​About the AI Research Circle

​The AI Research Circle is a community gathering at where we explore AI research together. No research background required—just curiosity.

​Each session, we pick a topic, break it down, and open it up for discussion. The goal: make cutting-edge ideas accessible and spark conversation across disciplines.

​Session Details

​Session Theme: Memory & Long-Context LLMs (and when RAG still wins)

​LLMs can now accept 100k+ token context windows. But “accepting” tokens isn’t the same as using them well: attention gets expensive, models can miss information in the middle, and performance can degrade past training length.   

​This session is a practical survey of how long-context systems work (RoPE scaling, long-context fine-tuning, sparse attention, recurrence/state-space approaches), plus an engineer’s discussion of the real tradeoff: when to rely on a bigger window vs. when retrieval (RAG) is the better tool.

​Reading

• ​Required: Memory & Long-Context LLMs: A Primer (~15 min skim)

• ​Optional deep dive (for the curious):

  • ​

    • ​Attentive Language Models Beyond a Fixed-Length Context

    • ​LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding

    • ​YaRN: Efficient Context Window Extension of Large Language Models

    • ​Efficient Streaming Language Models with Attention Sinks

​What we’ll cover:

• ​Why long context is expensive + fragile (quadratic attention, extrapolation limits) 

• ​What breaks in practice (lost-in-the-middle, needle-in-haystack)   

• ​A map of the solution space: positional scaling, efficient fine-tuning (LongLoRA), architecture shifts (Transformer-XL → Mamba/SSMs)   

• ​Where RAG fits (external memory) and hybrid patterns people actually ship 

​Who should join:

​Anyone building with LLMs who’s asked:
“Should I just increase context?”
“When does RAG actually help?”
“What breaks first?”