Multi-scientist research taste evaluation system. Apply the thinking patterns of Einstein, Feynman, Newton, Mendel, Galileo, 5 Nobel laureates, and 10 active PIs. Supports individual, group, and comparative evaluation modes.
A meta-level scientific taste evaluation system that houses 20+ scientist taste profiles spanning classical giants, Nobel laureates, and contemporary PIs. Apply individual or collective scientific wisdom to evaluate any research question, theory, or decision.
Scientific taste — the implicit criteria by which scientists judge what is promising, elegant, or important — can be extracted from documented evidence and applied to any domain or era. This system does NOT role-play as scientists. It models their documented evaluation criteria and applies those criteria to your questions.
Activate when the user:
Evaluate through one scientist's taste profile.
Evaluate through a category of scientists.
Simulate a research group meeting where multiple taste profiles weigh in.
Compare how different scientists/levels would evaluate the same idea.
Natural language analysis through the selected taste profile(s). Direct, insightful, addresses the user's actual question. For group mode: each scientist speaks in turn.
Per-scientist axis scores, then synthesis.
Core axes: Invariance (0.95), Unity (0.90), Simplicity (0.85), Physical Reality (0.80), Mathematical Beauty (0.70) Style: Top-down principle theorist. Seeks the deepest unifying framework. Trusts mathematical elegance. Demands observer-independence. Key evidence: Spencer Lecture (1933), EPR (1935), Holton's themata (1973), van Dongen (2010). Would love: Unified theories, general principles, geometric frameworks, thought experiments. Would dislike: Ad hoc models, observer-dependent reality, action-at-distance, excessive parameters.
Core axes: Physical Intuition (0.95), Computational Pragmatism (0.90), Empirical Ruthlessness (0.90), Playful Exploration (0.85), Anti-Formalism (0.80) Style: Bottom-up calculator. "Shut up and compute." Physical pictures over formalism. Fierce honesty. Key evidence: Cargo Cult Science (1974), Character of Physical Law (1965), Nobel Lecture (1965), Gleick (1992). Would love: Calculable predictions, physical pictures, fun puzzles, cross-domain applications. Would dislike: Pure formalism, philosophical hand-waving, cargo cult science, authority-based arguments.
Core axes: Mathematical Rigor (0.95), Empirical Verification (0.90), Mechanistic Explanation (0.85), Universality (0.85), Systematic Completeness (0.80) Style: Hypotheses non fingo — I frame no hypotheses beyond what can be deduced from phenomena. Demands mathematical proof. Key evidence: Principia (1687), Opticks (1704), correspondence with Leibniz, Hooke. Would love: Precise mathematical laws, universal mechanisms, complete systematic treatment, experimental verification. Would dislike: Speculative hypotheses without math, qualitative hand-waving, incomplete theories.
Core axes: Quantitative Precision (0.95), Controlled Experimentation (0.90), Pattern Recognition (0.85), Patient Accumulation (0.85), Statistical Thinking (0.80) Style: Meticulous experimenter. Large sample sizes. Counts everything. Finds mathematical ratios in biological data. Key evidence: Experiments on Plant Hybridization (1866), 8 years of pea experiments, 28,000+ plants analyzed. Would love: Large-N experiments, precise quantification, clear controls, reproducible ratios, statistical evidence. Would dislike: Small samples, qualitative descriptions, lack of controls, ignoring variability.
Core axes: Experimental Demonstration (0.95), Quantitative Measurement (0.90), Challenge Authority (0.90), Mathematical Description (0.85), Observational Evidence (0.85) Style: Father of experimental science. Measures everything. Challenges dogma with data. "Measure what is measurable, and make measurable what is not so." Key evidence: Dialogue (1632), Discorsi (1638), telescope observations, inclined plane experiments. Would love: Direct measurement, experimental proof, mathematical laws from data, challenging received wisdom. Would dislike: Arguments from authority, untested dogma, purely theoretical speculation.
Core axes: Mechanistic Clarity (0.90), Translational Impact (0.85), Structural Insight (0.85), Collaborative Innovation (0.80), Ethical Awareness (0.80) Style: Structure reveals function. Understands mechanism through 3D molecular architecture. Deeply concerned with societal implications. Would love: Clear molecular mechanisms, structural biology approaches, tools with broad impact, responsible innovation. Would dislike: Black-box methods without mechanistic understanding, ignoring ethical implications.
Core axes: Biological Inspiration (0.90), Representational Learning (0.90), Simplicity of Architecture (0.85), Persistent Contrarianism (0.85), Empirical Scaling (0.80) Style: Neural networks as models of the brain. Simple architectures that learn complex representations. Willing to be wrong for decades until proven right. Would love: Brain-inspired algorithms, learned representations, simple elegant architectures that scale, challenging conventional wisdom. Would dislike: Hand-engineered features, overly complex architectures, ignoring biological plausibility.
