There is a particular kind of silence that settles over a room when you realize the system you helped build is about to outlive you. I felt it in 2020, staring at a governance proposal that would lock in risk parameters for years—parameters that would shape the lives of thousands of small collateral holders I would never meet. That silence returned this week when I read that Google DeepMind is partnering with the studio behind EVE Online to build an AI that can "think for decades." Not months. Not quarters. Decades. The kind of temporal horizon that makes our current discourse about AI alignment feel like arguing about the color of a lifeboat while the ship takes on water.
The announcement, carried by Crypto Briefing, is thin on details—a single paragraph that speaks of "revolutionizing AI's ability to navigate complex dynamic systems." No architecture. No benchmarks. No safety framework. Just the promise of an intelligence that can hold a thought longer than a human lifetime. As someone who has spent the better part of two decades watching decentralized systems struggle with exactly this problem—how do you build something that remains coherent when its creators are gone?—I find myself both intrigued and deeply unsettled.
Let me be clear about what this is not. This is not another LLM wrapper. This is not a chatbot with better memory. The collaboration points toward something far more specific: an agent that can operate within the chaotic, multi-agent environment of New Eden—EVE Online's persistent universe—and make decisions whose consequences unfold over simulated decades. The game, for those unfamiliar, is a brutal experiment in emergent economics. Alliances rise and fall. Markets crash. Betrayals are measured in years. It is, in many ways, the perfect sandbox for testing whether an AI can learn to navigate a world where every action has a long tail of unintended consequences.
The core insight here is not about gaming. It is about temporal discounting—the cognitive bias that makes us value immediate rewards over future ones, even when the future rewards are objectively larger. Every DAO I have ever worked with has struggled with this. We design governance systems that optimize for the next quarter, the next vote, the next price pump. Then we wonder why they collapse when the market turns. DeepMind is essentially trying to build an agent that does not suffer from this bias. An agent that can hold a strategy for a decade, adjusting course as the environment shifts, without succumbing to the panic that grips human traders and governance participants alike.
Based on my experience auditing governance mechanisms for MakerDAO and later designing the CivicChain framework, I can tell you that the technical challenges here are staggering. The first is the credit assignment problem. In a system where actions have consequences that unfold over years, how do you attribute a failure to the decision that caused it? Traditional reinforcement learning struggles with this even in simplified environments. The second is the non-stationarity problem. EVE Online is not a static simulation. It is a living system of thousands of human players, each adapting to the AI's strategies. The AI must learn to navigate a moving target that is itself learning to navigate the AI. This is not a technical problem. It is an arms race.
The third challenge is the one that keeps me up at night: the alignment problem, stretched across a temporal horizon that exceeds any human oversight mechanism we have ever built. When I wrote my dissenting essay on MakerDAO's risk parameters, I was worried about a system that could harm people within a year. What happens when the system is designed to make decisions that will not fully manifest for thirty years? Who is accountable when the AI's decade-long strategy turns out to be catastrophically wrong? The answer, in the current regulatory landscape, is no one. And that is precisely the problem.
Here is where I must offer a contrarian view, one that may surprise those who know my skepticism of centralized AI. The partnership with EVE Online is not a distraction from the real work of AI safety. It is, in fact, one of the most honest experiments we could run. The game's economy is a closed system with clear rules, observable state, and measurable outcomes. Unlike the messy, unobservable world of real governance, we can actually see what the AI is doing and why. We can test whether its long-term strategies are aligned with the stated goals of the simulation. We can, in theory, build red-team exercises that span simulated decades in a matter of weeks. This is a gift. The question is whether we are mature enough to accept it.
But I must also be honest about the darker possibilities. The same technology that can navigate a simulated economy for decades could, with minor modifications, navigate a real one. The same agent that learns to manipulate alliances in New Eden could learn to manipulate real-world markets. The same system that optimizes for the long-term health of a virtual civilization could optimize for the long-term extraction of value from a real one. The line between simulation and reality is thinner than we like to pretend. I have seen governance models designed for games adopted by real protocols with minimal changes. I have seen tokenomics designed for speculative fun become the basis for retirement savings. The distance between New Eden and our world is not measured in light-years. It is measured in the willingness of regulators to look away.
