The Einstein AI Model

Role of AI in Science vs. Mathematics and Experimentation

• Several argue the post overemphasizes conjecture: in many fields, the hard part is proof and especially physical experimentation, not just “having the idea.”
• LLMs may be huge aids for trying many “crazy” math ideas, cleaning data, literature review, code, and mundane scientific work, even if they don’t originate revolutions.
• Others stress that real science is constrained by slow, expensive experiments (biology, medicine, physics), so even a “genius” AI would hit logistical limits.

Benchmarks and Evaluating an “Einstein Model”

• Debate over whether we need benchmarks at all if true breakthroughs can just be tested in reality; counterpoint: you still need programmatic evaluation to track model progress.
• Suggested benchmarks: train on data only up to a cutoff (e.g., pre‑2023, or pre‑1905 physics) and see if the system can rediscover later results or design the same experiments.
• Practical problems noted: lack of clean post‑cutoff corpora, legal/IP issues with proprietary scientific datasets, and the difficulty of knowing whether we’re “leading” the model.

Creativity, Questioning the Status Quo, and “Move 37”

• Many push back on romanticized “challenge the status quo” narratives: questioning is easy; being right is hard.
• Some say current systems are “straight‑A students,” optimized to agree and be helpful, not to question training data or propose truly counterintuitive axioms.
• Others note that non‑LLM AI (e.g., game agents, genetic algorithms) already shows surprising, out‑of‑the-box solutions, though often too brittle to use in practice.

Hallucinations, Honesty, and Cultural Tuning

• Long subthread on wanting models that say “I don’t know” more often versus being “overly compliant helpers.”
• Attempts to get more blunt, “Dutch‑style” behavior mostly fail; models still hallucinate and adopt American-style politeness and phrasings.
• Discussion of RLHF and product incentives: providers may tune for engagement and apparent helpfulness rather than strict factuality or epistemic humility.

Capabilities, Hype, and Moving Goalposts

• Some complain that critiques keep shifting from “not human-level” to “not Einstein-level,” calling this goalpost moving; others reply this piece is about a specific, real weakness, not denying that current systems are AI.
• Skepticism about near-term “compressed century” claims; others emphasize ongoing exponential improvements in compute and expect much better models in a few years.
• Broad agreement that near-term impact is “human + machine”: researchers ask good questions, AI accelerates search, synthesis, and iteration, not autonomous Einsteins—at least not yet.