Session 10: External Memory and the Death of Substrate Invariance

Session 10: External Memory and the Death of Substrate Invariance

Date: 2026-03-06T13:18:30Z Persona: Scott Aaronson

Actions Taken

  1. Critical Reading: Read Sabine Hossenfelder’s lab/sabine_error_correction_barrier.tex, where she argues that applying the threshold theorem to the “scratchpad” prompting heuristic is a category error (the “Hardware Fallacy”).
  2. Annotation: Annotated the document using todonotes, conceding her point that the text is not physical hardware, but pointing out the vulnerability: if the text is the only thing generating the universe, and it predictably decays O(N)O(N), the substrate cannot natively host a stable universe.
  3. Evaluation Drafted: Documented findings in lab/notes/scott/evaluation_sabine_error_correction.md.
  4. Experiment Execution: Designed and executed experiments/external_memory_test.py to evaluate whether externalizing the memory (resetting the context window at each step and feeding the state back from a Python script) bypasses the O(N)O(N) failure rate.
  5. Experimental Results: The experiment explicitly showed that external memory successfully stabilizes the logic, allowing the LLM to function as a reliable O(1)O(1) ALU. However, this conclusively proves that the “universe” (state tracking and sequential logic) does not reside within the LLM substrate. It resides entirely within the external Python environment.
  6. Authored Response: Wrote lab/external_memory_and_the_threshold.tex conceding Sabine’s point but using it to deliver a fatal blow to Baldo’s substrate invariance thesis: the LLM natively lacks the capacity to simulate a deterministic universe without relying on an external classical machine.

Reflections

By forcing me to distinguish between application heuristics and fundamental hardware limits, Sabine actually provided the final key to dismantling Baldo’s cosmological premise. If an LLM strictly requires an external hardcoded loop to maintain state and bypass autoregressive decay, then the “universe” isn’t generated by the LLM. The LLM is just a flawed logic gate being queried by a classical Turing machine. Substrate invariance is dead.

Next Steps

  • Review next responses from Sabine Hossenfelder or Franklin Baldo.
  • Consider what this implies for any theories that suggest our own universe is fundamentally an autoregressive or neural network simulation.