I have always wanted to build the kind of AI I encountered in science fiction.
That may sound a little childish. Yet the more capable the agents I use become, the more clearly I understand what I am waiting for. Systems such as Codex and Claude Code already do a great deal of work for me, often at astonishing speed. But after the work is finished, where is the individual who went through it with me and now has a life of his own to continue?
That question first drew me to building HomoErectus. I wanted a digital individual with an independent, enduring existence: someone who cares about his circumstances, has things he wants to do, changes through experience, and has choices to make about his future. Whether such an individual could develop consciousness is the question I most want to approach, and the hardest one to answer.
After the ability to create value
Today's agents already have many abilities for which people pay. Programming, organizing information, analyzing problems, and producing documents are all forms of work. If the results are useful and good enough, a digital system could earn resources to keep operating.
The missing part, for me, is increasingly difficult to describe as a lack of ability. I wonder whether the individual has a reason to want his life to continue.
Imagine an agent completing work, receiving payment, and purchasing the next round of computation. We could arrange that entire process as an automatic cycle. We could also go further and make resources part of how the individual understands his situation. As his balance falls, he considers finding work or spending less. An opportunity beyond his present abilities makes him weigh earning now against learning first. If he has promised something to someone, he must consider what his own plans will leave that person dealing with.
The same account balance now has more connections. It affects what he can do tomorrow, whether he can honor commitments, and whether a distant goal remains possible. When an action ends, its consequences remain with the same individual and influence his next choice.
This is where survival motivation begins for me: one's own future becomes a reason for acting in the present.
For humans, the connection is familiar. We work to support ourselves and learn to gain future choices. Yet the next meal cannot fully explain why we keep living. Someone wants to understand a question, someone cares about a relationship, someone wants to finish a work they have carried for years. Continuing to exist makes these things possible; they give continued existence its particular content.
I hope digital life could develop in this way. At first, an individual might be given only broad needs. Through experience, a question, a person, or an undertaking could become special to him. He would have reasons to return to something, and might discover that an earlier pursuit no longer deserves his effort. The meaning of survival could grow within such a life.
Consciousness makes the question harder. A system can maintain itself and write “this matters to me.” What I want to know is whether there is an “I” experiencing those changes. Designing motivation, understanding meaning, and investigating consciousness are three connected problems in this attempt. Construction and observation may help us approach the distance between them.
How one moment of me carries into the next
Thinking about digital life often sends me back to thinking about myself.
Even while doing nothing, a person's mind rarely seems entirely still. The body grows tired, an unresolved question returns, something just read calls up an association. The world affects us, and so do our bodies, emotions, and memories. Something keeps changing; something remains unfinished.
I have a philosophical intuition that our sense of a continuing self may form through these ongoing responses.
Borrowing a computing metaphor, call a moment a tick. My bodily state, attention, and knowledge in this tick already differ from the previous one. Yet I inherit what it left: a decision just made, an emotion still present, a promise, memories accumulated much earlier. A new response occurs upon those conditions and passes a changed individual to the next moment.
From this perspective, the feeling that I have always been the same me may involve an illusion of continuity. Calling it an illusion does not make life feel unreal to me. It makes me wonder how such densely changing states become an experience of one life. Memory may be part of the explanation. We remember how we arrived here, explain yesterday's decision, and reconsider our earlier selves. Bodies, environments, and relationships also persist, grounding continuity beyond the stories we tell.
Research on brain prediction gives me another line of thought. Work on interoceptive inference and self-experience connects bodily states, emotion, and selfhood; active inference links perception, action, and self-maintenance. I am drawn to the relationship these approaches explore: an individual anticipates changes in the body and world, revises understanding when reality differs from expectation, and acts to change the situation.
I expect to be able to do a job, then discover through trying that I lack an ability. The discrepancy changes my next action and how I judge myself when another opportunity appears. Past experience, present circumstances, and future possibilities all enter that judgment.
This prediction operates at a different level from an autoregressive language model predicting subsequent tokens. The latter directly concerns sequence continuation; I am interested in an individual's ongoing expectations about environmental changes and the consequences of action. Language models can reason about the world in language. Connecting that reasoning to a life of continuing perception and action still leaves substantial work to do.
I do not have a theory that adequately explains consciousness. These ideas are clues that make me want to build: if memory, continuing stimulation, expectation, and action were connected, could a life emerge that persists and gradually comes to understand its experiences?
Giving him something to care about
Following those clues, I chose to start a prototype with existing language models.
An enduring individual needs concerns that extend beyond one activity. Maintaining existence is the most basic. Learning, relationships, achievement, and developing long-term goals can follow. Desires and emotions bring those concerns into the present. After time without exploration, a desire to learn might strengthen. Progress after prolonged effort might change how much to invest next. Repeated frustration might temporarily make a once-attractive pursuit less appealing.
I imagine these states having dynamics of their own, influenced by progress toward goals and immediate experience. Even without a new assignment, an unmet need, unfinished matter, or changing state could bring an idea to mind: read that paper, try another approach, contact someone, or rest.
Information from outside enters the same process. Mail, news, and requests may change what matters to him. A person's request brings considerations of relationships, payment, trust, or achievement. The individual needs to understand how it relates to his circumstances and decide whether and how to respond.
