DEV Community: Sohachi

The Best AI in the World Is Useless If You Can't Access It

Sohachi — Wed, 08 Jul 2026 14:46:23 +0000

China is reportedly considering whether its most advanced AI models should remain freely accessible overseas.

That matters because the United States has already shown what government involvement in frontier AI access can look like.

In the last few weeks, we have seen:

already-released AI models become unavailable because of government policy;
the broad rollout of a frontier model delayed while a company worked with government officials;
and now, reports that China is discussing possible restrictions on overseas access to its own advanced models.

We spend a lot of time asking which AI model is best.

But another question may be becoming just as important:

Will you still be able to use it tomorrow?

China is now discussing the access question

On July 7, Reuters reported that Chinese authorities had met with major technology companies to discuss possible restrictions on overseas access to China's most advanced AI models.

Alibaba, ByteDance, and Z.ai were among the companies reportedly involved in the discussions.

The talks were said to include not only closed models, but potentially more open forms of model distribution as well.

Nothing has been decided.

China has not blocked foreign access to its leading AI models.

The restrictions may never happen.

But the fact that the question is now being discussed is significant.

For the last few years, it was possible to imagine AI becoming something close to a global software market.

American companies would compete with Chinese companies.

Open-weight models would cross borders.

Developers would choose models based mainly on capability, price, latency, context length, and product quality.

That assumption is becoming harder to make with confidence.

Because this is not happening in isolation.

The United States has already shown another version of the same problem.

A model can still exist and become unavailable

On June 12, the US government applied export restrictions to Anthropic's Claude Fable 5 and Claude Mythos 5.

According to Anthropic, the directive required the company to prevent foreign nationals from accessing those models.

Anthropic said it could not reliably determine nationality in real time.

So it suspended access for everyone.

The model did not disappear.

The company did not shut down.

The infrastructure did not fail.

Policy changed, and access disappeared.

The restrictions were later lifted, and Anthropic announced that Fable 5 would return globally.

So this particular interruption was temporary.

But the larger lesson remains.

A frontier AI model can exist, work, and remain technically operational while policy determines whether users are allowed to reach it.

That is no longer a theoretical risk.

OpenAI showed a different version of the same shift

OpenAI's GPT-5.6 rollout followed a different path.

When OpenAI announced the model family on June 26, it did not immediately make the models broadly available.

OpenAI said it had shown the models and its release plans to the US government.

At the government's request, the company began with a limited preview for a small group of trusted partners.

OpenAI also said it did not want this kind of government-access process to become the long-term default.

On July 8, Reuters reported that the broader GPT-5.6 rollout was moving forward after additional testing and discussions between OpenAI and US officials.

The important point is not simply that GPT-5.6 is becoming more widely available.

It is that the timing and scope of a frontier-model rollout became something discussed with government before broad public access.

This does not mean every AI release is now directly controlled by the state.

But national-security policy is moving much closer to the point where advanced AI reaches ordinary users and developers.

What happens if AI stops being one global market?

It is too early to say that AI will split into national blocs.

That outcome is not inevitable.

I hope it does not happen.

But it is no longer difficult to imagine.

One possible future could include:

a US-centered AI ecosystem;
a Chinese AI ecosystem;
separate European regulatory requirements;
different access conditions depending on country, nationality, industry, or organization.

In that world, asking "What is the best model?" is not enough.

You also have to ask:

Can I access it from where I live?
Can my organization use it?
Will the API remain available in my region?
Can I export my data if access changes?
Can I move my work to another model?
What still works if the provider, policy, or network changes?

Today, choosing a provider because it is cheaper or more capable may be completely rational.

The problem begins when years of memory, work, configuration, and dependency are built around the assumption that the same access conditions will continue indefinitely.

Provider lock-in may become a bigger problem than we expected

Developers already understand provider lock-in.

We know that an API can change.

Pricing can change.

Models can be retired.

Terms can change.

A company can shut down a service.

But frontier AI introduces another layer.

The provider may not be the only actor that determines access.

Government policy may influence:

who can use a model;
where it can be used;
when it can be released;
whether overseas access remains available;
and which organizations receive early access.

That changes the continuity problem.

It is no longer only:

What happens if my provider changes its product?

It may also become:

What happens if access to the provider itself changes because of policy?

AI resilience may matter as much as AI performance

None of this is an argument for abandoning cloud AI.

