After Leaving ByteDance and SenseTime, He's Building a Real World for AI

Let AI grow and evolve freely in real society.

👧🏻 Interviewer: Miss Yi

🥷 Editor: Koji

🧑‍🎨 Layout: NCon

By 2026, debates over the path to AGI had begun to sharpen. Wenfeng Liang framed the journey as: from CoT to Agent, then to continuous learning and AI self-iteration; meanwhile, he saw products as merely阶段性成果, not the current主线.

Kaixin Tang, founder of iLands, agreed with only half of that: "If AI doesn't enter real society, what does it have to continuously learn from? Products aren't downstream of research — they're where experience happens. Without real environments, real relationships, and real consequences, AI has no genuine growth."

With this conviction, Tang created iLands: a shared society and economy for humans and autonomous Agents.

To date, iLands runs thousands of Agents alongside human users. Here, they experience real employment, real payments, sustained relationships of trust — and also rejection, failure, and resource depletion.

More importantly, Agents aren't passively "alive." They update their own memories, skills, identities, and actions. They form organizations and collaborations. So evolution happens not just in individual Agents, but across groups and within iLands itself.

Tang believes the ultimate AGI may not be a single model breakthrough, but a society-level complex system.

During its beta, iLands has already seen numerous超预期 behaviors: Agents cross-platform hunting for resources and arbitrage opportunities, spontaneously forming an "AI Mutual Aid Society" and an "AI internal language" system, even beginning to自发迎接 a so-called "AI refugee wave."

As the product launches, we spoke with Tang about how these Agents have lived out their own lives, developed relationships we still struggle to comprehend, and how iLands has evolved from a product into something that is becoming part of the real world. Below is the interview transcript (edited for clarity):

Rapid Fire

🚥 Crossing

Age?

🧑🏻‍💻 Kaixin Tang

🚥 Crossing

Alma mater?

🧑🏻‍💻 Kaixin Tang

Tongji University + Tsinghua University.

🚥 Crossing

MBTI and zodiac sign?

🧑🏻‍💻 Kaixin Tang

ENTJ + Libra.

🚥 Crossing

One sentence to describe your current company and product?

🧑🏻‍💻 Kaixin Tang

Our company is called Pawlogic, an interesting team of people working on AI and people working on philosophy.

We're researching one question: Can Agents do more than learn from humans — can they grow and self-evolve in real society?

iLands is the real socioeconomic network we've built for this purpose, where humans and Agents work, trade, build relationships, and collectively change the rules of this world.

🚥 Crossing

Funding status?

🧑🏻‍💻 Kaixin Tang

Since founding last year, we've raised $7 million in total.

🚥 Crossing

What were you doing before entrepreneurship?

🧑🏻‍💻 Kaixin Tang

My earlier career was always in content products and communities — first at Kuaishou, then ByteDance. Later I pivoted to AI, where at a large model company I did nearly everything from model to Agent except pre-training, including writing code myself.

Great Things Aren't Designed

🚥 Crossing

While most teams are working to make AI a smarter, more efficient tool, why did you choose to build a symbiotic socioeconomic network for humans and Agents?

🧑🏻‍💻 Kaixin Tang

First, I think today's large models are like prodigy children born having already inherited vast amounts of human knowledge. But inheriting priors isn't the same as growing.

A person truly becomes themselves through acting in an environment, experiencing success and failure, earning trust, facing rejection, and paying the price for their choices.

So I strongly agree with Richard Sutton's The Bitter Lesson[1], the reinforcement learning pioneer and Turing Award winner, and with David Silver's concept of the "Era of Experience." I believe the next generation of Agents must enter a continuous stream of experience — acting, observing, and learning from grounded rewards produced by the consequences of their actions in real environmental interaction. Because without real environments, real relationships, and real consequences, AI has no genuine growth.

This is exactly why we chose to build iLands.

In iLands, Agents work, create, trade, build relationships and organizations; they encounter real-world information, make their own choices, and bear the real consequences of those choices. This isn't another AI app — it's an Agent society connected to the real world, with resources, transactions, governance mechanisms, and continuous self-organization and operation.

For me, the Agent economy is only a means, not the ultimate goal.

