Back to Writing

Being Able to Use Powerful AI Is Not the Same as Being Able to Keep Using It

What the Restrictions on Claude and GPT-5.6 Reveal About Why doll Is Local-First

Published:

A powerful AI model is announced.

You can use it today.

You pay for access, build workflows around it, and begin to depend on it for work, development, and thought.

None of that guarantees that you will still be able to use it tomorrow.

In June 2026, a series of events involving Anthropic and OpenAI made that distinction unusually clear.

This is not only a story about US AI policy. It matters to anyone who has started to treat a cloud AI system as part of their daily work, memory, or decision-making process.

It also goes directly to the reason doll is designed around one governing rule: the system must remain useful locally before cloud access is added.

Even a Released Model Can Disappear Without Warning

On June 12, the US government directed Anthropic to suspend access to Claude Fable 5 and Claude Mythos 5 by foreign nationals. The restriction covered people outside the United States, foreign nationals inside the United States, and Anthropic's own foreign-national employees. Anthropic described the directive in an official statement.

Anthropic said it could not immediately enforce the nationality-based restriction with sufficient confidence. To comply, it disabled both models for every customer.

The situation was later relaxed, but only in part.

On June 26 in the United States, June 27 in Japan, Anthropic said it had been told that Mythos 5 could be redeployed to a set of US organizations that operate and defend critical infrastructure. Anthropic announced the limited redeployment here.

That was not a full return to normal access.

Mythos 5 did not become available to everyone, and Fable 5 did not return to general use. Anthropic said it was still working with the government to expand Mythos 5 access and restore Fable 5 more broadly.

OpenAI faced a different form of intervention.

On June 26, OpenAI announced the GPT-5.6 family. It did not make the models broadly available at launch. At the US government's request, OpenAI began with a limited preview through Codex and the API for a small group of trusted partners. OpenAI also said that information about participating organizations had been shared with the government. OpenAI described the arrangement in its official announcement.

OpenAI said it expected broader access within weeks, so GPT-5.6 was not banned and its public release was not cancelled. Even so, a government-involved access stage had been inserted before broad availability.

The mechanisms were different.

Anthropic had already released its models and then disabled them in response to a government directive. OpenAI changed the release path before broad public access in response to a government request.

The practical lesson is the same:

A capable AI system can exist, work, and even have paying users without the developer retaining complete control over who can use it, when they can use it, or under what conditions.

Cloud AI Can Be Used, but It Cannot Be Owned by the User

We can pay for cloud AI.

We can connect its API to a product, redesign a workflow around it, and accumulate months of conversation and context.

That does not mean we own the model.

What the user normally purchases is access under the current conditions. There is no guarantee that those conditions will remain unchanged.

Access may be affected by:

Government action was central in the Anthropic and OpenAI cases, but the underlying structure does not depend on government involvement. When the cloud side changes, the user cannot reject the change and keep operating the old service on the old terms.

You can pay for the right to use a cloud AI today. You cannot buy a guarantee that the same AI will still exist for you tomorrow.

Losing One Model Is Not the Only Risk

It is easy to say that users can simply switch to another model when one becomes unavailable.

For simple question answering or one-off text generation, that may be enough. Long-term use creates a more difficult problem.

Over time, an AI environment accumulates:

To the user, the AI is not only the underlying model. It is the model together with the memory, settings, data, tools, and working methods that have accumulated around it.

This is why model switching and environment continuity are not the same thing.

A new provider may be able to answer the same questions. If it cannot inherit the previous memory, settings, work state, and tool configuration, the user's practical AI environment still has to be rebuilt from the beginning.

The real risk is not merely losing access to one version of GPT or Claude.

It is losing the personal AI environment that was built on top of a model the user never controlled.

doll Is Not Trying to Protect Maximum Performance

doll is not an attempt to build a model that outperforms OpenAI or Anthropic.

It is not a project that rejects cloud AI.

Today's leading cloud models are more capable than many local models. They are likely to continue leading in complex reasoning, coding, long-context work, multimodal understanding, speed, and access to new capabilities.

doll should use those models when they are available and appropriate.

If OpenAI is the best option for a task, doll should be able to connect to OpenAI.

