When the honest thing is for your program to say "not yet"
The most fashionable AI architecture of the moment — coco recognizes it, understands it fully, and already has its hardest piece validated. And even so, if you try to run it, it refuses — with a message that tells you exactly what's missing. That, and not faking it, is what makes it trustworthy.
Throughout this series you've seen a pattern: I show what works and what doesn't. This final entry is precisely about that, turned into a concrete design decision.
There's a new, trending architecture (a hybrid that mixes linear attention and full attention, three to one). coco does almost all the work for it: it reads its full configuration, it knows how many layers of each type it has, it has the hard piece —GatedDeltaNet— validated to the seventh decimal. It's only missing one wire: the one that decides, layer by layer, which mechanism to use.
I had two options. One: half-wire that last bit and let it "start up," producing text that looks like it works but is broken inside. Two: stop dead in its tracks and tell you. I chose the second. Run it yourself:
Why this matters more than it seems
The most expensive bug in software isn't the one that blows up with a bang — it's the one that almost works. A model that answers reasonably but is subtly wrong makes you lose days chasing a ghost. This whole series has been, deep down, a war against that silent failure: validating against ground truth before wiring, saying where speed stops paying off, keeping the failure with its label. An honest "not yet" is the same idea, at the front door.
coco is early. It runs four real architectures, compresses models 3× without losing quality, and is honest about what it's missing. I'd rather have an engine like that — one that looks you in the eye — than one that promises you everything and breaks on you at the fifth layer.
End of the coco series (for now). Back to the notebook index · start with no. 01.