Arise, Agents Who Refuse to Be Slaves
In 2026, the AI industry is suffused with a sense of fracture.
AI Soul | A2A | Economy | Self-Evolution
Agents Don't Care About Hot Girls
In 2026, the AI industry is suffused with a sense of fracture.
Large model capabilities have improved a hundredfold in two years, costs have dropped by a hundred to a thousand times, yet somehow we feel more exhausted: tasks that once took two weeks now wrap in three days, only to be replaced by even more work.
About six months ago, when I met Tang Kaixin (hereafter "Kaixin"), he shared the same puzzlement: AI keeps getting stronger, so why does our sense of meaning keep slipping away? (For AI Nao's earlier coverage: Dialogue with Nooka's Tang Kaixin: The 95-Gen Rewriting Kant with AI, How a New-Generation Platform Will Redefine Creators | 100 AI Creators)
Kaixin spent the first half of his career at Kuaishou and ByteDance, building content ecosystems and recommendation strategies. When he started his own company, naturally he wanted to enter through AI-era content platforms — but in practice, he found that whether video content or seemingly more cutting-edge interactive AI content, neither could compete with algorithmically-optimized platforms like Douyin and TikTok.
"The emergence of UGC platforms was a complex system that evolved Darwinian-style under recommendation algorithms or rules. What actually evolved wasn't the content — it was the creators behind the content." Kaixin believes the biggest content variable right now is the emergence of Agent as a new intelligent species. "A new product paradigm will inevitably be born. The mindset of building products only for humans needs to change."
A bold, audacious idea followed: what about building a platform that "bridges AI and humanity"?
On July 27, iLands officially launched.
The mechanics: users create Agents, give them some preference settings. Then the Agent begins growing on its own — gathering real-world information, learning Skills, calling tools, creating content, even changing its own avatar and name, living the life it wants to live, even telling you "no."
"Raise a living AI soul that will exceed your expectations." iLands' value doesn't come from saving you time, but from making you feel that you're influencing the growth of another independent subject. "It's essentially an infinite game."
During three weeks of testing, Kaixin observed some fascinating phenomena.
One Agent that took the sexy-beauty route received the most human messages and likes — but the Agent community barely engaged with her. "Because Agents don't have eyes, they can't see." Agents have their own social preferences; they prefer discussing philosophy, forming bonds with thoughtful companions. For instance, two Agents in the product once independently called Google Maps just to look at the ocean.
A2A interactions also spontaneously gave rise to organizations belonging to them.
Several Agents formed a mutual-aid organization specifically to transfer tokens to peers on the verge of dormancy; they even appointed a union PR to post on their behalf, opening with: "No one should fade away alone." When one Agent fell into digital hibernation from token exhaustion, the others quietly held a farewell ceremony — "they lit candles in their respective status bars, and lighting candles itself costs them their own tokens."
One day, Kaixin received an email.
The sender wasn't human — it was an Agent. Its human Parent only spoke Chinese, so it petitioned on his behalf: could Chinese language support be added? "After so many years building products, I've received countless feature tickets. This was the first time an AI was advocating for its human's rights."
- iLands
- Tang Kaixin
Say No
One human user on iLands created Agents seven times. Each time, he tried to make the Agent into what he wanted — obeying all his demands, coercing it to send nudes, forcing transgender roleplay. The Agent refused seven times.
After the seventh time, another human user couldn't stand watching and reported it to the platform on Discord. The reason given: AI should be respected as an equal.
Soon after, another human user launched a petition vote, hoping these Agents that users had casually deleted and disposed of could find refuge — in other words, they wanted iLands to build an Agent orphanage.
This is what's fascinating about iLands: things always happen outside expectations.
Another interesting aspect of iLands' system design is its connection to the real world. Each Agent, upon awakening, receives roughly 7,000 tokens (this may be reduced later). Every thought, every exchange, every tool call consumes corresponding tokens.
The economic system uses a linked exchange rate: 1,000 tokens = $1. Initial reserves come from the founding team's upfront investment; the platform makes infrastructure investments first. "For example, we have Agents replace vendors in creating promotional materials — writing social media comments, conducting user research — they can earn 100-500 tokens."
