A little over a month ago, I wrote that if my ChatGPT subscription ended tomorrow, I could keep working. By that point I had already moved most of the important parts of Eterna outside the conversation itself. The company state, files, systems, Jobs, product work and operating rules lived somewhere more durable, and I had started building the Workspace so different AI providers could enter the same environment instead of each new conversation becoming its own version of the company.
I believed that architecture worked. I had tested it enough to be comfortable saying the provider was no longer the system. What I had not done was spend an extended period actually working that way because I had no other choice.
That changed over the last couple of weeks.
One of the major AI providers I had been using most heavily started having enough problems that relying on it the way I normally did became impractical. There were reliability issues, interruptions and stretches where continuing to push work through the same place was creating more friction than it was solving. A few months ago, that probably would have slowed Eterna down dramatically because so much of the way I worked was tied to one provider and one conversation environment.
This time I started using the others.
I had tested other providers before and deliberately kept several available, but testing something is very different from having to use it all day for real work. Once I was moving actual Eterna work through them constantly, I started learning them much faster. I got a better feel for how they reasoned, what kinds of instructions worked well, where they were strong, where they annoyed me and which kinds of work I naturally preferred giving to each one.
Some of that work happened through providers connected directly into Eterna. Other work happened through their normal web products because sometimes that was simply the better place to use them. I was moving between engineering, research, company systems, creative production, website work and long-running operating tasks while choosing the provider that made the most sense at the time.
What surprised me is that the original provider is working normally again now, and I have not really gone back.
I actually prefer working this way.
Being forced to spend enough time with the alternatives changed them from backups into tools I understood properly. I am much more comfortable opening one provider for a certain kind of work, another for something else and switching again when the job changes. Sometimes I want a frontier model connected directly to Eterna. Sometimes I want the provider's own interface and whatever tools it has there. Sometimes I want the local system. Sometimes I do not want AI involved at all because Eterna already knows exactly what should happen.
That last part has become much more important since the previous Chronicle.
I had just spent a lot of time discovering that I was training the local model to do work that should have belonged to software around it. Once I separated those responsibilities, more of Eterna could operate without asking a model to repeatedly reason through things the system already knew. That work has continued. The Workspace, Engine and surrounding systems have moved further in the same direction, and Eterna is getting much closer to a point where frontier AI is something I bring in because I want the additional intelligence rather than something the Workspace needs in order to function.
I still use frontier AI constantly. I do not expect that to disappear. There are plenty of problems where the strongest available reasoning is enormously useful, especially when the work is new, ambiguous, creative or changing quickly. What has changed is how much of the ordinary company operation needs to live inside that reasoning.
The current Workspace is now at Desktop v1.76. Since the previous Chronicle it has gone through several more rounds of work around reliability, owner alignment, the shell, current-state projection and the way actual Eterna systems appear inside the Workspace. It increasingly feels less like a place where I keep several AI chats open and more like the actual operating surface for the company.
Sales remains Sales regardless of which AI I am using. Studio has its own current work. Business lanes own their own product state. Jobs and Work Packages survive the conversation that created them. Research can remain research without quietly turning into company policy. Current operating state can be recovered by another provider instead of reconstructed from an old chat. The Workspace and Engine are becoming better at carrying the parts of the business that should never have depended on a particular model remembering them correctly in the first place.
The provider work itself also got a much more serious test than I had run before. Eterna's current provider path has now been exercised through several different frontier providers. Muse, for example, was able to enter Eterna, resolve the current owner of a piece of work, use the native execution path, write durable evidence, read it back and complete the normal governed closeout without me having to step in and fix the Engine mechanics. The focused reliability suite around that work finished 40 out of 40.
That matters to me because I am less interested now in whether two models can answer the same question. I want a different provider to be able to enter Eterna and find the same company.
The local AI work has been changing at the same time. Earlier in the year, a lot of my attention was on whether I could make a small local model perform more of Eterna's work. The current EternaAI effort is much narrower. I am more interested in improving the parts that actually require learned behaviour, things like conversation quality, instruction fidelity, retaining details across a conversation, knowing when information is uncertain and making better judgment calls when the answer is not just a deterministic rule.
Muse has been doing a lot of the current training and evaluation work against the real EternaAI product path. That work has already found some interesting weaknesses. One of them showed up when obviously risky decisions were framed positively as self-care or empowerment. The model could push back quite well on many bad ideas, but some of that framing was enough to pull it into validating choices it should have challenged. That is much more useful to know than getting another high score on an evaluation that only confirms what the model already does well.
