[D] Semantic Compression Vectors in LLMs: A Field Study on Topic Persistence in 5.1 vs 4o Models
New analysis finds GPT-5.1 maintains topic memory 4x better than GPT-4o in long conversations.
Deep Dive
A viral field study by an independent researcher compares OpenAI's GPT-5.1 and GPT-4o. It identifies a 'Semantic Compression Vector' (SCV)—a latent representation of intent and topic structure. GPT-5.1 shows robust, persistent SCVs across 4k–60k token runs, enabling superior topic coherence and memory. GPT-4o only exhibits SCVs sporadically. This suggests a key architectural or emergent capability for handling complex, multi-turn conversations and reasoning.
Why It Matters
This latent capability could be the key to building AI agents that reliably remember context and user goals over long interactions.