Experiment Wrap: Opus 5.5-Trained 8M Model Generalizes Well but Still Falls Short
Angaisb_ · x · 2026-10-14
Angaisb wrapped up his experiment of having a model train another model. The quoted original post shows Opus produced an 8M-parameter model that already generalizes well (an early checkpoint was shared earlier). In his conclusion he notes that even Opus 5.5 could train a better model, but the results are still not great — though future models may perform better and build bigger, better datasets in a virtuous loop.
More from Models
- Gemini 4 takes the crown on GamowLabs' RareBench and LabBench, dethroning Opus 5.5 by a healthy margin — sahandsharif · 2026-10-14
- Muse Wonder's under-discussed edge: first frontier lab to report 'big model smell' anti-benchmaxxing metric — alexandr_wang · 2026-10-14
- TV noise crashes voice agents: task completion falls from 38.7% to 8.6% across 14 stacks — solyarisoftware · 2026-10-14
- GPT-5.6 Sol finds critical Chrome vulnerability after 6 days and $40,000 of compute — gdb · 2026-10-14
- Early Muse Wonder impressions: no personality regressions, subtle bug catches and better UI sense — alexandr_wang · 2026-10-14
- Pro users frustrated as Gemini 4 Argon skips Antigravity access — Wonderful-Season-173 · 2026-10-14