Boltzbit's BAST paper claims LLMs can learn 1000x faster by generating weights from live data
ahuja_priyank · x · 2026-09-24
Boltzbit released a preview of its paper "Infinite-Parameter LLMs — Generating and Adapting Weights from Live Data", introducing Bayesian Self-Learning Transformers (BAST).
- Instead of adding more context, BAST turns feedback from live interactions into targeted weight updates: a correction becomes a learning signal that changes the model, without replaying full history
- The company claims up to 1000x faster learning than human-engineered SOTA training algorithms (unverified by third parties)
- Motivation: pretraining text supply may run out within 2-5 years, so static-weight scaling is hitting a ceiling; the goal is a compounding intelligence layer owned and controlled by the customer
More from Models
- Claude Opus 5.5 Designs a Buildable LEGO Microduck: 1,113 Real Parts, Zero Collisions — soleio · 2026-09-24
- Sarvam Vision 2.1 launches with 94.97% on OmniDocBench and OCR for 22 Indian languages — itsOmSarraf_ · 2026-09-24
- Why Persimmon builds human-behavior models on NVIDIA Nemotron 3 Ultra base instead of an aligned assistant — CShorten30 · 2026-09-24
- Jev-as-a-Judge: cheap judge cascade keeps 99% of GPT-6 accuracy at 57% of the cost — dair_ai · 2026-09-24
- Odyssey Launches Agora-2, a Multi-Agent World Model Simulating Up to 20 Humans and Agents in Real Time — damianplayer · 2026-09-24
- Even an Elon-endorsed take: inferior models in better products now beat better models in worse apps — elonmusk · 2026-09-24