OpenAI: GPT-4

GPT-4 processes both text and image inputs through a unified transformer architecture, using vision encoders to embed images into the same token space as text, enabling joint reasoning across modalities. The model performs end-to-end training on interleaved image-text sequences, allowing it to answer questions about images, extract text from screenshots, analyze diagrams, and reason about visual content without separate vision-language alignment layers.

GPT-4 implements implicit chain-of-thought reasoning through its training on reasoning-heavy datasets, allowing it to generate intermediate reasoning steps before producing final answers. When prompted to 'think step by step', the model allocates more compute tokens to exploring solution paths, backtracking when needed, and validating intermediate conclusions before committing to outputs. This is achieved through instruction-tuning on datasets where reasoning traces precede answers.

GPT-4 classifies text into sentiment categories (positive, negative, neutral) or custom categories by learning classification patterns through instruction-tuning on labeled examples. The model uses transformer attention to identify sentiment-bearing words, context, and implicit meaning, enabling nuanced classification that handles sarcasm, mixed sentiment, and domain-specific language. Classification can be zero-shot (no examples) or few-shot (with examples), with few-shot improving accuracy.

GPT-4 extracts structured information (entities, relationships, attributes) from unstructured text by learning extraction patterns through instruction-tuning on examples where text is paired with structured outputs (JSON, tables). The model uses transformer attention to identify relevant spans of text, map them to schema fields, and format outputs according to specified schemas. Extraction can be guided by providing a target schema or examples of desired output format.

GPT-4 improves task performance through few-shot learning by conditioning on examples of input-output pairs provided in the prompt. The model uses transformer attention to recognize patterns in the examples and apply them to new inputs, enabling task adaptation without fine-tuning. Few-shot learning is particularly effective for custom tasks, domain-specific language, and non-standard output formats. Performance typically improves with 2-5 examples; diminishing returns occur beyond 10 examples.

GPT-4 generates code across 50+ programming languages by learning patterns from public code repositories and documentation during pretraining. It uses transformer attention to track variable scope, function signatures, and import dependencies across files, enabling it to generate syntactically correct and semantically coherent code snippets. The model can complete partial functions, generate boilerplate, refactor existing code, and explain code logic through instruction-tuning on code-explanation pairs.

GPT-4 supports structured function calling by accepting a JSON schema of available functions and returning structured JSON objects specifying which function to call and with what arguments. The model learns to map natural language requests to function calls through instruction-tuning on examples where user intents are paired with function invocations. This enables deterministic tool orchestration without parsing natural language outputs, as the model directly outputs structured data conforming to the provided schema.

GPT-4 answers questions across diverse domains (science, history, law, medicine, programming) by leveraging knowledge learned during pretraining on internet text, books, and academic papers up to April 2023. The model uses transformer attention to retrieve relevant knowledge from its parameters and synthesize coherent answers, combining multiple facts and reasoning steps. Knowledge is implicit in weights rather than retrieved from external databases, enabling fast inference without retrieval latency.

GPT-4 follows complex, multi-step instructions by decomposing tasks into subtasks and executing them sequentially. Through instruction-tuning on datasets where complex instructions are paired with correct outputs, the model learns to parse task specifications, identify dependencies, and generate outputs that satisfy all constraints. This enables it to handle nuanced requests like 'write a poem in the style of Shakespeare about machine learning, exactly 14 lines, with AABB rhyme scheme'.

GPT-4 maintains conversational context across multiple turns by processing the entire conversation history (user messages and prior assistant responses) as input to each new generation. The model uses transformer attention to track references, pronouns, and implicit context from earlier turns, enabling coherent multi-turn conversations where it can refer back to previous statements, correct itself, or build on prior reasoning. Context is managed by the client; the model itself is stateless.

GPT-4 generates creative content (stories, poems, marketing copy, dialogue) by learning patterns from diverse text sources during pretraining and refining them through instruction-tuning on writing tasks. The model can adopt specific writing styles, tones, and genres by conditioning on style descriptors in the prompt (e.g., 'write in the style of Hemingway'). Generation is controlled through temperature and top-p sampling, enabling trade-offs between creativity (high temperature) and consistency (low temperature).

GPT-4 translates text between 100+ languages and generates content in non-English languages by learning multilingual patterns during pretraining on internet text in diverse languages. The model uses shared transformer parameters across languages, enabling transfer learning where knowledge from high-resource languages (English, Mandarin) improves performance on low-resource languages. Translation quality is improved through instruction-tuning on translation pairs and multilingual instruction-following.

GPT-4 summarizes long documents, articles, or conversations by extracting key information and condensing it into shorter text. The model learns summarization patterns through instruction-tuning on document-summary pairs, enabling it to identify salient information, maintain factual accuracy, and adapt summary length based on prompts (e.g., 'summarize in 2 sentences' or 'provide a detailed summary'). Summarization can be extractive (copying key sentences) or abstractive (paraphrasing and synthesizing).