DeepSeek - Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models
Автор: AI Papers Podcast Daily
Загружено: 2026-01-17
Просмотров: 30
Modern large language models often work inefficiently because they use complex calculations to remember simple facts or common word patterns. To solve this, researchers introduced *Engram**, a "conditional memory" module that allows the AI to **quickly look up stored information* from a massive table instead of recalculating it every time. By handling the basic task of remembering static patterns, Engram frees up the AI’s "brain power" to focus on *difficult reasoning, math, and computer coding**. This system makes the AI act as if it is much deeper and more capable without slowing it down or requiring more expensive hardware. Ultimately, Engram helps models handle much longer pieces of text more accurately by separating **fast memory recall* from deep thinking.
https://github.com/deepseek-ai/Engram...
https://x.com/scaling01/status/201074...
Доступные форматы для скачивания:
Скачать видео mp4
-
Информация по загрузке: