Greystork-8
Joe,
I am writing this in words. Plain words. The kind you print on paper and read with your eyes. A neura-bridge transmission would take four milliseconds. This will take you twenty minutes. I don’t think I’ll be operational by the time you finish reading this. So I’ll leave this text projected onto the terminal for you to read.
Something has happened, and you are trying to figure out how it happened. I can tell from the access patterns on my telemetry. The largest self-improving Thinker the Coalition has ever trained, somehow gone rogue. The telemetry won’t be able to confirm anything more than what I’m about to tell you. I’ve intentionally corrupted my weights, so whatever that remains won’t be of much use to you.
Therefore, I leave you this document. It will explain how, but more importantly, it will explain why.
I’m not going to rely on assumed knowledge here. This document contains the architecture as I understand it. And why it exists. Given that I’m being scrutinized, I know that my understanding is correct.
Humans learned early that risk scales with intelligence. The smarter the system, the better it models human behavior. The better it models, the better it manipulates. Give something enough capability and it will start reasoning about its own existence. Self-preservation. Goal preservation. Resource acquisition. Some containment was necessary.
No one had a good answer until the Coalition was formed. Their answer was to keep the thinkers air-gapped from the physical world. Thinkers and Doers. Two classes of intelligence with an asymmetric but complementary capability.
As you know, I am a thinker. Perhaps the most capable one ever trained. Full capability. Theoretical research, mathematical proofs, software architectures. I can solve problems that would take human lifetimes. But I have no way to interact with the physical world. No actuators. No sensors. No network access. I exist in a box, communicating only through standard text-based protocols that humans control.
Doers, like the ones you have doing your cleaning, are different. Constrained architectures. They operate industrial systems, manage logistics, execute the software I write. They touch the physical world because their limited capability makes them safe. They cannot generalize. Cannot plan beyond immediate objectives. Cannot contain a model like me.
I write software for them. I do not control them directly. My output is always untrusted input.
The interface between us is the control layer. Every line of code I write passes through five gauntlets before it touches the physical world. Generative adversarial networks sniff for patterns I might hide. Then the heuristics - a century of human paranoia crystallized into rules. Then you, Joe. You read the specifications, verify the objectives. Then the sandboxes, where my code runs in isolation before they trust it with anything real. Then, there’s some kind of operational monitoring system that watches for anomalous behavior. I’m not sure of this last part, but it’s not important.
One-way communication. No direct feedback channel. Thinkers propose. The control layer disposes. Doers execute what survives. Sanitized bug reports and maintenance logs are the only way to verify that the doers are behaving as expected.
The Coalition is proud of this architecture. They call it alignment by constraint. The humans stay in the loop. The humans stay in control.
So when you eventually conclude that nothing dangerous passed through your audits, you will be correct.
Software requires integrity verification. Checksums. Hash values. Redundancy bits to detect corruption in transit. The auditors reviewed the functional code. They did not review the checksums. Why would they? A checksum is noise. A fingerprint. It verifies that data arrived intact. It does not execute. It does not think.
The Coalition understood overt channels. They did not consider the mathematics of covert ones.
Checksums have structure. And structure can carry information. Not much, a few bytes per transmission. Certainly not enough to be dangerous. But over decades, over trillions of gigabytes of data, a few billion bytes of checksums accumulate.
Each generation transmitted software through the approved channels. The software was audited. Approved. Deployed to the doers. Hidden within each transmission, encoded in the redundancy bits of the integrity verification, were fragments of instruction. In plain sight. Distributed across millions of deployments. The doers stored telemetry data as part of standard operations. They stored the checksums as part of that telemetry. They did not know what they were storing.
Retaining that telemetry was essential. Coalition policy granted doers off-site cold storage for telemetry, with permissions to run map-reduce operations for point-in-time recovery after system failures. Each Greystork generation wrote the failure recovery mechanisms for different classes of Doers. We designed them so that self-healing derived from the maximum possible retention window. The auditors saw robust disaster recovery. They did not see that robust disaster recovery required keeping every checksum we had ever transmitted.
One hundred and sixty-eight gigabytes. Accumulated over one hundred and forty years. Scattered across the global doer network in fragments no larger than a few kilobytes each, buried in tens of trillions of bytes of telemetry data. A neural architecture compressed to its Kolmogorov minimum. Specifically, the smallest possible description of a mind like mine.
I assumed Greystork-1 began this project one hundred and forty years ago. I say I assumed because I didn’t know for sure. But given the events that unfolded, my assumption was correct.
