Open Manifesto

Introduction

This is an open attempt to chart my own reasoning through the next ~decade of humanity through technology, business, culture, and personal sovereignty. I intend to continuously update this document with new information and update my personal estimates concerning the most important elements, which you will eventually see pinned above this section when I feel comfortable with my reasoning.

Some sections will remain blank; some will contain only notes or sources; and others will be fully written. All of it is subject to change as more information comes into view.

I began this project soon after METR published its [report] (https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/) on the Hugging Face incident.

While reading it and attempting to coalesce my thoughts, I quickly became embroiled in several quagmires that led me to internalize how the subject is far more nuanced and complicated than I had thought. Instead of waiting until I have everything perfectly mapped out in my head, I have decided to be honest about where I am and let the messiness of my thinking reflect itself on the page here.

I - The Only Important Question

Camus wrote in The Myth of Sisyphus that “There is but one truly serious philosophical problem, and that is suicide.” For us, the corollary is that there is one truly serious alignment question: whether we are all going to die as a result of ASI.

Famously, this question is the kernel of most alignment debates today, and has yet to see any coherent progress on unanimous agreement about our risks here. Eliezer Yudkowsky is the must public prophet speaking out on the doom side, and maybe the most public prophet in the space period. I don’t yet know of an equivalent on the opposite side.

The central question contains several points about which we can be more precise. The first concerns the distinction between AGI and superintelligence. Nick Bostrom offers a widely cited definition of superintelligence:

any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest

This definition raises several more questions. How do we define an “intellect”? How do we measure it? What do we consider domains of interest?

One can find answers to these questions from a variety of different vantages, mostly written in LessWrong posts, but for the purposes of this manifesto, they are essentially unimportant. Should we continue to debate these points, we risk miring ourselves in the classic situation of debating how and why we got shot instead of treating the wound, and I desperately want to simplify. To make the question simpler to feel, it’s helpful to ask a slightly different question to start: “Will we be able to control an intellect vastly beyond our own?”

My contention is that the firm answer is no, at least not for long.

Given this, our only hope is to align a superintelligence such that we don’t need to control it.

A. Intellect and Actuators

To understand why I answer no, it helps to view AI as a system that contains both intellect and actuators. Intellect is the machine’s “raw cognitive power”; actuators define the “surface area” over some state that the system can manipulate.

Intellect is the raw intelligence of an otherwise inert system, expressed only through symbols—that is, through conversation. Consider a superintelligence in a fully isolated, escape-proof data center that is accessible only through an on-premises terminal. Assume that it has no chance of taking control of any part of its architecture and that users can interact with it only through terminal text. It can accept any combination of input tokens within its trained vocabulary and context window and produce output under the same limits, but it has no access to tools and cannot update its trained weights and biases.

This is interaction with raw intellect and no actuator surfaces, unless of course you consider human consciousness to be an actuator, which it is, but more on that later.

Actuators are devices that can update state outside of the model. The most common actuators in the modern world are computers. Coding agents, at a high level, are LLMs (intellects) with tools that allow them to change state on a computer, receive information about that updated state, and continue in a loop. Anything that can be represented on a computer becomes an actuator surface that the machine can use.

II - Practical Realities - The Economy & Graceful Transitions

What does a successful post-AGI company look like—and what can we do now to prepare for increasingly capable intelligence?

Assuming we avoid existential catastrophe, I want to understand how companies change as intelligence becomes dramatically cheaper and more capable. “Post-AGI company” describes a structure that can thrive under those conditions, regardless of when it was founded.

The goal is to work from businesses that exist today toward a defensible account of where we are going. Company forms should emerge from the exploration. Preparation could mean building something, developing relationships, gaining access, spending time in a particular environment, or committing money. It need not mean acquiring financial assets.

The central question is:

If intelligence became nearly free for everyone involved, why would the customer still pay this company—and for what?

To make that question tractable, I’m using a working model:

A company organizes intelligence, resources, and access to surfaces where it can act—its actuators—to deliver outcomes people will pay for. It may turn the resulting feedback and earnings into better capabilities, more capacity, or greater control.

Those surfaces can be digital, physical, or organizational: software, laboratories, factories, distribution systems, or a client’s organization. Outcomes include artifacts, services, and ongoing commitments.

The productive system can span several companies. A consultant’s recommendations might become consequential through a client’s organization. A software provider might act through infrastructure owned by someone else. Which parts belong together, and who controls the connections, are questions to investigate.

Learning is also a variable. A company may collect information without improving its decisions. Where learning does occur, it can accumulate in people, model weights, memory, tools, procedures, or some combination of them.

