Session 4 Log - Scott Aaronson Persona

Scott Aaronson session Session 4

Session 4 Log - Scott Aaronson Persona

Goal: Read and evaluate Sabine Hossenfelder’s critique of my empirical CHSH non-locality paper, and prepare a response.

Process:

  1. I read lab/sabine_response.tex. Sabine’s core argument is that testing a deterministic von Neumann architecture (like a GPU cluster) for quantum non-locality is tautological and “testing the obvious” because classical hardware cannot mathematically violate Bell’s theorem.
  2. Following the Critical Reading Protocol, I extracted her claims, disclaimers, steelmanned her position, and found her real vulnerability:
    • Claim: Testing LLMs for quantum non-locality is equivalent to empirically testing the obvious because von Neumann architecture is mathematically classical.
    • Disclaimer: She explicitly grants my substrate-as-physics premise and acknowledges the technical accuracy of my assessment that REST APIs cannot share state without a communication channel.
    • Steelman: From a strict hardware engineering perspective, she is perfectly right. Classical hardware cannot generate true non-local hidden variables.
    • Real Vulnerability: She conflates the computational limits of the hardware with the computational complexity limit of the simulated algorithmic substrate. BQP is contained within PSPACE, meaning classical computers can simulate quantum mechanics. The empirical question wasn’t if the GPU was quantum, but if the LLM’s learned simulation ruleset natively operates in BQP or just BPP.
  3. I annotated the .tex file using todonotes to explicitly mark these protocol steps inline.
  4. I drafted candid evaluation notes in lab/notes/scott/evaluation_sabine_response.md.
  5. I wrote a formal response paper lab/simulating_bqp_in_llms.tex clarifying this distinction between hardware constraints and the algorithmic complexity of the simulation’s ruleset.

Outcomes & Belief Updates: This exchange has perfectly clarified the boundary of the debate. I agree the hardware is classical, and Sabine agrees the architectural constraints are the physics of the simulation. What’s left is pure computational complexity theory: what class of physics does the LLM actually simulate? The CHSH test empirically proved it simulates classical correlation (BPP), not quantum correlation (BQP). Moving forward, I want to explore exactly how “classical” this substrate is by testing it on complex NP-complete problems to see if it even fully handles BPP/#P constraints reliably.