MIT CSAIL Professor Hal Abelson emphasizes the importance of free software for securing our agency in our digital worlds. His project, App Inventor, enables anyone to create apps without coding. He also discusses free software freedoms, the book 'Blown to Bits', and the importance of privacy and data provenance in today's digital age.
This paper introduces Cross-Layer Attention (CLA), an extension of Multi-Query Attention (MQA) and Grouped-Query Attention (GQA) for reducing the size of the key-value cache in transformer-based autoregressive large language models (LLMs). The authors demonstrate that CLA can reduce the cache size by another 2x while maintaining nearly the same accuracy as unmodified MQA, enabling inference with longer sequence lengths and larger batch sizes.
MIT CSAIL researchers have developed three neurosymbolic frameworks - LILO, Ada, and LGA - that use natural language to help large language models (LLMs) build better abstractions for coding, AI planning, and robotics tasks.
