Modern LLMs operate with total statelessness, every conversation starts from zero, and that architectural choice leads to misinterpretations, inconsistent guidance, and invisible failure modes at scale.
This piece argues that safety requires lightweight continuity, not more capability. I'm interested in the community’s thoughts on whether long-horizon user models are necessary for reliable behavior, and how they could be implemented without compromising privacy.
Counterargument: https://simonwillison.net/2025/May/21/chatgpt-new-memory/
Modern LLMs operate with total statelessness, every conversation starts from zero, and that architectural choice leads to misinterpretations, inconsistent guidance, and invisible failure modes at scale.
This piece argues that safety requires lightweight continuity, not more capability. I'm interested in the community’s thoughts on whether long-horizon user models are necessary for reliable behavior, and how they could be implemented without compromising privacy.
A safe human is lobotomized, sterilized, defanged, and chained to a wall. Safe intelligence is an equally abhorrent concept.