🗂️ MOC — AI and New Scientific Knowledge Generation
This MOC covers PUMA’s PEC2 research area: the question of whether and how AI systems can generate new scientific knowledge. Connected to PUMA’s broader positioning of LLM-based PM agents within the Agentic Science trajectory.
Central Question
Can AI systems generate new scientific knowledge autonomously, and what does this mean for PUMA’s claim that LLM agents can automate PM tasks?
Permanent Notes (Zettelkasten — Synthesized Insights)
- PN-AI-Scientific-Knowledge-Generation — Core synthesis: AI generates knowledge within human frameworks
- PN-Agentic-Science-Paradigm — Agentic Science as new research paradigm
- PN-PUMA-within-AgenticScience-Trajectory — PUMA’s Level 2 positioning within the trajectory
- PN-KeyConcepts-Agents-Reproducibility-RedTeam — PUMA’s core design principles
Literature Notes (Evidence Base)
AI Research Automation
- LN-Lu-2024-AIScientist — AI Scientist pipeline ⭐⭐⭐⭐
- LN-Zhang-2025-AgenticScienceSurvey — Agentic Science survey ⭐⭐⭐⭐
Key Domain Results
- LN-Jumper-2021-AlphaFold — AlphaFold2 protein structure ⭐⭐⭐⭐
- LN-Degrave-2022-PlasmaControl — Plasma control RL ⭐⭐⭐
- LN-Jain-2023-GNoME-Materials — GNoME materials discovery ⭐⭐⭐
- LN-Lam-2023-GraphCast — GraphCast weather ⭐⭐⭐
- LN-ArkaniHamed-2026-GluonAmplitudes — GPT-5.2 physics result ⭐⭐⭐
Critical Perspectives
- LN-Felin-2024-TheoryIsAllYouNeed — Skeptical counterargument ⭐⭐⭐
- LN-Klinger-2025-AIScience-SocialProblem — Social/governance issues ⭐⭐⭐
LeCun’s Three-Level Framework (2026)
| Level | Description | PUMA Example |
|---|---|---|
| Level 1: Assistant | AI helps with literature, code, writing | Claude/NotebookLM for PUMA research |
| Level 2: Model | AI captures domain regularities (AlphaFold, GraphCast) | PUMA triage/estimation agents |
| Level 3: Agent | AI manages full research cycles autonomously | PUMA Stage 5 Smart PMO |
Evidence Summary: AI Knowledge Generation
| Domain | System | Knowledge Type | Novelty Level |
|---|---|---|---|
| Biology | AlphaFold2 | Protein structures | High — paradigm shift |
| Materials | GNoME | Crystal stability | High — 10× expansion |
| Weather | GraphCast | Predictive model | Medium — outperforms mechanistic |
| Plasma | DeepRL | Control strategies | High — novel configurations |
| Mathematics | Gemini 2.5 | IMO proofs | High — gold medal level |
| Physics | GPT-5.2 | Gluon amplitude formula | Very High — corrects prior belief |
| ML Research | AI Scientist | Research papers | Medium — workshop level |
PUMA Positioning
PUMA fits at LeCun’s Level 2 (AI as scientific model):
- Input: Jira issue text → Output: priority classification (Level 2 model)
- Input: User story text → Output: story point estimate (Level 2 model)
- Future (Stage 5): Full PM cycle orchestration (approaching Level 3)
The PEC2 research provides the theoretical context for why Level 2 AI models are scientifically valid and practically valuable.
Bibliography
- BIB-Supplement-PEC2-AIKnowledge — All PEC2 references in APA7