PUMA Vault

Literature Note — Theory Is All You Need: AI, Human Cognition, and Causal Reasoning

Se lee en 3 min

LN — Theory Is All You Need (Felin & Holweg, 2024)

Full Reference: Felin, T., & Holweg, M. (2024). Theory is all you need: AI, human cognition, and causal reasoning. INFORMS Studies in the Service Economy, 4(1), 1–23. https://doi.org/10.1287/stsc.2024.0189

Pass 1 — Bird’s Eye

Type: Theoretical / Position paper Main Claim: Data-driven AI is fundamentally retrospective and cannot generate novel scientific theory. Human cognition is prospective and causal — it can formulate hypotheses that go beyond existing data. AI lacks this creative, theory-driven capability. Relevance to PUMA: ⭐⭐⭐ Medium — provides the key counterargument to Agentic Science claims; essential for PUMA Section 2 to present a balanced state of the art.

Pass 2 — Key Arguments

Core Thesis

  • Human cognition = prospective, theory-first, causal: generates experiments designed to test specific causal mechanisms
  • AI cognition = retrospective, data-first, correlational: finds patterns in existing data but cannot generate truly novel causal models
  • “Theory is what allows scientists to know what data to collect and how to interpret it”

Evidence Used

  • Historical science: major paradigm shifts (Copernicus, Darwin, Einstein) required theory-first thinking, not data-driven discovery
  • Current AI: AlphaFold predicts structures but didn’t generate the theory of protein folding
  • Large language models: generate plausible text from patterns but cannot distinguish causal from correlational relationships

Key Limitation of the Argument

  • Written in 2024, before: single-minus gluon result (2026), First Proof project (2025-2026), Gemini IMO gold medal (2025)
  • These subsequent results suggest some LLMs can go beyond interpolation in formal reasoning domains

Pass 3 — PUMA Re-implementation

PUMA’s position: Felin & Holweg are partially correct about AI limitations in open-ended scientific creativity. However, PUMA’s domain (PM task automation) does NOT require the kind of causal theory generation that Felin & Holweg argue AI cannot do.

PUMA’s claim is more modest:

  • LLM agents can classify issue priority better than naive baselines using historical patterns (correlational is sufficient)
  • LLM agents can estimate story points more accurately than team averages using similar historical issues

This is precisely the “retrospective pattern matching” that Felin & Holweg describe — and for PM automation, that is exactly what is needed.

MIT Critical Questions

  1. How can I use this? → Cite as the strongest counterargument in PUMA Section 2 (literature review); then explain why PUMA’s tasks don’t require the causal creativity Felin & Holweg require.
  2. Does it really support its claims? → The theory is sound but increasingly challenged by 2025-2026 empirical evidence.
  3. What if Felin & Holweg are right for PM? → Even if AI cannot generate PM theories, it can still improve classification/estimation accuracy — which is PUMA’s actual claim.

APA7 Citation

Felin, T., & Holweg, M. (2024). Theory is all you need: AI, human cognition, and causal reasoning. INFORMS Studies in the Service Economy, 4(1), 1–23. https://doi.org/10.1287/stsc.2024.0189

MOCs

Vista Gráfica

Retroenlaces