Sources

Cook's blog post makes the case concisely: "Sometimes it takes a low-tech tool to find problems with a high-tech tool." Regex runs in O(n) time using a finite-state automaton. An LLM runs through billions of parameters to approximate the same computation, slower, more expensive, and less reliably. The gap between deterministic string processing and probabilistic generation is not marginal; it is orders of magnitude in every dimension that matters for production systems.

Huyen frames the core tension of hybrid systems: "It's easy to build something cool with LLMs; very hard to build production-ready." The production-readiness comes from the deterministic wrapper, not from the LLM itself. Input validation, retries, fallback logic, output schema enforcement, and logging are all conventional engineering that makes the probabilistic step reliable. The KreditFlow system at Maryville University is a concrete example: 330 Lambda functions, four Claude models, twelve prompt templates. If you stripped out the LLM calls, you would still have a functioning data pipeline. The LLMs add the intelligence; the code provides the skeleton.

OpenAI's own cookbook demonstrates the prompt-to-train-to-distill pipeline. GPT-4o generates labeled completions on real-world inputs. Those completions become supervised training data for GPT-4o-mini. The distilled model achieved 79.3% accuracy compared to 64.7% for the non-distilled baseline: a 22% relative improvement. The key principle: teacher output quality is the ceiling for student performance. Simon Willison publicly questioned whether anyone has a fine-tuning success story that beats "prompting existing hosted models or waiting for next-gen models." For most tasks, the answer is no. The exception is when you need the cost structure, not just the accuracy. At 63 million requests per year, the math does not care about model quality parity.

Predibase's guide to distilling smaller models emphasizes diverse, non-repetitive, balanced datasets. The more scenarios covered, the better generalization. Quality of the teacher's output serves as an upper limit for the student's performance, which means maximizing teacher quality before distillation is a non-negotiable prerequisite. The guide also addresses a common failure mode: teams that distill from a narrow subset of production data end up with a model that performs well on the common cases but fails on the edge cases that matter most.

Aggarwal et al. benchmark embedding-based classification against LLM prompting across multiple datasets. The results are unambiguous for classification tasks: embeddings achieve 44.1% accuracy versus 29.5% for prompting (+49.5% relative), with 81x faster text classification latency (15ms vs 1,220ms) and 10x lower annual cost at 63 million requests ($9,179 vs $92,325). Crucially, embeddings also produce well-calibrated probability distributions suitable for confidence thresholds, while LLM prompting returns concentrated scores that are "mostly uninformative" for real-world decisions. This means embeddings not only classify better but also tell you how confident they are in the classification.

Yan identifies evaluation as "a major differentiator between folks rushing out hot garbage and those seriously building products." His seven-pattern framework (evals, RAG, fine-tuning, caching, guardrails, defensive UX, user feedback) maps to the strategies in this article: evals sit inside every strategy, RAG is a variant of the hybrid pattern, fine-tuning is the distillation strategy. The vocabulary differs; the engineering principles converge. His emphasis on hybrid retrieval (traditional BM25 + semantic search) also supports the article's point that embedding-based approaches and deterministic code are not alternatives to LLMs but complements.

Meta's engineering blog articulates the lifecycle trajectory: start with prompting if baseline accuracy exceeds 85%, add RAG for knowledge-intensive problems, and fine-tune "only when the model's intrinsic behavior must be altered robustly and persistently." They note that hosting a fine-tuned model can cost more than the base model due to infrastructure overhead, making distillation to smaller models often more economical. Their recommendation of hybrid solutions combining fine-tuning and RAG as "often yield the best results" aligns with the article's lifecycle diagram showing systems migrating through strategies as they mature.

Husain reports that 60-80% of development time in real LLM projects is spent on error analysis and evaluation, not building features. His recommended workflow is striking in its simplicity: thirty minutes reviewing 20-50 outputs after each change, with a single domain expert as "benevolent dictator" evaluating pass/fail. This holds regardless of which strategy you choose. The evaluation discipline described in the companion Week 3 articles (Prompts Are Code, promptfoo walkthrough) implements this loop with tooling; Husain's point is that the loop itself matters more than the tooling.

Featherless AI's 2026 pricing survey documents the shift: competitive models like DeepSeek V3.2 charge $0.28 per million input tokens, compared to GPT-4's $30 per million in 2024. That is a hundred-fold reduction in two years. The report notes that "80% of applications work fine with $0.40-$2.50/MTok range; only coding agents, reasoning, multi-step planning benefit from $10+/MTok models." This price collapse recalibrates every decision in the framework. Tasks that were "too expensive to prompt" are now cheap. Tasks that were "worth fine-tuning" may no longer justify the upfront investment. The framework survives because it asks "what does the task require?" not "what does the model cost?" But the answers change. Revisit quarterly.

Shankar's data-flywheel pattern (evaluate, monitor, continually improve) describes the lifecycle within each strategy. Her SPADE framework for data quality assertions applies whether you are evaluating prompt outputs, embedding model accuracy, or fine-tuned model behavior. The PROMPTEVALS dataset (2,087 prompts, 12,623 assertions) is a useful benchmark for evaluating models' ability to generate guardrails. The tooling is strategy-agnostic; the practice is universal.