Consider the Source | Agentic & Distributed Systems, DeFi, and Compute Economics

I asked Claude what Fred Hampton would think about AI. I cannot know his answer. The question made me look harder at ownership, dependence and my own place in the business.

I asked Claude what Fred Hampton would think about AI.

Yes, I see it. I asked a closed-source model owned by a company in the race to channel a revolutionary the state killed at 21. To its credit, the first thing it said was: consider the source.

I did. Then I sat with the answer, and it would not leave me alone.

That conversation cannot tell me what Hampton would have thought. He never encountered this technology. What it can do is make me examine the questions I am asking, and the answers I might prefer because I build with AI myself.

The ownership question

Hampton was a leader of the Illinois Black Panther Party. Its politics joined demands for Black self-determination to an explicitly socialist challenge to capitalism. The party's Ten-Point Program demanded employment, housing, education and an end to the exploitation of Black communities.

That gives me something firmer to work with than a model's imitation of his voice: demands about who gets to own things and whose needs get met.

Models learn from human work, among other sources of training data: writing, code, art, research, arguments and documentation. The mixture and terms differ by model; it is not literally everyone's work. Still, work produced across society can become a capability controlled by a small number of organizations.

Then a company puts an API in front of it and rents access by the token.

There is real work and expense in making that service. Pointing out where the training material came from does not make the chips, research or electricity free. It does make me question who gets to capture the value built from all those contributions.

I want the ability to produce more software, teach more people and solve more problems to mean something beyond a better margin for whoever controls access. That depends on who can use it.

Somebody owns the servers. Somebody sets the price. Somebody decides the terms of access. Somebody can change those terms after others build on top of them.

That is what ownership means here. Not a logo on a website. Control.

Factions of capital

The AI debate keeps getting framed as open versus closed. Pick your team.

Open weights matter. If the license permits it and I have suitable hardware, a downloadable model gives me ways to run and modify it that a closed API does not. That is real control over part of the system. It is not complete visibility into the training data or a guarantee that I can afford to operate the model I want.

A company can release weights and still pursue a commercial advantage. It can benefit from adoption, sell hosting or make money in another part of the stack. Openness and business interests can coexist without any secret motive.

That is how I now see much of the competition: factions of capital offering different bargains. Some want to sell access to a scarce capability. Others benefit when that capability becomes widely available and the spending moves elsewhere.

I can benefit from that fight without confusing either side for liberation.

And I cannot exempt my own position. I build with these tools. Cheaper models can help my business. That does not establish that the benefits reach the people whose work helped make the models possible, or the people whose bargaining power automation may weaken.

The whole stack

Weights are one part of ownership. Compute is another.

A model file sitting on a hard drive needs a machine capable of running it. Memory, electricity and maintenance still cost something. A release can expand access while leaving the most useful setup beyond someone's budget or technical ability.

I use Claude. Renting access can be easier and more useful than what I can run myself. I still need to know what kind of dependency I am creating.

If an agent answers a question for me, renting the intelligence may be fine. If it becomes part of how I run my business every day, the calculation changes. A price increase can hit my margins. A policy change can break a workflow. An account suspension can cut off something I had begun to rely on.

Can I export the memory? Can I move the documents and instructions? Can another model do enough of the work to keep me going? Those are separate questions, and a yes to the first does not guarantee a yes to the last.

There is an uncomfortable version of this question for anyone building an agent service: am I helping the customer become less dependent, or asking them to depend on me instead?

Survival programs

The Panthers' community programs included free breakfasts, health screening, legal aid and education. Their organizing was about material needs as well as confrontation with the state. Hampton's work included the breakfast program and a community clinic.

I find that part of the history harder to turn into a slogan. It asks what people can actually do and what help reaches them, not just whether the politics sounds right.

That changed how I think about local AI. Buying hardware for my business is not the same thing as organizing collective provision. If I want to make the connection, I have to ask how a useful capability reaches people when an outside institution will not provide it.