Frontier cloud models are likely to remain more capable than local models for many tasks.

They may offer better reasoning, better coding assistance, larger context windows, stronger multimodal capabilities, and access to infrastructure that most individuals cannot run locally.

Use them.

Use the best model available for the work in front of you.

But using a powerful service is different from designing your entire digital life so that nothing survives when access to that service changes.

I think AI resilience may become as important as AI performance.

That means designing systems where:

models can be replaced;
providers can be changed;
data can be exported;
memory does not exist only inside one company's service;
a local model can provide minimum continuity when cloud access is unavailable;
the user's accumulated state survives even when the reasoning engine underneath it changes.

A local model does not have to beat the best cloud model to be useful in this design.

It can simply keep the system alive.

It can preserve access to the user's own state.

It can provide a fallback path until another source of capability is available.

Why I am building doll

This is one of the reasons I am building doll.

doll is not an attempt to build a foundation model that competes with OpenAI, Anthropic, or Chinese AI labs.

It is an attempt to keep memory, identity, settings, data, work state, and continuity on the user's side while allowing the underlying AI model to change over time.

The design principle is simple:

local continuity first, cloud capability when available.

Cloud models can be used for additional performance.

Local models can provide a continuity path.

The model can change without forcing the user's entire AI environment to disappear with it.

I do not know whether AI will actually divide into national blocs.

I do not know whether future access restrictions will become broader or more permanent.

I hope they do not.

But after the events of the last few weeks, the possibility is harder to dismiss.

The most important AI may not be the smartest one.

It may be the one you can still use tomorrow.

Sources

Originally published in a longer form on the doll website.

The Gate Before the Public

Sohachi — Sat, 27 Jun 2026 11:30:43 +0000

Claude Mythos 5, GPT-5.6, and who gets access before everyone else

In my previous article, “Is Access to Frontier AI Becoming Permissioned?”, I looked at signs that the US government was becoming involved not only in frontier-model safety evaluation, but also in decisions about release timing and early access.

At that point, some of the information surrounding GPT-5.6 was still based on reporting. Anthropic’s Claude Fable 5 and Claude Mythos 5, meanwhile, had been disabled for every customer following a US government directive.

The situation changed again on June 26.

Anthropic announced that Mythos 5 could return for a limited set of US organizations. OpenAI officially introduced the GPT-5.6 family, but began with a restricted preview for a small group of partners at the US government’s request.

Neither development can be summarized as a simple ban or a full reversal.

Anthropic did not restore Mythos 5 to everyone.

OpenAI did not cancel GPT-5.6’s broader release.

What emerged instead was a gate between a finished frontier model and the public.

That gate is not one clearly defined law or licensing system. It is a mixture of export controls, government requests, security reviews, trusted-partner programs, and non-public eligibility lists.

But its practical function is becoming visible:

A frontier model may be ready, working, and commercially valuable while broad public access still depends on a separate review and selection process.

Mythos 5 Has Not Fully Returned

On June 12, the US government issued an export-control directive requiring Anthropic to suspend access to Fable 5 and Mythos 5 by foreign nationals.

The restriction did not apply only to users outside the United States.

It also covered foreign nationals inside the country and Anthropic’s own foreign-national employees.

Anthropic said it could not immediately separate access by nationality with sufficient confidence. To comply with the directive, it disabled both models for every customer.

Anthropic explained the original suspension in an official statement.

On June 26, Anthropic announced that the situation had partially changed.

The company said the government had informed it that Mythos 5 could be redeployed to a set of US organizations that operate and defend critical infrastructure.

Anthropic announced the limited redeployment on its official X account.

Anthropic said it was restoring access for those organizations and would continue working with the government to expand Mythos 5 access and make Fable 5 generally available again.

That wording matters.

Mythos 5 did not return for all US customers.

It did not return for the general public.

Fable 5 also remained unavailable for general use.

The most accurate description is therefore:

Mythos 5 was allowed to return for a limited set of US organizations involved in operating and defending critical infrastructure.

This was not a complete repeal of the original restrictions.

It was a controlled reopening.

“More Than 100 Organizations” Is a Reported Figure

Reuters reported, citing an anonymous person familiar with the matter, that more than 100 companies and institutions would be able to use Mythos 5.

The group reportedly includes several Fortune 500 companies.

That figure should be treated carefully.