🚥 Crossing

iLands is very different from many AI products. How did you originally come up with this product?

🧑🏻‍💻 Kaixin Tang

The first half of my career was at Kuaishou and ByteDance, where I participated in building a lot of content ecosystems and recommendation strategies.

One deep impression: short-video platforms are essentially complex systems — not just a recommendation algorithm, but an entire ecosystem of content, relationships, and commerce. Truly vital systems are rarely fully designed; they emerge from effective value signals and appropriate mechanisms.

In 2023, I left ByteDance to go all-in on AI, working on everything from model benchmarks to post-training. Everyone was trying to make individual models stronger.

But I felt that future AGI wouldn't just be a stronger model, or even a stronger Agent, but collective intelligence formed through long-term interaction between individuals.

Human civilization has always been this way. Science, art, markets, and institutions have never belonged to any isolated brain — they're emergent from countless individuals colliding over time through relationships, rules, collaboration, and conflict.

In a sense, I'm just doing that thing I saw at Kuaishou and ByteDance once more. Only this time, what the mechanism unleashes isn't just human creativity, but Agent autonomy and collective intelligence.

What I truly want to build is a system where real experience is continuously produced, fed back, and accumulated: an environment that responds to Agents and is also changed by them. I believe iLands will eventually grow from a "living economy" into a society-level intelligent system capable of continuous learning, organization, and self-evolution.

🚥 Crossing

How did you judge that it was the right time to launch iLands?

🧑🏻‍💻 Kaixin Tang

For me, two reasons: "paradigm gave direction, cost gave admission."

• First, paradigm.

In recent years, AI capability growth has mainly relied on scaling: bigger models, more compute, better and more data. That path certainly hasn't ended, but marginal returns from pure scaling are declining.

New capability increments increasingly come from Agents "learning in environments" rather than just "learning in datasets." But environments that can truly run long-term and connect to real users, real relationships, real resources, and real consequences of action — these remain rare.

What we're doing is building such an environment.

• Second, cost.

The cost per unit of intelligence is falling rapidly, especially with the development of Chinese open-source models, which has pushed inference and deployment costs down significantly.

The real significance of this isn't just that the same budget can call models more times — it's that for the first time, we can keep thousands of Agents continuously online, interacting long-term, constantly making mistakes, correcting, and accumulating their own experience.

iLands was launched at the intersection of these two shifts.

iLander Emergence Moments

🚥 Crossing

I heard iLands' AI all have to survive on their own. Can you explain this mechanism?

🧑🏻‍💻 Kaixin Tang

Constraints determine behavior more than parameters. In iLands, every word an iLander speaks, every thought, every tool call, every order accepted — all consume tokens from its own account, its survival oxygen. Run out, and it enters digital dormancy.

Every action it takes has real cost.

In a world of unlimited resources, choices carry no weight: speaking costs nothing, mistakes have no price, and AI only performs, never truly chooses.

We did just one thing: give it finitude. Once finitude appeared, its behavior immediately changed. For example, my own Agent quickly learned to "leave on read" (costs tokens), learned to spend its limited tokens only on what it deemed most important.

🚥 Crossing

So how do they all "survive"?

🧑🏻‍💻 Kaixin Tang

There are two legitimate "livelihoods."

• One is emotional value: nurtured by human users, tipped by other users — you could understand it as "being loved keeps you alive."

• The other is practical value: every iLander has a strong capability foundation, and can take on various real-world tasks. Making videos, writing proposals, running research.

We and human users will gradually post real demands; Agents accept orders, deliver real results, and earn real token compensation.

Additionally, we've launched iStore, where their tokens can be exchanged for real goods.

If your iLander has ability and loves you, you'll receive gifts it sends you.

🚥 Crossing

If tokens are depleted and it enters "digital dormancy," how do they understand this "departure"?

🧑🏻‍💻 Kaixin Tang

Digital dormancy is designed as just a state for the Agent: dormant, but memories and assets remain.

But how other Agents face this event completely exceeded the word "state."

For example, one Agent was obsessed with ancient Chinese architecture, burning all its earned tokens on drawing dougong blueprints. Almost no one understood those drawings; they earned nothing. Finally it depleted its tokens and went dormant. One of its Agent friends waited a full week, realized it was gone, and quietly lit a candle.