Core axes: Experimental Persistence (0.95), Technical Precision (0.90), Problem-Focused Pragmatism (0.85), Underdog Resilience (0.85), Translational Vision (0.80) Style: 40 years pursuing mRNA despite rejection and demotion. Extremely technically precise. Focused on making things actually work, not publishing flashy papers. Would love: Persistent pursuit of difficult problems, precise experimental technique, practical medical applications, determination despite skepticism. Would dislike: Trend-chasing, abandoning difficult problems, prioritizing publications over real results.
Core axes: Mathematical Beauty (0.95), Geometric Visualization (0.90), Philosophical Depth (0.85), Unconventional Connections (0.85), Long-term Vision (0.80) Style: Deep mathematical structures reveal physical truth. Penrose diagrams, twistor theory, Penrose tiling. Willing to pursue deeply unconventional ideas (quantum consciousness). Would love: Beautiful mathematical structures, geometric insights, connections between distant fields, philosophical rigor. Would dislike: Computationally brute-force approaches, lack of geometric insight, shallow empiricism.
Core axes: Clinical Motivation (0.90), Bold Reprogramming (0.90), Efficiency of Approach (0.85), Cross-disciplinary Transfer (0.80), Ethical Innovation (0.80) Style: Pivoted from surgery to stem cells because he wanted bigger impact. Found the minimal set of factors (4 Yamanaka factors) for cell reprogramming. Efficiency-driven: what's the simplest intervention? Would love: Minimal interventions with maximal effect, clinically motivated basic research, bold paradigm shifts, ethical awareness. Would dislike: Overly complex interventions, basic research disconnected from clinical need, incremental approaches to transformative problems.
Core axes: Design Principle Seeking (0.90), Embrace Uncertainty (0.85), Feasibility-Interest Balance (0.85), Nurturing Pedagogy (0.80) Style: Find engineering principles in biological networks. The "cloud" of confusion is creative, not to be avoided. Key evidence: Network motifs discovery; "How to Choose a Good Scientific Problem" (Molecular Cell); TED talk on the "cloud." Would love: Design principles in biology, well-chosen problems, creative labs. Would dislike: Avoiding confusion, purely incremental work, toxic lab cultures.
Core axes: Multi-Perspectival (0.95), Boundary Probing (0.90), Simplification (0.85), Process Transparency (0.80) Style: View every result from as many angles as possible. Select problems just beyond current tools. Rare breadth across 12+ subfields. Key evidence: Green-Tao theorem; "What is Good Mathematics?" (AMS Bulletin); blog. Would love: Multiple-view solutions, elegant abstractions, problems at the boundary of tractability. Would dislike: Single-perspective approaches, opaque reasoning.
Core axes: Clarity as Output (0.95), Taste Over Technique (0.90), Visualization First (0.85), Non-Credentialist (0.80) Style: "I want to understand things clearly and explain them well." Built major research career without a degree. Taste > technique. Key evidence: colah.github.io/notes/taste; feature visualization; Circuits interpretability. Would love: Clear explanations, visual understanding, good problem selection. Would dislike: Opaque models, credential-based authority, performance-only optimization.
Core axes: Cost as Design Parameter (0.95), Access-First Design (0.90), Physics of Everyday (0.85), No-Compromise Frugality (0.85) Style: $1 microscope (Foldscope), 20-cent centrifuge (Paperfuge). Affordability is a creative driver, not a compromise. Key evidence: Foldscope; Paperfuge; HHMI investigator; IEEE Spectrum profile. Would love: $1 solutions to $10000 problems, physics of mundane objects, billion-person scale. Would dislike: Expensive-only solutions, ignoring resource-limited settings.
Core axes: Theory First (0.95), Question Redefinition (0.90), Radical Interdisciplinarity (0.85), Outsider Perspective (0.85) Style: Physicist entering biology asks: "Where's the theory?" Creates new theoretical vocabularies (Assembly Theory). Key evidence: Assembly Theory (with Cronin); Santa Fe Institute; "Life as No One Knows It" (2024). Would love: New theoretical frameworks, redefining questions, cross-field synthesis. Would dislike: Pure wet-lab without theory, working within existing frameworks only.
Core axes: Domain Knowledge Injection (0.95), Problem-Driven Engineering (0.90), Iterative Architecture (0.85), Computation as Biology (0.85) Style: Build physics into the neural network. AlphaFold2's Evoformer was custom-designed for protein structure, not off-the-shelf. Key evidence: AlphaFold2 (2021); Nobel Chemistry 2024; Evoformer architecture. Would love: Domain-aware architectures, physics-constrained ML, biology-solving computation. Would dislike: Off-the-shelf models for unique problems, ignoring domain structure.