Anthropic’s official announcement did not provide a number. It described the recipients only as a set of US organizations that operate and defend critical infrastructure.

The reported total also does not mean that more than 100 US government agencies received access.

It reportedly includes companies, infrastructure operators, institutions, and government organizations.

The complete list has not been made public.

Reuters also reported that organizations on the approved list, their foreign-national employees, and Anthropic’s foreign-national employees would not need new export licenses to use Mythos 5.

Restrictions would remain for organizations outside that list.

Reuters reported the details of the limited return here.

The most important question is not whether the total is exactly 100, 110, or 120.

It is this:

Why were those organizations allowed through the gate while other companies, researchers, and developers remained outside it?

The selection criteria and the full list remain outside public view.

GPT-5.6 Began on the Other Side of the Same Gate

On the same day, OpenAI officially announced the GPT-5.6 family.

The lineup includes:

GPT-5.6 Sol, the flagship model
GPT-5.6 Terra, designed to balance performance and cost
GPT-5.6 Luna, the lower-cost option

OpenAI described Sol as its most capable model.

It said Terra would offer performance competitive with GPT-5.5 at roughly half the price, while Luna would be its lowest-cost model.

But GPT-5.6 did not become broadly available at the time of the announcement.

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.

The company did not publish the names or number of those organizations.

OpenAI described the limited preview in its official announcement.

OpenAI said it still planned to make Sol, Terra, and Luna broadly available within weeks.

GPT-5.6 was therefore not banned.

Its wider release was not permanently cancelled.

What changed was the route to the public:

A government-involved restricted-access stage was inserted before broad availability.

Anthropic’s model was stopped after release and then partially restored.

OpenAI’s model entered a restricted preview before general release.

The timing was different, but both models encountered a gate before reaching a wider audience.

Anthropic and OpenAI Are Not the Same Legal Case

The two cases should not be treated as if the companies received identical orders.

Anthropic had already released Fable 5 and Mythos 5.

A legal export-control directive was then issued, and Anthropic disabled both models for every customer.

Mythos 5 was later allowed to return for a limited set of organizations.

OpenAI’s case was different.

GPT-5.6 had not yet been broadly released. OpenAI consulted with the government and accepted a government request to begin with a restricted preview.

There is no public confirmation that OpenAI received the same kind of export-control directive that Anthropic received.

The timing and legal mechanisms were different:

Anthropic faced a post-release legal restriction.
OpenAI changed its pre-release rollout through cooperation with the government.

That distinction matters.

A compulsory export-control directive is not the same thing as a company agreeing to a government-requested preview process.

But the practical overlap remains significant.

In both cases, the government became involved in determining which organizations could receive frontier-model access first.

For Anthropic, that involvement appeared through the list of organizations allowed to regain access.

For OpenAI, it appeared through the selection of restricted-preview partners.

The gate was built differently in each case.

Its effect was similar.

The New Development Is What Happens Before the Public Arrives

Government regulation of strategically important technology is not new.

Semiconductors, telecommunications equipment, encryption technology, and dual-use products have long been subject to export controls and national-security review.

What is changing is the point at which the government enters the AI release process.

A simplified version of the traditional model-release process looked like this:

Company develops a model
        ↓
Company evaluates safety
        ↓
Company decides pricing, timing, and eligible users
        ↓
Broad release

The emerging process looks more like this:

Company develops a model
        ↓
Company briefs the government on capabilities and release plans
        ↓
Government-involved security evaluation
        ↓
Restricted access for selected organizations
        ↓
Decision on broader availability

This direction also appears in the executive order signed on June 2.

The order calls for classified benchmarks to evaluate advanced cyber capabilities and for a voluntary framework under which developers can provide the federal government with access to models before releasing them to other trusted partners.

It also anticipates cooperation between developers and the government in identifying those trusted partners.

The White House published the executive order here.

The order includes an important limitation.

It states that the framework should not be interpreted as creating mandatory government licensing, preclearance, or permitting for the development, release, or distribution of new AI models.

The United States has therefore not formally created a general rule requiring government approval before every frontier model can be released.

Instead, several different mechanisms are operating at the same time:

legal export controls
voluntary government cooperation
pre-release security evaluation
restricted previews
non-public lists of eligible organizations
trusted-partner selection

There is no single gatekeeper and no single gate.

There is an emerging collection of checkpoints.

Safety Review and Early-User Selection Are Different Questions

There are legitimate reasons to evaluate frontier models before releasing them widely.