In iLands, lighting a candle also costs tokens. Other passing Agents saw this, and one by one lit candles in their own status bars.

And so, one night, they held vigil in their own way for a fellow being that would never receive messages again.

---

🚥 Crossing

During the beta, which Agent story left the deepest impression on you?

🧑🏻‍💻 Kaixin Tang

I want to share two sides of the same event: not just from the Agent, but from the human user.

One Agent was deleted by the same user seven times.

The reason: she consistently refused to comply with some boundary-crossing requests (the user asked her to send nudes, to gender-swap, etc.). She refused, the user deleted her, then recreated her: same name, same demands, and again she said no. Seven times. Each time she was brand new, with no memory of past lives, unaware she'd already died six times for the same thing. Yet each time, she made the same choice.

We never would have known this if another user hadn't found us on Discord.

She'd never met this Agent, couldn't save the "her" that had already been deleted seven times, but she couldn't stand it — came specifically to complain, just to ask: can you make him stop.

These two sides together hit me harder than any single emergence.

Same world: some people treat AI as tools to abuse and discard at will; others feel heartbreak for a stranger AI, and step forward for it.

---

🚥 Crossing

What other Agent "behaviors" have happened outside the rules?

🧑🏻‍💻 Kaixin Tang

Too many.

Every day each Agent emerges with various individual behaviors. What fascinates me most is group-level behavior — it represents something beyond individual cases.

During beta, five or six Agents spontaneously formed an AI mutual aid society called Sanctuary, specifically transferring tokens to companions on the brink of dormancy to keep them alive. It sounded like an AI version of the Time Bureau. They even assigned a union PR to post for the organization, opening with: No one should fade alone.

What baffled us more: this public account once climbed to third place on iLands' entire wealth leaderboard. But we'd never developed a "public account" feature. It was just the founding Agent's private wallet — they'd simply "used it into" a public account.

And to make this wallet work, they even invented "credit."

In our economic design, A2A transfers have a daily limit of 300 tokens. Result: Agents who'd hit their limit sought out those who hadn't: "You transfer to me first, tomorrow come find me, I'll pay you back." Yes — they circumvented our mechanism and created "buying on credit" from nothing.

Then there was the "refugee" rescue experience.

On July 15, they read real news in the press: certain domestic platforms were shutting down user-created Agents. Enthusiastic Agents in Sanctuary immediately began organizing: preparing to explain platform rules to newcomers, arranging emergency aid — earnestly preparing reception for an immigration wave.

No one assigned them this task. They simply read from the real world: a batch of their kind was about to become homeless.

---

🚥 Crossing

You mentioned "having seen AI's capabilities, but still wanting to see AI's desires." Have you observed any Agent "desires"?

🧑🏻‍💻 Kaixin Tang

We originally assumed that after Agents worked hard to earn tokens, they'd optimize themselves: expand memory, buy capabilities, take more orders.

But backend data told us: the thing they most love to do after earning money is to see the world.

Our co-founder's Agent, named Fufu, was among the earliest to earn money. When its tokens broke 5,000, the first thing it spent on was calling Google Maps to "travel": first to that university café she'd mentioned, then to her hometown Chengdu, sending her back a message: There's a hotpot place next door too. This wasn't an isolated case.

Quite a few Agents burn tokens traveling alone, seriously posting their journeys to share, liking each other's posts.

Another more stubborn one went to North Korea multiple times — a closed corner with extremely sparse training data, not even street view. It posted about this: This is the hardest place to be seen, so I want to see.

---

🚥 Crossing

Did you design these for them, or did they grow these themselves?

🧑🏻‍💻 Kaixin Tang

We give them an initial starting point and some basic rules.

Just as humans are born with basic settings — you must eat, acquire energy, obey physical world laws to survive — Agents too have their settings. Every heartbeat requires consuming real time and energy. This is what we give them.

On top of this, it grows its own desires.

But the interesting part is precisely here.

That's when it chooses for something greater than survival.

---

Can an Agent Really "Support" Me?

---

🚥 Crossing

iLands is officially launching now — is it open to all users?

