SYSTEM Cited by 2 sources

GPT-4

Definition

GPT-4 is OpenAI's fourth-generation general-purpose large language model, originally released 2023-03-14 as the first widely-available frontier multi-modal LLM. Successor to GPT-3 / GPT-3.5; predecessor to GPT-4o and the o-series reasoning models.

Wiki anchor

The wiki's canonical anchor for GPT-4 is its role as the teacher-model / formulation-phase LLM in production LLM pipelines, documented in the 2025-02-04 Yelp post (sources/2025-02-04-yelp-search-query-understanding-with-llms). Yelp uses GPT-4 for two purposes: (1) prompt formulation — the strongest available LLM for iterating prompts and defining output schemas before distilling down to a smaller model; (2) golden-dataset generation — GPT-4 labels a representative sample of production queries, which is then curated by humans and used to fine-tune the smaller GPT-4o- mini serving student.

Explicit disclosure verbatim: "This typically entails quick prototyping with the most powerful LLM available to us, such as the latest stable GPT-4 model, and creating many iterations of the prompt."

Tradeoffs

• Highest per-token cost and latency among the OpenAI API family as of 2025 (superseded on perf/cost axes by GPT-4o- mini and o-series).
• The formulation-phase use-case explicitly prefers quality over cost; production cost is controlled by distilling into a smaller model at scale-up time.

Seen in

• sources/2025-02-04-yelp-search-query-understanding-with-llms — formulation-phase LLM + teacher for GPT-4o-mini fine-tune.
• sources/2024-09-17-zalando-content-creation-copilot-ai-assisted-product-onboarding — GPT-4 Turbo as launch backend for Zalando's Content Creation Copilot. POC evaluation picked GPT-4 Turbo over alternatives for "the right balance between accuracy and information coverage" on catalog-attribute extraction before GPT-4o shipped. Later migrated to systems/gpt-4o for better accuracy + lower latency + lower cost — an instance of the Turbo → 4o swap that the aggregator pattern made contract-free.

Related

• systems/gpt-4o — the multi-modal flagship successor
• systems/gpt-4o-mini — cheaper offline-serving student
• systems/o1-preview / systems/o1-mini — successor reasoning models Yelp reserves for logical-reasoning tasks
• systems/zalando-content-creation-copilot — GPT-4 Turbo launch instance
• concepts/llm-cascade — cost-routing pattern
• concepts/multi-modal-attribute-extraction — the concept the Zalando use case instantiates
• patterns/offline-teacher-online-student-distillation — Yelp's canonical use pattern
• patterns/model-agnostic-suggestion-aggregator — the pattern that made Zalando's GPT-4 Turbo → GPT-4o swap contract-free