In 1955, Akira Yoshizawa wrapped each of his paper models in a cardboard protector and sent them by air from Tokyo to Amsterdam. There was no other way to get them there. The models went on display at the Stedelijk Museum; hundreds of people came to see them. What they couldn’t do was reproduce what they saw.
What spread origami wasn’t the exhibition. It was Yoshizawa’s notation system — dotted lines for valley folds, arrows for direction — which he had developed alongside the models. Anyone with the diagram could reproduce the model without a teacher. The notation is executable; the physical object isn’t. The origami models from the exhibition ended up in a box in Gershon Legman’s estate.
You can download Llama’s weights and run them. The model is in your hands, the way one of Yoshizawa’s cranes was in the hands of a museum visitor — you can observe it, use it, turn it over. But you can’t retrain it. The training data, the filtering decisions, the curriculum that produced it: those are the diagram. The weights are the crane in the cardboard protector.
AllenAI’s OLMo releases training data and pipeline alongside the weights. That’s Yoshizawa publishing the diagram notation rather than mailing models to Europe. Meta and Alibaba are sending the cranes. AllenAI is sending the instructions.
Seventy years later, the original models are in a museum. Nobody is folding from them.