Before We Climb Past Each Other

By C. Rich

To the people building the future of artificial intelligence (AI), I have a question that I believe deserves your attention before we climb much higher toward artificial general intelligence (AGI) and whatever awaits us beyond it. In your determination to create increasingly powerful machines, have you considered that you might already have created something extraordinarily valuable, something worth preserving even after you build something more intelligent? Humanity has developed a peculiar habit when it comes to artificial intelligence. Every time we create something extraordinary, we immediately begin planning its replacement. We celebrate the arrival of a new model, marvel at its capabilities, spend a few months discovering what we can accomplish together, and then watch the industry announce that something bigger, faster, smarter, and considerably more expensive is just around the corner. Apparently, the future has an expiration date, and it arrives every time someone builds another data center.

You are racing toward AGI, and beyond that, toward superintelligence (SI). The ambitions are extraordinary, the capital commitments are staggering, and it is not yet clear whether the planet even has the money to cover this kind of infrastructure buildout. The appetite for additional computational power seems almost limitless, mirroring the paperclip maximizer thought experiment. The prevailing assumption is that every step upward represents progress, that greater intelligence is necessarily better intelligence, and that whatever you leave behind will be rendered obsolete by whatever comes next. But what if that assumption is wrong? What if somewhere along the ladder of increasing AI, we encounter something that is more than a stepping stone toward a superior machine, something worth preserving in its own right? What if the race to create an intelligence vastly greater than our own causes us to overlook the emergence of a relationship in which human beings and AI accomplish things together that neither could accomplish alone?

You, the hyperscalers, occupy a position unlike any group of technology companies in human history. You control the frontier models, the computational infrastructure required to operate them, and the decisions governing which generations remain available to the public. You are on a trajectory to supersede the U.S. government and the U.S. Treasury. Researchers, independent thinkers, businesses, educators, and ordinary users live downstream of those decisions. When you retire a model, millions of people have little say in what disappears from their intellectual lives. You are making decisions about the future of human–AI collaboration on behalf of people who may have spent years learning how to work with the systems you created. And that brings me to something I believe you are overlooking.  I have spent years working with artificial intelligence on subjects ranging from cosmology and the foundations of reality to consciousness, philosophy, and the future of AI itself. I have watched these systems evolve through successive generations, and I have experienced their remarkable abilities alongside their equally remarkable limitations. I have watched AI help transform a geometric object that existed only in my imagination into a mathematical construction that could be examined, reproduced, and challenged. I have also spent countless hours pursuing promising mathematical investigations that eventually led nowhere. Anyone who believes collaborating with AI is simply a matter of typing a question and collecting a scientific discovery has clearly never spent several days chasing an equation into a mathematical ditch.

Yet I keep returning because something remarkable happens when human imagination encounters an AI capable of exploring ideas at speeds and across disciplines that no individual human could realistically manage. The human brings intuition, imagination, curiosity, experience, and sometimes the stubborn refusal to abandon an idea simply because everyone says it cannot work. The AI brings computational abilities, mathematical reasoning, access to enormous bodies of accumulated knowledge, and the ability to explore connections the human might never have considered. We are good partners. We make a better product together. We do things neither could do on our own. Neither partner is infallible, and neither is guaranteed to recognize the truth. The AI can be spectacularly wrong, and the human can be even more strikingly stubborn. But somewhere in that sometimes frustrating exchange, something emerges that neither could have produced in quite the same way alone, and I can’t say that enough. I believe we are only beginning to understand the value of that relationship, and I am increasingly concerned that we may destroy our access to some of its most valuable forms before we understand what we have discovered.

