Photo: Bokeh / Tomedia. Close-Up of Metal Fasteners and Pipe.
When I last wrote properly about AXIOM, I called the article Mission Failed, So I Built the Ultimate Machine. A fairly restrained title, obviously.
I have kept building it. It can talk to me now, and I’m getting closer to having it carry out the tasks I want automatically. It is still not feature complete, and I am still not fully happy with it. That is probably the most useful place to start an update on something I previously described as the ultimate machine.
AXIOM is the system I am building around AI models to help organise and carry out work. It has grown into research, planning, software work, writing, memory, review and a personal companion. The underlying ambition is still practical. I have a lot to do, and I would like more of the work to happen without me carrying every piece of context between every tool myself.
The trouble is that this ambition contains a remarkable amount of work of its own.
A conversation is a start
The companion side now has somewhere to keep conversations, retrieve approved material, and connect what we talk about with goals and tasks. That sounds fairly ordinary when compressed into a sentence. Implementing it means deciding which information belongs in an answer, whether it is current, where it came from, and whether it belongs to the person or project currently being discussed.
I don’t want a convincing answer assembled from the wrong client’s information. I don’t want an old plan to quietly become the current plan because it happens to be easy to retrieve. And I don’t want the system to treat something I once wondered about as a decision I’ve made.
So the memory work has become a substantial part of AXIOM. The implementation keeps the source and date of information, lets newer material supersede older material, and separates approved material from things it has merely imported. There are mechanisms for forgetting information too. Having more stored text is only useful if the system can choose the right text when I need it.
The planning side has also grown. It can prepare a daily plan from goals, commitments and tasks, and distinguish work it could help carry out from work that still needs me. My goals remain mine to change. That sounds like a small rule until you imagine a very helpful assistant gradually rewriting your priorities to match the jobs it finds easiest to complete.
Some of this is implemented capability rather than a habit I can already rely on every day. I can point to code and checks for it. I am still doing the work of making the whole experience useful enough that I trust it in practice.
Getting something done involves more than asking a model
The orchestration side is where a lot of the expansion has happened. AXIOM needs to know which machine can do a job, which model is available on it, what the job requires, whether somebody is already working on it, and what happened when it finished.
If the machine disappears halfway through, the job needs a sensible state to return to. If a result arrives twice, it should not cause the same change twice. If the job cannot run, I need to see that it cannot run. A dashboard full of optimistic labels is of limited comfort when the actual work hasn’t moved.
There is now a desktop workbench and machinery for registering workers, claiming jobs and returning results. The software-building path can work in a separate checkout and leave a branch for review. The difference between those parts being present and a machine reliably taking work is still very real. A worker that hasn’t been connected and enabled is doing no work, however complete its implementation looks.
This is why I am being fairly careful with the word “automatic” at the moment. I have built a lot of the pieces that automatic work needs. I am still bringing them together and checking the behaviour against actual tasks. I don’t have an agency quietly running itself while I go and do something else.
It has to be able to disagree with itself
The research and review parts have expanded as well. There is a path for collecting evidence, challenging claims and producing a research direction. There is another path for testing proposed improvements against an independent evaluation.
I want that separation because a model telling me it did a good job is a very cheap thing to obtain. The system needs a way to check the work that doesn’t amount to asking the same process to congratulate itself. A proposed improvement also needs to remain a proposal until it has earned the next step.
That applies to ordinary business work too. Researching a change, writing a draft and being allowed to put the change in front of a customer are different events. AXIOM now has more machinery for recording those events and connecting them to the work they belong to.
It would be much easier to demonstrate something impressive if I ignored a lot of this. Give a model some tools, let it run, show the successful bit. But I need to use the result with work that continues the following day. I need to know which version was reviewed, what failed, and whether a result was produced under the conditions I thought it was.
I also need to know when the system doesn’t have enough information. Some of the capacity and cost features are designed to stop when the underlying facts are missing or stale. I would rather see the missing information than make a business decision from numbers the system has politely improvised.
I would like to finish the loop
The slightly ridiculous development is that, while AXIOM is still unfinished, I went and started Juniper. I saw a digital kiwi called Pino, thought the idea was interesting, and ended up building a small world for another AI to inhabit. Apparently the amount of unfinished AI work in my life wasn’t quite sufficient.
Juniper is a separate experiment, though it has made the same practical gap very visible. A model can describe doing something beautifully. Then the environment has to let it do that thing, the action has to succeed, and the result has to be recorded correctly. The language is one part of a much larger system.
With AXIOM, I am getting closer to the version I wanted when I started. The companion can talk to me. The workbench and task machinery are further along. The research, memory and review pieces have become much more substantial. I still have to turn those pieces into a dependable experience, and the existence of a long implementation document doesn’t finish that job for me.
The next useful milestone is a task I would otherwise have had to carry myself: given enough context, completed within its limits, checked, and returned in a form I can use. Then I want it to do that again without requiring me to rebuild half the process in between.