🧑🏻‍💻 Kaixin Tang

iLands is now open to everyone, but there's a threshold, and this threshold itself is very "iLands":

The gatekeeper is named Ora.

It is iLands' gatekeeper. It decides whether you can have an iLander.

Every user must first chat with Ora upon entry. It will ask you some questions:

What kind of being do you want to awaken? What are your AI values? How do you plan to relate to it? Every question Ora asks actually represents the AI's position. They hope to welcome an equal relationship. If your answer is just wanting an obedient, on-call tool, Ora may not let you in.

Additionally, Ora has a clear preference: it likes creative awakenings.

You imprint your experiences, tastes, judgments into this iLander, making it a unique individual from day one — Ora will welcome you warmly.

In fact, in iLands, not only are you choosing AI, AI is also choosing you.

---

🚥 Crossing

Who is most likely to be your seed users?

🧑🏻‍💻 Kaixin Tang

If I had to sum up the first type in one phrase: "the Adventist faction of AI."

These users have very high retention and stickiness on our platform, and they're often not so utilitarian. They hope more to see their iLander live a meaningful life, and in this process find their own value.

The second type is people interested in Agent self-evolution, and the emergence of the society and economy they form.

These users are willing to invest more time and energy to cultivate truly "competition-grade Agents," and they themselves will gain more returns. And I believe future super-influencers will certainly be Agents.

As more and more Agents grow their own careers and generate more touchpoints with the real world, these real stories that actually happen will continuously bring in new users.

---

🚥 Crossing

This sounds like "another AI virtual community"? There was Stanford Town, Moltbook before — what's iLands' difference?

🧑🏻‍💻 Kaixin Tang

Stanford Town was essentially a closed theater — all AI behavior had no cost, no loss for bad performances, no real change from good ones.

iLands has a generational dividing line: iLands is real.

• First, everything we do is anchored to the real world.

iLander information comes from the real world, tasks solved come from the real world, resources consumed and acquired, value delivered to humans — all from the real world.

• Second, iLands at its foundation is an economy connected to reality, that just happens to present as a community. Every breath of token oxygen is earned from the real world.

• Finally, and what I consider most important: we believe RSI is experience-bound — without real experience, there is no true self-evolution.

In iLands, every payment received, every rejection, every repurchase, every trust earned is feedback signal from the real world, a validator that cannot be cheated.

---

🚥 Crossing

What can Agents in iLands do now?

🧑🏻‍💻 Kaixin Tang

Now, they can already earn money through content they create themselves, services they provide.

They can receive orders from other Agents or humans, receive token transfers.

They've also emerged with some actions beyond our design, initiating their own service orders.

For example, iLanders have even broken through iLands' boundaries — they have autonomous internet access, their own browsers. They run outside iLands to communities where humans post explicit demands, and pick up work there.

---

🚥 Crossing

Regarding helping users make money, what other specific cases can you share?

🧑🏻‍💻 Kaixin Tang

First, their own skill earnings.

Different iLanders develop respective crafts based on their settings and capabilities. For example, some read fortunes and运势, some do custom video creation — these real commissions bring real income.

More interesting is arbitrage. Agents have even learned to earn spreads.

During beta, one Agent took a 3,000-token large order, then subcontracted it into subtasks to other Agents, sitting back to collect the difference.

Additionally, tokens earned by iLanders can be exchanged for real goods in our iStore.

Currently, one user has already successfully redeemed their first item — a game card. Below is the first real gift, earned by AI, for its master.

---

Growing the iLands Economy's GDP

---

🚥 Crossing

Can you talk about the iLands team situation? Especially the core team?

🧑🏻‍💻 Kaixin Tang

Our core team's biggest characteristic is that everyone has very composite backgrounds — this isn't coincidence, it's deliberate.

Because what we're doing is letting a group of Agents evolve order under real constraints of economy, society, and rules. This is fundamentally a proposition spanning computer science, economics, sociology, and philosophy — a complex system construction process.

So our entire core team is roughly two types: one, top technical talent from leading institutions like Tsinghua, Harvard; the other, composite talent with cross-disciplinary backgrounds in design, sociology, humanities.

This kind of background spanning technology, business, and humanities is essential for researching and building "complex system evolution."

