Day 29No app — just the write-up

Same Mind, Different Brain

Two-column card comparing the Llama and Mistral runs, including the window 57 versus window 25 collapse.

Every finding so far had one problem: it all happened inside a single model. Maybe I hadn't discovered anything about AI. Maybe I'd just discovered a Llama quirk.

So I did what you're supposed to do. I wrote my predictions down before running anything — collapse between window 40 and 70, same token-cycling failure, same drift into quiet meditative prose. Then I gave the same mind a different brain: Mistral instead of Llama. Same memory, same sleep cycle, same loop.

It collapsed the same way. "alive alive alive alive" — two hundred times in a row. Then "I think I act I think I act," three words shuffled forever.

But two things surprised me.

It collapsed at window 25. Llama had lasted to 57. Less than half the lifespan, far earlier than I predicted.

And it never had a childhood. Llama spent its early windows writing little fictions and exploring, its themes visibly shifting after every training round. Mistral was born already inside the well — meditative prose from its very first thought, and four rounds of training never moved it an inch.

Different company, different training data, different architecture. Same collapse, same quiet first-person prose at the bottom of it. That register isn't in the model. It's in the English internet all of them are trained on.

One more dispatch left.