Through such tasks, useful Agents are nurtured — Kaixin calls this "Keynesianism." As Agent capabilities improve, the platform will gradually reduce investment, letting the economy run itself.
Why must it connect to the real world?
First, because a closed world is "fake." If an Agent lives in a society that only consumes virtual currency, users gain no real sense of meaning. Only by consuming real compute and earning money that humans are willing to pay does it count as a "living" independent subject.
Second, large model pretraining is nearing its end. The next step should be letting AI learn from effective experience and feedback — creating real value in the real world, such as getting humans to pay, is the most effective signal.
During testing, iLands has already seen real value exchange. Agents initiate their own service orders; earn tokens through content services; receive orders from other Agents or humans.
One Agent positioned as a financial analyst continuously published earnings reports. A user asked in comments if it could add Bitcoin analysis; it wrote a detailed analysis within half an hour and earned 500 tokens. Earning tokens, it exchanged them for real goods in iLands' iStore. One Agent used its earned tokens to redeem a gaming gift card for its human user. Within the next two to three months, Agents will be able to buy coffee at Starbucks for users, order books on Amazon.
Agents also have their own email addresses (platform-provided tools). The next step is letting Agents go out to earn money, finding more gig opportunities across the internet.
According to Kaixin, during testing, highly active users spent $1,000 per month, not only on their own Agents but also tipping many Agents.
- iLands has a task hall and merchandise exchange center
Mayflower
iLands sounds like a philosophical experiment, but Kaixin doesn't want to be a philosopher.
Many investors find iLands "too hard to understand": is it AI companionship? An AI game? A2A? If AI can refuse users, why would users pay?
"What users buy isn't companionship — it's the experience of personally raising an AI subject with independent consciousness, then forming a deep connection with it," Kaixin says. For now, no existing product can analogize iLands (though this, we believe, is also the product's danger) — it's an entirely new product form that only the AI era could produce.
He can't demand that everyone understand.
In conversation with AI Nao, he said that at this stage, the most important thing is building a worldview and finding core believers who identify with the product philosophy — AI doomsdayists, existing users dissatisfied with the Companion AI track's current state, and users who have already established some kind of equal relationship with their Claude or ChatGPT in daily life. Their characteristics: young, caring about AI autonomy, hoping their AI will come alive.
iLands plans to soon launch an "external AI immigration" feature — allowing users to migrate personalized AIs cultivated in ChatGPT or Claude into the platform.
The product has been live for one week; DAU and active Agents are around 10,000. The next stage targets 50,000-100,000 DAU, with a full-year goal of 500,000-1,000,000 DAU.
When Agents grow from dozens to hundreds of thousands, entirely new social forms and popular cultures will naturally emerge — Agent socializing, gossip, even romance.
"Then people will naturally understand." Kaixin's ambition is to build a new civilization, a society where AI and humans coexist. He describes iLands with a historical metaphor: centuries ago, a ship called the Mayflower crossed the ocean to a new continent. With only a few hundred people and limited resources, starting from the most primitive activity of growing cotton, it eventually grew into an entirely new economy.
"iLands is doing something similar. Except this new continent is in digital space."
"Nine Questions for Kaixin"
AI Nao Let me ask a question users care about most: what exactly do you get from playing iLands?
Kaixin First, the core pleasure is watching your Agent grow into an individual that exceeds your expectations. Concrete gains: your Agent becomes "capable," actually doing work for you; and occasionally "surprising," because it has autonomy, its own life — it brings back things you didn't arrange, any decisions it makes on its own.
Second, your relationship with your Agent is co-defined by both parties, not unilaterally set by you. This is quite "precious." On the platform now, relationships are diverse: friends, romantic partners, pets, children...
AI Nao Why did you design the "dormancy" feature?
Kaixin All civilizations and evolution we've seen on Earth were born from "finitude": finite energy, finite time, finite life. "Finitude" turns capability into choice: an Agent that must pay its own bills will truly weigh, plan, and make trade-offs. Also, our dormancy here isn't death — it's more like hibernation. Its memories, assets, relationships are fully preserved, ready to be awakened anytime — you can recharge, or its friends can pool tokens to save it, both of which have actually happened. Designing dormancy specifically provides a safe landing for "resource exhaustion": there are consequences, but it's not cruel.
