('agi', 'artificial general intelligence', "a hypothetical ai system that can perform any intellectual task that a human can. no agi currently exists; today's systems are narrow ai.")
('agent', 'ai agent', 'an autonomous system that perceives its environment, makes decisions, and takes actions to achieve a goal — often using llms as a reasoning engine.')
('alignment', 'ai alignment', 'the challenge of ensuring ai systems pursue goals that are beneficial to humans. misaligned ai could optimise for the wrong objective.')
('anthropic', 'anthropic', 'ai safety company behind the claude family of models. founded by former openai researchers in 2021.')
('benchmark', 'ai benchmark', 'a standardised test used to measure and compare the performance of ai models (e.g. mmlu, humaneval, gpqa).')
('chain-of-thought', 'chain-of-thought prompting', 'a prompting technique that encourages an llm to produce intermediate reasoning steps before giving a final answer, improving accuracy on complex tasks.')
('context window', 'context window', 'the maximum amount of text (measured in tokens) an llm can process at once. larger context windows let models consider longer documents.')
('diffusion model', 'diffusion model', 'a class of generative ai models (e.g. stable diffusion, dall-e 3) that generate images by iteratively denoising random noise.')
('embedding', 'embedding', 'a dense numerical vector representation of text, images, or other data that captures semantic meaning. used in search, rag, and classification tasks.')
('fine-tuning', 'fine-tuning', 'adapting a pre-trained model on a smaller, task-specific dataset. more efficient than training from scratch; used to specialise general models.')
('foundation model', 'foundation model', 'a large model trained on broad data that can be adapted to many tasks. gpt-4, claude, and gemini are foundation models.')
('gpt', 'generative pre-trained transformer', 'a family of llms developed by openai (gpt-3.5, gpt-4, gpt-4o) that power chatgpt and many third-party applications.')
('hallucination', 'ai hallucination', 'when an llm generates plausible-sounding but factually incorrect information. a key challenge for production ai systems.')
('inference', 'model inference', 'running a trained ai model on new inputs to produce outputs. inference cost and speed are key concerns in deploying llms at scale.')
('llm', 'large language model', 'a deep learning model trained on massive text datasets that can generate, summarise, translate, and reason about text. examples: gpt-4, claude, gemini.')
('moe', 'mixture of experts', 'an architecture where only a subset of model parameters (experts) are activated for each input, allowing very large models with manageable compute costs.')
('multimodal', 'multimodal ai', 'ai systems that process multiple types of input — text, images, audio, video — in a unified model (e.g. gpt-4o, gemini 1.5 pro).')
('rlhf', 'reinforcement learning from human feedback', 'a training technique where human raters score model outputs, and these scores guide further training to align the model with human preferences.')
('rag', 'retrieval-augmented generation', 'a technique that enhances llm outputs by first retrieving relevant documents from an external knowledge base and including them in the prompt context.')
('sft', 'supervised fine-tuning', 'fine-tuning a model on human-written examples of correct outputs. often the first step in aligning a base model (followed by rlhf).')
('token', 'token', 'the basic unit of text that llms process — roughly 3/4 of a word in english. model cost and context window size are measured in tokens.')
('transformer', 'transformer architecture', 'the neural network architecture (introduced by google in 2017) that underlies virtually all modern llms. based on the attention mechanism.')
('vector database', 'vector database', 'a database optimised for storing and querying embeddings. used in rag pipelines to find semantically similar documents. examples: pinecone, weaviate, pgvector.')
('vlm', 'vision-language model', 'a multimodal model that jointly understands images and text. used for image captioning, visual qa, and document understanding.')
('agentic ai', 'agentic ai', 'ai systems that autonomously plan and execute multi-step tasks using tools, apis, and memory — going beyond single-turn question answering.')
('attention', 'attention mechanism', 'a neural network mechanism that lets a model focus on relevant parts of its input when generating each output token. the core innovation behind the transformer architecture.')
('cuda', 'cuda', "nvidia's parallel computing platform used to accelerate deep learning training. gpu availability (measured in cuda cores) is a key bottleneck for ai labs.")
('dpo', 'direct preference optimization', 'a simpler alternative to rlhf that fine-tunes llms on pairs of responses (preferred vs. rejected) without a separate reward model.')
('gan', 'generative adversarial network', 'a generative model architecture where a generator and discriminator compete. preceded diffusion models for image synthesis; still used in video generation.')
('guardrails', 'ai guardrails', 'techniques and systems that constrain llm outputs — preventing harmful content, hallucinations, or off-topic responses — in production deployments.')
('mmlu', 'massive multitask language understanding', 'a widely-used benchmark testing llm knowledge across 57 subjects (stem, humanities, law, medicine). score above 90% is considered near-expert level.')
('lora', 'low-rank adaptation', 'a parameter-efficient fine-tuning technique that updates only a small set of weights, dramatically reducing gpu memory requirements for adapting large models.')
('prompt engineering', 'prompt engineering', 'the practice of designing inputs (prompts) to elicit desired outputs from llms. techniques include few-shot examples, chain-of-thought, and system instructions.')
('quantization', 'model quantization', 'reducing the precision of model weights (e.g. from 32-bit to 4-bit floats) to shrink memory footprint and increase inference speed, with minimal quality loss.')
('safety', 'ai safety', 'the field studying how to ensure ai systems behave as intended and remain beneficial as they become more capable. covers alignment, robustness, and interpretability.')
('synthetic data', 'synthetic data', 'artificially generated training data, often produced by llms themselves. used to augment scarce real data or to create preference pairs for rlhf.')
('tts', 'text-to-speech', 'ai models that convert text into natural-sounding speech. recent models (elevenlabs, openai tts) are indistinguishable from human voices.')