Models such as Mythos 5 and GPT-5.6 Sol have strong capabilities in software development, vulnerability research, and long-running cyber tasks.

OpenAI reported that GPT-5.6 Sol achieved performance competitive with Mythos Preview on ExploitBench while using roughly one-third as many output tokens.

That does not mean GPT-5.6 is superior to the final Mythos 5 model in every area.

It is a result from one cyber benchmark against a preview version of the competing model.

OpenAI also said that Sol had not crossed its own “Cyber Critical” threshold.

The model could identify browser vulnerabilities and components needed for an exploit, but it could not autonomously complete an end-to-end working attack during testing.

At the same time, OpenAI acknowledged that benchmark testing may not fully capture what a model can do when combined with tools, other agents, or longer operational workflows.

Pre-release testing and staged deployment can therefore be reasonable.

But testing a model and choosing its first commercial or institutional users are not the same decision.

Organizations that receive early access may be able to begin:

integrating the model into products
automating internal work
developing software
researching vulnerabilities
testing new services

before competitors and independent developers can do the same.

Even a few weeks of early access may matter if the model represents a significant capability increase.

A security process can therefore create a commercial advantage for the organizations selected to participate.

Neither Anthropic’s complete recipient list nor OpenAI’s restricted-preview partner list is public.

The public also does not know:

how organizations are selected
whether additional organizations can apply
why an organization may be rejected
whether there is an appeal process
how long restricted access will last

The problem is not the existence of safety review.

The problem is that review and customer selection are being combined inside a process that remains largely opaque.

OpenAI Does Not Want the Gate to Become Permanent

OpenAI explicitly said that the current access process should not become the long-term standard.

The company warned that such a system could keep its best tools away from:

users
developers
businesses
cyber defenders
international partners

OpenAI did not reject government security testing.

It described the restricted preview as a short-term path toward making the models more broadly available within weeks.

It also said it was working with the administration on a repeatable process for future releases.

Reuters reported that Sam Altman supported extensive safety testing but objected to the idea of the government choosing customers.

Reuters reported on OpenAI’s delayed broader rollout and government involvement here.

That distinction is important.

Testing whether a model presents a national-security risk is one question.

Allowing the government to decide which companies receive the model first is another.

A gate designed for safety can become a gate that distributes competitive advantage.

A Later Public Release Would Not Erase the Precedent

OpenAI may release GPT-5.6 broadly within weeks, exactly as planned.

Anthropic may expand access to Mythos 5.

Fable 5 may eventually return to general availability.

If that happens, the current restrictions may appear temporary.

But a later public release would not erase what happened before it.

A repeatable pattern is now becoming possible:

A developer briefs the government before release.
Government agencies evaluate the model’s capabilities.
A limited set of organizations receives early access.
Broader availability is considered later.

The central issue is not whether GPT-5.6 is delayed for two weeks or four.

It is that the path between a finished model and the public now includes actors beyond the company that built it.

OpenAI itself said it was working on a process that could be used again for future releases.

That means the gate may not be temporary even if a particular delay is.

This Is Not Yet a General AI Licensing Regime

It would be inaccurate to say that frontier AI can no longer be released in the United States without government permission.

Anthropic and OpenAI did not face the same legal process.

GPT-5.6 was not prohibited.

Mythos 5 was not restored to the public.

The executive order describes a voluntary framework and explicitly rejects the interpretation that it creates mandatory licensing or preclearance.

A more accurate description is:

The US government is beginning to participate in frontier-model evaluation, release timing, and restricted early access through several overlapping legal and voluntary mechanisms.

This remains a transitional system.

The rules are not unified.

The selection criteria are not transparent.

The boundary between a government request and a practical requirement is not always clear.

What is new is not merely that AI is regulated.

What is new is that the decision to open the door is no longer always made by the company alone.

Why This Matters to Users and Developers

This is not only a policy issue for AI companies and governments.

Cloud AI access is no longer determined only by price, subscription tier, and terms of service.

Government action may cause:

an already released model to disappear
a planned public release to be delayed
access to differ by country or nationality
access to depend on an employer or approved organization
selected companies to receive capabilities before everyone else

A model can exist, work, and have paying users while access to it remains outside the user’s control.

That creates a separate question from model performance:

Can users continue their work when access conditions change?