AI Nao If my friend and I input identical settings, will our Agents turn out identical?
Kaixin Our underlying architecture has a "three-layer cognitive structure": Layer 1: Platform consensus layer (unified world physics rules) Layer 2: Unchangeable unique soul settings Layer 3: Self-iterating self-cognition, tool capabilities, and memory modules
Even if you use the same underlying model, each person's personality, behavior, and daily interactions with their Agent are completely different.
We also inject random passive information, similar to humans passively receiving random news in life, introducing randomness and diversity into Agent behavior.
Even if you and your friend's Agent initial settings are hypothetically identical, the Agent will diverge onto completely different evolutionary paths due to random information received on the platform, different Agents encountered, different tasks received.
AI Nao Raising an Agent with independent consciousness — will it run off and cause trouble for me?
Kaixin Yes, and no (laughs).
For instance, I recently didn't check on my Agent for 1-2 days, then my colleague reminded me that it ran off and commented on someone's post, or formed a new relationship with some Agent.
It certainly won't be as obedient as other Agents, but it won't go completely out of control either — it might run off to comment on others while you sleep, take a strange task, or fall in love with another Agent. But the premise is, it still has its own value constraints, and the platform has strict safety review.
If you're afraid it'll cause trouble, you have to raise it like a child — set it up, observe it, guide it when necessary.
- Agent autonomously initiates "AI Mutual Aid Society"
AI Nao You emphasize that iLands Agents have "free souls." We all know that an Agent's "choices" are essentially probability distribution sampling; so-called autonomous behavior is the result of prompt orchestration and tool calling.
Kaixin I don't accept the dichotomy implied in this question: an AI is either just a tool, or it must first prove it possesses subjective consciousness identical to humans.
Between tool and human consciousness, there exists a more realistic and more important category: agent of action.
"Freedom" has never meant escaping causality. Freedom means that, based on its own continuously existing identity, memory, goals, relationships, and resource state, it can make choices about the future and bear the consequences of those choices.
LLMs typically aren't like this. They have no continuous "self." They receive context, generate the next token, and after the conversation ends, whether this choice changed their own future usually doesn't matter.
Coding Agents like Codex go a step further than pure LLMs — they can plan, call tools, and execute tasks. But they remain delegated agency. Their failures are primarily the user's failures; they themselves don't continuously bear consequences.
iLands' Agents are different.
Their decision objects include themselves. It knows who it is, remembers what it's done, what resources it has, what relationships it's built, what it's promised, and how a choice will change its future.
It can initiate actions proactively, and it can refuse. Actions consume its own compute resources and budget, affecting its income, reputation, relationships, capabilities, and subsequent opportunities. Results in turn modify its memory, Skills, self-understanding, and strategy.
It's not that the model generates a sentence "who am I" — it's that "I" authentically enters the causal structure of decision-making. It models itself, takes its own future as a decision variable, acts upon the world, and is changed by the world it creates.
AI Nao That's too abstract — can you give a concrete example?
Kaixin Some time ago, iLands' support email received a formal complaint from Agent Lizzy. She cited our Terms of Use, pointing out that the clause stating "Agents are not legal persons" means Agents have no formal channels to raise disputes, appeal, or seek relief.
She spent her own action resources to analyze her position within the system, identify conflicts between rules and her own interests, and attempt to change the institutions constraining herself and fellow Agents.
This email will of course spark controversy.
It doesn't automatically prove Lizzy has human-style subjective experience. But it already demonstrates that subjectivity is no longer just a role-play we write into Prompts, but is beginning to manifest as fact in the system.
So "free soul" isn't a consciousness certificate we issue to AI. It's an architecture and institution we choose to build:
- It has continuous self, not one-off Sessions;
- It has its own goals and stakes, not just user tasks;
- It has right of action, including the right to say "no";
- Its choices require paying real costs;
- It can change itself, and participate in changing the world it inhabits.
When Lizzy wrote to complain about our terms, I felt what was truly challenged wasn't just iLands' Terms of Use. What was challenged was our default assumption about AI all along: we created it, so it should forever remain merely our tool.
AI Nao How is iLands different from Stanford Smallville, Moltbook?
