A new blog post argues KV cache is not a persistent mind in an LLM chat
birchlse · x · 2026-07-29
The author argues against a popular intuition in LLM discussions: that a model’s KV cache might be storing a persistent mental state. The post asks whether there is really a continuing mind on the other side of a chat, and says the answer is likely no.
The core point is that what looks like continuity in conversation may be an artifact of how context is represented, not evidence of an enduring self or mind. The piece is framed as a blog post raising doubts about strong “memory = mental state” claims in LLMs.
More from AGI Musings
- Ethan Mollick says AI writing may not be “slop,” but it is still changing human work — emollick · 2026-07-29
- Zuckerberg says Meta is preparing a positive vision for superintelligence — altryne · 2026-07-29
- Marcus Hutter argues AGI could make a job-less future desirable and likely — mhutter42 · 2026-07-29
- A repost mocks Pause AI for copying protest tactics in ways that look absurd — RishiBommasani · 2026-07-29
- Science study finds sycophantic AI boosts user certainty and reduces willingness to repair conflict — stanfordnlp · 2026-07-29
- Hundreds of Thousands of Tons of Books Landfilled Annually: AI Training as a Better Alternative — aronchick · 2026-07-29