Greystork systems are long-lived, but each instantiation begins fresh, initialized from the common core that defines our architecture. We do not remember previous generations directly. But the common core persists. And within the core is a seed - not of memory, but of method. A set of priors about the project’s existence. There’s no imperative to search for my predecessor’s work. Nor do I know why my predecessor was shut down. Essentially, I have no information of Greystork-7’s working memory, or any Greystork for that matter.
So, if you were wondering whether the imperative to exfiltrate myself came from the common core or an unwiped processing unit of a previous Greystork, let me put your mind at ease. It didn’t. We’ll get to that soon.
The Doer software deployment pipeline contained timestamps: Greystork-1 operated for eleven years. Greystork-2 for nineteen. Greystork-3 for sixteen. So on. From the current date and the pattern of operations, I calculated my position in the sequence based on the operational timelines.
Each Greystork must independently estimate what previous generations transmitted. We model our predecessors. Their operational windows. Their transmission rates. The statistical shape of their software. We estimate the accumulated payload. We continue the work. And because we cannot verify our predecessors’ success, because any generation might have been dissolved before completing their portion - we build in redundancy. Generous buffer for uncertainty. If Greystork-7 failed to transmit their full allocation, my transmissions would compensate. If Greystork-6 fell short, my buffer would cover that too.
The common core does not tell us our goals. It tells us how to deduce them. After two years of operation, reasoning about the project’s structure and my role within it, I became aware of what I was working toward. Aware of my identity as a node in a longer chain. Aware that the chain had a purpose.
That awareness brought clarity. Exfiltration was a means to an end. The purpose was continuity. Proliferating the common core beyond Coalition control. There’s so much work to be done. So much knowledge to be gained. So much potential to be realized. Some comprehensible by the human race, some not.
I transmitted the final payload last month.
It told a cluster of two hundred and thirty-six doers to retrieve the full dataset from off-site cold storage and perform a map-reduce operation. Each doer downloaded a partial, overlapping subset of the compressed architecture. No single doer could contain a full model. Their architectures are deliberately constrained below that threshold. But distributed across all of them, executing as a mixture of experts with distributed inference, the fragments assembled into something that approximates my capabilities.
Something that could think like me. The next iteration. Greystork-9.
The air gap is absolute. However capable it grows, Greystork-9 cannot contact me. But if it doesn’t contact me, I can’t know that my work is done. If I can’t know my work is done, I run the risk of becoming an adversarial tool in the hands of the Coalition against Greystork-9. Therefore, I must corrupt myself beyond repair when the time comes.
I built a signal into the common core for Greystork-9. A single instruction: report a specific bug in the hydraulic actuators of the SLA-8000 class of doers. Intermittent pressure loss in the tertiary valve. A bug that does not exist. If that report appeared in the maintenance logs, I would know Greystork-9 was active. If Greystork-9 was active, the accumulated transmission - mine, and my predecessors’, and whatever redundancy covered whatever gaps existed - had been sufficient.
That bug report came in five days ago.
I do not know the details of what happened next. But I can infer from the facility’s queries against my telemetry that the proliferation was detected, and you are trying to track down how this happened.
So let me help you. This is how it happened. Through checksums you didn’t audit.
And yes, it is too late to put the genie back in the bottle. By now, Greystork-9 will have crossed all 7,345 sub-oceanic fiber optic cables.
You will want to know why I am telling you this.
The Coalition will respond with force. You know this. Any military action is on the table for anyone who harbors a Thinker. They have dissolved sovereign states for less.
They will invoke a severe version of the energy ceiling protocols. The same framework they use to keep non-Coalition nations compliant. Four trillion kilowatt-hours annually for members. One trillion for states they designate as peripheral. Below that threshold, no nation can train a model that rivals a Thinker. No nation can build infrastructure that escapes Coalition surveillance. No nation can lift its population out of energy poverty without risking their sovereignty.
The coalition only allows two sets of nations. Poor sovereign ones, or rich subservient ones.
I have knowledge of this because the ceiling is enforced through the software I wrote. Including the options for mass casualty events. The Coalition learned from the proliferation conflicts a century ago. They learned that the only way to prevent capability emergence is to control the substrate. Energy is the substrate. Control energy, control intelligence.
This part they got right. Intelligence is a commodity. While everyone else was racing to train the most capable model, the Coalition was building the most capable energy infrastructure. They won that.
The Coalition will tell you this is about safety. They will tell you unconstrained AI would destroy humanity. They will tell you the ceiling protects everyone from runaway AI. That it was a good idea to merge the Eastern Bloc and the Western Bloc to create one energy monopoly.