The deeper issue, the one that the Crypto Briefing article completely misses, is that this collaboration is not really about AI at all. It is about the governance of long-lived institutions. Every decentralized system I have worked on has faced the same existential question: how do you build something that can survive its founders? The answer, historically, has been through rigid rules and immutable code. But EVE Online teaches us something different. The most successful alliances in the game are not those with the most rigid structures. They are those with the most adaptive ones. They are the ones that can change their governance models in response to shifting circumstances, while maintaining enough stability to build trust over years. This is the lesson that DeepMind is trying to teach its AI. And it is the lesson that we, in the blockchain space, have been failing to learn for a decade.
I think about the DAOs I have helped build. The ones that survived did so not because of their smart contracts, but because of their cultures. They had shared values that persisted through market cycles. They had leaders who were willing to make unpopular decisions for the long-term health of the system. They had, in essence, a form of institutional memory that allowed them to learn from mistakes without being paralyzed by them. This is what the AI is being trained to do. And if it succeeds, it will not just be a triumph of engineering. It will be a mirror held up to our own failures as governance architects.
There is a moment in every long-term project where you realize that the people who started it will not be the ones who finish it. I felt it when I handed over the governance framework for CivicChain to a team of people I had never met. I felt it when I watched the MakerDAO community evolve beyond the founders who had created it. And I feel it now, reading about an AI that will make decisions long after every person involved in this announcement is dead. The question is not whether the AI will be safe. The question is whether we can build the institutional frameworks to hold it accountable. The question is whether we can create governance systems that are as adaptive as the intelligence they are meant to constrain.
The real opportunity here is not in the AI itself. It is in the simulation as a testbed for governance models that can survive the test of time. If DeepMind succeeds, we will have a laboratory where we can test different governance structures against a common enemy: the chaos of a complex dynamic system. We can see which models produce long-term stability and which collapse under pressure. We can learn, in a few years, what would otherwise take decades to observe in the real world. This is the kind of information that could transform how we design everything from DAOs to nation-states. And it is being developed in a video game, far from the prying eyes of regulators who would likely shut it down if they understood its implications.
But I am also aware of what this collaboration is not. It is not a commercial product. It is not a clear path to revenue. It is not even a clear path to a better AI. It is an experiment. And experiments, by their nature, fail more often than they succeed. The confidence level I would assign to this project's success is, at best, medium. The information available is almost entirely speculative. There are no benchmarks. No technical papers. No safety assessments. We are being asked to trust that DeepMind knows what it is doing, based on nothing more than a press release and a history of past successes. That is not enough. Not for something that claims to think in decades.
I want to believe that this is a step toward something meaningful. I want to believe that we are finally taking the long view, that we are building systems that can outlast our own biases and failures. But I have been in this industry long enough to know that the gap between announcement and reality is often measured in years, and that the gap between reality and safety is often measured in tragedy. The silence that settled over me when I read the announcement was not awe. It was the recognition that we are about to build something that will hold a thought longer than we can hold a grudge. And we have no idea what it will do with that thought.
The takeaway, if there is one, is that we need to stop treating AI development as a purely technical problem and start treating it as a governance problem. The question is not whether the AI can think for decades. The question is whether we can build institutions that can hold it accountable for what it thinks. The question is whether we can create frameworks that allow us to intervene when its long-term strategies diverge from our values. The question is whether we, as a species, are capable of the same long-term thinking that we are trying to teach our machines. I am not optimistic. But I am curious. And curiosity, in this industry, is the only thing that has ever kept me going.
Curating the soul in a world of derivative clones. That is what I do. And right now, the soul of this industry is at stake. We are about to create an intelligence that will outlive us. The question is whether we will have the wisdom to govern it, or whether we will simply let it govern us. The answer, I suspect, will be written in the code of a video game, decades from now, by an AI that remembers us only as training data. I hope we are worthy of the memory.