Alex Kessinger

I got rage baited by "Did RLMs just kill RAG (again)?" So, I'm catching up on this trending paper: What are recursive language models (RLMs)? RLMs are an inference strategy to allow an LLM to process "infinite" context. How it works: 1. Variable space: Instead of pasting a large document into the context window, the document is loaded as a variable inside a coding environment. (The key here is that the root LM doesn't see the content of the document, it only sees that there's a document) 2. Symbolic code: The root LM writes code to process the large document into manageable pieces (e.g., grep and chunk) 3. Recursion: The root LM spawns sub-agents to process smaller slices of the original large document and only returns the response to the root LM. This means the root LM's context window is never polluted with irrelevant information (avoids context rot) Online I've seen discussions around different topics: - How RLMs are different from coding agents? - How RLMs are different from RAG? - How RLMS are different from grep with extra steps? I think there's a key mental shift to understand the difference: Variable space vs token space Old Mental Model: You have a text. You tokenize it. You stuff as much as possible into the prompt (the "Context Window"). If it doesn't fit, you cut it or summarize it. The model "sees" the text. New Mental Model: The text exists only as a variable in a Python environment (REPL). The model never sees the text unless it explicitly writes code to print a snippet of it.

Obsidian Sync now has a headless client, so you can sync vaults to a server without using the desktop app. Try the open beta:

Rewatching many of my "favorite " movies from my youth (with my teenagers). Surprisingly, many of them don't stand up, more surprising is that Con Air is the gold standard. We. All. Put. Our. Phones. Away. Very few have met that standard so far.

I recently saw a graphic that was about making things possible vs making things easy. It had an image of a bridge vs a tight rope walk. I loved it but can not find it anymore. Does anyone know what the image is?