More Pretraining Data Reduces Diversity But Boosts Alignability, Argues Thread
minhnguyen · x · 2026-11-15
Synthesizing a dozen papers (e.g. "When Bad Data Makes Good Models"), minhnguyen proposes: more pretraining data does cause mean reversion given the same/minimal specialized RL, but it also increases alignability. Alignability becomes a force multiplier if you properly craft RL objective heuristics to maximize alpha—an objective that's currently uncommon in RL practice.
Related event: AI yields consensus, not alpha: two perspectives(2 posts)→
More from Research
- Flint 40B self-improves into Flint-Law, gaining 15% on legal tasks over 11 steps — RichardSocher · 2026-11-15
- Sebastian Raschka publishes from-scratch RoPE tutorial and survey of positional embeddings in modern LLMs — rasbt · 2026-11-15
- Training your own LLM judge for ~$5 in 1h40 on one GPU, 89.5% accuracy — MaziyarPanahi · 2026-11-15
- Ivy Grad Trading Analogy: Reading More Doesn't Help Without Deliberate Practice — minhnguyen · 2026-11-15
- LeCun's CoRL slide: 'If you are interested in human-level AI, don't work on LLMs' — Heavy-Ad9232 · 2026-11-15
- Dataset Distillation via Drifting compresses 1,300 ImageNet images per class into just 10 — kastnerkyle · 2026-11-15