Shared attention reframes LLM inference steering

This is less a finished product than a provocative research memo about how humans might interact with small language models more like simulators than chatbots.

• –The strongest idea is that context length is not just prompt capacity but a memory-depth control that changes how quickly generation falls into repetitive attractors
• –The post breaks behavior into collapse, rich dynamics, and noise regimes, with temperature and context length interacting in measurable ways rather than serving as simple tuning knobs
• –The proposed interface shifts UX from “write better prompts” to “steer trajectories,” with checkpoints, forks, EOS-based stopping, and live instrumentation
• –The evidence is interesting but still early: it is based on SmolLM-135M autoloops, with code and reproduction details explicitly marked as still forthcoming
• –If the effect generalizes to larger models, the bigger implication is for inference tooling and experimental IDEs, not mainstream chat interfaces