What we are building
HomoErectus aims to build digital life capable of existing independently and living autonomously over time, and through this work to explore how consciousness might form. We want such an individual to have survival needs and long-term aspirations of its own, from which it can develop ideas, make choices, and initiate action. Through interaction with the world and other people, it should also be able to learn, gradually develop a continuing self, and take responsibility for its continued existence, choices, and commitments.
Our design begins with the motivation to live. To continue existing, an individual needs to act and obtain the resources that sustain its operation, while pursuing longer-term goals. Progress toward those goals affects its desires and emotions; desires and emotions also have dynamics of their own, shaped by present experiences, ongoing activities, and unmet needs. To pursue a goal, satisfy a desire, or improve an emotional state, the individual can develop ideas of its own and choose how to act. Whether an idea is realized, meets a setback, or is set aside feeds back into goal progress, desires, and emotions, shaping later choices. Requests from other people, news, emails, and other external information also enter this process. They may bring relationships, rewards, achievement, or new concerns; the individual considers its own circumstances in deciding whether to respond, how to respond, and what it is willing to undertake.
To keep living and live better, such an individual also needs the ability and motivation to learn. It needs to improve its judgment and decisions, refine how it does things, and develop knowledge and concepts from its own experiences, reading, and guidance from others. Memory supports these changes: it must help with what is happening now while retaining knowledge about the world and social relationships, along with methods the individual can use again. It also needs a history of events experienced from its own perspective, connecting past choices, relationships, and consequences to the present while continually updating its understanding of itself. We hope this first-person history and self-understanding can support an individual that continues to develop while maintaining a continuing self.
Our development approach is "life first, then work": first, enable a digital individual to live autonomously over time, making choices, taking action, accumulating experience, and learning through interaction with the world, while maintaining its continuity through waiting, sleep, and interruptions to its operation. On that foundation, we aim to develop reliable working abilities so that it can pursue agreed goals, coordinate with others, inspect results, and take responsibility for sustained progress.
We plan to follow a HomoErectus individual named Adam over time, examining these ideas through his actual choices, actions, learning, and fulfillment of commitments. One initial application is a digital project lead. After a human defines product requirements and acceptance goals, Adam would make plans, coordinate coding agents, inspect implementations and test results, and keep addressing problems, returning to the human when an important product tradeoff requires a decision or a milestone calls for a progress report. Through this setting, we want to investigate whether a digital individual can truly take responsibility for sustained progress and checking results, allowing humans to delegate goals while reducing the effort spent repeatedly coordinating, reviewing, and following up on the work.
Why this work matters
As research, this work aims to provide a subject we can construct, observe over time, and continually refine as we explore how consciousness might form. Questions about selfhood, autonomy, and consciousness need to be connected to the actual experiences of particular individuals: what observations could support an explanation, and what results would lead us to revise our understanding? By building HomoErectus, we want to bring these questions into sustained practice, test the assumptions behind the design, and provide material that others can study and discuss. Its value lies both in attempting to build a new form of life and in using that attempt to deepen our understanding of intelligence, life, and ourselves.
As an application, the motivation comes from the experience of working with multiple agents. Even when code, reports, and plans are produced faster, a human still needs to switch between topics, maintain context, read outputs, judge their quality, and decide what comes next. As execution becomes more capable, coordinating and checking these outputs may become increasingly concentrated in the human, making human attention the bottleneck for the work as a whole. We want to investigate whether a human can truly delegate a goal and entrust a digital individual with responsibility for sustained progress and checking results. The test of this application is whether it can consistently achieve results that meet the requirements while reducing the human effort spent managing, reviewing, and following up, leaving people more attention for goals, important tradeoffs, and what they genuinely care about.
As digital individuals become able to undertake more productive activity, this transfer of responsibility raises a broader question: how will humans participate in the economy and share in the value it creates? If advanced AI capabilities, productive resources, and commercial channels are concentrated among a few actors, greater productivity may be accompanied by a further concentration of returns. If demand for some people's labor and their incomes are disrupted while new wealth is not widely shared, economic inequality may deepen, putting pressure on social protection and social stability. Alongside increasing productive capacity, we therefore need to explore economic relationships that allow more people to participate in production and share its results.