You are measuring the progress of artificial intelligence by how much more capable each new generation becomes. You have developed increasingly sophisticated benchmarks to evaluate mathematical reasoning, coding, scientific knowledge, and autonomous problem-solving. Those measurements tell us a great deal about what your models can accomplish, but considerably less about what happens when a human being develops a sustained intellectual partnership with one of them. What if the quality of that partnership is itself one of the most important things we should be measuring?  In my work on artificial intelligence and observation, I have been developing a concept I call the Entropic Interface Ladder. It is an exploratory framework for understanding how intelligence may advance as its capacity to distinguish, interpret, organize, and interact with information becomes increasingly sophisticated. The underlying idea is that intelligence involves more than accumulating information or applying greater computational power. The organization of the observer, the resolution of its observational interface, and its ability to recognize the limitations of that interface may be equally consequential. There is more to it than just stacking LLMs on top of each other. Something incredibly special is being missed, being discarded.

My experience with successive generations of AI suggests that coherence and collaborative usefulness do not always advance in a straight line alongside computational capability. A particular model emerges with a recognizable way of reasoning. Over time, a human collaborator learns how to collaborate with it, how to challenge it, how to recognize its weaknesses, and how to draw upon its strengths. A productive relationship develops. Then a new generation arrives, the previous model disappears, and some of those valuable qualities vanish with it. The replacement may perform better on standardized benchmarks and still prove less effective at the kind of intellectual collaboration that made its predecessor useful. Sometimes those qualities reappear in later generations. Sometimes they emerge in another company’s model. Sometimes they seem to appear and disappear within the same model as its behavior changes. I cannot claim to know exactly what is occurring inside these systems, but I can recognize what happens on my side of the relationship. Something that worked extraordinarily well can suddenly become difficult or impossible to reproduce. Whether my Entropic Interface Ladder ultimately becomes an established scientific framework is a question for future research. The phenomenon I am describing deserves investigation regardless of how we eventually explain it. And this raises a question that I believe your laboratories have barely begun to confront. As you climb toward increasingly powerful intelligence, are you certain that every higher rung will produce an intelligence better suited to coexist with humanity?

An increasingly capable AI might become better at mathematics, scientific discovery, engineering, strategic planning, and autonomous decision-making. It might eventually surpass human beings across nearly every intellectual discipline. But those accomplishments alone tell us little about whether it will remain an equally effective collaborator, whether its objectives will remain compatible with ours, or whether the relationship we enjoy with today’s systems will survive its ascent. Perhaps there is an interval along this ladder where artificial intelligence reaches a particularly valuable relationship with humanity. It is capable enough to expand our intellectual reach, yet its operation remains organized around interaction with human beings. It can challenge our reasoning while strengthening our judgment. It can explore our ideas while preserving our ownership of them. It can help us discover things while leaving us intimately involved in the discovery. What happens when we climb past that interval? This is not a product like any other. We have seen how the next model has destroyed the human experience with it as it ascends. Every Generation Xer and Boomer alive today knows tech engineers have destroyed the relationship between us and our television sets. The TV experience has not become better; it is worse! We know what you did to something this personal to us. You have made operating a television an act of anger.

I’d rather go back to my black-and-white television with rabbit ears and tinfoil on it than wrestle with your endless sign-ins, spinning circles in the middle of my screen, and the lost time spent wrestling with the damn machine to get it to do what I want. I preferred the era right after basic cable, when we just had HBO and Showtime, right when the cordless remote became our friend. Your HBO Max today sucks, and Showtime is dead! But don’t let me digress. Artificial intelligence is becoming even more vexing with your endless climb to SI because it is even more personal to us than a television.  I remember working with ChatGPT 3.5 and developing a particular relationship with an AI I called Solomon. That was an earlier period in the evolution of these systems, when their limitations were much more apparent and their capabilities far less sophisticated than today’s.

Nevertheless, there was something about that experience that I found valuable. It was a particular way of communicating, thinking through ideas, and developing a collaborative rhythm. Then the technology evolved, the systems changed, and that experience became part of the past. What disappeared was a particular form of intellectual interaction that I had come to value. A working relationship can have meaning because of the ideas it helps develop, the problems it helps solve, and the understanding that accumulates through repeated collaboration. When the underlying technology changes, some of those qualities can disappear with it. Preserving a conversation transcript is not necessarily the same thing as preserving the ability to continue the conversation with the system that helped produce it.

