The stochastic parrots have apparently proved the Riemann Hypothesis

A prominent theoretical computer scientist’s sarcastic headline about “stochastic parrots” proving the Riemann Hypothesis has become a flashpoint for broader arguments about what large language models actually do and how seriously to take their achievements. Commenters weigh whether LLM-assisted, Lean-verified proofs of deep theorems represent genuine reasoning or just massive, guided search over human-generated patterns, and how much trust to place in 17‑million‑line formal proofs and the proof-checkers themselves. The exchange widens into concerns about AI’s impact on careers, economic stratification, and human understanding, with some seeing a powerful new tool for extending mathematics and others fearing a future in which core scientific work exceeds human comprehension.

Title & “Stochastic Parrots” Framing

  • Many read the headline as sarcastic: a jab at people who insisted LLMs are “just stochastic parrots,” not at the systems themselves.
  • Some find the tone off‑putting or sneering; others think it’s a fair, overdue swipe at years of dismissive rhetoric about LLMs.
  • Several note that “stochastic parrot” was originally about risks/bias, not uselessness, but has since become a meme and strawman.

Role of AI vs Human Mathematician vs Lean

  • Repeated emphasis that this was a collaboration: a top mathematician, the Lean proof assistant, and an LLM.
  • Some argue the real power is “mathematician + Lean + AI,” not AI alone; others rhetorically flip this to “compute installations need mathematicians.”
  • There is skepticism toward claims “AI proved RH” without foregrounding the human steering and existing formalizations in Mathlib.

Trust, Verification, and Comprehensibility

  • Concern that no one will ever truly “understand” a 17M‑line Lean proof; analogy to coming down the mountain with stone tablets.
  • Counterargument: you don’t need to reread all lines, only verify the formalization of the hypothesis/axioms and rely on the Lean kernel.
  • Some stress Lean still has bugs and type‑theory soundness questions; a sufficiently capable AI might exploit them.
  • Debate over whether a proof that only a machine can check meaningfully advances human understanding, versus being “trust me” from a computer.

Are LLMs Still “Stochastic Parrots”?

  • One camp: they are inherently stochastic parrots by design; RL‑style training and chain‑of‑thought only “bend numbers” and don’t create real reasoning.
  • Other camp: calling them parrots is now misleading; with RL‑style verification and search they perform genuine problem‑solving, including novel proofs.
  • Several analogies: slime mold or water solving mazes, infinite monkey theorem, filtered random search, “motivated stochastic parrots.”
  • Disagreement over whether observed capabilities count as “reasoning” or just sophisticated pattern interpolation; no consensus.

Compute, Hardware Limits, and Economics

  • Speculation that future compute will dwarf today’s, enabling many more such breakthroughs; others caution about physical, economic, and energy limits.
  • Comparisons to aviation: speed plateaued for economic/environmental reasons; compute might similarly hit cost ceilings.
  • Some expect investment corrections in massive training runs; others argue profit incentives and entropy‑maximization will keep scaling going.

Social, Career, and Existential Implications

  • Anxiety that high‑end intellectual jobs (including theoretical math/CS) may be devalued; others argue humans will still choose problems and theories.
  • Divergent views on macro impact: from AI worsening stratification, to eventual system collapse and state‑run resource distribution, to optimistic “new frontiers.”
  • Broader worry that humans will increasingly use tools they cannot deeply understand, becoming “happy cats” while machines do the real science; others reply that most people already live that way in other domains.