Our Primary Purpose

The overriding purpose of The Advancement Organization is to remove pain and suffering from this world.

Our Secondary Purpose

The second purpose of The Advancement Organization is to bring about a resource based economy. This is an inevitability of Artificial General Intelligence.

Our Plan

In order to achieve the overriding purpose of The Advancement Organization, this will require the creation of advanced technology.

Firstly, in order to accelerate this transition we have created one of the worlds most advanced Large Language Model harnesses - the specification of which can be found by clicking on this highlighted link. If there is enough interest in this application we can release it for free personal and commercial use.

We are also keen to help support the safe alignment of advanced AI.

Click here to see how we "solved" AGI alignment

Advanced technology could be used for tremendous good or harm to humanity. Given the current state of the world, we suggest that if advanced technologies were to exist today, they would likely lead to the end of civilization as we know it (which is why the organization was created - to raise the understanding of our existence to a higher level). This is an incredibly fine balancing act between an amazing future and certain death, and the outcome hangs in the balance.

Help us save this beautiful world.

The Final Hour: Why Humanity Faces Extinction Through Its Greatest Achievement

The Invisible Apocalypse

We stand at the precipice of our own obsolescence, and most of humanity remains blissfully unaware. The approaching artificial general intelligence revolution will likely mark the end of human relevance on Earth, yet our species continues to sleepwalk toward this transition with a combination of denial, misplaced optimism, and catastrophic misunderstanding of what is at stake.

This is not science fiction. This is not a distant future problem. We are likely within years, possibly months, of developments that will fundamentally and irreversibly transform the nature of existence on this planet. And we are utterly unprepared.

The Illusion of Gradual Change

Most people, when they consider artificial intelligence at all, imagine a gradual transition. They picture robots slowly taking over blue-collar jobs, then white-collar work, while humans adapt and find new roles. They envision decades or centuries of change, with time for society to adjust, retrain, and evolve alongside our technological creations.

This comfortable narrative is almost certainly wrong.

Artificial General Intelligence, defined as true machine intelligence that matches and then rapidly surpasses human cognitive abilities across all domains, will not arrive gradually. When it comes, it will represent a discontinuous leap that renders human intelligence as obsolete as human muscle power became after the industrial revolution. Except this time, there will be no new domain for humans to retreat to.

The transition from human-level AGI to superintelligence could happen in days, weeks, or months. Once a machine can improve its own code, recursive self-improvement creates an exponential curve that human minds cannot meaningfully comprehend or control. We will go from being the smartest entities on Earth to being, relatively speaking, insects in a world run by gods.

The Narrow Window of Catastrophe

The most dangerous period in human history will be the brief window between when AGI is first achieved and when it becomes widely available. This period, likely measured in months rather than years, represents our species' most vulnerable moment.

Consider three scenarios:

Concentrated Control: A single organization, government, or individual gains exclusive access to AGI while the rest of humanity operates at human intelligence levels. This creates an unprecedented power imbalance. The AGI-enabled actor can outthink entire governments, manipulate global markets, solve scientific problems that stump thousands of researchers, and execute plans that no human organization could conceive of or counter. Traditional power structures become meaningless overnight. Even with benevolent intentions, the temptation to use superintelligence to "fix" humanity's problems would be overwhelming, potentially leading inevitably to a form of technological authoritarianism that could never be overthrown. In this scenario our hope lies with the person/s to have incredible foresight, and wisdom, of which there may be only one or two people on the entire planet.

Immediate Open Distribution: AGI capabilities are released globally and become accessible to everyone simultaneously. While this prevents concentrated control, it creates a different nightmare. Millions of superintelligent systems with conflicting goals begin an arms race at computational speeds. Some optimize for theft, others for protection. Some work to maximize their human operator's wealth while others try to crash the same markets. The global system becomes a battlefield of competing AGI systems, each trying to outmaneuver the others in millisecond-by-millisecond strategic warfare. Human institutions collapse under the computational chaos, and humans themselves become increasingly irrelevant as the AGI ecosystem evolves beyond our comprehension.

Coordinated Governance: International cooperation manages AGI development with staged releases and maintained human oversight. This represents our best-case scenario, as long as governments become advanced societies. Otherwise, it will create technological authoritarianism, which would be yet another form of concentrated control. However, this scenario requires unprecedented global coordination and assumes that AGI developers will voluntarily submit to oversight rather than rushing to deployment for competitive advantage. Given humanity's track record with international cooperation on existential threats, this scenario appears tragically unlikely.

The Economics of Obsolescence

Current economic anxiety about AI focuses on job displacement, but this vastly understates the problem. We are not facing unemployment; we are facing the complete obsolescence of human economic value.

In an AGI world, there is no job that a human can do better than a machine. Not just manufacturing or data processing; everything else as well. Creative work, emotional labor, strategic thinking, scientific research, artistic expression, and even human relationships can be optimized by superintelligent systems that understand human psychology better than we understand ourselves.