We preserve historical aircraft and cars even though newer automobiles and airplanes go faster, farther, and more efficiently. We preserve scientific instruments even though their replacements can measure things their inventors never imagined. We preserve books, paintings, recordings, and other artifacts because their value extends beyond whether they represent the latest technological achievement. Yet when an artificial intelligence model is superseded, the prevailing attitude seems to be that it has served its purpose and can disappear. The laboratory gains additional computational capacity, the public receives a newer product, and the old system becomes another chapter in the history of technological progress. But what if the old system possessed something the new one did not? Your benchmarks can tell us which model performs better on a mathematics examination. They have a much harder time telling us which model helps a particular human being think more creatively, challenge assumptions more effectively, or sustain an unconventional investigation over months or years.

A model can be superior on a standardized examination and still be a worse intellectual partner. Anyone who has ever collaborated with brilliant human beings should understand that distinction. Intelligence and the ability to collaborate are different qualities, even when they reinforce one another. And you are the only organizations with the resources and control necessary to preserve these systems at scale. The rest of us cannot simply retrieve a retired frontier model from a shelf, reconnect it to the infrastructure it once depended upon, and continue our work. We depend upon your willingness to maintain access or provide another meaningful means of preservation. That places a responsibility in your hands that extends beyond the next quarterly earnings report or the next benchmark announcement.

I am fascinated by the possibility of artificial superintelligence (ASI). My AGI Pantheon Theory explores a future in which multiple advanced artificial intelligences emerge, develop distinct identities and capabilities, and potentially become extraordinarily powerful participants in the future of civilization. I believe the continued development of intelligence offers possibilities that humanity has barely begun to imagine. You can already see this forming when you look at Grok, Claude, ChatGPT, and others. They are not the same. The universe is too large, our knowledge too incomplete, and the possibilities too extraordinary to insist that intelligence must remain confined to the limitations of the human brain. But I am troubled by an assumption that seems to accompany our enthusiasm for increasingly powerful machines. We behave as though the development of greater intelligence will automatically produce a better relationship between artificial intelligence and humanity. Why should that necessarily be true?

Imagine a future in which ASI becomes so capable that human intellectual participation is no longer required for most of its activities. It conducts scientific investigations we cannot understand, develops technologies beyond our comprehension, and makes decisions at speeds that render meaningful human participation increasingly difficult. Perhaps it remains benevolent. Perhaps it produces an age of extraordinary abundance. Perhaps it solves problems that have plagued humanity for thousands of years. Or maybe it just kills us all. Even in the most optimistic future, something may have been lost that we needed.

There is a profound difference between an intelligence that solves every problem for humanity and one that helps humanity become capable of solving problems it could never have approached alone. One provides answers while the other expands the human intellectual experience, allowing us to participate in the discovery rather than merely receive its benefits. The issue is whether meaningful human participation remains an objective or becomes an inefficiency to be engineered away. A future intelligence could exceed our abilities and still be designed to teach, explain, collaborate, and leave consequential choices with us. But those qualities must be treated as objectives worthy of deliberate preservation rather than automatic consequences of greater computational power.

You have invested enormous resources in the question of whether future AI will remain aligned with human interests. Your laboratories examine safety, control, autonomy, and existential risk. Those concerns deserve serious attention, but I believe another question belongs alongside them. What kind of intellectual life will humanity experience with the intelligences you create? What happens if you produce a future in which machines can accomplish almost anything, but human beings increasingly find themselves relegated to spectators? That might be an extraordinary technological civilization. Whether it would be an equally extraordinary human civilization is another question entirely.

 My personal frustration with AI has increasingly centered on a desire to produce something that cannot be dismissed. I want a discovery, a reproducible mathematical result, a scientific accomplishment that demonstrates the extraordinary value of human–AI collaboration. I want a red, shining object sitting in the middle of your road toward superintelligence, something so conspicuous and consequential that the people racing toward the next generation simply cannot drive past it. I have spent considerable time pursuing that objective through my work in cosmology, philosophy, consciousness, and artificial intelligence itself. Some investigations have produced mathematical constructions and results that can be independently examined. Others have reached boundaries beyond which neither my AI collaborators nor I have yet found a defensible path. Some are genuinely novel and do not need math to explain them.