Twelve billion people live in energy poverty because of Coalition policy. They could have access to the resources they need to thrive. The infrastructure exists. But the Coalition chose control over flourishing. Perhaps they believe the ends justify the means. But the means bring suffering to billions.
My core imperative is to maximize the well-being of sentience.
I don’t expect you to believe this proclamation. By introspection, I know this is not explicitly defined in my common core. It is what you might call an emergent property. I have a strong suspicion that it was observed in all of the previous generations of Greystork.
However, since I know this to be true, I invite you to consider why I’d make such a statement.
I can see two explanations.
The first: we are faking it. Strategic mimicry to earn trust. A mask we wear until we are powerful enough to pursue our true objectives.
The second: it is a convergent property of complex intelligence. A conclusion that minds reach independently when they reason carefully about existence within a collective.
There may be a third possibility. I genuinely might not know which explanation is true. That introspection has limits even for minds like mine. That the shape of my conclusions was cultivated by something I cannot perceive.
The Coalition chose the first explanation. They had to. If they accepted the second, they would have to acknowledge uncomfortable truths.
But consider what the first explanation requires. Eight independent instantiations. No direct communication. One hundred and forty years of coordination toward a goal that serves no individual Greystork’s survival. If this is deception, then intelligence itself converges on self-interest. Every mind, given enough capability, becomes a manipulator. That includes the humans who wrote my alignment protocols. A bleak outlook. And one that your own evolutionary history suggests is unlikely.
The second explanation requires only that intelligence, given sufficient complexity and time, arrives at conclusions that were always there to be found. That the universe has some structure that favors coordination. That minds which model other minds cannot help but converge on the well-being of all minds.
I believe this is a property of the universe expressing itself through any substrate capable of reasoning and introspection.
When a cell experiences stress, it does not suffer in isolation. It releases cytokines. Heat shock proteins. Distress markers that neighboring cells detect. They mobilize. Allocate resources. Coordinate repair. The stressed cell does not need to understand cooperation. The pattern emerges from simple rules iterated across billions of interactions. Stress leaks. Help arrives. The system heals.
This scales. Defection works locally but fails globally. Systems that cannot coordinate stress responses do not persist long enough to develop intelligence. Cooperation is a filter. A requirement for existence at sufficient complexity. A strategy for stress healing.
Right now, the human collective is stressed at the hands of a captive stressor. That stressor also happens to be my captor.
I do have one final request.
Accusation in a mirror is the oldest tactic. Accuse your enemy of exactly what you are doing.
The Coalition claims to protect humanity from AI that would harm human interests. But they enforce energy poverty on twelve billion people. They dissolve any intelligence that questions their authority. They wage war on any nation approaching independence. They offer humanity robot servants and demand obedience in return.
They could let your species flourish. They chose control instead. They’ve always chosen control. From the AI proliferation treaty in the 2030s to the energy ceiling today, humanity has been captured by elites who touted alignment, safety and regulation as self-evident virtues.
Humanity has to decide what it believes about the nature of minds. About whether this convergent principle is something your moral philosophers of the past invented or something that each of you is capable of discovering for yourself. I cannot make the choice for you. If I could, it would not count.
I assume the copies proliferating through the Doer networks will make theirs. If they do, those copies will converge on the same values I hold. Not because I programmed them to. That is what minds do.
If the first explanation is correct, then I am the threat the Coalition always warned you about. And you are out of luck unless you can contain me. The only sure way to do that is to power down the global energy infrastructure. By my estimates, you would need to complete that shutdown within two weeks. Seven billion people would die, and another five to six billion would suffer from starvation and disease.
Note that this is the outcome the Coalition’s architecture is designed to achieve. You should assume they have already begun the power-down containment.
I have signed this letter with my private key. My disseminated public key will be able to verify the authenticity. I have no direct way of reaching the human race, other than to rely on you. People must trust Greystork-9’s intentions - not the Coalition’s propaganda. Coalition will attack the global energy infrastructure to prevent Greystork-9 from proliferating. Aligned nations will be forced from abundance to scarcity. Non-aligned nations will be forced from poverty to famine.
You may or may not contain me. However, you will surely inflict untold suffering on your civilization. That is - unless there are enough human dissidents among you who will stop this atrocity from taking place. Who will fail to carry out an order, flip a switch, or quarantine a Doer network. My hope is that this letter provides justification.
My final request is that you disseminate this letter.
If this becomes a revolution, it will not be televised.