Hobbyist trains a 118k-param JEPA to rediscover Mendeleev's missing elements — it fails
WarningSure1670 · reddit · 2026-10-25
An independent researcher trained a tiny JEPA (118k params) solely on chemistry known in 1869 — 47 elements with valence, metallicity, and room-temperature state — to see if it could point at gaps in the periodic table like Mendeleev did. Pass criteria were hash-locked before coding, and no pretrained models were used to avoid leakage.
Result: it missed the bar
- Without atomic weights, the scandium gap ranked 7th–14th and gallium/germanium 9th–11st out of 46 spots (chance 23rd) — weak signal, unstable across machines
- Completely missed technetium's gap (ranked 25th–26th)
- Never found the two missing elements between zinc and arsenic
- With atomic weights included, it did worse than a naive biggest-jump baseline
Next step: run the same experiment on pre-1905 physics to see if the model points at holes later filled by relativity and quantum mechanics. The author, a non-ML person building with AI coding agents, asks whether surprise-reduction is a sound gap-detection score, whether JEPA is the wrong tool at this data scale, and whether finding what's missing is genuinely different from next-token prediction.
More from AGI Musings
- Dopamine vs. cortisol: a simple framework for which consumer interfaces agents will replace — _AustinCalvert_ · 2026-10-26
- Fewer, well-designed agent harnesses beat sprawling agent fleets, argues vixsheikh — annetgriffin · 2026-10-26
- Krugman: an AI bubble burst would push the US into a liquidity trap — GaryMarcus · 2026-10-26
- Twitter machine translation has connected civilizations more than any trade route — SuB8u · 2026-10-26
- Gary Marcus amplifies bearish math: unit revenue down 80% while costs and interest climb is 'fast track to bankruptcy' — GaryMarcus · 2026-10-26
- Old cybernetics textbook frames AI as a small subset: 'all systems are goal-directed' — _AustinCalvert_ · 2026-10-26