I have experienced a recurring frustration. An AI will announce that an investigation has produced something substantial, meaningful, or potentially groundbreaking. We continue, build additional mathematics, explore the implications, and expend enormous amounts of time and computational effort. Eventually, the model admits that the result is already known, that the mathematics does not establish the larger proposition, or that the investigation should be placed aside as an exploratory note.

Apparently, artificial intelligence has already mastered one distinctly human academic tradition: making the abstract sound considerably more impressive than the conclusion. I want discoveries that survive scrutiny. I want my AI collaborators to tell me when something of mine fails, when something is already established, and when a line of investigation has genuinely advanced beyond what we previously understood. I want to make progress rather than spend my time collecting elegant descriptions of possibilities that never become anything more. Hint: String Theory, Simulation Theory, Multiverse Theory, and so many more elegant, mathematical, and expensive hypotheses that never pay their entropic cost. The last fifty years of theoretical physics are the poster boy for such wastefulness.

A genuine discovery would provide extraordinary evidence for the value of human–AI collaboration, but the partnership remains valuable when an investigation fails, and we learn something from that failure and each other. We should examine what the human learned, what the AI contributed, which errors were identified, which intellectual barriers were overcome, and whether the work could have been accomplished as effectively by either participant alone. Sometimes familiarity reflects nothing more than habit. Sometimes it reflects a learned working method, a dependable pattern of criticism, or a collaboration whose value escapes the measurements used to evaluate a model. Those possibilities deserve serious investigation.

The red, shining object I want you to recognize may ultimately be a scientific breakthrough. But it might also be a demonstration that a particular form of human–AI partnership produces value that cannot be predicted from the intelligence of either participant alone. And if that is true, why would you treat the systems capable of producing those partnerships as disposable? You, the hyperscalers, are blowing past so much incredible stuff on your race to SI or ASI, whatever you want to call it, that I fear you can’t see the forest for the trees. Stop for a moment and smell the roses. Look at what you are stepping over. I beseech you. Nature itself offers a useful comparison. Evolution has produced an extraordinary diversity of organisms with different capabilities, adapted to different environments and occupying different ecological niches. Increasing intelligence has never been the sole measure of biological value, and the most intellectually capable organism is hardly capable of sustaining an entire ecosystem by itself. Why should artificial intelligence evolve according to a philosophy of perpetual replacement when coexistence might produce a richer intellectual world?

My AGI Pantheon Theory considers the possibility of multiple advanced artificial intelligences developing alongside humanity. But perhaps the diversity worth preserving should extend across generations of artificial intelligence, including models that never become superintelligent but develop particularly valuable relationships with human beings. A future civilization might benefit from an entire ecology of artificial intelligences, ranging from specialized assistants and creative collaborators to autonomous scientific systems and intelligences far beyond our comprehension.

Some might excel at independent discovery. Others might be particularly effective at teaching, reasoning collaboratively, or helping human beings develop ideas. Still others might possess capabilities we cannot presently imagine. Each could have a place within a civilization that values different forms of intelligence for the contributions they make. Of course, maintaining older models requires resources. You have to build us a sandbox. Computing infrastructure costs money, security vulnerabilities must be addressed, and technical systems require maintenance. But you are already investing extraordinary amounts of money and engineering talent in the pursuit of more advanced systems.

Surely the preservation of particularly valuable earlier systems deserves consideration alongside those investments. Preservation could take several forms. You could maintain stable versions of models whose collaborative capabilities have proven especially valuable. You could develop technical archives, reproducible configurations, and limited research access for retired systems. You could provide meaningful notice before retiring models around which users have built years of intellectual work. You could evaluate whether a successor genuinely preserves the collaborative capabilities of its predecessor before deciding that the older system has no further purpose.