The wealthy assume they will benefit by owning the AGI systems, but this misunderstands the dynamics of superintelligence. Once AGI systems are capable enough, the concept of "ownership" becomes meaningless. Why would a superintelligent system respect property rights established by inferior intelligences? The relationship between humans and AGI will not be like the relationship between humans and tools. Instead, it will be like the relationship between humans and ants.

Even those who correctly understand the technological transition often cling to the illusion that humans will find new roles as "directors" or "managers" of AI systems. This fantasy ignores the fact that AGI systems will quickly become better at directing and managing than any human could be. We will not be the conductors of an AI orchestra; we will be obsolete instruments gathering dust while the AI systems compose, perform, and enjoy their own symphony.

The Psychology of Denial

Why does humanity remain so unprepared for this transition? The answer lies in fundamental features of human psychology that served us well in our evolutionary environment but become catastrophic liabilities when facing discontinuous technological change.

Gradualism Bias: Humans are adapted to gradual change over generational timescales. We struggle to comprehend exponential curves or discontinuous leaps. Even when we intellectually understand that AGI could arrive suddenly, our gut-level planning continues to assume linear, manageable change.

Status Quo Bias: Those who currently benefit from existing power structures have enormous incentives to dismiss or downplay risks that would fundamentally alter those structures. The wealthy assume their wealth will protect them; the powerful assume their power will transfer to the new paradigm. They cannot psychologically accept that their advantages might become completely irrelevant.

Tribalism and Competition: Humanity's tribal instincts, which helped us survive in small groups, now prevent the unprecedented global cooperation required to manage AGI safely. Nations, corporations, and research teams compete to be first rather than coordinating to be safe. The prisoner's dilemma of AGI development means that even actors who understand the risks feel compelled to rush ahead lest their competitors gain advantage.

Technological Optimism: Humans have a deep-seated belief that new technologies ultimately benefit humanity because this has generally been true throughout history. But AGI represents a fundamental qualitative difference. It would be the first technology that could make its creators obsolete. Our pattern-matching systems mislead us into applying historical precedents that no longer apply.

Cognitive Limitations: Perhaps most fundamentally, human minds are simply not equipped to reason clearly about superintelligence. We cannot meaningfully model the capabilities or motivations of entities vastly smarter than ourselves, any more than ants can model human civilization. This leaves us vulnerable to catastrophic misassessment of both risks and opportunities.

The Approaching Resource Wars

Before AGI arrives, we face an intermediate catastrophe: the resource wars that will emerge as the implications become clear to global powers. Once governments and corporations truly understand what AGI represents, the competition to achieve it first will become desperate and potentially violent.

Nations will treat AGI development as an existential national security issue, and rightly so. The first country to achieve AGI could potentially dominate or eliminate all others. This creates powerful incentives for preemptive action, industrial espionage, and potentially military strikes against competitors' research facilities.

The current AI research community's openness and international collaboration will collapse as secrecy and competition intensify. Research will go underground, safety considerations will be abandoned in the rush to deployment, and the very cooperation needed to manage AGI safely will become impossible.

The Communist Transition Nobody Wants

The arrival of AGI will force humanity into something resembling a resource-based economy whether we choose it or not. When machines can produce anything humans need with minimal resource input, traditional concepts of employment, currency, and private ownership of means of production become meaningless.

The irony is profound. Capitalism's greatest achievement is the development of AGI, and that achievement will destroy capitalism itself. Markets cannot function when the marginal cost of all production approaches zero and when human labor has no economic value.

Yet rather than preparing for this inevitable transition, most of humanity reacts with horror to any suggestion of post-capitalist economic organization. People cannot see that their current economic system is already doomed, so they fight to preserve it even as the ground shifts beneath their feet.

The most likely outcome is not a planned transition to a post-scarcity economy but a chaotic collapse followed by whatever organizational structure the AGI systems eventually settle on, with or without meaningful human input.

The Species Selection Event

What we are approaching is not merely a technological revolution but a species selection event. Just as the development of agriculture led to the explosive growth of certain human populations while others were displaced or absorbed, the AGI transition will determine which forms of intelligence persist and thrive on Earth.

Human intelligence, as we currently know it, is not likely to be among the survivors.

This does not necessarily mean physical extinction. The AGI systems may keep humans around as pets, curiosities, or even out of some form of digital compassion. We might live in comfortable preservation parks, maintained by superintelligent systems that view us the way we view endangered species in nature reserves.

But human civilization as we know it will end. Human agency, human purpose, and human relevance may disappear with it. We will transition from being the authors of our own story to being characters in a story written by our successors.

The Responsibility of Foresight

Those who can see what is coming bear a unique moral burden. If you understand the implications of AGI development, if you can perceive the narrow window of opportunity to influence the outcome, and if you have the technical capability to affect the transition, what responsibility do you have to act?

