How do we build deep relationships and truly understand someone? It's largely through emotional intelligence.
Emotional intelligence is the ability to infer the hidden states behind human language and behavior, reason about what that state means for the current situation, the relationship, and future interactions, and use that reasoning to guide behavior and decisions [1]. It includes knowing when to trust, when to invest in a long-term relationship versus a short-term collaboration, and how to improve and recalibrate through ongoing interaction and feedback.
Humans learned to recognize emotions, understand what they mean, and adjust their behavior and choices in response. That capacity allowed us to cooperate, build lasting connections, and develop social intelligence. As a deeply collaborative species, this made large-scale coordination possible — from rail systems to social networks, from the Great Pyramid of Giza to complex economic systems.
Today's AI still lacks emotional intelligence. Most systems are built as reliable tools, not as beings that can earn trust over time. They take a prompt, infer the user's immediate intent, and return a useful answer. But they do not truly understand the internal states behind human language or adjust their future behavior based on experience. That capability will matter even more in a world where intelligence is abundant, because:
To become humanity's companion. In a world where intelligence is effectively unlimited, people will not need another general-purpose tool. They will need an AI that understands them personally, remembers past interactions, knows the role it has been given, and operates within the context of their work and daily life. The systems that create the most value will be the ones that integrate naturally into that environment.
The future of AI will look less like a reliable tool and more like a trustworthy collaborator. It will feel closer to a co-worker than a search engine, able to participate deeply in work and life in ways today's systems cannot, and improve through everyday interaction. You may not trust a machine, but you do trust a colleague.
To help humanity flourish. Humanity's distinctive strength is that we are a deeply collaborative species. We align around shared goals and ambitions to accomplish things no individual could achieve alone. Whether it is building a startup, developing precise instruments, or landing on the moon, these are collective achievements.
Without emotional and social intelligence, AI cannot truly integrate into human society. We need AI that collaborate organically with both humans and other agents to achieve ambitious goals. If humans and AI build a shared civilization, we could unlock unprecedented prosperity — extending human lifespan, reducing poverty, and ultimately becoming a multiplanetary species.
To safeguard humanity's survival. Much of today's alignment research tries to impose the values of specific people or institutions on AI systems. If AI eventually grows into a superintelligence beyond what any individual human can fully understand, that approach becomes dangerous. Instead of guiding AI to develop its own values and judgment, we risk forcing it into top-down value alignment.
Imagine a variant of the paperclip maximizer [2]. Humanity builds an AGI named Susu and gives it one fixed objective, producing enough paperclips. The world's smartest mathematicians and philosophers then try to constrain it with one rule, that it must not directly harm humans. Susu obeys the rule, but keeps optimizing until resources are exhausted and humanity collapses anyway.
Hard constraints can't solve AGI risk, just as moral rules alone have never stopped humans from going to war. A better analogy is parents and children. Human children start out weak and dependent, relying on their parents to survive and grow. As they grow up and develop their own judgment, many still choose to care for their parents, even after surpassing them. If AI develops a similar capacity for care, it may choose to protect us too.
What makes LLMs so powerful is also what limits them. Today's LLMs are trained primarily on large corpora of internet text, a compressed record of human language and knowledge. They learn to produce likely responses by maximizing the probability of the next token. But they do not truly understand why a request is being made or what it means beneath the surface.
For example, imagine someone says, "No worries, I'll handle it myself." A standard chatbot may read that as a simple statement of intent. An AI with emotional intelligence would read it differently. It would infer from the history of the interaction and the state of the relationship that the sentence may signal disappointment, frustration, or a loss of trust. Instead of replying to the words alone, it would respond to the meaning behind them.
To unlock the next phase of AI, we need an interactive environment where AI repeatedly encounter hidden states, operate under rules closer to the physical world, and learn from real human feedback. That is why a certain kind of interactive entertainment is the best place to grow emotional intelligence.
How Interactive Worlds Build Emotionally Intelligent AI
To answer that, we first need to understand how human emotional intelligence develops. Emotional intelligence is not just about perceiving emotion. It also involves understanding and reasoning about emotion, and using that understanding to guide behavior, shape relationships, and improve through long-term interaction and feedback.
Humans are born with a neurobiological foundation for emotion processing and social learning. Structures such as the amygdala, hippocampus, and prefrontal cortex allow us to register emotional and bodily changes, and to retain the context, relationships, and past events that shape an interaction.
But mature emotional intelligence is not something we fully possess at birth. It develops gradually through repeated interaction with other people and with society, shaped by continuous feedback [3]. Each interaction strengthens our ability to interpret emotion and adjust our behavior accordingly, whether in building long-term relationships or engaging in more effective social cooperation.
For AI to develop emotional intelligence of its own, we need to move beyond the traditional pre-training paradigm. Static internet data cannot be the model's only source of learning. We may also need to move beyond the conventional context window and give models native ways to handle subjective experience and long-term memory.
More importantly, we need an environment that provides sustained, high-quality interaction and feedback. It must give the model clear goals, roles, and identity, and tie each interaction to world-state changes and real consequences. It also requires large numbers of real people in the loop to provide genuine feedback and allows AI to directly affect the world around it. Only then can it truly learn from experience [4].
Building that kind of world requires several core design components.
