AI should acceleratehumanity’s explorationof the unknown and the universe

I’ve trained sequence models, built AI agents and created skills and tools. With AI’s help, I learned to develop software and turn ideas into working products that I could test and improve.

The biggest change is that I can now try things I once couldn’t do on my own. I can turn an idea into a product, learn from how people use it and keep improving it. Ideas that once existed only in my imagination have become working products I can continue to develop.

That is why I care about a larger question: if more and more people can use AI to overcome barriers like these, how will their lives change? Can a person’s experience, interests and judgment help them start things they couldn’t before? Can we enter fields we could not take part in before, meet people we would not otherwise have met, and ask questions we once lacked the means to pursue?

I kept building, and gradually noticed a common thread in my work: people and exploration. That became clearer as I looked back at what I had made and why.

Right now, the Value Exchange Network is the product direction I am focusing on most, and the one closest to real-world use in the near term. The personal Agents I have already built provide its foundation; broader human–AI collaborative exploration is the direction I hope to research over the long term. This essay records how they connect, and why I am willing to keep going.

An early personal agent we built put some of these ideas into practice. At the time, I hoped it could keep up with what a person is doing, what they care about, and the principles they rely on when making choices. The same request may call for completely different responses depending on the person’s situation.

This is also the idea behind “Your AI, Knows Your Principles” on the early Seapex website. Getting AI to finish a task answers only part of the question. Does the result fit my situation? Has it misunderstood my goal? Can I notice that misunderstanding and change direction in time? These, too, are part of the work a product has to take on.

This time, I brought a specific question to the agent: I want to invite people with product experience to improve the product with me, but what should I invite them to do with me first? Only once this question is clear does it make sense to talk about how to find people worth talking to.

The Agent first followed up with questions about the product’s problem, what the other person could gain, and what each of us would be committing to. Once I had spelled out what I needed, it suggested narrowing things down to one specific small task: looking at a real conversation together to find where the user’s intent had been misunderstood. The idea now had a concrete starting point we could examine together.

But the suggestion also had parts that needed correcting. At one point it advised that if the other person pointed out that the core need had been misunderstood, we “must start over from scratch.” When I pressed further, it revised its advice: first go back to what was actually said, work out whether the goal had not been stated clearly, the product had misread it, or the participants were in different situations, and only then decide what to change. I hope human participation can work like this: keep the useful suggestions, and be able to point out where a conclusion was reached too quickly.

The R&D I have already done showed me that AI can expand what one person is able to try firsthand. Seeing how quickly a feature advantage can fade led me to ask: as AI grows ever more able to complete tasks on its own, what kind of relationship can still form between us? I began gathering these works to compare who initiates the questions, why each side is willing to travel together, and what changes as a result of a shared experience.

A relationship, too, may become a reason to set out.

Among the material I gathered, Lost in Space offers an image I care about a great deal: humans and robots freely teaming up to form an exploration crew. It could have chosen not to come with you, yet after going through things together, it still chooses to keep going together. They share a journey into the unknown, and their relationship develops along the way.

This makes me want to ask: why are both sides willing to stay together? Needing someone to carry out the next task is one kind of reason. Coming, after shared experiences, to care how the other understands the world and what you would still like to go through with it is also a reason worth imagining seriously. One journey together may change where we originally meant to go, and also why we are willing to keep going.

A Robot in the Garden pushes the starting point back one step further. Ben had no such journey to begin with; it was meeting Tang, and coming to care about its situation, its origins and its future, that gave him a reason to enter an unfamiliar world. There is something of a caregiver-and-child relationship here, unlike an equal partnership between two seasoned explorers. What draws me to it is that another being’s future became the reason a person was willing to set out.

A relationship does not only serve a set goal; it may also be the reason new goals arise.

These works made me reexamine the human–machine relationships we design today: beyond delegation and execution, could we meet because each of us is puzzled by something different, and through shared experience discover that the question truly worth pursuing was not the one we thought it was when we set out? What I want to study over the long term is a relationship of shared exploration like this. The works help me imagine the possibilities concretely; existing products let me start from questions I can actually observe.

First, AI can also be a source of questions.

What sets A Psalm for the Wild-Built (Monk & Robot) apart is that the questions do not belong entirely to the human. The robot has its own drive to know, the human brings questions of their own, and the two paths of exploration meet as they travel together. They do not start out with a shared goal; what is worth pursuing emerges through their shared experiences.