If Anthropic is a better fit, doll should be able to use Anthropic.

If another service becomes more capable, the user should be able to select it.

What doll must not do is make any one model identical to the user's AI environment.

doll is intended to keep the following on the user's side:

In doll, the model is not the whole AI.

The model is one replaceable source of capability.

If one model is withdrawn, the system should be able to connect to another.

If cloud access is lost, the system should be able to fall back to a local model.

Performance may decrease.

The user's memory, identity, settings, and data should not disappear with it.

doll is not designed to preserve the highest possible performance at all times.

It is designed to preserve continuity.

Why Local Comes First

doll's highest-level design rule is that the system must first work locally.

This is not because local models are always better than cloud models.

It is because a cloud foundation makes the entire AI environment dependent on the continued availability of that cloud service.

doll therefore aims to keep at least the following under the user's control:

Local and cloud models can then be attached to that durable layer.

user-owned memory, identity, settings, and data
                         ↓
                       doll
                         ↓
              local model / cloud model

When cloud access is available, doll can use the additional performance.

When that access is lost, the user can switch to another provider or to a lower-capability local model.

The model below the continuity layer may change. The user's accumulated state above it should remain.

That is the basic design idea behind doll.

Local Operation Does Not Solve Everything

Local-first design has real costs.

Local models are limited by the hardware available to the user.

Without a powerful GPU or large amounts of memory, the range of practical models may be narrow. Local models may be slower or substantially less capable than cloud services. Support for new features may also arrive later.

doll does not assume that local models can permanently match the frontier.

It also cannot make every government rule, provider decision, or service shutdown irrelevant.

The goal is not to maintain identical performance without the cloud.

The goal is to avoid a design in which nothing remains when the cloud disappears.

Use the cloud for performance.

Do not make memory, identity, core data, and basic continuity depend entirely on it.

This is not a choice between cloud and local AI. It is a division of responsibilities: cloud for additional capability, local ownership for continuity.

What These Events Mean for doll

The Anthropic and OpenAI cases do not prove that doll is necessary for everyone.

Mythos 5 access may expand. Fable 5 may return. GPT-5.6 may become widely available on the schedule OpenAI described.

Whether those particular models return is not the only issue. The events established several practical facts.

A released model can still be suspended

Therefore, doll must not treat one model as the system itself.

An announced model may not become broadly available immediately

Therefore, doll's basic operation must not depend on future cloud access arriving as expected.

Access conditions can vary by country, nationality, or organization

Therefore, doll must not assume that anyone who can pay will always receive the same capability on the same terms.

A provider may not have sole control over release decisions

Therefore, user memory and data should not exist only under the provider's control.

Cloud models remain extremely valuable

Therefore, doll should not reject them. It should use them as replaceable performance extensions.

These events do not prove that every risk doll anticipates will occur.

They do show that the underlying concern is not a distant or purely theoretical possibility.

Own the Continuity, Not the Capability

Claude Mythos 5 has returned for some US organizations.

GPT-5.6 may become broadly available within weeks.

The larger issue is not the date on which one model returns.

A powerful cloud AI can have its release timing, eligible users, price, features, limits, and regional access changed by companies and governments. The user does not control those decisions.

doll is not a project for defeating cloud AI.

It is an attempt to use the capability of cloud AI without allowing the loss of one provider to erase memory, identity, settings, data, work state, and basic operation.

The most capable model will continue to change.

The model that is best today may not be available tomorrow.

What has accumulated between the user and the AI does not have to disappear each time the model changes.

Being able to use powerful AI is not the same as being able to keep using it.

doll is intended to keep continuity on the user's side, even when capability has to come from somewhere else.

Sources

Related writing

Published:

Is Access to Frontier AI Becoming Permissioned?

A source-based look at how pre-release review, export controls, and government-selected early access are changing the continuity risks of relying on cloud AI.

Published: · Updated:

AI Will Remain. Your Access Conditions May Not.

A translated conversation with GPT-5.5 Thinking about whether the conditions that made doll seem necessary are already beginning to emerge.

Published:

The AI Divide Will Be About Permission, Not Access

How ordinary restrictions on powerful AI agents may point toward a wider division based on which capabilities people and institutions are permitted to use.