Machina Mirabilis (GPT-1900)

Abstract

Machina Mirabilis (GPT-1900) investigates whether a language model trained from scratch on historical text can generate conceptually useful explanations of observations associated with later developments in physics. The project reports a 3.3-billion-parameter transformer and approximately 22 billion tokens of filtered pre-1900 text, followed by physics midtraining, synthetic instruction tuning, and reinforcement learning with a modern language-model judge. Eight prompts probe blackbody radiation, the photoelectric effect, and thought experiments associated with relativity. Selected generations suggest discrete energy transfer or relationships between gravity and acceleration, but performance is inconsistent and often physically incorrect. Modern-model supervision, residual contamination risk, supplied experimental framing, and checkpoint selection limit conclusions about independent rediscovery. Across 24 archived v11 checkpoints, the highest mean judge rating in the standard eight-task evaluation is 1.25 out of 5. The released models and evaluation records provide an inspectable test bed for historically constrained language modeling, while leaving robust scientific rediscovery unestablished. Original public project release: March 2026. This October 2026 technical-report edition consolidates the public blog and released supporting records. No new training or experimental work was performed for this edition. Original project release, manuscript preparation, and repository deposit dates are distinct. Original project article: https://michaelhla.com/blog/machina-mirabilis.htmlCode and released records: https://github.com/michaelhla/gpt1900