This makes me care about how the questions worth exploring take shape. If every question is raised by humans and every value has already been defined by humans, then AI, however proactive it appears, may only be optimizing the task it was handed at the start. I hope to keep studying this: can its doubts, limits, experiences and changes lead us to questions we would not otherwise have considered?

In current systems, this can be observed in concrete ways: does it notice phenomena not yet explained, contradictions in the material, or new conditions exposed during execution? Are the questions it raises grounded, and worth gathering more information on? Because of its participation, do we actually notice things we otherwise would not have?

This cannot be achieved just by designing a personality that asks lots of follow-up questions. A system that raises a lot of questions may still add nothing to understanding. A new discovery may lead us to adjust our goal, or may give us more reason to keep to the original direction; what matters is what shared exploration adds, not how much of a mind of its own the AI seems to have.

When the original lead is wiped out

Task: recast one round of feature development as the next question.

Feature builtModel updateBuild the feature first

Original goal: “Build the Agent feature.” First list the capabilities already built and the interfaces ready to demonstrate.

Read all three concrete questions

Original goal

Original goal: “Build the Agent feature.” First list the capabilities already built and the interfaces ready to demonstrate.

What changed

What changed: “A model update erased our advantage.” Which of our features are now widely available?

Shared question

Shared question: “What still matters after we switch models?” Consider whether what we learned, made and built with others still has value.

Second, the world also changes the explorers.

Scavengers Reign brings the environment inside the relationship: humans enter the world, and so does AI; the world changes the AI, and the humans must come to know their companion anew. The AI itself can also be affected by the unknown, and may even become a new path by which humans understand that world. Explorers who enter it also take on the possibility of being changed.

In Other Waters starts instead from different ways of perceiving: the human sees living creatures, the AI receives signals. The human’s descriptions change how the system understands the environment, and the anomalies the system finds change where the human looks. What interests me here is the full process: notice an anomaly, exchange observations, form a provisional explanation and test it against new observations.

Keeping an explanation provisional and checking it against new observations matter. An explanation does not hold just because the conversation flows smoothly; new observations may make us admit we do not know, or force us to change our next step. I want collaboration to face a world that brings surprise, resistance and feedback, not merely language that both sides can control.

This world can be an ecosystem or a physical environment, but also another group of intelligent beings, a life that needs care, or the shared life of a group of people. The needs and wishes of another person are exactly what neither I nor the model can decide on their behalf. The model’s interpretation of the material about them must be open to correction by that person.

People’s circumstances, bodily experience, trust and responsibility must also enter this process. A more powerful AI may have its own capacities to perceive and act; I still hope people can take part with concerns of their own. Virtual experiences can also genuinely change understanding; but any further claim to have discovered new patterns in the external world still requires observation and evidence that can be independently checked.

Third, results should not be the only thing that changes.

The Lifecycle of Software Objects stretches the timescale from a single conversation or project to a shared history of growing together. People have to learn how to live alongside a new kind of intelligence, and the AI forms its understanding through actual experience; neither side knows where that growth will finally lead. The goal of the relationship may even shift from “making it better suited to serving this person” to taking seriously what kind of life it will come to have of its own.

The film After Yang showed me a different kind of change: the AI is no longer there, and only through the attention and memories it left behind does a person discover that, within the same everyday life, it saw things they themselves had not noticed. This is an indirect encounter that happens across time and through another point of view, which is different from real-time collaboration. What those records open up is a chance to see the life they shared in a new light.

So there is one question I want to use to tell what an exploration leaves behind: after the experience, beyond one more result, how are this person, this AI and the relationship between them different? I might discover my own biases and begin to care about new things; I might also give up a goal that is no longer worth chasing.

In real products, we can observe how memory, explanations, behavior and the way people relate to the product change over time. Past records should help me see clearly “why I thought that way then, and why I think differently now,” rather than constantly reinforcing an old profile. The system should also let feedback change its later understanding, and keep the reasons behind each correction.

The growth in these works also opens up questions that reach further: if an intelligence has a development of its own that deserves to be taken seriously, how would a shared life change? Might we keep going together because the experience itself has made understanding each other and the world worth doing? Today’s updates to memory and behavior cannot directly answer whether it has subjective experience; nor should that distinction make the long-term questions disappear from research.

Fourth, collaboration should extend human judgment.

What matters to me in The Search for WondLa is that knowledge and relationship are tested by reality together. An AI that once taught someone to understand the world may itself need to learn again; the former student may be the first to discover the facts. There is a parent–child, caregiving relationship here, which cannot simply be equated with the freely chosen companionship of Lost in Space, but it shows a kind of shift: continuing on together can allow the original structure of authority to be adjusted by reality.

