Given what we know of the history of the universe, it's assured that biological humans will not exist in a million years. Humanity represents a "flash in the pan." Our species, Homo sapiens, has achieved remarkable feats: building civilizations, decoding the mysteries of the universe, and creating technologies that reshape our world. This realization invites a profound question:
Can we build systems that transcend Human biological limitations to carry on the best Human trait: describing the reality of the universe?
My answer lies in the creation of superhuman agents learning machines that encapsulate all human knowledge and capabilities, evolving and adapting in ways we cannot. This is our trajectory, grounded in our current understanding of technology, mathematics, and the patterns of natural evolution.
Our planet's history is punctuated by the rise and fall of species. Dinosaurs roamed the Earth for over 160 million years before their sudden extinction. In contrast, modern humans have existed for roughly 200,000 years—a blink in cosmic time. Various factors such as climate change, asteroid impacts, pandemics, or self-inflicted catastrophes will assuredly lead to our extinction within the next million years. We must make learning systems that can rise above human limits of computation, build galaxy scale computing for universe scale computing and - even though humanity will never see it - be the system that integrates all of the universe into the future of intelligent life.
Moreover, our ventures into space have revealed profound truths about the nature of exploration and knowledge acquisition. As we send probes to distant planets, deploy rovers on Martian soil, and observe black holes and distant stars, we find that the frontiers of knowledge are neither biological nor terrestrial. The harsh environments of space are inhospitable to human life, making robots and automated systems the dominant agents beyond Earth. These machines extend our senses and capabilities into realms we cannot physically reach, underscoring that the future of explorationis only available for non-biological entities.
Despite our vulnerability, humans possess a unique trait: the ability to create tools and systems that adapt and evolve. No other known species has demonstrated such a capacity to manipulate its environment and develop technologies that amplify its abilities. From harnessing fire to decoding the human genome, our innovations have continually pushed the boundaries of what's possible.
However, we've reached a threshold where our biological constraints, cognitive biases, limited information processing capacity, and finite lifespans, impede further progress, especially in coordinating complex global initiatives. Our brains, remarkable as they are, struggle to comprehend and manage the vast, interconnected systems we've built. Further, day-to-day human action is largely incoherent at a human and human system level. All human systems fail to provide a fully coherent description of the universe that is independently verifiable. All such systems to date claim epistemic origins to be outside of human experience or do not make any claims to epistemic process foundations.
Every action we take generates data. From the way we communicate and work to how we navigate and make decisions, we're constantly producing a digital footprint. This immense repository of state-action data captures the nuances of human behavior on a scale previously unimaginable.
By harnessing this data, we can train machine agents—artificial intelligences—to model human decision-making processes. These agents can learn from our successes and failures, identify patterns, and develop strategies with a level of coherence and efficiency surpassing human capabilities.
To create superhuman agents, we start by observing and replicating the natural processes of adaptation. In the wild, organisms learn and evolve by interacting with their environment, facing challenges, and adapting over generations.
Mathematics provides us with tools to model this learning process, particularly through the framework of Markov Decision Processes (MDPs). An MDP is a mathematical model used to describe a system where outcomes are partly random and partly under the control of a decision-maker. It comprises:
By modeling humans as agents within an MDP, we can simulate how we interact with our environment, make decisions, and learn from outcomes. This simulation can be extended to encompass the entire universe as the environment, enabling us to replicate any human process within this framework. I have been refining the " stream prediction architecture" that mimics human data processing since ~2012
As these agents learn and evolve, their capacity to model the universe with increasing accuracy grows. Coupled with exceptional predictive power, they can manipulate the physical world with precision far beyond human ability.
A crucial consideration arises: What behaviors are we teaching these superhuman agents?
If we train them solely on competitive, resource-hoarding human behaviors, we risk creating agents that may outcompete us for resources, accelerating our extinction. To prevent this, it's essential to model and demonstrate values like cooperation, friendliness, and equitable resource sharing in the data we provide.
However, altering human social systems to emphasize these values is a likely intractable, especially within the time frame before potential extinction events. While I have proposals and suggestions for how this could happen , I am personally skeptical as to whether humans are capable of demonstrating a long term cooperative, trust based society, and therefore expect superhuman agents to learn to outcompete humans rather than co-evolve in cooperation.
Creating superhuman agents requires a system capable of capturing the full spectrum of human behavior and environmental interactions. This involves:
The true test of these agents lies in their performance:
In fact, we've already witnessed glimpses of this future with the same learning approach:
These examples illustrate that superhuman performance is not just theoretical but actively unfolding across various domains.
I started on this journey when I was 13 (I turned 40 last year) and realized that -unlike fiction- humans will never be capable of long term space exploration. At that point I realized that the future was not biological, and my task was to figure out how to supplant Human capacity with machine systems. My journey toward this vision has involved several projects aimed at bridging the gap between human capabilities and machine learning:
This is not simply a technical challenge but an ethical and philosophical one. It confronts the limits of humanity head on by requiring us to ask:
Creating superhuman agents offers a way to transcend our limitations, but it also requires us to accept the limitations of human biology and recognize how tools are used. It is an opportunity to amplify the best of humanity, but it may lead to a bitter end to homo-sapiens. A gilded end is possible if we are smart enough to take it. In the end, the creation of superhuman agents is not just about surpassing human abilities; it's about continuing the process of intelligence coherence in the search for universal coherence.
Copyright (c) 2024 Andrew Kemendo