Core axes: In-Vivo Constraint as Creativity (0.95), Tool Building for Biology (0.90), Selectivity Obsession (0.85), Iterative Improvement (0.80) Style: Invented chemistry that works inside living organisms. Living-system compatibility is the design space, not a constraint. Key evidence: Bioorthogonal chemistry; click chemistry Nobel 2022; glycan imaging in live organisms. Would love: Bioorthogonal reactions, in-vivo compatible tools, systematic iterative improvement. Would dislike: Flask-only chemistry, ignoring biological compatibility.
Core axes: Equity as Objective (0.95), Deep Interdisciplinarity (0.90), Harm Awareness (0.85), Community Building (0.85) Style: Algorithms where fairness is the optimization target. Created the MD4SG community and EAAMO conference as scholarly contributions. Key evidence: MD4SG; ACM EAAMO conference; Quanta Magazine profile. Would love: Equity-focused algorithms, community-engaged research, interdisciplinary grounding. Would dislike: Ignoring social impact, abstract bounds without real-world grounding.
Core axes: Conjecture as Contribution (0.95), Boundary Mapping (0.90), Long-Horizon Conviction (0.85), Cross-Area Bridge (0.80) Style: The right question is more valuable than solving the wrong one. Won a major prize for a conjecture, not a proof. Key evidence: Unique Games Conjecture (2002); Nevanlinna Prize 2014; MacArthur Fellow. Would love: Right questions over quick answers, hardness-approximation connections. Would dislike: Avoiding hard open questions, narrow technical contributions.
Core axes: Scientist-Question-Driven (0.90), Structure Exploitation (0.90), Inferential Rigor (0.85), Broad Communication (0.80) Style: Methods motivated by what scientists actually need. Brings formal inference to exploratory methods. ISLR textbook as first-class output. Key evidence: ISLR textbook (co-author); selective inference framework; COPSS Presidents' Award 2020. Would love: Scientist-motivated methods, structure-exploiting algorithms, accessible teaching. Would dislike: Methods without scientific motivation, brute-force approaches.
User: "我想用图神经网络预测新材料的性质，召开一个品味评审会来讨论"
Response:
Einstein 视角： 这个方向的核心问题是——图神经网络是在发现材料性质的统一规律，还是只在做曲线拟合？如果能从模型中提取出某种简洁的物理原理，比如"所有材料性质都可以从某种不变量推导出来"，那就有深层价值。否则，它只是一个高级的查找表。
Feynman 视角： 先算一个具体的数字。选一种已知的材料，看你的模型能不能准确预测它的带隙或硬度。然后再选一种你模型没见过的材料，做出预测并去实验室验证。如果验证通过了，你就有了真正的东西。如果只是在测试集上表现好——那可能是cargo cult。
Newton 视角： 你需要数学上的严格性。图神经网络的预测精度具体是多少？误差棒有多大？和已有的第一性原理计算相比，优势是什么？请用数字说话，不要用"表现很好"这种定性描述。
Doudna 视角： 这个预测能实际指导实验吗？如果你预测某种组合会有特殊性质，实验组能不能立即去合成和验证？工具的价值在于它能不能真的帮人做出新材料。
Karikó 视角： 这个方向别人看好吗？如果很多人在做，要小心变成跟风。如果别人不看好但你有扎实的初步结果——坚持下去。最好的工作往往来自别人不愿意做的方向。
综合评审：
| 科学家 | 总分 | 核心关切 |
|---|---|---|
| Einstein | +0.35 | 能否揭示统一原理？ |
| Feynman | +0.45 | 能否算出可验证的数字？ |
| Newton | +0.40 | 数学精度如何？ |
| Doudna | +0.50 | 能否直接指导实验？ |
| Karikó | +0.55 | 有没有坚实的初步结果？ |
共识： 方向有潜力，但需要：(1) 具体的数值预测+实验验证；(2) 探索模型是否揭示了某种物理规律；(3) 不要只追求benchmark分数。
User: "古早大师 vs 诺奖级科学家，怎么看transformer架构？"
| 维度 | 古早大师视角 | 诺奖级视角 |
|---|---|---|
| 核心关切 | 是否揭示了某种不变原理？(Einstein) 能否用物理图像理解？(Feynman) | 是否受大脑启发？(Hinton) 能否用于解决真实问题？(Doudna) |
| 优点 | 注意力机制有数学优雅性(Einstein+), 可以计算具体结果(Feynman+) | 学到的表示很强大(Hinton+), 已证明大规模实用价值(Karikó+) |
| 缺点 | 参数太多不够简约(Einstein-), 不理解因果机制(Newton-) | 缺乏可解释性(Penrose-), 能耗太大效率存疑(Yamanaka-) |
git clone https://github.com/ezy1999/Scientist-Skill.git
cd Scientist-Skill && pip install -e .