AI Models Now Outperform Humans at Experimental Research Taste
Summary
研究显示 AI 模型在科研实验的『品味』（taste）判断上已超越人类，能够选择更有价值的研究方向与实验设计。
Similar Articles
We're breaching the point where the "taste" of a model's decisions is becoming as or more important than its raw intelligence
作者认为模型能力已达到临界点，选择 AI 模型的关键正从原始智能转向设计品味与风格契合度，未来将出现面向不同用户品味的模型分层。
TasteVal - Experimental Research Taste of Frontier AI's is Doubling Every 3 Months and the Best Model Now Exceeds Our Human Baseline
TasteVal reports that frontier AI models' experimental research taste has doubled roughly every 3 months since December 2025, with Opus 5.5 now surpassing an expert human baseline — a key predictor of how quickly ASI could be reached once coding is automated.
@VraserX: The human advantage inside frontier AI labs is increasingly described as research taste. Choosing the right problem. Kn…
The human advantage in frontier AI labs is increasingly described as research taste, which involves choosing the right problem and knowing which results to trust. If AI models develop this ability, it could transform the research loop completely.
Why I don't think AI will take our taste: we get bored of whatever is common and models make what's common
The author argues AI won't erase human taste: brains reward surprise while models gravitate toward the statistically common, so audiences will always move on before models catch up. The post surveys supporting research (Artificial Hivemind, RLHF diversity reduction, Billboard typicality studies) and honestly presents counter-evidence like LLM-generated research ideas and AI-preferred poems.
@li9292: https://x.com/li9292/status/2089739889906688235
This article explores the meaning of 'human touch' in text, analyzes statistical features of AI detection methods such as perplexity and burstiness, and argues that human writing can be diagnosed but cannot be optimized without creating a paradox.