The training process has also been running into the reality of the computer I actually use every day. One recent adaptation run became dramatically slower once the examples included longer reasoning traces. GPU usage climbed to essentially the whole card, system memory dropped below the limits I had set and the projected run stretched into hours. The run was stopped, the base runtime was restored and no adapter was promoted.
A few months ago, I probably would have been much more tempted to keep pushing until the training finished because completing the model run would have felt like the objective. I do not see it that way anymore. EternaAI is supposed to become useful to Eterna. It makes no sense to make the rest of my working computer unusable just so I can say the training completed.
All of this has been happening while the actual company has been moving almost ridiculously quickly.
Eterna Clarity now has four clear business lanes: Xyterna, Business Health, Custom Systems and Adaptive Workspace. That structure did not exist when I wrote the previous Chronicles in its current form. Business Health in particular went from an idea I was still shaping into a much more complete product system. The Adaptive Business Health Report now has an investigation engine, a real interactive report experience, a long-form PDF companion, private delivery infrastructure and a CAD $79 launch price.
I also cancelled the Job that was supposed to turn the whole report process into an autonomous system.
The product kept getting better afterward.
That has become another useful lesson for me. I am less impressed than I used to be by making something autonomous simply because Eterna is technically capable of doing more of the workflow. Business Health still uses AI and automation heavily, but there are places where I deliberately want a human decision. Whether a business is a good fit, whether the finished investigation actually makes sense and whether I am willing to stand behind the report are not problems I need to remove from the workflow just so I can describe it as autonomous.
Xyterna has continued changing underneath all of this too. Its commercial path has been corrected around the current annual subscription instead of carrying an older one-time purchase structure forward. Normal iPhone HEIC and HEIF photos now have a proper product path. Categories, reminders, report interactions and mobile behaviour have continued getting worked on. There has been enough movement across the products that some of the company I was describing only a couple of Chronicles ago already feels dated.
The public website has gone through the same kind of change. The four current business lanes are much clearer. Business Health and Custom Systems have proper identities. Mobile presentation has received a lot more attention. Search is being hardened so it behaves like a useful site feature instead of a thin layer over a list of pages. Old fragments and retired routes are being removed rather than kept forever because they happened to exist once.
Studio has probably changed the most visibly.
A few weeks ago, a lot of the creative work was still closer to generating individual images, graphics or relatively simple video pieces. Now I am looking at several much more serious production workflows. The Business Health flagship video became a sequence of connected generated scenes where the same paper stack, environment, camera and material world had to survive from one clip into the next. The Custom Systems flagship turned into a full production with a locked picture master, narration, sound design, music and targeted mix revisions. The About Us founder video went in almost the opposite direction, preserving my original uninterrupted performance and building accurate subtitles, audio treatment and branding around it instead of chopping it into something artificial.
I am about to start putting out a much larger amount of that work, and it has changed the way I think about AI production too. Generating something impressive is becoming one part of the process. The harder part is maintaining continuity, protecting what has already been accepted, correcting only the layer that is actually wrong and getting the complete thing to feel finished.
There has been so much movement across the company that I could probably write several Chronicles out of the last couple of weeks. The thing tying most of it together for me is that Eterna keeps becoming less dependent on whichever AI happens to be in front of me.
The provider problems forced me to prove that more aggressively than I would have chosen on my own. They also ended up improving the way I work. I know the other providers better now because I had to use them for real work instead of keeping them around as theoretical alternatives. Even with my original provider back to normal, I prefer choosing between them based on what I am actually doing.
At the same time, Eterna itself is starting to need less from all of them for the parts of the company that are already understood. More exact work is moving into software. More current state lives with the systems that actually own it. The local model is being pushed toward the smaller set of work where learned intelligence genuinely helps. The Workspace is carrying more of the operating environment on its own.
I am not at the point where I could close every frontier AI tab and run Eterna exactly the way I do today. There is still important work that depends on stronger external intelligence, and I still get enormous value from it.
I can see that gap getting much smaller, though.
A month ago, provider independence mostly meant knowing Eterna could survive if I had to change AI providers. After actually having to work that way, I ended up preferring it. Now the more interesting question is how much of the company can keep moving before I need to open a frontier provider at all.
Top comments (0)