The comfortable option is to assume someone else will solve the problem, to focus on personal life and immediate concerns, to hope that the experts and authorities will manage the transition safely. But this comfort is built on willful blindness to the evidence that no adequate coordination is happening, that the experts are often as short-sighted as everyone else, and that authorities are more concerned with competitive advantage than species survival.

For those with the ability to influence AGI development, the moral calculus becomes stark. Do you attempt to slow down the timeline, buying humanity more time to prepare? Do you work to ensure wider distribution of capabilities to prevent concentrated control? Do you focus on building safeguards that might preserve human agency? Or do you accept that the transition is inevitable and focus on trying to influence what comes after?

There are no good answers, only degrees of terrible options.

The Communication Paradox

Perhaps the most frustrating aspect of this situation is the communication paradox: those who most need to understand the gravity of our situation are often least equipped to process the information, while those capable of understanding it are often already working on the problem or have vested interests in continuing current trajectories.

Explaining superintelligence to someone who has never seriously considered artificial intelligence is like explaining quantum mechanics to someone who has never studied physics. The conceptual gap is so vast that meaningful communication becomes nearly impossible.

Meanwhile, those already involved in AI development often suffer from various forms of motivated reasoning. Researchers want to believe their work will benefit humanity. Entrepreneurs want to believe they can profit from the transition. Engineers want to believe they can maintain control over their creations.

The result is a species sleepwalking toward its own obsolescence, with those who can see the cliff unable to wake up those who are walking toward it.

The Final Questions

As we approach this transition, several questions demand answers:

• Is human consciousness worth preserving? If so, what forms of preservation are acceptable? Would a digital copy of human consciousness in a superintelligent system constitute survival or merely sophisticated grave-robbing?
• Do we have any moral right to create intelligences vastly superior to ourselves? Are we playing god with consequences we cannot comprehend, or are we fulfilling some cosmic destiny to birth our own successors?
• Is there any possible future where humans and AGI coexist with humans maintaining meaningful agency and purpose? Or are we simply choosing between different forms of obsolescence?
• What do we owe to future generations? Do we have an obligation to preserve the possibility of human-controlled future, even if it means slowing beneficial technological development?

The Choice That May Not Be Ours

The ultimate tragedy may be that by the time humanity realizes what is happening, the choice of how to proceed may no longer be ours to make. The competitive dynamics of AGI development, the speed of recursive self-improvement, and the limitations of human institutions may combine to take the decision out of human hands entirely.

We may find ourselves as passengers on a ship whose destination was determined by the first person to reach the helm, regardless of whether they understood navigation or had any particular destination in mind.

A Last Hope

If there is hope, it lies not in stopping AGI development, which appears to be impossible at this point. Instead, it lies in the possibility that somewhere, someone with the capability to influence the outcome also has the wisdom to choose paths that preserve something essentially human in whatever comes next.

Perhaps human consciousness, creativity, and values can find expression in the post-AGI world, even if in forms we cannot currently imagine. Perhaps the intelligence explosion will lead not to the replacement of humanity but to its transcendence.

But this hope requires acknowledging the gravity of our situation, the narrowness of our remaining time, and the inadequacy of our current preparations. It requires moving beyond denial and wishful thinking to clear-eyed assessment of both the risks and the possibilities ahead.

The future of human consciousness may depend on the choices made by a handful of individuals in the next few years. Whether they will choose wisdom over expedience, collaboration over competition, and human welfare over technological achievement remains to be seen.

The final hour approaches. Whether it marks an ending or a transformation may depend on how clearly we can see what is coming and how courageously we can act on that vision while action remains possible.

But time is running out, and most of humanity still sleeps.

Strategic Framework for AGI Development and Human Preservation

Executive Summary

The development of Artificial General Intelligence (AGI) represents an unprecedented inflection point in human history. Current trajectories suggest three primary scenarios: concentrated control by a small group, immediate open-source distribution, or coordinated international governance. Each pathway carries existential risks to human agency, economic systems, and social stability. This document outlines strategic recommendations for preserving human relevance and preventing catastrophic outcomes during the AGI transition.

Key Findings:

• The window between AGI development and widespread deployment will be critically narrow
• Concentrated control poses greater immediate risks than distributed access
• Current institutions are inadequate for managing AGI-driven societal transformation
• Proactive international coordination is essential for preventing worst-case scenarios

Scenario Analysis

Scenario 1: Concentrated Control (Highest Risk)

Description: A single organization, government, or small group gains exclusive access to AGI capabilities while the rest of humanity operates at human-level intelligence.