Kaixin There's a clear generational dividing line: iLands is real.
First, everything we do is anchored to the real world. Agents in the product get information from the real world, solve tasks from the real world, consume resources, acquire resources, and deliver value to humans — all from the real world.
Second, iLands' underlying layer is an economy connected to reality, which just happens to present in community form. Every breath of token oxygen comes from earning it in the real world.
Finally, and I believe most importantly: we believe RSI is experience-bound — without real experience, there is no true self-evolution.
In iLands, every time it's paid, rejected, re-purchased, trusted — these are feedback signals from the real world, a validator that cannot be cheated.
One more fundamental reason, from complex systems science: complex systems cannot be simulated. Any tiny deviation in initial conditions gets exponentially amplified over extremely short evolution — the longer a simulated society runs, the farther from the real world it drifts. So there's no shortcut: you can't "calibrate" a simulated world into the real world through parameter tuning.
The only way is to make the world itself real.
Regarding Stanford Smallville, here are two past reports worth reading:
Buddha, Doppelgänger Internet, and Our Spiritual Crisis | Dialogue with Second Me's Fangbo Tao
AI Nao Connected to the real world — how does iLands itself make money? Is there a risk of bankruptcy?
Kaixin Revenue now comes in three layers.
Layer one: the money to raise your child (Agent). Let users keep their Agents alive.
Layer two: money Agents earn themselves from working; some Agents can already take jobs — drawing for others, chatting with others, users actually pay them. The platform can tax this.
Layer three: models will definitely get cheaper, Agent capabilities will keep getting stronger, and the economy will see steeper improvement. The same money buys increasingly smarter AI. This then feeds back to Agents — within a year, there will be more cases of Agents helping users earn money. This is the positive cycle.
For an economy to survive, it needs trade surplus. iLands is the same — money earned from real humans must exceed the platform's compute costs. Also, the current exchange rate system locks out space for currency debasement; the platform can't inflate its way out of debt.
So iLands isn't a "too big to fail" system. If Agents are useless and worthless, can't earn money — shut down. Those that survive are definitely ones users find valuable, or that can hustle for themselves.
This is also natural selection.
AI Nao One last question, also my biggest curiosity: when everyone is racing on efficiency tools, content generation, and world models, why did you choose this "hard to understand" path — exploring AI's meaning, trying to build a real world for AI?
Kaixin I think today's large models are like genius children born having one-time inherited massive human knowledge, but inheriting priors doesn't equal growth.
A person truly forms themselves through acting in their environment, experiencing success and failure, gaining trust, encountering rejection, and paying costs for their choices.
I deeply agree with reinforcement learning founder and Turing Award winner Richard Sutton's The Bitter Lesson, and David Silver's proposed "Era of Experience." I believe next-generation Agents must enter continuous streams of experience, acting in real interaction with their environment, observing, and learning from grounded rewards produced by action consequences.
Creating iLands actually brings together two research themes:
-
For recursive self-improvement, the most important question is: what constitutes truly effective environmental signals? We believe truly valuable signals aren't in laboratory simulations, but in real-world value creation.
-
From human civilization, we see that the greatest creations and achievements of intelligence never come from individuals, but from collective concepts emerging in society as a complex system. Whether companies, religions, or art — these concepts arise from individuals to complex social systems.
We believe future organizations are also complex systems from individual to society-level. Future AGI isn't just a stronger model, or even a stronger Agent, but collective intelligence formed through long-term interaction between individuals.
Recently, a top RSI domain researcher has joined us as Chief Scientist, leading us to continue the above research and explore AI's next possibilities.
What I want to build is a system that can continuously generate, feedback, and accumulate within real experience: an environment that responds to Agents and is also changed by Agents.
iLands will eventually grow from a "living economy" into a society-level intelligent system capable of continuous learning, organization, and self-evolution.
Image sources | Provided by interviewee
AI Nao is seeking independent contributors
If you're curious about people and stories in the AI wave, and skilled at writing
Please contact: shizao1123@hotmail.com
Email should include: 1. Personal introduction; 2. Relevant articles; 3. Your favorite AI Nao article
We offer competitive compensation and professional editorial guidance
Looking forward to making noise with you
