Since the rise of generative AI in late 2022, there’s been a growing familiarity with its basic concepts and how it facilitates natural language interaction with computers. Some of us can even casually discuss terms like “prompts” and “machine learning” over coffee. However, as AI progresses, so does its terminology. Are you familiar with the distinctions between large and small language models, or the significance of “GPT” in ChatGPT? We’re here to offer an advanced exploration of AI terminology to bring you up to date.
Reasoning/Planning
AI systems now exhibit problem-solving capabilities, utilizing learned patterns from historical data to make sense of information, akin to human reasoning. Advanced systems go beyond this, planning and executing sequences of actions to achieve objectives.
Training/Inference
The development and utilization of AI involve two phases: training and inference. Training involves educating an AI system with a dataset, enabling it to perform tasks or make predictions based on that data. Inference is the subsequent application of learned patterns and parameters to generate predictions or responses.
SLM/Small Language Model
Small Language Models (SLMs) are compact versions of Large Language Models (LLMs). Both employ machine learning techniques to recognize patterns and produce natural language responses. However, SLMs, like Phi-3, are trained on smaller datasets with fewer parameters, making them more compact and suitable for offline use on devices like laptops or phones.
Grounding
Generative AI systems excel in generating stories, poems, and answers but may struggle with discerning fact from fiction or outdated information. Grounding involves anchoring AI models with tangible examples and current data to enhance accuracy and produce contextually relevant output.
Retrieval Augmented Generation (RAG)
RAG integrates grounding sources into AI systems to enhance accuracy without requiring retraining. By incorporating additional knowledge, such as product catalogs for chatbots, RAG improves responses and efficiency.
Orchestration
The orchestration layer guides AI programs through tasks, ensuring coherent responses. It manages chat history to maintain context and can utilize RAG patterns to enrich responses with fresh information from the internet.
Memory
While AI models lack traditional memory, orchestrated instructions enable them to “remember” information by retaining context from previous interactions or incorporating grounding data. Developers explore methods to enhance AI’s ability to retain information temporarily or permanently.
Transformer Models and Diffusion Models
Transformer models, exemplified by ChatGPT, excel in understanding context and generating text rapidly. Diffusion models, primarily used for image creation, employ a gradual approach to distribute pixels, creating coherent images.
Frontier Models
Frontier models are cutting-edge AI systems with broad capabilities, pushing the boundaries of what AI can accomplish. These advanced models are subject to safety standards and knowledge-sharing initiatives to ensure responsible development.
GPU
Graphics Processing Units (GPUs) play a crucial role in AI by facilitating parallel processing, essential for training and inference tasks. Massive clusters of interconnected GPUs power today’s most advanced AI models, enabling their remarkable capabilities.