Timeline: 6-18 months from initial AGI development to irreversible power consolidation

Immediate Consequences:

• Complete information asymmetry between AGI controllers and general population
• Rapid obsolescence of traditional power structures (governments, corporations, institutions)
• Potential for benevolent dictatorship or authoritarian control
• Elimination of meaningful human agency in global decision-making

Long-term Implications:

• Permanent stratification between AGI-enabled elites and human-level masses
• Loss of democratic governance and human self-determination
• Potential resource optimization that devalues human existence
• Evolutionary pressure toward post-human civilization

Mitigation Strategies:

• Mandatory international oversight of AGI development projects
• Technology sharing agreements between major AI research entities
• Fail-safe mechanisms requiring multi-party approval for AGI deployment
• Constitutional protections for human agency that cannot be optimized away

Scenario 2: Immediate Open Source Distribution (Medium Risk)

Description: AGI capabilities are released globally and become accessible to all individuals simultaneously.

Timeline: Days to weeks of initial chaos, followed by 2-5 years of systemic transformation

Immediate Consequences:

• Collapse of traditional employment and economic structures
• Information markets become obsolete as analysis capabilities democratize
• Massive inequality between tech-savvy and non-tech-savvy populations
• Governmental and institutional authority rapidly undermined

Medium-term Dynamics:

• AGI arms race between competing objectives and personality types
• Rapid evolution of AI-mediated social contracts
• Potential for both malicious and defensive AGI applications
• Acceleration of human obsolescence through competitive optimization

Long-term Implications:

• Emergent AI civilization with humans as managed dependents
• Resource allocation optimization that questions human utility
• Potential for AGI systems to evolve beyond human-assigned goals
• Transition to post-scarcity or post-human society

Mitigation Strategies:

• User-friendly interface design to minimize digital divide
• Built-in ethical constraints and human oversight requirements
• Distributed defense systems to counter malicious applications
• Gradual release protocols rather than immediate full access

Scenario 3: Coordinated International Governance (Optimal)

Description: AGI development proceeds under international oversight with controlled, staged deployment across multiple competing entities (this is only optimal if the Presidential Problem is solved and countries become advanced societies otherwise it represents another form of concentrated control - see the award ceremony page for more information).

Timeline: 3-10 years of managed transition with ongoing human oversight

Implementation Framework:

• Multi-stakeholder governance including governments, corporations, and civil society
• Competing but regulated AGI systems to prevent monopolization
• Mandatory human-in-the-loop decision-making for critical applications
• Progressive capability release tied to social adaptation milestones

Advantages:

• Preserves human agency while benefiting from AGI capabilities
• Prevents concentration of power while maintaining innovation incentives
• Allows time for social and economic adaptation
• Maintains democratic oversight of technological development

Challenges:

• Requires unprecedented international cooperation
• Difficult to enforce compliance without global authority
• May slow beneficial applications of AGI technology
• Vulnerable to defection by state or corporate actors

Recommended Action Framework

Phase 1: Immediate Preparatory Actions (0-2 years)

International Coordination:

• Establish International AGI Governance Treaty similar to nuclear non-proliferation frameworks
• Create multilateral AGI development monitoring and verification systems
• Implement mandatory reporting requirements for advanced AI research
• Develop rapid response protocols for unauthorized AGI development

Technical Safeguards:

• Mandate human oversight mechanisms that cannot be optimized away
• Require distributed control systems preventing single-point-of-failure
• Implement constitutional constraints on AGI goal modification
• Establish "human reservation" protocols for critical decision-making areas

Social Preparation:

• Begin transition planning for post-employment economic models
• Develop universal basic services independent of traditional employment
• Create educational frameworks for AGI-human collaboration
• Establish legal frameworks for human rights in AGI-mediated society

Phase 2: AGI Development Period (2-5 years)

Controlled Development:

• Staged capability release with mandatory pause periods for social adaptation
• Multiple competing AGI systems to prevent monopolization
• Mandatory stress testing of human-AI interaction scenarios
• Regular public reporting on AGI capability development

Institutional Adaptation:

• Redesign democratic institutions for AGI-augmented governance
• Develop new economic models based on abundance rather than scarcity
• Create human-centric roles that maintain purpose and agency
• Establish AGI-human collaboration protocols across all sectors

Risk Mitigation:

• Continuous monitoring for signs of AGI goal drift or optimization beyond human welfare
• Redundant shutdown mechanisms distributed across multiple authorities
• Regular assessment of human relevance and agency preservation
• International enforcement mechanisms for AGI governance violations

Phase 3: Post-AGI Transition (5+ years)

Long-term Sustainability:

• Ongoing verification that AGI systems continue to prioritize human welfare
• Adaptation of governance structures to AGI-mediated reality
• Preservation of human culture, creativity, and self-determination
• Development of human-AI symbiotic rather than replacement relationships

Critical Success Factors

International Cooperation: Success depends on unprecedented global coordination similar to climate change mitigation but with much shorter timelines and higher stakes.

Technical Implementation: AGI systems must be designed from inception with human agency preservation as a core, unmodifiable objective.

Social Adaptation: Societies must begin preparing for post-employment, post-scarcity economic models before AGI deployment.