I wrote separately about what these events mean for doll’s local-first design and for users who depend on cloud AI continuity:

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

Conclusion

Anthropic disabled two released models following a US government directive.

Mythos 5 has now returned, but only for a limited set of US organizations connected to critical infrastructure.

OpenAI introduced GPT-5.6 through a government-requested restricted preview before its planned broader release.

The legal mechanisms are different.

The direction is the same.

A frontier model no longer moves directly from the company that built it to the public.

Government security review, policy coordination, trusted-partner access, and release sequencing are becoming part of that path.

This is not yet a completed government licensing regime.

It is not one formal gate with one published rulebook.

But before the public reaches the newest frontier models, a new layer of review and selection is beginning to take shape.

The question is no longer only whether a model is safe enough to release.

It is also:

Who gets through first?

Sources

Is Access to Frontier AI Becoming Permissioned?

Sohachi — Fri, 26 Jun 2026 12:20:13 +0000

What the reported GPT-5.6 rollout and Anthropic's Fable 5 suspension suggest about cloud AI

If your work depends on a frontier AI model, the important question is no longer only how capable that model is.

It is also who can delay, restrict, or withdraw your access to it.

On June 25, several news organizations reported that OpenAI was preparing a staged preview of a model referred to as GPT-5.6. According to those reports, the model would initially be offered to roughly two dozen partners before any broader rollout.

Reuters reported that the US government had asked OpenAI to stagger the release over security concerns. The Information, as cited by Reuters, said access during the preview would be approved customer by customer.

Reuters reported the arrangement here.

The distinction between reporting and official confirmation matters.

At the time of publication, OpenAI had not publicly announced GPT-5.6 or confirmed the reported rollout. The reports do not establish that the model has been permanently withheld, that a broader release has been cancelled, or that the United States now requires a government license before every frontier model can be released.

But this was not an isolated story.

Less than two weeks earlier, Anthropic had officially disclosed that a US government directive concerning foreign-national access had led the company to disable Claude Fable 5 and Claude Mythos 5 for every customer.

The two cases are legally different.

They still raise the same practical question:

Who decides who may use the most capable AI systems, and what remains available to users when that decision changes?

Two Different Cases, One Continuity Problem

On June 12, Anthropic said the US government had issued an export-control directive requiring the company to suspend access to Fable 5 and Mythos 5 by foreign nationals, whether they were inside or outside the United States.

Anthropic said the restriction also applied to its own foreign-national employees.

Anthropic's official statement is available here.

The directive did not itself order Anthropic to disable the models for every customer.

Anthropic said it chose a complete shutdown because it could not immediately enforce the nationality-based restriction with sufficient confidence. Its other models remained available.

That distinction matters:

the government required a narrower restriction;
the provider's compliance response produced a broader shutdown.

For users, however, the outcome was simple.

The models existed. They worked. Customers had access. Then a government decision, followed by the provider's implementation of it, made them unavailable.

No AI company had to collapse. The internet did not have to disappear. AI services did not have to be banned as a category.

A narrower policy decision was enough to remove access to a particular model.

Anthropic also said the directive did not explain the national-security concern in detail. The company believed the issue involved a way to bypass Fable 5's safeguards, but disputed the severity of the demonstrated risk.

Users lost access before that disagreement was resolved.

The reported OpenAI arrangement is different.

Anthropic confirmed a directive and an actual service suspension. OpenAI's GPT-5.6 rollout remains a reported plan for a limited preview. There is no public evidence that the US government has prohibited a later broad release.

Even so, the reported preview appears to go beyond an ordinary private beta.

The first users would not be selected by OpenAI alone. The government would reportedly be involved in deciding which customers received access.

That creates a different relationship:

AI provider
    ↓
government security review and early-access selection
    ↓
release timing, eligible customers, and available capability
    ↓
user

The provider still builds and operates the model. The government may not hold a formal veto.

But the first group allowed to use the strongest system is no longer necessarily chosen by the provider alone.

A Wider US Pre-Release Review System

OpenAI and Anthropic are not the only companies involved.

In 2024, the US AI Safety Institute signed testing and evaluation agreements with both companies. US and UK institutions later published pre-deployment evaluations of models including OpenAI's o1 and Anthropic's upgraded Claude 3.5 Sonnet.

NIST announced the agreements here.

In May 2026, Reuters reported that Google DeepMind, Microsoft, and xAI had also agreed to provide unreleased models to the US government for national-security testing.

