OpenAI: GPT-4 (older v0314)

Processes multi-turn conversations using transformer-based attention mechanisms with an 8,192 token context window, enabling coherent dialogue across multiple exchanges. The model maintains conversation history within the context window and applies causal masking to prevent attending to future tokens, allowing it to generate contextually appropriate responses based on prior turns. Architecture uses decoder-only transformer with rotary positional embeddings to handle sequential dependencies in dialogue.

Generates syntactically valid code across 50+ programming languages by leveraging transformer patterns trained on public code repositories and documentation. The model applies language-specific formatting rules learned during training and can generate complete functions, classes, or multi-file solutions based on natural language descriptions. Uses in-context learning to adapt to coding style and patterns provided in the prompt.

Accepts a system prompt parameter that establishes role, tone, and behavioral constraints for the model, enabling fine-grained control over response style without retraining. The system prompt is prepended to the conversation context and influences token generation probabilities across all subsequent user messages through learned associations between instructions and output patterns. This is implemented via the OpenAI Chat Completions API's system role parameter.

Performs chain-of-thought reasoning by generating intermediate reasoning steps before producing final answers, leveraging transformer attention patterns to maintain logical consistency across multiple reasoning hops. The model can decompose complex problems into sub-problems, track variable states across steps, and validate intermediate conclusions. This emerges from training on mathematical proofs, scientific papers, and structured reasoning examples.

Synthesizes information from multiple sources or long documents by identifying key concepts, extracting relevant details, and generating coherent summaries that preserve essential information. The model uses attention mechanisms to weight important tokens and generate abstractive summaries (not just extractive) that reorganize information for clarity. Trained on news articles, academic papers, and web content with human-written summaries.

Generates original creative content (stories, poetry, marketing copy, dialogue) by sampling from learned distributions of language patterns associated with different genres and styles. The model uses temperature and top-p sampling parameters to control output diversity, and can adapt to specified tones, genres, and narrative constraints provided in the prompt. Trained on diverse creative writing from the internet and published works.

Translates text between 100+ languages and understands semantic meaning across linguistic boundaries by leveraging multilingual token embeddings and cross-lingual attention patterns learned during training. The model can preserve tone, formality, and cultural context in translations, and can answer questions about text in languages different from the query language. Supports both direct translation and back-translation for quality validation.

Answers factual and conceptual questions by retrieving relevant knowledge from training data and generating coherent responses. The model explicitly acknowledges its knowledge cutoff (September 2021) and can indicate uncertainty when asked about events or developments after that date. Uses attention mechanisms to identify relevant context within the question and generate targeted answers rather than generic summaries.

Extracts structured information from unstructured text by mapping natural language content to predefined schemas or JSON structures. The model uses instruction-following to generate valid JSON or structured output that conforms to specified field definitions and data types. Leverages learned associations between natural language patterns and structured representations from training data.