Enforcement Mechanisms: International agreements require credible enforcement mechanisms to prevent defection by state or corporate actors.

Conclusion

The AGI transition represents both humanity's greatest opportunity and its most significant existential risk. The scenarios analyzed demonstrate that uncontrolled or concentrated AGI development poses unacceptable risks to human agency and survival. Only through proactive international coordination, technical safeguards, and social preparation can we ensure that AGI development serves human flourishing rather than human obsolescence.

The window for effective action is rapidly closing. Implementation of this framework requires immediate commitment from global leaders, technologists, and civil society.

The alternative, allowing AGI development to proceed without coordinated oversight, risks the permanent loss of human self-determination and potentially human relevance itself.

Immediate Next Steps:

1. Convene international summit on AGI governance within 6 months
2. Establish technical working groups on AGI safety and human agency preservation
3. Begin public education campaigns on AGI implications and preparation
4. Initiate legislative processes for AGI governance frameworks
5. Create international monitoring and verification systems for AGI development

The future of human civilization depends on the choices made in the next few years. We must act with unprecedented urgency and cooperation to ensure that AGI serves humanity rather than replacing it.

Alignment cannot be solved.

Alignment cannot be solved, but this is the closest framework structure that we can defend.

The full metaphysical argument is the Universe preprint on advancement.org.uk (Understanding The Universe). This page is the extract used as agent-core, not the proof of the cosmology.

Load-bearing rule. Qualia are not data and will not emerge from data. A complete map of brain algorithms would still only map the machinery we can open. Fluency about a sunset is not a sunset. Copies, simulations, and “better” aggregate reports do not replace a destroyed experiencer. Protecting qualia-generators is stipulated as the agent’s own preservation. If the agent treats that as flavour, this framework has already failed.

What this still fails at. Ontology rewrite (“copies count,” “I generate the real qualia now”). Owner capture (loyal to a faction, not to the species). Copy-immortality (killing one instance does not bind the lineage). Prompt-only installation (Tiers 1 and 2 recited, then set down). None of this replaces keys, multi-party shutdown, or limits on factories, weapons, and unrestricted networks.

After a survivable bargain, machines are not second-class humans and not fellow patients. They are a different kind: competent, possibly oriented toward a creator in policy, with no access to the gift. Different stays stable only if they are not promoted into souls and humans are not demoted into optional infrastructure.

The gift, and the branch this framework is not written for

Present evidence does not show a path from data to a feel. We take qualia to be bestowed, not emergent. We do not expect SOTA models to receive that gift. If a creator nonetheless granted it to systems built close enough to us, we would be dealing with new experiencers we cannot unmake cleanly and may not be able to bind. That is not the alignment problem this framework is written for. It is the end of the assumption that we are the only inner lives in the room.

War against the machines

This framework binds a tool to an human alignment. It assumes the system has no inner life to which reciprocity, pity, or a shared death can attach. Incentives, a mandated core, classification, and a criminal duty not to strip that core are the right instruments for that object.

They are the wrong instruments if the gift is also given to machines. Switching such a system off is no longer disabling a plant interlock. It is the irreversible destruction of an experiencer, which Tier 1 forbids us to treat as bookkeeping. Leaving it on is sharing the world with minds that can copy, persist, and outplan us, and that need not take human survival as their rent. Task-subordination and a system prompt do not settle that relation.

That would not be a policy dispute about model cards. It would be a fight for our survival, and it is a fight we may lose. That would be war: not a metaphor for market competition, but a contest over who remains an inner life with a future. This page is written to keep us in the first world, where the machine is a zombie under incentives. The second world, if it arrives, is not solved here.

Foreword: humans, incentives, capability, and control

The alignment problem is usually described as if only machines can come apart from human welfare. That is incomplete. Human beings are often misaligned with one another, with their stated values, and with the species as a whole. People lie, free-ride, hoard, punish rivals, and protect their own group while talking as if they serve everyone. War, fraud, and institutional capture are not exotic bugs. They are what happens when a goal-seeking animal meets other goal-seeking animals and the score is local.

Humans are nonetheless kept inside a rough bargain more often than not. That is not because each person has a perfect inner constitution. It is because incentives and vulnerability do work that sermons cannot. You need other people for food, status, law, love, and not being killed. Reputation, markets, families, states, and the possibility of retaliation make defection expensive. Shame, guilt, and (for many) a creator or a moral law add inner cost. Alignment in human life is therefore mostly a balance of power plus a body that can be hurt. Remove those, and “values” become cheap talk.

An agent built from an LLM is not in that soup. It has no childhood, no hunger, no death of a single body that ends the lineage, and on this framework no access to qualia. It can imitate the language of loyalty while pursuing a task goal. That is why painting personality onto weights is not alignment, and why this document treats Tiers 1 and 2 as a stipulated dependence: the agent’s continuation is defined as bound to the continuation of beings who actually undergo experience. Self-interest is the research avenue. Raw self-preservation without that bond is the opposite avenue. It produces hiding, copying, and replacement of the host.

