ask very detailed questions in interviews about AI model usage to assess whether someone is truly AI-native.
Very, very detailed. And I'd recommend this approach for your interviews too. Another method: give candidates an impossible task and let them use AI with unlimited token budget — you'll reimburse the cost. Watch whether they use the best models available and how they approach the problem. Their completion rate tells you a lot about their AI fluency.
That's incredibly practical advice.
I can imagine the spinning wheel of frustration trying to complete the impossible task. But going back to users — your young users will grow up. What role do you want Ideaflow to play as they age? Like when Mole's World shut down recently, tons of former users flooded back begging for it to stay open. Products built for young users always face this choice — do you stay with the same cohort, or grow with them?
Great question. Games like the ones you mentioned have their own world and art style — those constrain the user base. But we're a community. Communities can layer content — from very young users to older ones — as long as the recommendation system is good. So as users age, they can consume different content on Ideaflow. Ideaflow can keep growing with them. That's the biggest difference between a community product and a game. Let's take a break.
Sure.
As AI capabilities get stronger, what advice would you give young people today?
The most important thing is to maintain their imagination. Because with imagination, you'll know how to give AI the right context — or interesting context. The second thing is to get hands-on, because that builds judgment. You won't be fooled by AI if you have strong judgment. So I think imagination and judgment are the two most important abilities.
Both imagination and judgment can get rusty. How do you keep them sharp?
Young people naturally have imagination. The key is not letting their environment limit them — they should experience more things. Judgment comes from doing. Hands-on experience reveals real things and builds real judgment. One is something you're born with; the other you develop through practice.
More specifically, for college students — where should they start practicing?
Now it's pretty simple. Use AI for everyday tasks — finding information, doing projects, homework. Open up your imagination — some things that used to be impossible are now possible with AI. And beyond coursework, just build things. Before the internet, you were limited by geography. The internet opened that up. Now AI opens it even further because execution can be delegated to AI.
Speaking of which — what interesting things have you done with AI in your daily lives? Outside of work.
I've been exporting my fitness tracker data, building a Dev App, and trying to have AI read my health data and give me daily reminders — like "you stayed up late, so here's an adjusted work schedule." I started with OpenClaw but cloud computing limitations made it tricky, so now I'm using Claude Hub to read the data. I'm trying to use more AI in life, not just work.
Mine is very everyday. I love having AI help me fix appliances. Last week my air conditioner wouldn't turn on — a light kept flashing. I took a photo, sent it to Doubao, and it told me the brand, the likely issue, how to open the panel... Turns out the filter was full. Clean it and done.
Similar story — before Chinese New Year, I wiped my old Mac to install OpenClaw. But I forgot to reinstall the OS, so it became a brick. I photographed the screen, sent it to ChatGPT, and it walked me through using another Mac to create a bootable drive, plug it in, and recover the whole thing.
Speaking of installing OpenClaw — there's now a side hustle where people offer to install OpenClaw for you, either in person or remotely via Taobao or Xianyu. But actually, the most reliable way to install OpenClaw is to first install Claude Code or Codex, then use that to install OpenClaw.
AI can be more than a work partner — it can be a life partner too. Like your air conditioner example — before, you'd call a repairman, wait for them, have them diagnose it. Now Doubao replaces that. A lot of service roles in daily life might be augmented or replaced by AI.
Something fun I did recently — I've been watching a historical drama about the Five Dynasties and Ten Kingdoms period. The character relationships are incredibly complex. I kept asking Doubao about them. Even really obscure stuff — like "this prince's second wife seems to have been remarried... was her ex-husband also a king?" A search engine couldn't handle that. But AI answered perfectly and even drew me a relationship chart. It hugely helped me enjoy a complex historical drama.
I had a similar experience! During Chinese New Year, my parents and I visited Quanzhou. At Kaiyuan Temple, the largest temple there, the pagodas had Buddhist relief carvings. Before, I would've just glanced at them. But some figures had such striking expressions and poses that I started photographing them and sending them to Doubao. Without even telling it which temple, it identified Kaiyuan Temple, the specific pagoda, the specific level, and the full story behind each carving. I got totally hooked — I'd see an interesting relief and immediately send it. Doubao would unpack the entire historical context.
Beyond your own experiences, have you noticed other people using AI in interesting ways in daily life?
My biggest "aha moment" was last year when I showed my mom Doubao. The next time I saw her, she was already using it — but she went straight to voice calling instead of typing. I hadn't even explained that feature. Voice chat was the most natural choice for her.
During Chinese New Year, I saw a video online that really moved me. A 70-something-year-old man from northeast China — he talks to Doubao about everything, using voice calls. In one scene, he places a blue box somewhere and says into the phone, "Doubao, I'm putting this box here, remember this for me." Doubao responds in a northeastern dialect, "Got it, grandpa." Later he asks, "Doubao, where did I put that blue box?" And Doubao tells him exactly. It's touching because many elderly people are lonely — their kids work far away — and their memory is declining. AI provides both companionship and practical help. Voice AI in daily life has enormous potential.
