I've been changing something fundamental in AIRUNCODE.
I don't think the frontier model should do all the work.
Right now we have Claude, GPT, Chinese models, API endpoints, local models, different subscriptions, different context windows and completely different pricing.
But most software treats the model as the product.
I'm trying to make the model replaceable.
In AIRUNCODE, different models can operate inside the same environment and use the same tools. A frontier model can solve the difficult part of a problem, while a smaller or local model continues the workflow instead of burning expensive tokens doing work it is perfectly capable of doing locally.
Basically: don't pay the architect to carry every brick.
This becomes much more interesting when the environment itself is programmable.
I've now connected AIRUNCODE's graphical engine deeply enough that models can build, compile, inspect the result, modify it and compile again.
That's how I think generative software should work.
Not:
prompt → generate demo → congratulations.
But:
build → compile → inspect → correct → rebuild → repeat.
A sample should be something the system can continue working on until it becomes an actual product.
The easiest analogy I have is LEGO.
I don't want to give the model a finished LEGO set and permission to move three pieces.
I want to give it the box, the pieces, the tools and enough understanding of the environment to build what it needs.
And coding is probably only the first environment where I'm testing this idea.
The next thing I want to explore is simulation.
Imagine a robotics lab where thousands of simulated attempts aren't evaluated only by a reward function.
A frontier model can observe a robot performing a complex procedure, understand what it is trying to accomplish, identify where the procedure went wrong and provide corrective supervision.
Not just:
failed → -1
succeeded → +1
But:
you grabbed the correct component, your position was correct, but step 4 happened before the safety verification required in step 3.
That becomes especially interesting for procedural skills where the order, context and rules matter as much as reaching the final state.
Frontier intelligence becomes the teacher.
Smaller models become workers.
Simulation provides experience.
And the environment keeps the knowledge instead of forcing every new model to rediscover everything from zero.
That's the direction I'm exploring with AIRUNCODE.
The model shouldn't own the workflow.
Top comments (0)