REDDIT · REDDIT// 6d agoOPENSOURCE RELEASE

GPT-1900 intuits modern physics from pre-1900 data

Michael Hla's 3.3B parameter LLM was trained exclusively on text published before 1900 to test "out-of-distribution" reasoning and scientific discovery. The model showed provocative glimpses of intuition, suggesting light consists of discrete energy packets and the local equivalence of gravity and acceleration when prompted with period-contradictory observations.

// ANALYSIS

A fascinating experiment in "scientific time travel" that challenges current assumptions about LLM reasoning versus data memorization.

–Aggressive decontamination (22B tokens filtered for "future leaks") represents the project's most significant technical hurdle and achievement.
–While exhibiting "glimpses" of intuition, the model remains a small-scale prototype prone to period-appropriate "bullshitting" using 19th-century prose.
–The results suggest LLMs can synthesize breakthroughs from contradictory data, though they remain heavily tethered to historical biases.
–Open-sourcing the dataset and weights provides a unique benchmark for testing "true" reasoning in larger, modern models.

// TAGS

llmreasoningopen-sourcephysicsmachina-mirabilis

DISCOVERED

6d ago

2026-04-05

PUBLISHED

6d ago

2026-04-05

RELEVANCE

8/ 10

AUTHOR

Primary-Track8298