← Notes·The Physics of AI Engineering
Dojo Note · Beta

Key Concepts

FikAi ·

Core Thesis

Production AI systems fail not because LLMs are unreliable, but because system architectures ignore the physical constraints of how transformers actually work. Treat LLM engineering like aerospace engineering: respect the physics, or crash.

The Three Fundamental Laws

1. Finite Attention — Information recall follows a U-curve. First 10% of context: ~90% recall. Middle 40-60%: ~50% recall. Final 10%: ~85% recall. Critical data buried in long contexts is effectively invisible.

2. Stochastic Accumulation — Errors compound exponentially across steps. A 2% per-step error rate yields only 36% success over 50 steps. Formula: P_success = (1-p)^N. Multi-step chains require checkpointing and retry mechanisms to be reliable.

3. Entropic Expansion — Context grows linearly with time; capacity is fixed. Without compression, every long-running workflow eventually overflows. Active summarization and priority-based eviction are mandatory, not optional.

Key Terms

TermDefinition
TokenAtomic text unit (sub-word, not word). JSON is ~2x less efficient than plain text.
State AmnesiaLoss of accumulated knowledge when a process terminates. LLMs are stateless—persistence must be explicit.
Poisoned WellContext contaminated with error traces biases the model toward failure patterns.
Cognitive OffloadingDelegating deterministic tasks to code so the LLM focuses only on high-level reasoning.
Priority StackContext architecture placing critical data at attention boundaries (first/last 10%), not the middle.

What Managers Need to Know

  • "Better prompts" won't fix production failures. Architecture determines reliability.
  • Checkpoint every turn. Pod crashes should be invisible to workflow completion.
  • Measure token utilization and hallucination rates. These are leading indicators of system health.
  • Cognitive offloading reduces hallucinations by 60-80%. Let code handle parameters; let LLMs handle decisions.
  • Long contexts aren't free. Attention is zero-sum—more tokens means less precision per token.

Bottom line: LLMs are probabilistic text generators. Production reliability comes from deterministic scaffolding that respects their physics.


Key Concepts — FikAi notebook for The Physics of AI Engineering.