AI self-improvement gets models decent fast, but topping Kaggle is still hard
subinium · x · 2026-10-13
Subinium shares hands-on lessons on RSI (using AI to improve AI): getting a model to decent performance is fairly doable now, but reaching the top — even on Kaggle — remains hard. The attached screenshot shows linking Colab Pro to Kaggle unlocks extra GPU compute hours. He encourages continued experimentation.
More from Research
- RLVR produces "evolved modularity" AIs; 3D reasoning gains confirmed as RL-trained, not emergent — jessi_cata · 2026-10-13
- Cosma Shalizi's notebooks read Attention and Transformers as kernel smoothing and Markov models — tak3sh8 · 2026-10-13
- Nat Lambert: the next frontier research lab might be a data company — joecole · 2026-10-13
- Schema agent harness hits 99% RHAE on ARC-AGI-3, paper and code released — xiuyu_l · 2026-10-13
- Google's AMIE publishes first prospective clinical study in The Lancet, 90% diagnostic match — demishassabis · 2026-10-13
- Brain-inspired recurrent networks explain perceptual grouping dynamics in natural scenes — tserre · 2026-10-13