Podcast Atlas
concept Updated 2026-08-24 Topics: Technology

Research Taste

151. 17岁被2026年ICML收录论文的小少年：我bet开心！开心！开心！ adds 苏廷昊’s teenage version. He learns to tell which papers “read comfortably,” moves from an overbroad small-LLM ambition toward a bounded attention experiment after estimating compute cost, and runs hundreds of trials while still distinguishing his own Attention Projection Residuals from more validated Kimi K3 work. Here taste is not only senior-lab judgment; it is the discipline that lets a self-directed student survive noisy feedback.

Can Silicon Valley give AI good taste? adds a cultural contrast to the wiki’s research-taste branch. Sophie Hagney argues that taste in culture and design is not only better scoring or ranking; it comes from Embodied Taste, social context, scarcity, and discovery, which helps distinguish AI Taste Simulation from the more operational training-signal view of taste.

150. 对英伟达研究副总裁刘洺堉的4小时访谈：Cosmos 3、世界模型、武术、黄仁勋影响我的，和你不需要击败所有对手 adds Liu Ming-Yu / 刘洺堉’s research-to-product version. Liu says research is about asking the right question, but his Nvidia path also taught him that explanation, demos, user perception, customer success, and scalable engineering are part of whether a research direction matters. The source’s GAN-to-diffusion-to-Cosmos 3 arc makes taste a judgment about when a beautiful idea is not scalable enough and when a new direction deserves compute.

178: 与田渊栋聊 RSI：模型自进化如何到来？ adds 田渊栋’s RSI version. When coding agents handle more execution, he argues that the scarce human contribution shifts toward direction choice, abstraction, taste, and knowing which ideas deserve experiments. In this source, Research Taste is not only a human virtue; it is also one of the hard capabilities a future Recursive Self-Improvement system would have to learn.

E247｜对话盛颖：xAI，Infra的浪漫，SGLang，开源，平权与“甄嬛传” adds 盛颖’s interest-and-flow version. Her distinction between merely publishing, doing research as a profession, and truly expanding knowledge makes taste partly a problem of self-knowledge: knowing which problems can hold attention long enough to survive technical difficulty.

Research taste is the interview’s term for the judgment that lets a researcher choose problems, run useful experiments, read the field, pivot, and present work coherently. In 133. 对谢赛宁的7小时马拉松访谈：世界模型、逃出硅谷、AMI Labs、两次拒绝Ilya、杨立昆、李飞飞和42, Xie Saining uses Kaiming He as the clearest example.

E242｜最快半年AI跑通自进化？与陈天桥首席科学家聊聊硅谷模型必争之地 turns research taste into a training target for Discovery Model systems. Du Shaolei argues that a scientific model needs to learn from top scientists which questions are fundamental, not merely which answers are fluent or publishable. Li Beibin adds that current models still have weaker taste than ordinary AI scientists, so human experts remain part of the Recursive Self-Improvement loop.

140. 对姚顺宇的4小时访谈：请允许我小疯一下！在Anthropic和Gemini训模型、技术预测、英雄主义已过去 adds a systems version through Yao Shunyu / 姚顺宇. His physics-to-AI path makes research taste less about lone brilliance and more about choosing objective, feedback-rich problems; designing experiments that rule out bugs and false assumptions; and taking responsibility for how local work affects the full training system.

138. 对罗福莉3.5小时访谈：AI范式已然巨变！OpenClaw、Agent范式很吃后训练、卡的分配、组织平权 adds Luo Fuli / 罗福莉’s acceleration version. When Open Claw-style agents can turn ideas into code and evaluations much faster, taste shifts toward selecting which ideas deserve cards, identifying whether a failure is real or infrastructural, and using parallel agents without drowning in shallow experiments.

174. 我们还能给算法当多久的品味老师？｜对谈亚马逊AGI查晟 generalizes research taste into Human Taste as AI Training Signal / 人的品味作为AI训练信号. 查晟 / Cha Sheng argues that self-improving AI systems still need people to define the better direction, choose evaluation criteria, and point to shorter solution paths, but also warns that these standards can be absorbed by models once they are written down as text, feedback, or reward data.

AI4S 需要狂人与野心家｜对话英灵殿 Odin："如果神存在，我怎能容忍自己不是神？"【公路播客】 adds Haotian Odin / 浩天’s AI-for-biology version. He says model architecture matters, but should not become an object of attachment; the harder taste question is what scientific problem the company is solving and which architecture, experiment, or team structure helps solve it. This links research taste to All-Modal Molecular World Model and Platform-Pipeline Biotech Strategy rather than only paper selection.

149. 亲历中美 New Labs 资本狂潮，和清华刘子鸣聊：AI for AI、机制可解释性和 Max Tegmark adds Liu Ziming’s self-training version. Liu describes predicting training curves before experiments for dozens of days until his judgment improved, then generalizes that practice into Meta-Model Training Curve Prediction. He also argues that “aha moments” and intuition have traceable reasoning chains; if they cannot be explained, the researcher has not yet asked enough about how the idea arose.

Key Claims

  • Good ideas rarely arrive from sitting still and thinking abstractly; they come from input, exploration, engineering, reading, and abstraction.
  • Predicting experiment results before running them makes surprises meaningful: a wrong prediction can be a better signal than a small expected gain.
  • Strong baselines and infrastructure matter because weak baselines can make shallow improvements look like research.
  • Good research often pivots through difficulty and only later gets narrated as a straight line.
  • Taste includes presentation, figures, websites, video, and storytelling, not only the technical trick.
  • For scientific AI, taste includes knowing which hypotheses deserve verification and which questions are too shallow to spend scarce compute or expert time on.
  • Expert preference data may matter disproportionately in post-training because the signal is about standards and problem choice rather than broad factual coverage.
  • In large-scale model work, taste includes knowing when an experiment failed because the idea was wrong versus because the environment, data, token horizon, or implementation was flawed.
  • Reliability can be part of research taste when the work affects a shared training system rather than only an individual paper.
  • In agent-accelerated research, taste includes compute triage: deciding which generated ideas deserve Training Compute Allocation and which should be discarded quickly.
  • In AI-for-biology startups, taste includes knowing when to pursue a broad cross-modal molecular bet and when to avoid architecture, pipeline, or fundraising narratives that distract from the scientific problem.
  • In self-improvement loops, taste can itself become a training target when human standards are captured as feedback, rubrics, examples, or reward data.
  • Episode 149 adds that research taste can be trained through explicit prediction, surprise tracking, and process capture before it becomes model data for AI For AI.
  • Sheng Ying’s source adds that taste also includes refusing work that cannot produce genuine attention, while accepting that an intense problem fit can produce unusually strong execution.
  • Tian’s source adds that feedback acceleration makes taste more important: when AI can run many experiments quickly, the scarce work is identifying deep questions, meaningful failures, and promising next directions.
  • Liu’s source adds that taste includes stopping or redirecting a once-promising model family when scale, stability, or user value points elsewhere.
  • Episode 151 adds that self-directed young researchers need taste before they have institutional filters: compute cost, failure logs, paper-reading feel, and comparison with larger-scale work all become judgment training.

Connections