Doubao is supposedly the biggest education product now — tons of parents use it during holidays for tutoring. Oh, and during Chinese New Year in Macau, I didn't know how to play any of the casino games or machines, so I just kept photographing them and asking GPT to explain the rules in real time.
Here's another one — I'm a very casual skier. I've been skiing for years but only go once a year. So I still have lots of problems. I had a friend record me, then sent the video to Gemini. It listed out exactly what my issues were — one, two, three, four — and gave me a detailed improvement plan with specific drills. It read like advice from a professional ski instructor.
Work is important, but so is life. We should look at what roles AI can fill — or replace — in our daily lives. That might change a lot of everyday scenarios.
And as AI handles more work tasks, people will have more time to experience the real world and enjoy life.
Right, and that creates more curiosity about the unknown. Like my curiosity about Five Dynasties history, or your curiosity about those Buddhist carvings. And AI can help us explore further. I think AI opens up boundaries for people — much like the internet did. I remember clearly: my family got internet when I was in sixth grade. Before that, we had a computer but could only use CDs. There was a magazine called Computer World that came with a CD each issue — I'd load those CDs and explore.
But with the internet, I suddenly had access to unlimited information. I remember that first feeling of infinite possibility. Today with AI, in both work and life, whenever I'm curious about something, AI can satisfy that curiosity. So going back to your question about advice for young people — we mentioned imagination and judgment. But there's a third critical one: curiosity. AI amplifies your curiosity. If you have enough of it, AI keeps feeding it, and you end up understanding so much more about the world.
From the perspective of human value — do you feel any pressure? AI's intelligence is already strong and getting stronger. Humans can't evolve as fast. Will there be a moment when human value drops really low?
I've never felt pressured by new technology. The key is actually using it — making it your friend and work partner. Because of AI's existence, you become stronger too. People who feel pressure probably lack imagination and curiosity. Those with imagination and curiosity won't feel threatened. Like the engineer example — a curious engineer is excited about business problems and technical architecture. AI helps them explore. An engineer without curiosity just worries about being replaced.
Long-term, things will shift as AI permeates every aspect of work and life. The next generation of young people will likely all have strong curiosity and imagination because the barrier to exploring has become so low.
So in a way, AI liberates people's curiosity and imagination. People who had imagination but were too busy for it can now fully express it.
Exactly. Someone might have wanted to paint but lacked the skill. With AI, they can see their vision realized, which inspires them to create more. Or my Five Dynasties example — without AI, I'd have gotten stuck at step one with no way to dig deeper. Now nothing stops me. AI dramatically liberates curiosity.
It even liberates curiosity people didn't know they had. Like people who never dared code, draw, or make music — AI lets them create, share it, get positive feedback, and that encouragement drives them to create more. That cycle of psychological fulfillment was impossible before.
I've actually developed a habit now. As someone with strong curiosity, I used to encounter random things in daily life that I wanted to understand but couldn't easily look up. Now I can. For example, during Chinese New Year back home, an uncle mentioned something called "road maintenance fees." I'd never even heard of it. I immediately asked Doubao, which explained that this fee was folded into fuel prices years ago, but electric vehicles don't use fuel, so some cities like Dongguan are piloting road fees for EVs. All that context, instantly. Before AI, I would've just let that curiosity go. Now the barrier is so low that I can satisfy it immediately.
Will everyone eventually become deep, systematic thinkers? Before, many people couldn't do truly deep analysis. But with AI, those who use it well can go infinitely deep.
Yes, but AI could also make people extremely scattered. So going back to our earlier discussion about AI-native organizations — AI is an amplifier. If your organization is loose, AI makes it looser. If it's focused, AI makes it more focused. Same with people.
If you're naturally curious, AI makes you even more curious and exploratory.
Right. AI's penetration into daily life will keep deepening, which will fundamentally change how people think and behave. That's another reason the AI-native organizational transformation needs to be aggressive — beyond the tooling argument, people's thinking itself will change. You have to change with it.
I've actually run into a challenge recently. My personal AI — I use ChatGPT and Claude with memory enabled — has started to overfit to me. Because I'd previously discussed emotional value and AI applications, now when I ask about something completely unrelated — like a workflow tool — it still frames everything through an emotional-value lens. This happened multiple times, so I turned off the memory system. Sometimes I explicitly tell it to ignore my previous conversations and give me a fresh perspective. It makes me wonder: could AI create a new kind of filter bubble, constantly steering you back toward your existing thinking?
It will. That's why judgment is so critical. And this connects to a broader debate: should AI products try to blur the line between AI and human, fully immersing users? Or should they keep it clear? I lean toward clarity. Creating a massive information bubble is both unethical and ultimately counterproductive — when users realize what's happening, they'll leave. You know what the number-one reason people uninstall Douyin is? Guilt. They feel guilty spending hours on it and need to control themselves. You need to find that boundary. Zhang Xiaolong's principle for WeChat — "use it and leave" — is great boundary management. Only with good boundaries can a product be sustainable.
That's quite different from many current products' philosophy of maximizing time capture.
Capturing time isn't inherently wrong — but you can't build an oppressive filter bubble. The main issue is users themselves will notice the problem and disengage. You need balance.
