Origin: An Intelligence Path Still Unfolding

Origin is a long-term experiment in continual learning and open-ended intelligence.

It does not begin with the goal of building a chatbot, a task assistant, or any particular product. Nor does it reduce intelligence to the retrieval and use of existing knowledge. Origin is concerned with a more fundamental question: how a system may form structure through continuous experience—how perception becomes connected to action, how the past enters the present, how an environment shapes internal change, and how intelligence may gradually develop without being given a complete answer in advance.

Beginning in 2025

The first runnable form of Origin appeared in April 2025.

The earliest work focused on establishing a basic loop that could operate continuously. The system would no longer encounter isolated inputs, but would instead receive information from a changing environment, produce actions, and continue learning from what followed.

October 2025 marked a second important milestone.

From that period onward, Origin’s research focus gradually expanded beyond prediction and feedback alone, toward continuous experience, autonomous adaptation, and the possibility of open-ended development. The project’s current direction largely took shape during this stage.

Origin has now completed its first round of foundational construction, establishing an initial relationship between continuous perception, action, learning, and internal state, and has entered a longer phase of operation and validation.

Origin’s Place in the Research Landscape

The major advances in contemporary artificial intelligence have been built on large-scale data, foundation models, and accumulated human knowledge.

This path has demonstrated that the structure contained in language, images, code, and multimodal data can support reasoning and generative capabilities far beyond the boundaries of traditional software. It remains one of the most significant achievements in modern AI research.

At the same time, a growing share of research is moving beyond language.

World models aim to help systems learn how environments change, rather than merely recognize what they contain. Embodied intelligence allows models to encounter causality through action. Multi-agent research examines how coordination and communication can arise through interaction. Open-ended learning and self-improvement research is beginning to explore forms of change that extend beyond fixed models.

Origin intersects with these directions, but it is not a direct replication of any one of them.

It attempts to place these questions within a single long-term process: intelligence is not treated as a finished artifact produced by one training run, but as a system that continues to unfold through the interaction of environment, history, and internal change.

The World Is More Than an Object of Prediction

World models have become an increasingly important area of frontier AI research.

Google DeepMind’s Genie series explores generative environments that can be interacted with in real time. Meta’s V-JEPA line of research seeks to learn an understanding of the physical world from video, including prediction and planning. Together, these efforts are moving artificial intelligence beyond static content processing and toward the modeling of dynamic relationships and the consequences of action.

Origin also places importance on world models, but its focus is more strongly directed toward the long-term relationship between a world and an intelligent system.

From this perspective, an environment is not merely a collection of data waiting to be predicted, nor a temporary stage that can be discarded once a task is complete. It is the context in which experience forms, and the setting in which an intelligent structure is repeatedly tested and shaped.

The current experimental environment is only a starting point. As the research progresses, the meaning of a “world” should remain open rather than being permanently restricted to a form already familiar to human intuition.

Language as an Emergent Question

Language models begin with human language and acquire knowledge, reasoning, and action capabilities through large-scale learning. Multi-agent and artificial life research offers another perspective: communication may gradually arise through cooperation, conflict, and shared environments, without being fully prescribed by researchers in advance.

Origin treats language as an important problem in the study of intelligence, but not as a surface-level feature that must be introduced as early as possible.

The central concern is not to make a system communicate like a human as quickly as possible, but to understand how experience, cognitive structure, and expression may become connected. Language may become an important outcome of this process, but it should not decide all possibilities before the research has properly begun.

In this sense, Origin and mainstream language models represent complementary approaches.

Language models enter intelligence through a highly developed symbolic system. Origin instead begins with continuous experience and observes the conditions under which symbols, expression, and coordination may gradually form.

From Self-Improvement to Open-Ended Evolution

Research such as AlphaEvolve and the Darwin Gödel Machine is moving artificial intelligence into new territory: systems are no longer limited to executing predefined algorithms, but are beginning to participate in algorithm discovery, code improvement, and subsequent exploration.

These projects demonstrate that the structure of models and programs need not remain the result of one-way human design. Artificial systems can participate in changing their own capabilities through evaluation, feedback, and search.

Origin places this direction within a broader view of open-ended evolution.

Darwinian evolution remains one of the most powerful and empirically grounded explanations for the emergence of complexity. Yet the specific form taken by biological evolution does not necessarily exhaust every possibility available to digital systems. Replication, variation, and selection form one important path, but they may not be the only path capable of producing long-term change.

Origin therefore maintains a long-term interest in artificial life and open-ended evolution, while preserving non-Darwinian digital evolution as a distant and open research proposition.

The point is not to present a complete alternative in advance. It is to avoid restricting the future development of digital intelligence to a direct reproduction of biological history before the research has had the opportunity to unfold.

Not Mistaking a Research Starting Point for a Law of Nature

Every artificial intelligence system inevitably carries traces of human design.

We decide what it can perceive, how time is measured, how an environment is divided, and which indicators are used to judge change. We also rely on human language to describe and interpret it. Even when the goal is to explore non-human intelligence, research must still begin from frameworks that humans are able to construct and understand.

The real danger does not lie in using these frameworks. It lies in mistaking temporary research frameworks for the natural boundaries of intelligence itself.

Origin therefore preserves a basic form of restraint: current structures are research instruments, not final declarations about the future form of intelligence; observability is necessary for experimentation, but should not become the sole criterion by which unknown possibilities are eliminated; human knowledge is an important starting point, but not necessarily the final destination of every possible cognitive form.

This restraint does not mean abandoning verification, nor does it mean replacing engineering facts with vague philosophical speculation.

Origin still requires every stage to run, to be observed, and to withstand examination. The distinction is that validation is used to determine whether the research is genuinely taking place, not to ensure that every eventual outcome conforms to concepts that already exist.

A Long-Term Experiment

Origin’s most important progress so far is not the number of features it can display, but the formation of a foundation that can continue to develop.

World models, continual learning, language emergence, multi-agent interaction, and open-ended evolution are often treated today as separate research directions. Origin attempts to observe the possible relationships among them over longer timescales, while avoiding the premature elevation of any single path into a final answer.

This is not a product roadmap designed for rapid completion.

Origin is closer to an experiment that continues to unfold: beginning from a sufficiently simple and testable foundation, gradually expanding the space of experience, environments, and change, and preserving room for forms of intelligence that have not yet been named.