Click any tag below to further narrow down your results
Links
This article discusses the unique difficulties in hardware design for large language model inference, particularly during the autoregressive Decode phase. It identifies memory and interconnect issues as primary challenges and proposes four research directions to improve performance, focusing on datacenter AI but also considering mobile applications.
The article analyzes Claude's memory system, highlighting its use of on-demand tools and selective retrieval compared to ChatGPT’s pre-computed summaries. It details the methodology used for reverse-engineering Claude's architecture and outlines key differences in memory and conversation history management.
This article critiques the performance of LLM memory systems like Mem0 and Zep, revealing they are significantly less efficient and accurate than traditional methods. The author highlights the architectural flaws that lead to high costs and latency, arguing that these systems are misaligned with their intended use cases.
This article presents the Titans architecture and MIRAS framework, which enhance AI models' ability to retain long-term memory by integrating new information in real-time. Titans employs a unique memory module that learns and updates while processing data, using a "surprise metric" to prioritize significant inputs. The research shows improved performance in handling extensive contexts compared to existing models.