Ship of Theseus
I swapped the 31 parts of one language model, Qwen, one at a time for the matching parts of another, DeepSeek. After every swap I asked the hybrid who it was.
“Hello! I'm an AI assistant created by {{ company }} to help with your tasks.”A hybrid with 26 of its 31 parts from DeepSeek
- The ships in the middle forget their name. In 1,745 of 1,920 answers they named neither model
- Identity sits near the output. Swapping from that end, the ship calls itself DeepSeek after about 5 parts, against about 22 from the input end
November 2026Theseus's ship had every plank replaced until none of the original wood was left. Both models come from the same Qwen pretraining run, so a part of one fits into the other. The design and five hypotheses were written down before the main run, which took 4.9 hours on two Kaggle GPUs.
