PUMA Vault

AI systems can generate new scientific knowledge, but only within human-defined research frameworks

Se lee en 3 min

AI systems can generate new scientific knowledge, but only within human-defined research frameworks

Core Insight

Atomic Claim

Current AI systems (LLMs, multi-agent pipelines, robot scientists) can contribute to scientific discovery across multiple domains — mathematics, physics, materials science, structural biology — but they consistently operate within frameworks, datasets, and evaluation criteria defined by human researchers. The key question is not whether AI can produce new knowledge, but to what degree it can do so autonomously.

Three Levels of AI Scientific Capability (LeCun, 2026)

Yann LeCun structures AI’s role in science into three levels:

  1. AI as scientific assistant: Literature review, code writing, experiment design assistance, figure generation. Agilizes routine tasks without changing the epistemological structure of science.
  2. AI as scientific model: Domain-specific models (AlphaFold, GraphCast, plasma RL controllers) that capture complex system regularities better than traditional mechanistic models, providing new predictive capabilities.
  3. AI as scientific agent: Systems that formulate hypotheses, plan and execute experiments, analyze results, and propose conclusions — approaching limited scientific agency.

Key Evidence of AI-Generated Knowledge (2023–2026)

Mathematics

  • Gemini 2.5 Deep Think: Gold medal performance at IMO 2025 — 5/6 problems solved with formal proofs. (Huang & Yang, 2025, arXiv:2507.15855)
  • GPT-5 + Malliavin-Stein: Extended fourth-moment theorems with explicit convergence rates — proof generation with human verification. (Arras et al., 2025, arXiv:2509.03065)
  • “First Proof” project: AI-generated complete demonstrations of novel group theory and combinatorics problems with lemmas not in the expert reviewers’ standard repertoire. (Aletheia + academics, 2025-2026)

Physics

  • Single-minus gluon amplitudes: GPT-5.2 Pro proposed a closed-form expression for n-point gluon tree amplitudes, correcting a belief that they were zero. Result formally verified against Berends-Giele recursion. (Arkani-Hamed et al., 2026, arXiv:2602.12176)

Materials Science

  • GNoME: Google DeepMind’s graph network discovered 2.2M stable crystal structures, expanding known stable materials by ~10×. Hundreds verified experimentally. (Jain et al., 2023, Nature)

Structural Biology

  • AlphaFold2: Structure prediction at near-experimental accuracy for proteins without homologous structures, enabling inference of folding pathways. (Jumper et al., 2021, Nature)

Automated Research Pipelines

  • AI Scientist v1/v2: End-to-end ML research pipeline — idea generation → experiment → paper → review. One paper accepted at ICLR 2025 workshop. (Lu et al., 2024/2025/2026)
  • Robot Scientist Adam: Discovered functions of orphan yeast genes autonomously via hypothesis generation and robotic experiment execution. (King et al., 2009)
  • Robot Scientist Eve: Identified anti-malarial and anti-tumor drug candidates via active learning. (Williams et al., 2015)

The Skeptical Counter-Argument (Felin & Holweg, 2024)

Rival Hypothesis

“Theory Is All You Need” argues AI is fundamentally retrospective:

  • Human cognition is prospective and causal — generates experiments driven by theory, not just data.
  • AI is retrospective and correlational — pattern-matches on existing data.
  • True scientific novelty requires formulating causal models not determined by prior data — which current AI cannot do spontaneously.

Assessment: This is partially correct but increasingly challenged. The single-minus gluon result and First Proof cases suggest some LLMs can generate results that correct prior human beliefs — going beyond interpolation.

PUMA Connection

PUMA’s own design reflects this three-level model:

  • Level 1 (Assistant): Claude, NotebookLM for literature synthesis and writing
  • Level 2 (Model): LLM agents as predictive models for issue priority and effort
  • Level 3 (Agent): Stage 5 Smart PMO agents approaching limited scientific agency in PM

The question of whether PUMA’s agents can discover new PM patterns (not just classify) is the direction of future work.

MOCs

Vista Gráfica

Retroenlaces