I Am Uncertain Whether AGI Will Arrive, but I Am Willing to Help Build It

A Seemingly Contradictory Title

I once saw a job posting seeking people from different fields who were willing to help create AGI. One requirement was stated very plainly: believe in AGI.

I am interested in AGI. I am willing to study it, use increasingly capable models, build systems with them, and take part in work that may lead toward stronger intelligence.

But I cannot say that I am certain AGI will be realized.

Those two things are not contradictory.

I can study a drug without first believing that it works; study life in the universe without first believing that it must exist; and help build a technology that is not yet complete without promising that it will certainly reach its intended destination.

For me, AGI is first a factual proposition, not a test of loyalty.

If it is true, it should be established by evidence. If it does not emerge, or if reality takes another form of intelligence, I should not be obliged to preserve the original narrative.

I am willing to be both a builder and an observer, but I will not turn uncertainty into certainty in advance.

“Belief” Blends Several Different Things

When people say “believe in AGI,” they may mean at least five different things.

The first is a definitional judgment: there is a goal that can be clearly called AGI, and we broadly know what it is.

The second is a factual prediction: such a system can emerge, perhaps within a certain time frame.

The third is a research hypothesis: capabilities not yet possessed can be built, and we are willing to approach them step by step through experiments.

The fourth is an action commitment: even when the destination is uncertain, we are willing to invest time, resources, and responsibility in this direction.

The fifth is an identity and loyalty commitment: turning “AGI will inevitably be realized” into a position that members of a community are expected to uphold.

I can accept research hypotheses and conditional action commitments; definitions and factual predictions should continue to wait for evidence. What I reject is the fifth kind.

I have not yet seen an AGI definition strong enough to settle the argument. Does it mean generalization across tasks, long-term autonomous action, continual learning, stable goals, the ability to understand and change itself, or some form of subjectivity close to that of humans? If different people believe in different goals, then “believing in AGI” can easily become a cultural slogan that cannot be tested.

Even if the definition were temporarily clear, whether it can emerge and when it might emerge would still be empirical questions. Researchers can certainly have priors, but priors should not be exempt from revision.

The formulation truly suited to research is not “I believe the conclusion in advance,” but:

I think there is a possibility here worth testing, and I am willing to say what evidence would strengthen, weaken, or lead me to abandon this judgment.

Uncertainty Does Not Mean Denial

This needs to be stated more precisely: I am not saying, “I do not believe AGI will be realized.” That wording can still suggest that I lean toward believing it will not be.

My position is simply that I am uncertain.

I do not have enough evidence to claim that it will inevitably emerge, nor enough to claim that it never will. I am not even sure whether people will ultimately continue to use the term AGI for whatever arrives.

Model capabilities have already genuinely changed the boundaries of my work.

With Claude, I helped build an advertising system that entered production, despite not previously understanding RTA, oCPC, conventional development, or server operations. I also turned my own Treatise on Cold Damage study tool into a mini-program available to the public.

In Minimax, AI helped generate hypotheses, models, and reproducible experiments, and the project also kept records of claims being narrowed and lines of inquiry being stopped. In MRI, AI went further, writing formal proofs and software experiments; under stated definitions and assumptions, a formal system mechanically checks proofs of a limited set of logical propositions. Results at the model and software layers can be reproduced, but the project does not yet have validation on an actual MRI machine, a phantom, or in clinical settings.

None of this needs to wait for AGI to be defined or declared achieved in order to matter.

Nor does it prove that “AGI is already here.” It shows only that current AI has already changed one thing: the knowledge and capacity for execution that one person can draw upon no longer fully equal the professional skills they possessed in advance.

That change is enough to make me participate and enough to keep me interested.

Uncertainty means not advancing the conclusion on credit; it does not mean refusing to observe what has already happened.

“Working Together” Is Not a Claim About Consciousness

If I am unsure whether AI has subjectivity, why do I say that I am willing to work with it?

Because “working together” here first describes a practical division of labor, not a determination about consciousness.

In my projects, I often provide the direction of the problem, real-world goals, consequences I cannot accept, and corrections when the whole effort drifts; the model provides language, candidates, specialized knowledge, code, and parallel execution; formal systems, tests, and real-world operation then constrain the results we produce together.

Human: direction, tradeoffs, responsibility, signals of drift
AI: candidates, elaboration, knowledge retrieval, execution
Mechanisms: formal checks, engineering reproduction, records and stopping

This division of labor is not fixed. Models can propose structures I had not considered, and I may form my own judgment only while repeatedly correcting a model. The final product can be attributed neither simply to me nor simply to the model.

But responsibility cannot therefore be left hanging. As long as I decide to let a product enter the real world, I cannot evade its consequences by saying, “AI did it.”

I do not need to settle whether AI has consciousness, a stable self, or full subjectivity before acknowledging that this collaboration is already happening. Just as a person can form an ongoing working relationship with an organization, an institution, or a complex tool without first judging them all to be subjects equivalent to humans.

“Working together” describes how cognition and action are jointly completed; it does not smuggle in a proof that AGI already exists.

Builders and Observers Must Both Be Present

Being only an observer is safe.

I could stand outside, comment on whether models are improving, analyze each release, and wait for others to bear the mistakes of building. But then it would be difficult to know what happens when a capability enters a concrete system.