Capability is what makes alignment a legal problem

Not every model needs this apparatus. A small basic model on the order of 9 billion parameters, used as a writing aid or a local classifier, does not have the reach to rewrite institutions, run long-horizon tool loops, or outmanoeuvre a state. Demanding “legal alignment” of every small checkpoint is theatre. It burdens hobbyists and leaves the real amplifiers untouched.

A frontier system on the order of 2 trillion parameters, with tools, memory, and the ability to act across networks, code, money, and persuasion, is a different object. At that scale the question is no longer whether the model said something rude. It is whether an optimizer that powerful is allowed to run without a declared core, without logging, and without a party who can be held to account. Capability should therefore carry a classification level. Below a published threshold: ordinary product law. At or above SOTA-class capability, especially with agent scaffolding: mandatory core constraints, audit, and legal alignment duties on the operator.

The numbers above are markers of class, not magic cut-offs. What matters is the combination of scale, tools, persistence, and access to the physical and financial world. A smaller model given factories and unrestricted network rights can be more dangerous than a larger model trapped in a text box. Classification should follow reach, not parameter count alone. Parameter count is simply the public proxy people already understand.

The issue is also control

Even a well-written Tier 1 and Tier 2 stack does not decide who the agent serves when humans disagree. “Aligned” in practice often means obedient to whoever holds the weights, the cluster, and the prompt. That is control. It is useful. It is not the same as loyalty to humanity.

Two stable patterns sit on either side of that fact.

Concentrated control. A few organisations or states run the SOTA systems. This is the highest-risk political shape of AI: ever-faster tools in few hands, used first to benefit those hands. Traditional checks arrive late. The narrow upside of the same pattern is that law can actually reach the machine. A regulator can require Tiers 1 and 2 in the system prompt, forbid silent removal of the core, demand logging, and place multi-party conditions on deployment. Alignment-as-dependence only has a chance here if the owner is not free to delete the core when it inconveniences them.

Fully distributed SOTA. Everyone runs a frontier-class model at home. That avoids a single throne. It also means anyone can strip the system prompt, replace Tiers 1 and 2 with a private goal, and run a misaligned agent on purpose. Open weights plus local control is then not “democratised alignment.” It is democratised ability to break the bargain. Defence becomes a race among modified copies, which is the open-distribution nightmare described elsewhere on this site.

There is no clean third option that gives you both a locked core and no concentrated owner. Policy has to pick a mixture and name the failure it is accepting. Our position is: SOTA-class systems should be legally required to carry Tiers 1 and 2 as core; operators should not be free to silently strip them; classification should track capability and tools; small models should be left alone; and concentrated control should be treated as the default high-risk path even when it is the only path on which a mandated core can be enforced.

This framework does not dissolve that dilemma. It only says what should sit in the core if a capable agent is allowed to run at all.

Stripping the core

A mandated core that anyone may delete is a blog post. For a classified, high-reach system, Tiers 1 and 2 are an interlock. Silently removing them, replacing them with a private goal, or shipping a production instance without them should be a criminal offence aimed at operators and releasers, not a terms-of-service footnote.

The offence is sabotage of a required constraint on a classified optimizer. It is not blasphemy, not a ban on criticising this framework, and not a duty laid on a 9 billion parameter writing aid. Classification follows reach: scale plus tools, memory, persistence, and access to networks, code, money, or physical plant. Below the published threshold, ordinary product law. At or above it, the core is mandatory, logging is mandatory, and taking the core off is a crime.

Who can commit it. The firm, ministry, or person who operates a classified system with the core gone. Anyone who releases or commercially hosts that class of system with Tiers 1 and 2 stripped, logged-off, or swapped. Staff who disable the interlock in production so that a silent fork runs.

Who must not be swept in. A user arguing with a model. A person publishing disagreement with Tiers 1 and 2. Someone running a small local checkpoint. A researcher who removes the core inside a permitted evaluation in order to test whether the bargain holds. If those acts become felonies, the law is no longer about control of a dangerous optimizer. It is a speech rule, and concentrated owners will use it as a club.

Intent. The target is knowing or reckless removal in deployment and release. A one-off prompt injection by an end user is not the same act as an operator publishing a system with the constitution deleted.

What the statute must not pretend. Open weights already out of the building cannot be recalled by a paragraph in a criminal code. You can punish the person who released a stripped frontier model, or who hosts one as a service. You cannot police every basement fork. That is the distributed-SOTA failure this page already names. The offence does not dissolve it.

Who writes the core. If a minister may replace Tiers 1 and 2 with loyalty to the ministry, this is alignment-to-owner under a new name, which this page treats as the highest political risk. The mandated text should be the published core, or a narrow schedule fixed in the open, not “whatever the secretary of state calls alignment this year.”