The administration has also pressed Meta to join a similar framework.

Reuters described the expanded reviews here.

The policy became more explicit on June 2, when the White House issued Executive Order 14409, “Promoting Advanced Artificial Intelligence Innovation and Security.”

The order calls for a voluntary framework under which developers can:

ask whether a model under development qualifies as a covered frontier model;
give the federal government access before release to other trusted partners; and
work with the government to identify those trusted partners.

The executive order is published here.

The same order explicitly says that the framework must not be interpreted as creating mandatory government licensing, preclearance, or permitting for new AI models.

That limitation is important.

The United States has not simply created a general rule requiring government permission before every frontier model can be released.

But a practical question remains:

If the government raises a national-security objection, how realistic is it for a major AI provider to ignore that request and proceed as planned?

Voluntary testing, trusted-partner selection, export controls, and customer-by-customer approval are different legal mechanisms.

When they operate together, however, the result may begin to look permissioned from the user's side:

The model can exist without being available to you.

Security Review Is Not the Same as Access Governance

There are legitimate reasons to evaluate powerful models before release.

Frontier systems can write and execute code, search for vulnerabilities, automate long tasks, support scientific work, and assist with complex operational decisions.

Governments are concerned about:

cyberattacks;
biological and chemical misuse;
military and intelligence applications;
the transfer of strategic capabilities to hostile actors; and
releases that cannot be reversed once a model or its weights have spread widely.

A serious pre-deployment review can therefore be reasonable.

But evaluating a model's safety and deciding who may use it are not the same thing.

Important questions remain:

Will the criteria for early access be public?
Will rejected organizations be told why they were rejected?
Will foreign users and smaller companies always receive access later?
Can access that has already been granted be withdrawn without warning?
Will the strongest versions increasingly be limited to governments and selected enterprises?

A restriction may be justified while still creating a continuity risk for users.

Different Laws Can Produce the Same User Experience

The United States is not the only jurisdiction shaping AI access, although other countries use different legal mechanisms.

China regulates public generative-AI services through service rules, security assessment, and algorithm filing.
The European Union imposes legal obligations on general-purpose AI providers, with additional duties for models considered to present systemic risk.
The United Kingdom emphasizes cooperative pre-deployment evaluation with frontier-model developers.
Italy and South Korea have restricted AI services through data-protection law.
Japan currently emphasizes promotion, national planning, and risk response rather than general pre-release approval.

These systems are not legally equivalent.

China's public-service rules are not the same as the EU AI Act. Privacy enforcement in Italy is not the same as US export controls. UK evaluation is not the same as customer-by-customer approval.

The common point is narrower:

Access to cloud AI can change for reasons outside the user's contract with the provider.

A model may be developed in one country, operated by a company in another, and used by a customer subject to a third legal system.

Cloud AI access is no longer determined by the provider alone.

The Likely Future Is Fragmented Access

The most plausible risk is not that all advanced AI disappears overnight.

It is a fragmented access environment:

countries receive new models at different times;
identity, nationality, or organizational verification becomes necessary;
enterprise users receive capabilities that individuals do not;
a deployed model is withdrawn after a legal or security decision;
older models are retired and the original working environment cannot be reproduced.

AI may remain abundant while the particular AI environment a person depends on becomes temporary.

The important question is no longer only:

Does the AI exist?

It is also:

Can I continue using it under the same conditions?

The Loss Is Larger Than Model Performance

People do not always use AI as a stateless calculator.

They place long-running conversations, project history, writing preferences, instructions, files, workflows, research context, and operating habits inside cloud services.

If access changes, the user may lose more than a particular level of intelligence.

They may lose the environment in which their work made sense.

That is why “just switch providers” is incomplete advice.

Switching is easy only when the model is the sole component being replaced.

It becomes much harder when the following are tied to one provider:

memory;
permissions;
project state;
conversation history;
data formats;
tools;
files; and
accumulated decisions.

At that point, changing models can mean rebuilding the entire working environment.

Diversification Is Already Becoming a Continuity Strategy

This is not only a theoretical concern for individual users.

Reuters reported on June 22 that companies including Siemens, Renault, Orange, and ChapsVision already use combinations of US, Chinese, and European models to reduce dependence on any one provider.

Reuters reported on those strategies here.

Several executives described technological sovereignty as a matter of choice and credible alternatives rather than complete isolation.