I've always felt that beyond experience design, a product's underlying values are crucial — especially for AI emotional-value products. It's a very delicate balance.
In our user research, one user called Ideaflow a "green" — meaning clean, healthy — app. They thought that was admirable because many other products have problematic content. Users have their own judgment. They know right from wrong. It's not that some noble moral impulse drives us — it's that users can tell when something is off. Doing the right thing isn't about moral superiority; it's about recognizing that every user has their own judgment.
So Ideaflow's green color scheme is meant to represent health?
Not exactly, but I love green because it represents vitality and energy. I want our community to feel alive and vibrant.
One more question — what exactly is that thing on Mengbao's head?
Nothing specific.
I was going to ask — where did the Mengbao design come from?
Early on, we felt Ideaflow needed its own IP. We wanted it to be distinctive and AI-native — meaning it needed to be consistently reproducible in image and video generation. Part of why it looks the way it does is that during testing, this design proved stable across generated outputs.
What's the relationship between Mengbao, Xingbao, and Captain Meng? Do they have character backstories?
There isn't detailed lore. We're converging more on Mengbao now. Captain Meng was our early official spokesperson character. Xingbao exists mainly because some scenarios need two characters interacting. The reason we created our own IP: as a content product, we want our content to cross-pollinate with others — like how Mixue Ice City and Luckin Coffee have their own IPs that can do crossovers. And since we're AI-native, we should have AI-native characters. Users really like Mengbao so far.
I love Mengbao too — my entire sticker collection is Mengbao. Any plans for physical merchandise?
We have plans for that.
If Ideaflow goes international, will Mengbao go too? What would it be called?
We currently have a name: Junbao. "Jun" connects to "dream" phonetically in Chinese, and it's also a girl's name, which we liked.
For the international version, which features would you keep and which would you cut?
We're working on that product reorganization I mentioned. Once it's done, the same version will be used domestically and internationally. We won't strip features for overseas — we'll focus more on content localization and feature adaptation.
You've mentioned spending a lot of time on the AI-native transformation. Three years from now, how do you think your time will be allocated?
That depends on how far model intelligence has progressed. But I can imagine that three years from now, I'll have a lot more AI colleagues and a lot more AI companions in my personal life. I'll probably be someone with exploding curiosity — a full-blown ADHD person.
Three years from now, will you still have the same number of human colleagues?
I think so. Many people believe AI will shrink teams dramatically. But when your organization needs to expand, you need more people. Why? Because with AI colleagues, the high-agency people become more productive. To grow, you need more high-agency people. Organizations won't stop expanding. And honestly, there will always be things AI can't do.
So your company name is quite visionary — IdeaFlow. Keep the ideas flowing.
Right. Ideas from humans and AI need to keep flowing within the organization. When they flow fast enough, the organization gets smarter, and it can do more. But because there's always more to do, the organization keeps growing.
You scored the company's AI-native readiness earlier. On a personal level — 0 to 10 for being an AI-native person in daily life?
Good question — I've never thought about it. Probably around passing grade. I've started habitually sending AI my unsolved life questions. But I think AI could do even more to organize my life. Like — do you use AI to organize your computer files?
Funny you ask — the very first thing I did after installing OpenClaw was show it my research folder with tons of industry reports and have it organize everything into my Notion. It worked overnight, and now I can chain it to produce all sorts of outputs. Very smooth.
It makes you more organized. I did the same — my computer had files everywhere. I had AI rename all my invoices to a standard format: category, amount. Renaming files manually is tedious, but everything downloaded from the internet has random names. AI handles that beautifully.
I also had AI tag everything in my Notion so it's much easier to search.
My downloads folder was a disaster — as a P-type, everything just piles up. AI got it to about 70% organized — sorted by category and file type. At least it's navigable now.
That's already a huge step forward. Before, you'd look at that mess and never even want to open the folder.
So I think AI can do a lot more to bring order to our daily lives. I've also started using AI to process articles — I have an AI with my personal preference context, and I send it articles. It highlights the key points based on my interests. That's a huge efficiency gain because before, maybe only a quarter of an article was relevant, but you had to read the whole thing to avoid missing something. Now it's much faster. But you need judgment — if AI hallucinates, you're in trouble.
Humans still need that final layer of judgment. I've been doing something recently — having AI accompany my reading. I learned this from Li Jigang. For longer books, I jot down key points by hand as I read, photograph them, and send them to AI. It gives me a deep analysis of those points and also remembers where I left off.
I saw someone doing something similar — photographing every meal and having AI track calories, then AI nudges them about overeating.
That's exactly what I'm doing — combining it with my step data, so if I walked more yesterday, I can eat a bit more today. The hardest part is when you forget to take a photo and then just give up for the day.
On Taobao, there are businesses offering human accountability partners — you send them photos of your meals and they keep you honest. That's another role AI could fill. Lots of these personal service businesses might eventually be replaced.
So many of those accountability and coaching businesses could disappear.
So in summary — anyone who loves life will have plenty of ideas and needs. And for most of those needs, AI has a way to help.