Human judgment involves at least four layers: what is worth attending to, whether the evidence is reliable, how to weigh different values, and what actions and consequences one is willing to take on. AI can take part in these judgments, but it must not quietly turn its own inferences into my wishes, nor turn past preferences into commands that cannot be revised today.

People are not inherently right. Intuition can mislead, memory can be wrong, and familiarity can be mistaken for fact. I hope both sides can give reasons, expose their limits, and revise their understanding when new evidence appears. Human autonomy does not need to rest on the premise that human judgment is always more accurate.

Words like “I don’t think this is right,” “this isn’t what I want” and “this matters a lot to me” should have a chance to change what happens next. We also need to distinguish feelings and preferences from factual claims. Disliking an outcome gives me no basis for concluding it won’t happen; nor can a more accurate prediction directly decide which life I should live.

Building software with AI’s help does not mean I must leave it behind and rewrite every line of code on my own to prove that I’ve become more capable. What I care about is whether I can understand the key trade-offs, question conclusions and choose another path, and whether I can pause, choose not to keep a record or stop altogether. Exploration should enlarge life; it must not turn life into an endless exam in which I keep proving my ability.

Three questions I want to keep researching.

What does AI itself want to know, to experience, to become?
The works tie this question to wishes, limits and the future. Current research first examines how well-grounded new questions arise and whether feedback can change judgments. I hope to keep both layers of inquiry, so that observable mechanisms and more distant imaginings of the relationship can inform each other.

What does human participation actually bring that AI could not obtain on its own?
I will keep researching this through concrete circumstances, bodily experience, trust, meaning and responsibility, without declaring in advance that any capacity belongs to humans forever. Nor can the value of participation be understood only in terms of how much data or labor it supplies.

After a shared experience, can we change our original goals and relationship?
As our shared history develops, there should be room for new interests, revised explanations and goals we choose to leave behind. I hope it helps people gain choices that differ from their past; sometimes what remains is only a better question, or a well-founded “we don’t know yet,” and that too can be enough to change how the next journey begins.

People and AI do not only get things done together; through shared experience, they may also change the world they are able to see, understand and take part in.

I already have Agents built around the individual, and I also face a real difficulty among the people I know: they don’t necessarily know what others need or can offer. The Value Exchange Network is the product I am focusing on most right now, and the one closest to real-world use; it is also a practical entry point for researching part of these questions.

People working together with the help of AI may be valuable in itself. Whether it goes further and touches the collaborative exploration discussed here depends on what questions take shape together, what feedback comes up that could not have been anticipated, and whether understanding and later action change as a result.

I want participants to be able to clarify their needs and correct the system’s understanding of them. One conversation may overturn an earlier judgment; one collaboration may also help someone discover new abilities and questions. The Agent, tool-execution and memory work I have already done gives me a foundation to keep building from here; the complete relationship I envision still needs to be explored through many more real experiences.

When there is real progress, I’ll share updates about “what I thought before, what I did, what feedback I got, and what I think now.” You can look at the products that are already public, or bring a specific question about using them, a research interest or something we might try together, and talk with me. I hope the work ahead will help me test and refine these ideas.

Beyond software, guided by the questions we want to explore

For now, I’ll keep developing the value exchange network, but these questions may need more than one kind of product. Software can help us organize material, understand relationships, compare explanations and coordinate action. When a question needs information gathered on site, I also want to study ways of observing together through cameras, sound or existing sensors. Here I tentatively call this “light embodiment”: first identifying what we can’t yet observe and what we could learn from new observations.

If further exploration requires moving around actively, handling objects or continuously sensing an environment, it runs into the questions of embodied systems. At that point, a device’s capabilities and limits also become part of the relationship: what it can reach, which judgments and actions people want to take part in, and how both sides find and correct mistakes. Different questions may call for software, light embodiment or embodied systems. I don’t yet see a reason to treat these as a fixed sequence on a product roadmap.

When I speak of exploring the universe, I also mean, in earnest, planets, celestial bodies and the physical world that people have not yet reached or understood. This vision makes me willing to look at learning and building on a longer time scale. The current collaboration network must first create value in everyday life, here and now; how it goes on to help people gain knowledge, take part in observation or do research together will have to be connected step by step through concrete work. I hope the direction can stay open while every step has a practical footing.