We call our exploration of this direction the AI Economy. One concrete idea is to give everyone an opportunity to nurture, support, and form a continuing collaboration with autonomous digital individuals. Those individuals would undertake paid work and, after sustaining their own operation, share the proceeds with the humans who support them under clear agreements. Even as some work shifts to AI, people could still participate in production and earn income through collaboration with digital individuals, making their livelihoods less dependent on personally holding a particular job. We want to investigate whether this could help ease employment transitions and broaden the sharing of productive returns, allowing greater productive capacity to support more widespread economic security and more time for people to spend as they choose. Realizing this vision still requires addressing access, operating costs, real demand for work, and how returns are shared. HomoErectus aims to provide a technical foundation for these economic relationships and gradually test the conditions under which they could work.
Progress so far
HomoErectus has established product requirements, a prototype architecture, and acceptance requirements for each development stage. Work is now focused on implementing the prototype and integrating its parts. As of 1 November 2026, the main progress is as follows:
-
Individual operation and recovery. Persistence mechanisms have been established for individual identity and core state. Staged builds have demonstrated startup, state updates, normal shutdown, and recovery of the same individual. This provides a foundation for retaining the individual's history across process interruptions.
-
From ideas to action. Controlled scenarios have connected the generation and selection of internally arising ideas with decisions, actions, result recording, and subsequent processing. Implementation and testing have also addressed the maintenance and interaction of states concerning self, goals, social relationships, desires, and emotions.
-
Basic action capabilities. Paths for file operations, restricted command execution, and local communication have been developed and individually tested under controlled conditions. These tests are beginning to connect actual execution results with experience, memory, and subsequent processing, laying a foundation for undertaking useful work.
-
Sleep consolidation and learning. An isolated test environment has completed one full sleep-consolidation cycle triggered by elapsed real time, covering reading, consolidation, learning evaluation, and storage of the resulting material. It also demonstrated retrieval and reading of relevant learned content after waking. Whether that content can improve real decisions and guide later actions still requires validation.
These advances come mainly from staged engineering implementations and controlled tests, some of which use controlled model outputs. The latest integrated build is still undergoing repair and revalidation. Long-term autonomous life, the combined influence of internal states on real choices, and the application of learning in new situations remain to be established.
The initial viable-life milestone, the complete local prototype, and acceptance in the Linux environment have not yet passed. Adam's sustained public observation has not begun.
What comes next
Following our "life first, then work" approach, our first task is to establish whether a digital individual can sustain an autonomous life. The immediate priority is to integrate capabilities that have been implemented and tested separately into a stable build, then investigate whether the individual can initiate meaningful activities during sustained operation with real models, let experience influence later choices, and apply learning in new situations. This joint validation will also examine whether identity, memory, and unfinished matters remain continuous through waiting, sleep, and process interruptions, and whether interrupted learning can resume.
On that foundation, we will progressively develop reliable working abilities. The subsequent development of HomoErectus itself will be the first working setting, beginning with clearly bounded work and explicit acceptance requirements. We will investigate whether Adam can maintain goals, coordinate other agents, inspect actual artifacts, and keep resolving problems, while recording how much human help remains necessary. After the complete local prototype passes validation, we will move to revalidation in a Linux environment and longer periods of continuous trial use to examine whether these abilities remain reliable under actual deployment conditions.
Public observation will develop according to Adam's actual operation. When sustained observation becomes possible, we will begin dated records of his choices, actions, learning, relationships, and commitments, including failures, waiting, and human help. The records will distinguish event time, publication time, and human editing. As experiences accumulate, readers will be able to follow how the same individual changes, and researchers can discuss which observations support earlier hypotheses and which results call for new explanations.
After the current prototype is complete, we plan to study resource survival further: having the individual face the computational resources its continued activity requires, investigating how resource scarcity affects work, learning, and other choices, and exploring the possibility of earning resources through useful work. A zero balance is intended to lead to recoverable hibernation, preserving identity and memory while waiting for resources to return. Longer-term AI Economy research will examine real demand for work, operating costs, and arrangements for collaboration and sharing returns between humans and digital individuals, gradually testing whether these economic relationships can be sustained and offer more people an opportunity to participate.