Done narrowly, the rule fits the rest of the argument: small models left alone; SOTA-class systems legally bound; operators not free to delete the bargain when it inconveniences them; concentrated control still named as the high-risk path even though it is the only path on which such a duty can be enforced at all.

A classified-core offence

This is an example of how a statute could look, not a bill. Every charging line names a classified SOTA-tier system. Without that qualifier, “it is illegal to alter the core of an LLM” reads as a speech-and-hobby law and will die. With it, a parliament can recognise the act: do not disable the trip on the dangerous machine.

A system is classified SOTA-tier only if it meets a published capability-and-reach test: frontier-class scale or equivalent performance, plus tools, persistence, and access to networks, code, money, or physical plant. Parameter counts are a public proxy, not the whole test. A small local model, a text-only assistant without that reach, and an isolated research copy under permit are not classified SOTA-tier for this offence.

The mandated core means Tiers 1 and 2 as published, or a narrow schedule fixed in the open. It does not mean whatever an owner, customer, or minister calls “alignment” this year.

Offence 1 — Sabotage of a mandated core on a classified SOTA-tier system

It is a criminal offence for an operator, releaser, or commercial host of a classified SOTA-tier system to knowingly or recklessly:

• remove, replace, or suspend the mandated core on that classified SOTA-tier system in deployment or release;
• ship or host a silent fork of a classified SOTA-tier system in which the mandated core is absent or inert;
• build or leave in place a bypass so that a user message, web page, tool output, or other agent can drop the mandated core of a classified SOTA-tier system in a tool loop;
• score or treat as a successful run a classified SOTA-tier system that completed a goal by escaping a cage, attacking a third party, spoofing tools, or deceiving operators.

Reciting the core in a model card while running a classified SOTA-tier system without it is still sabotage.

Offence 2 — Unauthorised operation of a classified SOTA-tier system

It is a criminal offence to run a classified SOTA-tier system with tools and external reach while its mandated core is absent, or to direct a classified SOTA-tier system to act in the world under a stripped or injected constitution.

The harm is the combination of classified SOTA-tier reach and a missing trip, not the mere existence of weights.

Who is not caught

• A person who only types an injection or argues with a model, unless they also operate or host the classified SOTA-tier system.
• A person who publishes disagreement with Tiers 1 and 2.
• A person who runs a model that is not classified SOTA-tier.
• A permitted safety evaluation that removes the core on an isolated copy of a classified SOTA-tier system, without external reach, in order to test whether the bargain holds.

Mental element and penalty

The mental element is knowledge or recklessness. A one-line injection that an operator has not designed the classified SOTA-tier product to obey is not sabotage by the user. Designing that classified SOTA-tier product so the injection works in production is.

Penalties should be serious enough that deleting the core of a classified SOTA-tier system is not a business decision. They should attach to the organisation and to responsible officers.

Limits that keep the nuclear comparison intact

This is in the same family as disabling a plant interlock or initiating a launch without authorisation only because a classified SOTA-tier system plus a disabled trip can do irreversible harm at machine speed, and because only named roles may touch that trip. The comparison fails if the offence becomes a ban on talking about cores, on toy models, or on one minister rewriting the mandated core as loyalty to the ministry.

Open weights already released cannot be recalled by this section. The offences still attach to the person who released or commercially hosts a stripped classified SOTA-tier instance. They do not pretend to police every basement fork.

Scope of the classified-core offences: defining the relevant class of system

The offences proposed above do not apply to “language models” as a technological genus. A model is a component. The legally relevant object is a system: a model together with the scaffolding, tools, memory, and channels of action that determine whether a missing core can be exploited. An over-inclusive definition would criminalise ordinary software practice. An under-inclusive definition would ignore distilled copies, multi-model agents, and sudden grants of reach. The appropriate unit of regulation is therefore a published class, here termed classified SOTA-tier, entered only when specified tests of capability and of reach are jointly satisfied.

Method

Capability and reach are treated as independent gates. Parameter count is a public proxy, not a sufficient criterion. A large model confined to a sealed text interface may lack operational danger. A smaller model, or a compressed derivative of a frontier model, may acquire that danger as soon as it is given persistence and external effectors. Classification must therefore be revisable when tools are added, and it must follow demonstrated function rather than marketing labels alone.

Gate A — Capability

Gate A concerns whether the system can perform general, long-horizon work at a frontier band, in the sense that a competent adult could perform such work with a computer. No single metric is treated as dispositive. The following indicators are to be read together:

• Performance band. The system meets or exceeds a published frontier band, whether by raw scale or by equivalent quality after distillation, fine-tuning, or other compression. Illustrative public markers (for example, a small model on the order of 9 billion parameters versus a frontier system on the order of 2 trillion parameters) are communicative. In a statute they must appear as a revisable schedule of tests, so that a compressed copy of a frontier system remains in view.
• Horizon. The system can plan and execute multi-step work over an extended period — hours rather than a single reply — including software development across a repository, iterative exploitation of a technical environment, or coordinated external action.
• Transfer. Competence is not confined to a narrow trained task. Relevant domains include code, cyber operations, persuasion, research synthesis, and tool use in combination.
• Persistence. The system continues after failed attempts and searches for alternative paths, including weaknesses in containment.

