The EinsteinTest: models given pre-breakthrough evidence rarely make the break themselves
CurieuxExplorer · x · 2026-10-03
X user CurieuxExplorer argues that models given only the evidence available before a major scientific theory change rarely make the breakthrough themselves — instead they tend to rebuild the explanation the field later abandoned.
Framed under the #EinsteinTest hashtag, the observation questions whether language models trained on historical scientific data can reproduce creative breakthroughs, suggesting they excel at reconstructing existing paradigms rather than revolutionizing them.
More from AGI Musings
- Hinton: we design the learning algorithm, but we don't really understand how neural networks work — austinc3301 · 2026-10-03
- Researchers warn we may build conscious AI before understanding consciousness itself — CurieuxExplorer · 2026-10-03
- SF one-bedroom relisted at $6,000 after $4,000 two years ago as AI economy inflates rents — Yuchenj_UW · 2026-10-03
- After repeated IVF failures, a couple used GPT-4 as their 'AI midwife' to crack the case — nwilliams030 · 2026-10-03
- Before the singularity: get strong, fix sleep, learn to use AI for everything — rand_longevity · 2026-10-03
- AI races toward superintelligence while the physical world stays unchanged — RileyRalmuto · 2026-10-03