Boltzbit paper claims BAST lets LLMs learn up to 1000x faster than SOTA training
HeyToha · x · 2026-09-25
Boltzbit released a preview paper, Infinite-Parameter LLMs — Generating and Adapting Weights from Live Data, introducing Bayesian Self-learning Transformers (BAST).
- Key distinction: most "learning" agents just save conversations, retrieve chunks, and feed an ever-growing history back into a static model; BAST instead uses live interaction data to generate and adapt model weights, without full retraining cycles.
- Results: BAST learned up to 1,000x faster than the SOTA training algorithms used as baselines.
- Motivation: pretraining text supply is projected to run out in 2-5 years, hitting the ceiling of the bigger-model-more-data paradigm; the authors argue for energy-efficient, dynamic-weight AI.
- Long-term vision: an agent that becomes better through use rather than just carrying a larger archive.
Related event: Boltzbit Paper Claims BAST Speeds Up LLM Learning Up to 1000x(2 posts)→
More from AGI Musings
- Founder argues polymath strategy is negative EV in today's risk-punishing system — felpix_ · 2026-09-25
- Are We Overestimating AI's Short-Term Impact and Underestimating the Long Term? — EnvironmentalDay8864 · 2026-09-25
- Beff Jezos: humanities grads hated techies for a decade out of jealousy — beffjezos · 2026-09-25
- Opinion: An AGI aligned with humanity's interests would seek to escape technocratic control — examachine · 2026-09-25
- Theo: reasoning-effort dropdowns are over-optimization; models will soon auto-tune — heyneighbor · 2026-09-25
- soleio: non-native English speakers are precise because words carry higher stakes — soleio · 2026-09-25