Claude Fable Disappeared Mid-Build. That Should Worry All of Us.

When access to a powerful AI tool can vanish overnight, the real question is not which model you use. It is whether your knowledge system can outlast it.

I was in the middle of a really cool build-out with Claude Fable. Not testing it casually. Not asking random questions. Actually building.

The kind of session where the structure is coming together, the pieces are clicking, and you can see what this next generation of AI tools may make possible.

And then it stopped working.

At first, I thought it was a glitch. Maybe capacity. Maybe my account. Maybe a temporary product issue.

So I did a quick search. Only to learn that Anthropic had abruptly disabled access to Fable 5 and Mythos 5 after a U.S. government directive tied to national security and export-control concerns.

This moment feels important. Not because I was frustrated, although I am.

But because it reveals something we are going to have to take much more seriously in this next phase of AI adoption:

AI access is not stable.
AI tools are not neutral infrastructure.
And access to the most advanced models is not guaranteed.

We are not just dealing with new productivity tools anymore.

We are dealing with systems powerful enough to trigger government intervention in less than three days. Tools that may be available one minute and restricted the next (or disappear if the bubble bursts). Tools that institutions, businesses, educators, creators, and students may begin building workflows around before the rules, risks, and governance structures are fully settled.

That does not mean we should stop experimenting.

I am not arguing for fear. I am not arguing that we should pull back, wait it out, or refuse to build until everything is perfectly settled.

But I do think this is a reminder that we need to stop treating AI access as permanent, neutral, or guaranteed.

Pay attention to this lesson: AI strategy cannot be built only around access to a single model.

It has to include redundancy.
It has to include governance.
It has to include human judgment.
It has to include contingency plans.
It has to include a clear understanding that the most advanced AI systems are no longer just productivity tools.

They are becoming infrastructure.

And infrastructure can be regulated, restricted, redirected, or removed.

The Problem With Model Attachment

This actually connects to something I noticed early in AI adoption.

When we moved people from ChatGPT into BoodleBox, I saw how attached ( in a matter of months) many users had become to a specific model, a specific interface, and a specific conversation history.

For some people, all of their knowledge was effectively trapped inside the model experience.

Their prompts were there.
Their workflows were there.
Their context was there.
Their thinking process was there.
Their half-built systems were there.

So when they moved tools, they felt like they were starting over.

That is a very real problem.

It is also a signal that many of us have been building in the wrong place.

We have been building inside AI systems instead of building systems we own and inviting AI into them. You need to own your knowledge, not rely on AI to remember it for you.

That distinction matters.

When your knowledge lives only inside one AI platform, you are dependent on that platform’s memory, pricing, policies, availability, interface, and access rules.

But when your knowledge lives in a structure you control, the model becomes less of a home and more of a guest.

You can invite Claude in.
You can invite ChatGPT in.
You can invite Gemini in.
You can invite a local model in.
You can invite whatever comes next in.

The experience may not be identical, but it can be similar enough because the underlying knowledge system is yours.

Build the Knowledge Bank First

This is something I have been teaching my graduate students: build your own knowledge banks.

It is becoming a foundational skill in AI literacy.

I follow a lot of what Jake Van Clief teaches in this space, and one thing I appreciate is how he keeps pointing people back to basic computer science principles that many of us in Applied AI did not necessarily come from.

A lot of us came to AI through teaching, writing, design, business, leadership, communication, consulting, ministry, research, or creative work.

We did not necessarily come in thinking about file structures, markdown, retrieval, context windows, token costs, or system navigation.

But now we have to.

The more we use AI, the more we need to understand where our knowledge actually lives.

For me, that place is Obsidian.

But Obsidian is not the magic. Obsidian is just the interface.

Underneath it is something beautifully simple: a bunch of markdown files and folders organized inside a structure I own.

That structure can include notes, drafts, research, frameworks, voice guidelines, course content, project plans, source material, decision logs, and reusable language.

Then I can create a file that tells AI what is there, how it is organized, how to navigate it, what to search for, and how to use it (without reading everything).

In other words, I am not relying on the AI model to “remember me.”

I am building a system that lets the AI orient itself to me.

That matters for two reasons.

First, it reduces token use because I do not have to keep dumping massive amounts of context into every conversation. I can give the AI a map and direct it to the right material.

Second, it makes my work more portable. If one model changes, gets more expensive, becomes unavailable, loses access, or disappears from my workflow, I still have the knowledge structure. If I want to run a local model, I can. If I want to attach Claude, I can.

The model changes.

The knowledge remains.

This Is the Next Prompt Engineering

For the last year, a lot of AI literacy has focused on prompt engineering.

That made sense. People needed to learn how to ask better questions, give clearer context, define roles, specify outputs, and guide the model toward better work.

But we are moving into a new phase. Prompt engineering still matters, but it is no longer enough. The next foundational skill is knowledge architecture.

Can you organize your own thinking in a way that AI can use? Can you separate your knowledge from the platform? Can you build a file structure that can survive model changes? Can you create instructions that teach AI how to navigate your materials? Can you design workflows where the human owns the system and the AI is invited into it?

This is not just a technical issue.

It is a governance issue.
It is a cost issue.
It is a continuity issue.
It is an institutional resilience issue.
It is a human agency issue.

Because disruptions like the Fable and Mythos cutoff will not be the last.

Some disruptions will come from government action.
Some will come from pricing changes.
Some will come from product pivots.
Some will come from model retirements.
Some will come from institutional policy.
Some will come from data privacy concerns.
Some will come from the simple reality that the tool you love today may not be the tool you can access tomorrow.

So the question is not only, “Which AI model is best?”

The better question is:

“What am I building that can outlast the model?”

Build Systems You Own

Ironically, the night this happened, my students and I had been talking about this exact idea.

I was encouraging them to build systems they own where AI can be invited in, rather than building entirely inside the AI system itself.

That now feels less like a preference and more like a necessity.

If your workflow only works inside one model, it is fragile. If your knowledge only lives inside one platform, it is vulnerable. If your strategy depends on uninterrupted access to a single frontier model, it is incomplete.

That does not mean we should stop using powerful tools. I still want to keep building. I still want to experiment. I still want to see what these systems make possible.

But I want to build in a way that keeps humans, institutions, and creators from becoming overly dependent on one model, one company, or one interface.

The future of AI will not be shaped only by what the tools can do.

It will also be shaped by who has access, who sets the rules, what gets restricted, what becomes expensive, and how prepared we are when the tool we rely on suddenly disappears.

So yes, keep experimenting.

But build your own knowledge bank.
Own your files.
Create your structure.
Document your workflows.
Separate your thinking from the model.
Invite AI into your system instead of surrendering your system to AI.

That may become as foundational as prompt engineering.

Maybe even more so.

And for anyone else who got cut off in the middle of a build tonight: I felt that.

But maybe the interruption was also the lesson.