Intrinsic Capabilities
Foundational competencies encoded during Pre-training, independent of downstream prompting, fine-tuning, or system-level Orchestration Layer. Determines baseline reasoning fidelity, language comprehension depth, pattern extraction efficiency, and emergent behavior thresholds.
Core Dimensions
- Scaling Alignment: Capacity correlates predictably with parameter count, compute budget, and training corpus quality (scaling-laws)
- Emergent Reasoning: Multi-step logic, code synthesis, and cross-modal alignment manifest non-linearly past critical training thresholds
- Constraint Boundaries: Inherent limits including context window saturation, hallucination propensity, and alignment friction points
- Task-Generalization: Baseline ability to transfer learned representations to unseen domains without explicit instruction tuning
Intrinsic vs. Extrinsic Performance
- Intrinsic: Fixed by model weights, architectural topology (e.g., attention mechanisms, mixture-of-experts routing), and data curation
- Extrinsic: Dynamic modifiers including prompt structure, tool-use delegation, RAG pipelines, and runtime control flow
- Performance delta increasingly decouples from raw parameter scaling; marginal gains now require systematic harness-engineering rather than architectural overhauls
Recent Insights
- Performance variance in modern LLMs is increasingly driven by orchestration code rather than base architectural modifications
- System-level harnessing (prompt routing, tool chaining, feedback loops) often yields higher ROI than marginal weight updates
- Engineering focus has shifted from raw parameter scaling to deterministic control flow around stochastic generators
- See detailed breakdown in Orchestration Over Architecture: Harness Engineering for Optimal LLM Performance
Related Concepts
- Emergent Abilities
- scaling-laws
- prompt-engineering
- System Architecture
- Tool Use & Function Calling
- Alignment & Safety