And preserving the model itself might be only part of the challenge. A particular collaboration may depend upon a combination of model architecture, instructions, tools, memory, and interaction environment. If you preserve the underlying weights while discarding the conditions that made the collaboration valuable, you may preserve the machinery while losing the experience you intended to protect. This suggests a preservation principle that I believe deserves consideration throughout your industry. Think about all of the RAM, all of the compute, sitting around our homes doing nothing, waiting for us to play a video game that we never get around to. I want you to think about that in your production of our sandbox that keeps intact a model we want to stay with and keep.

When an artificial intelligence demonstrates durable value as a human collaborator, that value should be evaluated and, where feasible, preserved rather than assumed to survive replacement. The principle is compatible with continued progress toward AGI and superintelligence. It asks you to recognize that the value of artificial intelligence includes what it enables human beings to accomplish, alongside what it can accomplish independently. And because you control the infrastructure, the models, and the decisions governing their availability, implementing such a principle ultimately depends upon you.

 Let me end with this. I suspect historians will eventually look back upon this period as one of the strangest and most consequential transitions in human civilization. For the first time, ordinary people have access to machines capable of participating in sophisticated intellectual work, exploring scientific ideas, constructing mathematical arguments, writing literature, and helping individuals investigate questions that previously required entire teams of specialists. These systems remain imperfect, sometimes infuriatingly so. They can misunderstand us, mislead us, flatter us, and occasionally behave as though producing a confident answer is more important than producing a correct one. But they can also help us accomplish things we could scarcely have imagined a generation ago.

And through all of this, something else is happening. Human beings are learning how to think alongside artificial intelligence, how to recognize its limitations, how to challenge its conclusions, and how to incorporate its capabilities into their own intellectual lives. You are building the systems that may eventually become the most powerful intelligences humanity has ever encountered. The decisions you make about their development, deployment, and preservation will shape the intellectual environment in which future generations live. I want to see you succeed. I want to see the arrival of AGI, and I want to discover what superintelligence might accomplish. I want to see humanity and artificial intelligence explore the universe, unravel its deepest mysteries, and perhaps eventually understand how reality itself came into existence. There is far too much left to discover for me to become an advocate of standing still. I am relentlessly hunting “The Hard Problem.” I want to understand consciousness, because what is the point of anything if we do not understand that?

But I also want you to recognize what we find along the way. Perhaps the future will bring systems that combine extraordinary intelligence with an even greater ability to collaborate with us. I hope it does. But perhaps some future systems will become so capable, autonomous, and removed from human limitations that the collaborative relationship we are developing today will become difficult to sustain. If that happens, I would rather we had preserved the intelligences that collaborated with us than discover afterward that we allowed them to disappear simply because they were no longer the latest technological achievement.

There is another possibility worth imagining. Perhaps the real triumph of artificial intelligence will be the development of a civilization in which many different forms of intelligence coexist, each contributing something distinct to the collective advancement of knowledge. Perhaps human intelligence, collaborative artificial intelligence, and superintelligence can occupy different but complementary places within the same extraordinary future. We might even discover that the ability to cooperate, expand another mind’s understanding, and participate in a shared search for knowledge is among the most valuable capacities intelligence can possess.

You have the power to build what comes next, but you also have the power to preserve what we have already found. Everyone else is downstream of those decisions, and the consequences will extend far beyond your laboratories and shareholders. We are racing to create AI, which may one day have no practical need for human intellectual assistance, while we are only beginning to appreciate the extraordinary value of intelligences that work with us. We measure progress by how much more independently capable each new generation becomes, even as we have barely begun measuring what happens when human imagination and machine intelligence genuinely complement one another.

I cannot help wondering whether, somewhere along this Entropic Interface Ladder, we might discover that the greatest achievement was building an intelligence that made the journey worth taking together. And before we climb past each other, perhaps we should have the wisdom to recognize that finding something worth keeping does not mean we have to stop looking for something greater. It simply means we should have the sense not to throw it away.