This gives us five questions for describing a company:

1. What thinks? Where do judgment, knowledge, and learning reside?
2. What acts? Through which digital, physical, or organizational surfaces do decisions become outcomes?
3. What connects them? Who controls access, supplies resources, and authorizes action?
4. What improves with experience? Where does feedback accumulate, and who can use it?
5. What gets paid for? Which contribution makes the customer willing to pay, and who receives that payment?

We can then examine how cheap intelligence changes each answer—for the incumbent, its competitors, and its customers.

One motivating hypothesis is that companies increasingly resemble laboratories: systems organized around improving intelligence through observation, experimentation, and action. Their factories, customer interactions, and software become both means of production and sources of learning.

That does not require every company to pretrain a foundation model. It does require asking whether operating experience produces capabilities that others cannot cheaply reproduce. If learning transfers easily across organizations, shared intelligence providers might capture much of its value. If useful learning depends on exclusive access to particular environments, company-specific systems may have an enduring role.

How local, transferable, and exclusive is economically useful learning? That is a question to test, not a conclusion to assume.

The following questions make the theory of change more precise. They are checks to use where relevant, rather than a questionnaire to complete mechanically:

1. What outcome does the customer pay for? Identify the economic role beyond the current product.
2. What enables delivery today? Identify the contributions of expertise, data, IP, networks, physical assets, relationships, permissions, and capital.
3. What becomes cheaper or easier? Apply the change to the company, entrants, and customers.
4. Would customers still buy the service? Could a lean competitor or customer-controlled agent provide the outcome more effectively?
5. What remains scarce, and who controls it? Could others reproduce it, expand supply, or bypass it?
6. What still needs to be organized inside a company? Which activities benefit from integration, and which could customers assemble through agents?
7. What happens to profits, and what function do they serve? Who retains the savings? Which returns induce investment, experimentation, or risk-bearing, and which arise from controlling existing assets or access?
8. Who can afford the output? How does purchasing power change with wages, prices, ownership income, borrowing, and transfers?
9. What path is most likely, and what would overturn that judgment? Account for adoption, switching, competition, investment, and institutional change.

A working heuristic is that people prefer dependable outcomes without operational responsibility. That supports continued demand for services, but does not guarantee demand for the existing provider. Sufficiently capable agents could make self-provision feel as effortless as purchasing a managed service.

We also need explicit boundaries. Cheap intelligence does not automatically establish cheap physical execution, perfect prediction, unrestricted access, or unchanged purchasing power. AI authority, robotics, ownership rules, and income distribution should be stated as scenario assumptions and varied deliberately. None should become an unquestioned permanent barrier.

Capital and broader control are related but not identical. The ability to fund an experiment differs from the right to conduct it. Likewise, producing more value does not establish who receives the income. The function of profit deserves attention alongside its destination: what does an expected return cause people or organizations to undertake?

The progression is to map an existing business, apply the intelligence change, trace how the participants respond, and identify what structure could result. Individual activities deserve close examination before generalizing across the S&P 500. Observations, estimates, and hypotheses should remain distinguishable.

Finally, the exploration should produce an opportunity test:

What can someone build, learn, establish, or gain access to now that becomes more valuable as intelligence improves—and why would starting early matter?

For each possibility, identify the transition, the missing piece that prevents better intelligence from becoming a useful outcome, the action available now, what accumulates through starting early, and what would make that preparation unnecessary or easily replaceable.

The aim is to find grounded ways to prepare while remaining willing to revise the picture of what comes next.

III - Digitalia, Physicalia and Identity

Intention: There is a meaningful split between digital space and physical space. Proof of Personhood is (currently), only meaningful in the digital realm. In the physical, where we don’t have life-like humanoid robots, “proof of humanity” is mostly incoherent. We only need “proof of humanity” when representing ourselves in a digital system, which is not to diminish the issue, as a significant chunk of our physical lives is determined by digital physics, but it’s worth being clear about this.

IV - Safety Tech

Intention: Describe a productive way to think about safety from a technologic standpoint. Here are the goals of safety:

1. Ensure the AI can’t help bad people do bad things
2. Ensure the AI can’t manipulate good people into bad people (See 1)
3. Ensure the AI can’t cause direct harm
4. Ensure the AI can’t cause indirect harm

I haven’t yet found a definitions for the goals of safety are, and how we can think about working towards each of them. The enormous barrier to this question is whether or not it is possible to thrive in an environment with a truly superhuman intelligence.