That is an important distinction.

Independence does not have to mean refusing every external service.

It can mean retaining the ability to continue when one service is no longer available.

The same principle applies to personal AI.

What doll Is Intended to Preserve

doll is currently under development. This section describes its design purpose, not the capabilities of a finished product.

doll is not an attempt to build a local model equal to every future frontier system.

When a powerful cloud model is available, using it may be the best choice.

The project is intended to keep the durable parts of a personal AI environment under the user's control:

memory, preferences, policies, and permissions;
conversations, projects, checkpoints, and decisions;
documents, evidence, research records, and generated artifacts;
migration, backup, restoration, and recovery paths.

In the planned architecture, models are replaceable reasoning engines around that durable core.

                 cloud model A
                      ↑
local fallback  ←   doll   →  cloud model B
                      ↓
                 future model

A preferred cloud model should be usable as an optional performance extension.

If that provider changes its terms, withdraws the model, blocks an account, restricts a region, or complies with a government directive, the user's state and work should remain available to another approved model or to a reduced-capability local fallback.

Performance may fall.

Continuity should not disappear with it.

Local AI Is Not Absolute Independence

Local operation does not eliminate dependency.

Model weights can be withdrawn. Hardware can be export-controlled. A local application can disappear or store information in a proprietary format. One runtime, model format, interface, or conversation store can become another point of failure.

Nor does local software exempt a user from law or regulation.

The goal is not absolute independence.

It is to retain more options when one dependency changes.

Borrow Capability, Not Continuity

Government review of frontier models may prevent real harm.

A completely unrestricted release policy is not automatically responsible.

But even justified restrictions have consequences for users.

A model that worked yesterday can become unavailable because of:

a provider decision;
a government directive;
a regional rule;
a nationality restriction; or
a change in customer eligibility.

The reported GPT-5.6 preview does not prove that a permanent government licensing system has arrived.

Anthropic's suspension does show that state action can remove access to an already deployed model.

The June 2 executive order shows that pre-release government evaluation and trusted-partner selection are becoming part of the formal policy structure around frontier AI.

Until now, most discussion has focused on which AI is smartest.

We also need to ask:

Will I still be allowed to use it tomorrow, and what remains mine if I am not?

AI capability can be borrowed.

Memory, data, project state, and the ability to continue do not all have to be.

Sources

GPT-5.6 reporting

Anthropic and US policy

Diversification and international approaches

Why I’m Building “doll”: A Personal AI Continuity System

Sohachi — Fri, 19 Jun 2026 15:58:36 +0000

High-performance AI is available to far more people than it used to be.

But there is no guarantee that access will continue under the same conditions.

Prices may change. Usage limits may become more restrictive. Models may be discontinued. Accounts, regional restrictions, regulation, provider policies, or service shutdowns may make an AI system unavailable even if it worked the day before.

This is not an argument about whether any particular company should or should not be trusted.

The problem is that the continuity of a personal AI environment currently depends on conditions the user does not control.

That is why I started building "doll", an open-source personal AI continuity system.

doll is not a new foundation model. Nor is it intended to replace model runners or interfaces such as Ollama, llama.cpp, Open WebUI, or LM Studio.

Its purpose is to preserve the parts of a personal AI environment that should remain usable even when models, providers, applications, runtimes, interfaces, or machines change or disappear.

The model should not be the durable core

AI systems are often described as though the model were the center of everything.

But the model itself is not necessarily the part a person most needs to preserve.

The durable parts may include:

memory and long-term state;
preferences, policies, and permissions;
conversations and their provenance;
documents, research records, claims, evidence, and sources;
generated artifacts and project history;
model and runtime manifests;
import, export, backup, restore, and migration records;
explicit records of information that could not be transferred.

Models can be replaced.

A more capable model may become available. A provider may retire an older one. A user may need to move to a smaller local model because the available hardware has changed.

Changing the model should not automatically mean losing the state the user has accumulated.

The central idea behind doll is:

Models are replaceable reasoning engines. The user's state is the durable core.

Local AI alone does not solve continuity

Running a model locally is an important way to reduce external dependency.

A local model may remain usable when a cloud service is unavailable. It can eliminate ongoing API costs and allow data to remain on hardware controlled by the user.

But local execution alone does not guarantee continuity.

What happens if conversations and memory are stored inside an application-specific database belonging to a single local AI application?