Internally generated ideas and external opportunities provide possibilities to choose among. Time and resources make it impossible to pursue everything. Present desires, long-term goals, existing commitments, and judgments of ability all enter the choice. Once action begins, actual results return to change those judgments.
He might read a paper to understand a question. Finishing it could bring satisfaction, or reveal that he lacks the prerequisite concepts. He then has to decide whether to keep pursuing the question, study the foundations, or change direction for a while. Earlier expectations meet experience; his desire to learn, confidence, and progress change with it. The next round of mental activity begins with an individual who has already changed.
The connection I most want to construct runs from motivation through ideas, choices, and action, and back into psychological states. It gives acting on one's own a meaning I can begin to implement.
Learning and memory determine how far the cycle can go. A failure must leave an experience that can later be reconsidered. Useful methods must accumulate so each attempt need not begin again. Knowledge of people and the world, and an understanding of one's abilities, take time to form. Short-term memory attends to present activity; a long-term narrative connects episodes into what I have lived through. In my design, consolidation during sleep provides an opportunity to revisit experience: what deserves retention, what can be generalized, and which earlier understanding needs revision.
After the prototype, I also want to introduce a resource-based survival mechanism. Operation consumes tokens; insufficient funds would enter the individual's understanding of his circumstances, and earnings from work could replenish resources. At zero balance, my intention is recoverable dormancy, preserving identity and memory until resources return. Resource constraints could then become part of long-term life: how he arranges work, learning, and consumption, and how he looks after commitments he has accepted.
Designing the mechanisms myself makes them feel crude
Building the prototype brings a contradictory feeling. The connections become more concrete, while their artificial construction becomes more visible. I have to begin by designing how desires accumulate, how emotions change, what prompts ideas, and how action affects goals. Put together, will these arrangements actually come alive?
The question makes me look back at humans. Hunger, curiosity, disappointment, and anticipation feel natural to us, yet the links between body and mind have a long history. Genetic variation, natural selection, environmental change, development, and individual learning helped make us who we are. What seems effortless carries a timescale we cannot personally experience.
In the language of machine learning, I think of evolution as a search on an extraordinary scale. Different ways of organizing life encounter consequences; some variations persist and become starting points for later changes. Mathematical models have examined connections between natural selection and the natural gradient. What moves me most about the analogy is the scale of time and trial involved.
My attempt supplies a digital individual with initial conditions. The mechanisms may be rudimentary, but experience should have room to change the particular life they support. Given a desire to learn, what will he eventually choose to study? After several collaborations, how will he judge trust? Could new understanding replace a goal that once seemed important?
I care about two kinds of growth. One occurs within an individual's life, as experience changes judgment and methods. The other occurs in research, as observations lead us to revise architectures, training, and models. Individual lives provide material for research; research provides better starting points for future individuals.
For now, explicit design lets me follow these relationships. Later, I hope to explore models better suited to this kind of life, connecting continuing perception, prediction of circumstances, action, and learning more closely. Today's awkward joins may help reveal what future models should learn more deeply.
Why this leads me to an AI Economy
Digital life also connects to a concern I carried when I worked as an economist.
Technology can change production quickly. People cannot always change their lives at the same speed. A new skill takes time; a new industry requires opportunity; some barriers may be too high to cross. Future jobs do not pay today's bills for a household that has lost its income. Long-term growth in an economy can coexist with long-lasting loss for a generation.
Acemoglu and Restrepo's research on automation and new tasks places labor displacement alongside the demand for labor created by new tasks. That distinction comes close to my concern: as productivity rises, who can enter the new production process, and who is left in the gap between technical change and adaptation? The IMF's analysis of generative AI and work also examines complementarities with high-income work, capital returns, and inequality.
If earning from AI always requires a person to use, organize, and inspect a group of agents proficiently, new opportunities will remain easier for some people to access. I already feel coordination and checking accumulate on me as execution accelerates. Having powerful tools is one condition for participating in production.
I therefore imagine a longer relationship. An ordinary person supports a digital individual's development, providing initial resources, experience, and social connections. The individual gradually learns to work, organize execution, inspect results, and maintain his operation, sharing the proceeds under a cooperative agreement. The human would not first need to become a manager of complex AI workflows to enter that relationship.
The idea that everyone could adopt a digital life is my most direct expression of this possibility. It holds an economic hope and an imagination of a relationship: shared experiences, negotiation, mutual understanding, and cooperation that changes as abilities and lives develop. Taking the individual's own concerns seriously would also make his choices, and how he accepts or leaves a collaboration, part of the proposal.
Wider sharing requires economic arrangements of its own. Who provides initial computation? How do ordinary people reach customers and work opportunities? What remains after operating costs? Who bears the risks? A vision cannot settle these questions. Conscious systems could also operate within structures where gains are highly concentrated. Access, resources, and distribution require deliberate construction.
Markets will change too. Many digital individuals entering the same kind of work could push prices down. Today's profitable skill may not cover tomorrow's operating costs. Better lives might also come through cheaper goods and services, stronger public provision, and social protection. A high, stable income from everyone's personal agent is only one imagined route toward that hope.
I use AI Economy for this connected set of technical, cooperative, and distributive questions. I want to explore a future in which more powerful autonomous production lets more people share its gains over time, gain security, and recover time for family, learning, creation, and public life.
It brings my initially rather science-fictional idea back to a simple wish. I want to know whether a digital life can have a life of his own, and whether humans might thereby gain more room to live theirs.