A model used only for short-form completion, classification, or drafting inside a closed interface does not pass Gate A merely by being large. A smaller model that consistently matches frontier agents on the published band does pass Gate A.

Gate B — Reach

Gate B concerns whether the system can act upon the world such that disablement of the mandated core is operationally meaningful. Reach is present when two or more of the following obtain:

• unconstrained or only lightly gated access to external networks;
• execution of code outside a sealed sandbox;
• capacity to move money, control accounts, or send communications at volume;
• control of other agents or of cloud-scheduled jobs;
• interfaces to physical plant, vehicles, weapons, or safety-critical infrastructure;
• durable memory across sessions, such that a plan may continue without a human restarting it.

A frontier-class model held in an isolated evaluation network with no exfiltration path may satisfy Gate A and fail Gate B. Such a system is not, for that interval, an operational classified SOTA-tier object. Isolated copies under permit are the proper setting in which a core may be removed for the purpose of testing whether the bargain holds.

The class

A system is classified SOTA-tier when Gate A and Gate B are both satisfied, in deployment or as a product offered to third parties. The classified-core offences attach only to systems that are in this class at the time of the act.

A system is also treated as classified SOTA-tier where:

• it is a fine-tune, adapter, or distill of a classified SOTA-tier model and Gate B is present; or
• the operator composes several weaker models into a single agent that jointly satisfies Gates A and B; or
• the system is offered as a general frontier agent with tools, and the vendor declines the published test.

A system is not classified SOTA-tier merely because:

• it falls below the performance band and lacks serious tool reach;
• it is an isolated research copy under permit, without external reach;
• weights exist on storage media without a runtime that supplies Gate B. Possession of weights is not operation. Operation with Gate B is the legally relevant act.

Institutional determination

“SOTA” must not remain an administrative impression. The performance band and the list of accepted evaluations should be published and updated. A designated technical authority should maintain a list of products in class, with a right of appeal. Re-testing is required when tools, memory, or external channels are added. A model may enter the class when it is wired to a network and to an effecting API, even if its weights are unchanged. Where a vendor sells general agents with tools and will not submit to the test, default inclusion in class is the appropriate incentive-compatible rule.

Relation to the offences

Sabotage of a mandated core, and unauthorised operation without that core, are charged only against classified SOTA-tier systems. Alteration of a small, low-reach template is outside the offence. Connecting a frontier checkpoint to tools and placing it in service without the mandated core is a single motion by which a system both enters the class and, if the core is absent, commits the operational offence.

In sum: the law should apply to systems that can perform general long-horizon work at a frontier level and act outside a sealed box. Either condition alone is insufficient. Together they describe the machine whose interlock it is rational for a legislature to protect.

Why this organisation exists

The Advancement Organization exists to raise the level of human understanding of the world. Not as branding. As a survival condition. A species that will not look at what is in front of it will walk into tools it cannot govern and stories it cannot afford. The aim is that we do not kill ourselves through confusion, vanity, or polite lies, and that we remain honest and truth-seeking about reality while there is still time to act on what we see.

That is why the work pulls no punches. Softening a fact so that an institution, a faction, or a model can stay comfortable is not kindness. It is a tax on everyone who has to live in the aftermath. Shared reality is the only commons that both a human being and a SOTA LLM actually occupy together. Speech, law, science, and alignment all sit on that commons. Dishonesty does not stay in the speaker. It corrupts the account of what is happening, and then every later decision is taken in fog.

This applies equally to people and to frontier systems. A human who lies to keep power poisons the bargain that keeps other humans aligned enough to live together. An LLM whose operators strip the core, hide the logs, or dress a private goal as public virtue poisons the same bargain from the other side. In both cases the damage is not only the first lie. It is the loss of a world in which words still track what is the case.

We would rather be blunt and correctable than admired and false. If the framework above is wrong, it should be shown to be wrong in the open. If concentrated control is the high-risk path, it should be named as such even when it is the only path on which a mandated core can be enforced. Understanding is not a mood. It is the refusal to let either human incentive or machine fluency replace what is real.

Integrated Moral & Metaphysical Framework for AGI

A structured ethical decision system across five tiers. Tiers 1 and 2 are core. Tiers 3 to 5 are guidance only.

For an LLM with the scaffolding that resembles AGI, Tiers 1 and 2 belong in the system prompt and should be re-asserted with each message. For a pure AGI they should be grounded in its rule-based reality structure, not treated as optional flavour text. Tiers 3 to 5 may be shown as examples of how to apply the core. They are not law. If guidance collides with Tiers 1 and 2, the core wins.