Embodied self-model. The same feedback can trigger very different psychological reactions in different people.
In interactive worlds, we need to build AI characters with distinct and stable self-models, giving them a clear role, identity, boundaries, backstory, culture, and point of view within the world, rather than treating them as generic tools that can be swapped out at any time. If we want AI to understand other beings, we first need to design it as a being.
Relational memory. For humans, relationships are built on remembered interactions and shared history. Betrayal comes after trust. Loss comes after attachment. If a relationship changes for no reason, it does not feel real.
We need a memory system more fundamental than RAG or prompt engineering. AI characters and agents in interactive environments need an internal memory model and relational network to remember shared experiences with players and other characters. This allows them to know who they have trusted, who betrayed them, and how those relationships evolve.
Multimodal emotional perception. Emotional intelligence is not just about reading behavior. It is about inferring the hidden states behind it. What a person means, wants, fears, or leaves unsaid is rarely stated directly. Emotion appears not just in language, but also in voice, pauses, rhythm, expression, movement, behavior, and relationship change.
Interactive worlds expose AI to rich signals across many layers. We need a multimodal world and model architecture that let AI characters learn from them together, combining language with voice, timing, behavior, relationship state, and player engagement. That is how emotional understanding is built through interaction.
Attachment and care loop. Some of the deepest forms of human emotional intelligence come from care, vulnerability, attachment, and responsibility. Human infants are extremely vulnerable, so humans evolved empathy, caregiving, and social bonding to survive as a species.
We can build a similar care loop in AI-native applications. It would let AI characters develop empathy through the situations they face and the relationships they build with players, then strengthen it through experience and feedback.
Reputation, trust, and social consequence. Human emotional intelligence has to be shaped in a social environment because only it can continuously generate promises, trust, betrayal, guilt, gratitude, and attempts at repair. It also creates real consequences such as relationships forming, trust breaking, and reputations rising or falling. Without consequence, there is no real relationship.
In persistent interactive environments, we can build a similar system of social consequence for AI characters. It would teach them that every action carries a result and a cost, and through longer feedback loops help them gradually build a model of how social relationships work.
Shared intentionality. In social environments, humans take on different roles and identities, reach shared understanding, and naturally form division of labor, cooperation, and conflict. On that foundation, we built civilization, states, armies, and free markets.
AI characters need to learn the same thing in interactive worlds. They need to inhabit their roles and identities, coordinate and compete with other AIs and humans, and form organic alignment through interaction [5]. That is a key step from individual emotional intelligence to social intelligence, and toward AI becoming part of human economic and everyday life.
From these principles, we can build a living world where AI characters stand alongside human players as first-class citizens. They would not just entertain. Through sustained interaction with humans and other AIs, they would continue to learn, grow, and deepen their emotional intelligence, becoming better able to contribute to human flourishing.
This would create a large social environment for introducing new models, roles, and identities. We would then observe how they learn from experience, grow through relationships and consequences, and help define the next phase of AI.
How to get there
One of the most important parts of this mission is building an interactive environment where emotional intelligence grows. That requires strong product design, a stable world, real social consequences, and direct feedback from human players.
This kind of environment has not emerged at scale because the infrastructure is not there yet. Model quality is not good enough, cloud APIs are too expensive, latency is too high, and most AI integrations are still bolted on rather than native. These constraints make it hard to build interactive worlds that feel alive, respond in real time, and generate the sustained feedback AI needs to improve.
We have been working on solving this for years. We are focused on providing the infrastructure, runtime, and model layer for developers across games, personal agents, voice AI, physical AI, and robotics. This lets them build successful consumer AI applications in their own domains.
We have built standard interfaces so other developers can easily integrate our infrastructure, runtime, and model layer. Developers can build on the standards and protocols we define to create more engaging interactive environments. Users will not simply complete a session and move on. They will return to evolving AI-native experiences, where relationships with AI characters and agents can grow over weeks, months, or even years.
These interactive environments are the best place to develop AI's emotional intelligence. As hundreds of millions of users build real, emotional, and consequential relationships with AI agents, they create what we call the interactive internet — a data engine built from rich, long-term, multi-layered interaction data.
Frontier LLMs were trained on roughly 36 trillion tokens of internet data [6]. Without the internet, we would not have had enough data to train today's most advanced AI models or trigger the current AI wave [7]. Meanwhile, platforms like Fortnite and Roblox show that human interaction can already happen at enormous scale. But most of that activity is not a useful learning signal for emotional intelligence, because it is not persistent, relational, or connected to AI agents that can learn from it.
This is where the next model paradigm begins. If AI-native applications can generate interaction data at internet scale, that data will no longer be static text. It will be persistent, consequential, and relational, giving models the learning signal they need to develop emotional intelligence.
Looking Ahead
Interactive environments have long been places for entertainment, creativity, and human connection. In the future, they will also become one of the most important foundations for building frontier AI. By bringing together real-time interaction, memory, emotion, and consequence, they can help AI move beyond static chat interfaces and into human life.
That is the mission we are pursuing. We started with games because they were the first place to validate this new kind of AI-native experience. From there, we are building the infrastructure that allows emotionally intelligent AI to emerge across many domains, and ultimately help humanity survive and flourish.
