Discovery Model

178: 与田渊栋聊 RSI：模型自进化如何到来？ adds 田渊栋’s stronger-discovery aspiration. In this source, low-level Recursive Self-Improvement can optimize code or operators, but a higher-order discovery system would need to act more like an Einstein or Newton: extracting deep structure from sparse evidence, not merely running many benchmarked experiments.

A discovery model is the source’s name for an AI system that can propose new hypotheses and verify whether they hold. In 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, Apodex uses this frame to distinguish itself from generative chat, image, or video products: the goal is not only to answer questions, but to solve hard scientific and technical problems.

The episode argues that the hard part is not producing a plausible hypothesis. A useful Discovery Model has to find questions humans have not already written down, search and reason across evidence, use code or simulation when possible, and pass AI Verification. For open scientific domains, it also needs Research Taste: the ability to prefer fundamental problems over shallow or merely publishable ones.

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 a wet-lab-facing version through Yinglingdian AI / 英灵殿. Haotian Odin / 浩天’s Scientific Discovery Automation goal is to automate data analysis, hypothesis formation, and path search in biological discovery, but the source also shows why discovery models need Domain Expert Alignment and experiment: molecular candidates are not scientific progress until they survive synthesis, binding, measurement, and application constraints.

Key Claims

• Discovery requires novelty, but novelty without verification is not enough.
• Out-of-distribution hypotheses are especially hard because the model cannot simply retrieve the training-set answer.
• Deep Research is a stepping stone because search, planning, and synthesis are needed before a model can make scientific proposals.
• Code, simulation, formal math, and agent-team critique are all possible verification surfaces.
• Top-scientist feedback is treated as a scarce post-training signal for teaching problem choice and research taste.
• Tian’s source adds that scaling up from low-order optimization to deep discovery requires abstraction and principle-finding, not only faster experiment throughput.

Connections

• Apodex, Chen Tianqiao / 陈天桥, Du Shaolei, and Li Beibin — company and people attached to the concept in the source.
• AI For Science, AI Materials Discovery, and Causal AI — adjacent scientific AI themes in the wiki.
• Research Taste and Problem Definition In Research — judgment about which scientific questions matter.
• Recursive Self-Improvement, Deep Research, and AI Verification — technical loop needed for discovery rather than answer generation.
• Yinglingdian AI / 英灵殿, All-Modal Molecular World Model, AI Drug Discovery Platform, and Scientific Discovery Automation — biological discovery branch added by the Shizilukou Crossing source.
• Tian Yuandong / 田渊栋, AI Research Feedback Compression, Mechanistic Interpretability, and AI For Science — LateTalk episode 178’s route from AI-for-AI toward broader discovery systems.