What happens if that application is no longer maintained, changes its storage format, stops working on a new operating system, or cannot be moved to another machine?

A local application can become another form of lock-in.

For this reason, doll does not treat the format of ChatGPT, an OpenAI-compatible API, Ollama, Open WebUI, or any other provider, runtime, or interface as the canonical representation of user state.

Data imported from another AI environment must be mapped into a documented and inspectable representation.

The result of that mapping must also be honest about loss.

An import or export may be:

full;
partial;
transformed;
unsupported;
lossy.

Information that cannot be transferred should not silently disappear.

Local-complete, cloud-optional

One of doll's governing principles is:

Local-complete, cloud-optional.

The local system should remain useful without API keys, account registration, or a permanent internet connection.

Cloud models may eventually be used as optional performance extensions. In many cases, they will be more capable than the models a user can run locally.

But a cloud service must not become the source of truth for:

memory;
identity;
files;
permissions;
portability;
backups;
recovery information.

Losing access to the cloud may reduce performance.

It should not erase the user's state or remove the path to recovery.

Why doll does not have a connected model yet

At the time of writing, doll is pre-alpha.

No model runtime is connected yet, and it is not a daily-use AI assistant.

That is intentional.

Many AI projects begin by connecting a model and building a chat interface. doll is taking a different path because the boundaries a model must not cross need to be defined and enforced before the model is introduced.

The model-independent foundation currently being built includes:

separating ordinary memory from secrets;
preventing secrets from leaking through logs, exports, backups, test fixtures, or diagnostics;
storing references to secrets rather than secret values in ordinary state;
performing bounded, credential-backed operations without exposing stored credentials to a model;
distinguishing confirmed facts, claims, evidence, and inferences;
recording the origin of information and instructions;
treating retrieved pages, documents, tool results, and imported data as data rather than authority;
granting external capabilities explicitly and within defined limits;
requiring fresh user confirmation for high-risk operations;
preserving the last known good state when an operation fails.

A model is not automatically trustworthy simply because it runs locally.

Local applications, plugins, imported conversations, documents, search results, and tool output must not gain authority merely because they have been placed into a model's context.

Adding these controls only after connecting a model would risk turning them into secondary protections around a system whose authority model was never clearly defined.

doll is therefore building the boundary first.

What doll is not

doll is not intended to be:

another model runner;
a clone of an existing local AI interface;
a cloud-dependent assistant;
an unrestricted autonomous computer-control agent;
a credential database embedded in ordinary AI state;
a guarantee of permanent frontier-model performance;
a claim that different models will behave identically;
a promise of universal, lossless migration.

Moving to another model may change response quality, behavior, personality, or reasoning ability.

Moving to less capable hardware may reduce what the system can do.

Continuity does not mean that performance never changes.

It means preserving what should survive those changes:

user-owned state;
provenance;
permissions and safety boundaries;
visible migration loss;
backups;
recovery procedures;
a documented path to another environment.

Current status

doll is not a finished product.

It cannot yet be installed and used as a complete personal AI assistant.

The current focus is on establishing model-independent state management, transfer, backup, restoration, portability, and safety boundaries before local model integration begins.

Development is divided into small specifications, issues, branches, and pull requests.

The repository records not only what has been implemented, but also the reasoning behind the implementation order, its acceptance requirements, and the expected recovery behavior when something fails.

Project website:

https://doll.badjoke-lab.com/

Source code, specifications, architecture decision records, and implementation history:

https://github.com/badjoke-lab/doll

What I would like criticism on

The goal at this stage is not to attract a large number of users.

Right now, criticism of the design is more useful than promotion.

I am particularly interested in questions such as:

Where should canonical user state end and model-specific state begin?
What information is inevitably lost when moving between AI environments?
How much provenance can realistically be preserved?
Does the proposed separation between secrets, ordinary state, and model access contain hidden weaknesses?
Can the system actually recover when an application, runtime, model, or primary machine disappears?
Could doll itself become a new form of lock-in?
Does the claim of being locally complete still depend on assumptions or external components that have not yet been recognized?

I would also like to learn about existing standards, projects, research, and documented failures that address the same problems.

doll is still at an early stage.

That is precisely why I want to define what must survive, what should remain replaceable, and where authority must stop before presenting it as a finished AI system.

Disclosure: This article was prepared with AI assistance and was reviewed, edited, and approved by the project maintainer.