Stenobird

Episode

Radically Better Reasoning: Elicit's Andreas Stuhlmüller & Jungwon Byun on World Models for Research

Podcast
"The Cognitive Revolution" | AI Builders, Researchers, and Live Player Analysis
Published
Jun 17, 2026
Duration seconds
6370
Processing state
processed
Canonical source
https://www.cognitiverevolution.ai/radically-better-reasoning-elicit-s-andreas-stuhlmuller-jungwon-byun-on-world-models-for-research/
Audio
https://pdst.fm/e/mgln.ai/e/1113/pscrb.fm/rss/p/traffic.megaphone.fm/RINTP9702631647.mp3
JSON
/v1/public/podcasts/the-cognitive-revolution/episodes/radically-better-reasoning-elicit-s-andreas-stuhlm-ller-jungwon-byun-on-world-models-for-research
Markdown
/podcast/the-cognitive-revolution/radically-better-reasoning-elicit-s-andreas-stuhlm-ller-jungwon-byun-on-world-models-for-research.md

Actions

  • POST https://stenobird.com/v1/public/podcasts/the-cognitive-revolution/episodes/radically-better-reasoning-elicit-s-andreas-stuhlm-ller-jungwon-byun-on-world-models-for-research/transcription-requests
    Idempotently request low-priority transcript generation for this episode.
  • GET https://stenobird.com/podcast/the-cognitive-revolution/radically-better-reasoning-elicit-s-andreas-stuhlm-ller-jungwon-byun-on-world-models-for-research.md
    Read the agent-friendly Markdown representation of this episode resource.

Summary

Elicit founders discuss moving beyond simple LLM outputs toward structured 'world models' that enable verifiable scientific reasoning. They explore how domain-specific reasoning primitives and process supervision can prevent the opacity of modern frontier models.

Topics

  • World Models
  • Process Supervision
  • Scientific Machine Learning
  • Reasoning Primitives
  • Life Sciences AI
  • Domain Specific Languages
  • Verifiable AI
  • Causal Inference

Highlights

  • Main idea: Using a Domain Specific Language (DSL) to define reasoning primitives allows frontier models to execute structured, guaranteed workflows
  • Practical takeaway: Implementing process supervision—rewarding the quality of steps rather than just the final answer—is essential for high-stakes decision support
  • Failure mode: Relying on 'neuralese' or uninterpretable chain-of-thought can lead to unverifiable claims in critical fields like toxicology or drug discovery
  • Technical approach: Developing 'world models' as heterogeneous, evolving representations of knowledge to enable causal and counterfactual analysis
  • Operational insight: Automating software engineering via systems like 'The Line' can enable high-velocity development, delivering dozens of code changes weekly

Chapters

  1. 1:00 Reasoning Primitives and Microservices: How Elicit uses a DSL to create structured, verifiable reasoning workflows using discrete microservices.
  2. 9:00 The Importance of Process Supervision: Moving from simple output evaluation to inspecting the entire reasoning path to ensure reliability.
  3. 17:00 AI in Life Sciences: Applying evidence-based reasoning to drug target ranking and clinical research workflows.
  4. 25:00 Handling Conflicting Evidence: Strategies for evaluating claims when scientific literature presents contradictory results.
  5. 33:00 The Need for Intermediate Reasoning Layers: Discussing why complex tasks require specialized layers of reasoning beyond standard LLM prompting.
  6. 41:00 Verifiable Conclusions in Biology: The challenge of providing proofs and traceable evidence for low-level biological insights.
  7. 49:00 Building Explicit World Models: Moving away from massive context windows toward structured, interpretable representations of knowledge.