Research progress

Often, when researchers disagree about what topics or approaches to prioritize, they are also disagreeing on what kinds of progress are more important. This stems from implicit preferences in research goals and aesthetics. One researcher might dismiss a theoretical model as impractical, while another sees a prototype as ephemeral.

These tensions reflect deeper contrasts in how progress is interpreted and pursued. ‘Progress’ can’t be decomposed cleanly, but it can still be probed through contrastive axes.

Contrasts of purpose and scope

These consider questions like: What kind of progress is being aimed at, and on what timescale? What is the intended breadth and function?

• Instrumental vs. Terminal: Treat the work as a ‘ride to the next stop’ (e.g., LLM-as-tool) or as a ‘vehicle toward the ultimate’ (e.g., LLM-as-pathway to AGI)?
• Immediate vs. Speculative: Does it address near-term needs (e.g., pandemic response tools) or invest in distant possibilities (e.g., artificial consciousness or fusion energy)?
• Exploratory vs. Exploitative: Does it venture into new paradigms (e.g., analog neural computing, non-symbolic reasoning) or refine the dominant one (e.g., transformer variants and scaling curves)?
• Demonstrative vs. Infrastructural: Is it meant to show what’s possible (e.g., AlphaFold proof-of-concept) or to lay stable ground others can build upon (e.g., PyTorch, COCO dataset)?
• Problem-solving vs. Synthesizing: Does it address targeted technical challenges (e.g., graph isomorphism algorithms) or unify diverse ideas into overarching theories (e.g., unified scaling laws or integrated cognitive architectures)?
• Incremental vs. Foundational: Does the work improve details within a paradigm (e.g., refining activation functions, training tricks), or does it reshape the field (e.g., making backpropagation viable, parallelizing attention)?
• Generalist vs. Specialist: Does it aim to span domains (e.g., general-purpose models like GPT-4 or CLIP) or achieve peak performance in a narrow domain (e.g., AlphaZero for Go, or Stockfish for chess)?
• Humanist vs. Machine-oriented: Is the goal to augment human reasoning (e.g., computer algebra systems, interactive theorem provers), or to automate cognition autonomously (e.g., autoformalization of proofs, or autonomous research agents)?

Contrasts of epistemic orientation

These consider questions like: How is knowledge generated, and what kind of evidence is considered valid? What is the work trying to know or demonstrate, and how does it relate to truth or utility?

• Empirical vs. Theoretical: Is knowledge advanced through observed behavior (e.g., emergent scaling trends in LLMs) or from formal reasoning (e.g., computational complexity classes, information-theoretic bounds)?
• Heuristic vs. Formal: Does it rely on practical intuition (e.g., designing deep architectures by intuition and trial) or on strict formalism (e.g., algebraic topology in data analysis)?
• Descriptive vs. Axiomatic: Does it model what is observed (e.g., descriptive linguistics or behavioral economics) or deduce consequences from assumptions (e.g., formal semantics, first-principles physics)?
• Results vs. Mechanism: Does success depend on performance (e.g., beating benchmarks with black-box models) or on explanation (e.g., understanding internal representations or algorithmic convergence)?
• Technological vs. Scientific: Is the goal to build powerful artifacts (e.g., large-scale language models, quantum processors), or to uncover explanatory principles (e.g., theories of language acquisition, quantum decoherence models)?
• Model power vs. Model fit: Do we value models for capability and coverage (e.g., overparameterized LLMs) or for tight, interpretable alignment with reality (e.g., classical mechanics)?
• Permissive vs. Rigorous: Does it allow for looser inference or anecdotal insight (e.g., case studies, ethnography) or demand strict reproducibility and control (e.g., pre-registered experiments, mechanistic modeling)?
• Abductive vs. Statistical: Does it infer plausible causes from limited data (e.g., historical linguistics, astrobiology), or validate conclusions via statistical rigor (e.g., randomized control trials or large-scale behavioral studies)?
• Agnostic vs. Committed: Is the aim to remain neutral and exploratory (e.g., anomaly detection or open-world hypothesis mining), or to decisively test a claim (e.g., falsifying a neuroscience model or evaluating a drug mechanism)?

Closing thoughts

These axes are contrived and may stem from misperceptions of how progress in discovery or innovation is achieved. It’s unclear whether some styles are more promising than others. History may favor pluralism since ‘what helps’ is difficult to anticipate. In effect, foundational discoveries have emerged from speculative or loosely-grounded work, and infrastructural tools can drive theoretical advances.

When researchers disagree, it’s not limited to facts, but also ‘styles of inquiry’, favored risks, and what counts as a contribution. Intellectual humility and mediation are thus critical. Style is also personal: prolonged mismatch between individual preference and project demands can lead to dissatisfaction—even when the work is “important.” Some like loose wide-open problems and others prefer tightly-framed, rigorous domains.

It’s important that we try to support ‘diverse epistemic personalities’ and to resist dogmatism. A healthy ecosystem protects the space for diversification and disagreement until clarity emerges.