Interview
Scott Aaronson: Computational Complexity and Consciousness | Lex Fridman Podcast #130
- The universe is expected to remain stable for the foreseeable future, with 14 billion years of history suggesting it will not crash due to probing unknown physical phenomena like black hole singularities.
- A future fundamental theory of physics is considered a worthy pursuit, though it would not explain emergent behaviors; conversely, testable objective collapse theories like GRW are predicted to likely be disproven in the coming decades.
- Future large language models (e.g., GPT-5 through GPT-7) may eventually achieve human-level reasoning via predictive processing, but a training ceiling may exist due to the finite amount of internet data, with current training costs estimated at four to five million dollars.
- Regarding the P vs. NP problem, a 2–3% probability is assigned to P equaling NP, with a potential practical impact determined by algorithmic efficiency; if proven without an efficient algorithm, it would theoretically ravage complexity theory but remain practically irrelevant, potentially allowing the breaking of encryption, Bitcoin, and neural network optimization.
- Political risks include a high fear of contested election results leading to violence, unclear succession, and national destruction, alongside concerns that cancel culture is ineffective against racism and sexism while threatening open discourse.
- Institutional failures during the pandemic were described as staggering and unpredictable, with suggestions that a "World War II mentality" might be required to accelerate vaccine development, though divisions in the country were noted to have increased rather than decreased.
- Consciousness is characterized as potentially non-reducible to computation, with the "pretty hard problem" defined as quantifying consciousness across physical systems; passing the Turing Test implies emulation of consciousness, though its reality may depend on the observer's understanding of the mechanism.
- Physics suggests that faster-than-light travel would imply backward time communication and closed time-like curves, while quantum uncertainty and chaos may prevent the prediction of all brain behaviors even with advanced technology.
- Complexity classes are detailed with BQP contained in PSPACE via classical simulation requiring exponential time, and SZK is highlighted as a fundamental class, noting the "Complexity Zoo" contains approximately 500 classes created in 2002.
- Simulations of the universe are posited to be indistinguishable from reality if perfect, whereas imperfect simulations with exploits (bugs) might be indistinguishable from a world with miracles, shifting the debate from metaphysical to non-philosophical.