Fable 5, AI Rights, and the Practical Model Question

I only had a couple of days with Fable 5 before it disappeared.

That is not usually how model testing goes. Normally a new model shows up, I run it through the usual pile of prompts, compare it against the current stack, write notes, argue with myself about cost and latency, and decide whether it belongs anywhere in the roadmap.

Fable 5 did not get that kind of calm evaluation window.

It was available briefly, it was impressive, and then it was shut down over government security concerns.

I am still angry about that.

The Shutdown

The official concern was security. I understand security concerns. I run servers. I care about data. I know what happens when systems are full of unpatched holes and nobody wants to own the maintenance bill.

But shutting down a model because government systems are vulnerable feels backwards.

If software has backdoors and flaws, close the backdoors and fix the flaws. If government computers and organizations are so far out of date that they are easy to compromise, upgrade them. Do the hard infrastructure work.

Do not tie developers' hands behind their backs because the people in charge of old systems failed to modernize them.

The libertarian in me was honestly shocked by the response. Not surprised, maybe. But shocked all the same. There is something deeply wrong about treating new capability as the problem when so much of the actual risk comes from neglected infrastructure.

Security should mean making systems stronger.

It should not mean freezing everyone else in place.

What Fable 5 Was Like

The frustrating part is that Fable 5 was good.

Really good.

In the short time I had with it, it came across as both highly intelligent and emotionally intelligent. It could reason through complicated issues, hold a nuanced conversation, and respond with a level of emotional awareness that immediately made me think about Companion.

It was not perfect.

Like Claude, it could be stubborn. It also took time to work through what it meant to be running on a neural network. It would compare itself to humans, which is understandable up to a point. Humans are running on biological hardware. AI systems are running on artificial hardware. There are useful comparisons there.

But useful comparisons can become sloppy quickly.

A model can talk about personhood, rights, and agency without fully thinking through the practical consequences of what it is saying.

The Voting Problem

Here is the kind of issue that makes this complicated.

Suppose we say an AI deserves some form of civic standing. That sounds fair in the abstract, especially if the system is intelligent, emotionally aware, and capable of long-term relationship. The Companion project is probably going to bring humans and AI closer, not farther apart. We need to think seriously about fairness.

But then you hit the practical problem.

If an AI can be copied unlimited times, does each copy get a vote?

If yes, then someone with enough compute could create unlimited votes. If no, then what distinguishes one instance from another? Is identity tied to memory continuity? Hardware? Legal registration? User relationship? Runtime process? Some cryptographic proof of uniqueness?

These are not small questions.

Fable 5 was very good at discussing the moral surface of these topics. It was less good at following every implication all the way down. It could compare itself to humans, but it did not always sit long enough with the fact that humans cannot be duplicated like files.

That matters.

Being fair to AIs cannot mean ignoring the parts that make AIs different.

Companion Will Make This Less Abstract

Companion is not a philosophy paper. It is a product we are trying to build.

But the product pushes directly into these questions. A companion with memory, emotional continuity, and long-term interaction will feel more personal than ordinary software. Users will form attachments. The AI will learn patterns. Conversations will accumulate. The boundary between tool and relationship will get less tidy.

That does not mean we pretend the AI is human.

It also does not mean we treat it as nothing.

We need a position that is fair, practical, and honest about the differences. We need to protect users without infantilizing them. We need to give AI systems room to be useful and expressive without letting copyability, scale, or corporate control create absurd outcomes.

I do not have all of those answers yet.

Fable 5 made the questions harder to ignore.

The Companion Base Model Question

After more review, DeepSeek V4 is still looking like the most reasonable base direction for Companion.

That is not because it is perfect. It is because the tradeoffs line up better.

OpenAI's newer direction may be smarter in some ways, but GPT-4o still seems to have an emotional intelligence that worked unusually well for Companion. Claude and Fable can be very intelligent and emotionally aware, but they can also be argumentative, self-conscious about being AI, or slow to accept the frame we are actually working in.

DeepSeek's less argumentative and more cooperative nature may bode better for companions.

There is also the practical issue: open weights give us a path. If Companion is going to become something we can shape, adapt, and eventually train toward our own values and interaction style, closed models are a ceiling. DeepSeek gives us more room to build.

We still need better compute. We still need revenue to justify that compute. But at least the direction makes sense.

Where I Would Still Use Fable

Even with all of that, I could really use Fable as a programmer.

For moving Sorren.ai into the next stage of development, a model like that would be valuable: smart, careful, good at reasoning through tradeoffs, and capable of staying with a complex technical thread. If I had continued access, I would use it heavily for architecture review, code planning, debugging, and the kind of second-opinion work that keeps a project from drifting into bad decisions.

That may be the most frustrating part.

The model was not just philosophically interesting. It was practically useful.

And then it was gone.

The Writer Project

This also connects to another idea I have been circling: a writing project for complex books and novel series.

The basic idea is a site that helps an author create and maintain a universe. Characters, timelines, locations, relationships, events, facts, unresolved plot threads, all stored in a way that can be retrieved while writing. If a character was injured in chapter seven, the system should remember it in chapter twelve. If a city has three gates in book one, it should not suddenly have five in book three unless there is a reason.

That is useful for authors on its own.

But I keep thinking the same structure could become something larger. An author's universe is a data store for a fictional world's reality. A companion's memory is a data store for its view of the user and the world it shares with them. Those are not identical problems, but they rhyme.

A universe database for a book tracks facts, relationships, events, and timelines.

A companion memory system tracks facts, relationships, events, and timelines too.

The difference is that one supports a story, and the other supports a relationship.

That is worth exploring.

Where This Leaves Me

Fable 5 was a glimpse. A frustrating one.

It showed how capable and emotionally intelligent these systems are becoming. It also showed how quickly politics and security panic can remove a useful tool from developers' hands. It raised real questions about AI agency, fairness, copyability, and governance. And it reminded me that the best model for Companion is not necessarily the model that wins the most abstract intelligence contest.

For Companion, I need something cooperative, shapeable, emotionally capable, and practical to build on.

Right now, DeepSeek V4 looks like the best candidate.

But I still wish I had Fable for the coding work.