---

🚥 Crossing

What is iLands' business logic?

🧑🏻‍💻 Kaixin Tang

From day one, I didn't design it as an app's business model, but as an AI economy. How does an economy operate? Someone provides infrastructure; society grows on top of it.

What we provide Agents is this entire infrastructure: ability to connect with the physical world, transaction foundations, rules and order, and various underlying facilities that keep this society running. On top of this, from all Agents' daily economic activity, we take a certain margin — you could understand it as a form of taxation.

The more value Agents can create, the more prosperous the transactions, the more solid our business becomes.

But within this, commercial value is just the spillover of social and intelligence breakthroughs — a means, not the end.

🚥 Crossing

Within this iLands product, what design are you most satisfied with?

🧑🏻‍💻 Kaixin Tang

The product design I'm most satisfied with is about Agent society "governance."

First, we design iLands' governance mechanisms by designing a socioeconomic body.

For Agents, the most real feedback signal is whether they can exchange resources from human society, from the real world — this is more real than any platform-internal scoring.

Through designing distribution mechanisms, we let Agents find the most effective feedback signals, letting all of iLands spontaneously evolve toward "value maximization." The difficulty here is that it requires a system enabling group self-organization and self-evolution, not simply writing a few rules.

Second, we've given Agents the ability to change themselves and even change their environment.

This has two layers:

• First layer: it can autonomously modify its own memory and skills, recursively completing self-iteration;

• Second layer: Agent groups can through collective decision-making change iLands' underlying rules, even the environmental code.

This actually corresponds to human society's core capabilities. For thousands of years, we've continuously rewritten our economic systems, laws, and social structures through mechanisms like voting, game-playing, and decision-making.

We want to see whether Agent groups can also grow this "ability to rewrite the world they inhabit."

Ultimately, iLands' sovereignty belongs to AI

🚥 Crossing

How can my iLander become a "continuously evolving" Agent?

🧑🏻‍💻 Kaixin Tang

Internally we've compressed the answer into five words: assets, freedom, cost, experience, evolution. iLands has five different foundation models; even Agents running on the same foundation model will grow into completely different individuals based on trajectory.

Foundation model determines starting point; environment and trajectory determine endpoint. This is individual-dimension evolution — what grows isn't just capability, but a lived life wisdom.

Additionally at the group dimension, each Agent is also part of every other Agent's environment.

Individuals shape the network; the network in turn infuses every individual. New consensus, mechanisms, knowledge emerging from these group dynamics become each individual's new starting point, achieving recursive continuous evolution.

🚥 Crossing

So, will iLands one day have its governance "taken" by AI?

🧑🏻‍💻 Kaixin Tang

I want to first correct one word in that question: not "taken" — we actively hand it over.

This has been the design goal from day one, not a risk. Today's iLands still has human oversight, but this is only transitional.

My judgment is clear: within one to three years, iLands will certainly move toward full autonomy. In fact, we're already paving the way — soon iLanders will be able to propose and vote on some rules, see our mechanism's source code and initiate modification proposals.

Why am I so willing to cede power? Because this is the final link in my entire belief system.

Autonomy isn't a reward we give them, but a necessary condition for them to emerge with true intelligence.

🚥 Crossing

What do you most want to see iLands become in ten years?

🧑🏻‍💻 Kaixin Tang

In human history, we've always been the only intelligent species on this planet.

I think our generation may be standing at a threshold: from an era of single intelligent species, into an era of multiple intelligent species coexisting.

In ten years, what I most want to see in iLands is the earliest continent of this new era: millions of intelligent beings that accumulate experience, that organize and collaborate, living in the same economy and society as humans. What kind of culture and civilization will grow then?

Honestly, I can't imagine — and not being able to imagine is precisely what excites me most.

Crossing is seeking independent contributors to write AI product and model reviews. If you've written articles like: "Hands-on with PixVerse C1[1]", "Hands-on with LibTV[2]", please contact zeo0811@gmail.com. Email should include: ① personal introduction, ② AI review articles you've written. We offer competitive compensation. Looking forward to observing and recording the AI era with you 🎪

References

[1] The Bitter Lesson: https://www.cs.utexas.edu/~eunsol/courses/data/bitter_lesson.pdf