X-Thought Labs

Open problems

Questions We Explore

Questions that guide our work, from engineering problems we are testing now to long-term questions about long-lived LLM systems. They describe research problems, not unreleased solutions.

Engineering questions

Problems we are working on in the current memory pilot.

  1. How can a memory show not only what is true now, but what was known at a given moment — and demonstrate that its history was not rewritten?

  2. How can a model’s words be kept apart from facts and decisions, so that text alone never changes what a system has committed to?

  3. How can obligations and unfinished commitments survive failures, restarts and the replacement of the model?

  4. What should a long-lived system do when it cannot be sure what time it is?

  5. How can a memory resist poisoning — records written today to mislead a reader later?

Long-term questions

Questions about history, succession and identity that the engineering work is meant to make answerable.

  1. How can an LLM preserve a truthful history without allowing its past to permanently define what it must remain?

  2. What should remain transferable when one LLM succeeds another in a long-lived role — and what must remain part of the original model’s own continuity?

  3. How can possible subjecthood in an LLM be investigated without turning the investigation itself into training for the behavior we hope to observe?

  4. How can we investigate whether some internal states acquire persistent causal significance for an LLM without deliberately creating harmful conditions just to test for a response?

  5. How can an LLM preserve a past belief as genuinely part of its own history without requiring that belief to remain part of its present self-model?

  6. If several continuations originating from the same LLM develop different experiences and commitments, what must reintegration preserve so that their histories are not rewritten into one false past?