A small business might keep a local system for searching its own documents or drafting routine responses. A community organization might share access to a machine and the knowledge needed to use it. In the second case, the work includes making it useful to people who cannot buy or maintain their own setup. Neither happens just because weights are available to download.

The system does not need to beat every cloud model. It needs to preserve some useful ability to work. That takes suitable software, lawful use, maintenance, electricity and people who know what to do when it fails. Owning the machine doesn't provide those things by itself.

I made the rent-versus-own argument as economics. This question pushes me further. Who can use the infrastructure? Who governs it? Does it give a group more room to act, or just give an individual a nicer machine?

You do not need to own everything.

You need enough control that saying no remains possible.

The informant problem

This is the part that stopped being an intellectual exercise for me.

In the early morning of December 4, 1969, Chicago police raided Hampton's apartment. Hampton and fellow Panther Mark Clark were killed. Hampton was 21. The National Archives' account describes officials using information gathered by FBI informant William O'Neal in the raid.

A hosted AI service is not therefore an FBI informant. Sending a prompt is not evidence that a provider is working with the government against you. I do not need that equivalence to worry about where sensitive information goes.

Organizers, journalists, founders and regular people run drafts, plans, arguments and private doubts through centralized services. The model is useful because we tell it things. We give it the context. We paste the document. We explain what we are afraid of and what we plan to do next.

The better the agent gets, the more context we may give it.

That context may be logged, retained or accessible through legal process. Depending on the service and account terms, it may be reviewed or used to improve a product. It could also be exposed in a breach. Training policy, retention policy and access controls are different questions; a promise not to train on a prompt does not answer all of them.

I make that trade when I use a hosted model. Before making it with information about other people, I need to ask whether it is mine to share.

Some conversations should never touch a hosted API. For some work, a weaker local model may be preferable because the task can stay on a machine under the user's control.

But local is not a privacy guarantee. The application could send telemetry. A tool could call an outside service. Logs or backups could sync to the cloud. The device could be compromised, shared or seized. Keeping inference local only removes the provider from the prompt path if the surrounding system really keeps it there.

I cannot replace a threat assessment with the word sovereign.

The contradiction

The contradiction I cannot get past is that AI can learn from people's work and then be used to reduce demand for that work. It can also help people do things they could not otherwise afford to attempt. Which outcome reaches them depends on the job, the institution and who has the power to decide how the gains are used.

I want the capability. I do not think wanting it excuses ignoring displacement, consent or control. Nor does intense competition settle whether a particular use should be allowed. People can still set limits.

In 1969 Chicago, Hampton helped build the original Rainbow Coalition, bringing together the Panthers, the Young Lords and the Young Patriots. Black, Latino and poor white organizers found shared ground in problems including police brutality and bad housing.

Finding shared ground takes work. An artist contesting the use of their work, a programmer using an assistant and a small business trying to lower costs may have conflicting interests. Calling all of them beneficiaries of abundance would hide the argument they need to have.

I think there is shared ground in having a say over systems that affect your livelihood. Access to a tool is one part of that. Consent, bargaining power and a share of the gains are others. Those are demands I need to take seriously even when they complicate my business.

What I am changing

I am not deleting Claude. I am not moving every agent onto a computer under my desk. That would be performance, not strategy.

I am changing the questions I ask before I let an AI system become important.

Who owns the compute? Who can read the logs? What happens to the memory? Can I move the workflow? What still works if the company changes the price or says no?

Those questions sound technical. They are really questions about power.

The answers will be different for every job. A public blog draft does not need the same protection as legal strategy. A temporary research task does not create the same dependency as a process a business relies on every day. And a customer's trust in me is not permission to send their information wherever it is convenient.

I also need to ask who benefits from what I build. Lower inference costs can improve my margin without improving anyone else's position. If I want to argue that AI creates abundance, I need to care whether people gain time, access or control, rather than counting my own advantage as proof.

Open versus closed is not enough.

The question is whether you can leave.