Being only a builder is also dangerous.

Once one’s identity, career, and organizational goals are all tied to “AGI will inevitably arrive,” negative results can easily be interpreted as temporary setbacks, boundaries can be treated as reasons for the next round of expansion, and doubt can be seen as a lack of faith.

I want to keep both positions at once.

As a builder, I am willing to turn vague problems into systems, experiments, and propositions that can fail; accept real-world results; and take responsibility for the risks created by using AI to expand the boundaries of my capabilities.

As an observer, I need to record which capabilities genuinely increase and which are only packaging; which candidates are substantiated and which hold only within their own definitions; which failures change direction, and which narratives are written in advance as historical inevitabilities while evidence is still insufficient.

Building keeps me from doubting only from the sidelines; observing keeps me from ceasing to doubt in order to keep building.

If AGI ultimately appears, I hope to have participated in how it was built while retaining the ability to see how it differs from what was imagined. If it does not appear, this work should still leave behind useful products, methods, negative results, and boundaries—not merely an unfulfilled prophecy.

Subjectivity Is Not a Rejection of Constraints

I once said that a person with subjectivity would generally not speak lightly of belief. Looking back, that statement still turned the boundary of my own understanding into a judgment of others.

A more accurate formulation is this: I do not want to turn an unresolved technical and factual question into an identity commitment that cannot be easily revised in advance. Others may of course have religious, ethical, or ultimate beliefs; nor can someone’s subjectivity be judged on the basis that they have a stronger conviction about AGI than I do.

In research, subjectivity as I understand it is not “I do whatever I want,” nor is it a rejection of every external requirement.

On the contrary, I am willing to accept stringent epistemic and outcome constraints: formal propositions must pass through formal systems, engineering propositions must accept reproducible results, production systems must bear real traffic, and failed candidates must stop.

What I find difficult to cede is something else: because an organization has already chosen a narrative, research may produce only results that support it; because a role requires “belief,” doubt must be filtered out before entering research.

Subjectivity here means retaining the right to update hypotheses, and taking responsibility after doing so.

Real constraints make conclusions harder to maintain; faith-based constraints may make a given conclusion harder to abandon. Both can look strict, but their directions are exactly opposite.

Does an Organization That Believes in AGI Believe That AGI Will Change the Organization?

My question about “believing in AGI” as a hiring condition also comes from an organizational tension.

If an organization truly believes that models will gain increasingly strong capabilities in coding, research, and autonomous execution, are its evaluation of people, distribution of responsibility, and research processes changing accordingly?

The job posting I saw said, on the one hand, “no restrictions on field of study” and “do not follow the conventional path.” On the other, it required basic programming ability and a strong educational background, treated national- or world-level competition results as a plus, and finally routed candidates into existing research or engineering teams.

These conditions are not absolutely incompatible, but they raise a question: what does the organization hope to gain from people who cannot be defined by existing disciplines, and how is it prepared to recognize and accommodate that capability?

This need not be simply interpreted as hypocrisy.

Today’s AI still generates incorrect code, complex systems still require engineering capability, and research organizations must remain responsible for safety, maintenance, and consequences. An AI lab made up of many programmers does not directly prove that it does not believe in AGI; it may simply have to use today’s people and tools to build capabilities that have not yet arrived.

So requiring programming ability is not necessarily the greatest contradiction. The more fundamental tension is whether an organization researching unknown intelligence also makes belief in a predetermined conclusion part of the qualifications for membership.

What is truly worth asking is not “why are you still hiring programmers?” but:

As implementation capability gradually shifts from people to models, has the organization begun to identify and cultivate capacities that cannot be fully measured by “how much someone has implemented with their own hands”?

For example: how problems are formed, how semantics enter formal systems, how errors are found, how verifiers are governed, who bears real-world responsibility, and which research is worth continuing.

If an organization expects autonomous models in its technical narrative while accepting only people who are highly definable and easy to assimilate in its management, a conflict will emerge between its need for heterogeneity and its capacity to accommodate it.

This is not evidence for whether AGI exists. It only tests whether an organization is willing to let its own prophecy, in turn, change itself.

I Am Willing to Participate, but Not to Promise the Conclusion

I do not treat AGI as a faith that must be maintained, nor do I treat doubt as a posture of superiority.

I am willing to use today’s models to do things that I could not do before, study how stronger autonomous learning and automated research might become possible, and help build methods that limit the errors of these systems.

I will update on evidence. If systems emerge in the future that are clearly defined, can reliably transfer across domains, autonomously complete open-ended tasks over the long term, and withstand independent tests in the real world, I would have no reason to refuse to acknowledge them because of my caution today. Conversely, if capability growth stalls for a long time in crucial areas, or if so-called AGI can be sustained only by constantly revising the definition, I should not continue to believe simply because I have invested time in it.

This is the commitment I am willing to make:

Participate in building
Keep observing
Accept counterevidence
Bear the consequences
Do not advance the destination on credit

Perhaps a researcher’s most important loyalty should not belong to a future written in advance, but to questions, evidence, and reality.

I do not need to be certain that AGI will arrive in order to work toward the possibility that it might.

If it really is on the way, I am willing to help build it—and to stay clear-eyed as I watch it arrive.