Adapters are an advanced method for making pre-trained AI models adaptable to new tasks without complete retraining. These modules save time, money, and resources by efficiently repurposing existing models for different tasks in areas like natural language processing, computer vision, and robotics.
Artificial General Intelligence (AGI) refers to an AI system that possesses a wide range of cognitive abilities, much like humans, enabling them to learn, reason, adapt to new situations, and devise creative solutions across various tasks and domains, rather than being limited to specific tasks as narrow AI systems are.
A chat interface built on top of GPT-3.5. GPT-3.5 is a large language model developed by OpenAI that is trained on a massive amount of internet text data and fine-tuned to perform a wide range of natural language tasks. Example: GPT-3.5 has been fine-tuned for tasks such as language translation, text summarization, and question answering.
The cost of large language models primarily stems from their size and complexity, which demand significant computational power, storage, and resources for training and deployment. These factors can result in substantial expenses for building, maintaining, and using such models, sometimes amounting to several dollars per conversation or thousands of dollars per month.
Data Augmentation is a technique used to artificially increase the size and diversity of a training set by creating modified copies of the existing data. It involves making minor changes such as flipping, resizing, or adjusting the brightness of images, to enhance the dataset and prevent models from overfitting.
The process of adapting a pre-trained model to a specific task by training it on a smaller dataset. For example, an image classification model trained on all intersection pictures can be fine turned to detect when a car runs a red light. At Moveworks, we’ve been fine-tuning LLMs for enterprise support for years.
Foundation models are a broad category of AI models which include large language models and other types of models such as computer vision and reinforcement learning models. They are called "foundation" models because they serve as the base upon which applications can be built, catering to a wide range of domains and use cases.
Grounding is the process of anchoring artificial intelligence (AI) systems in real-world experiences, knowledge, or data. The objective is to improve the AI's understanding of the world, so it can effectively interpret and respond to user inputs, queries, and tasks. Grounding helps AI systems become more context-aware, allowing them to provide better, more relatable, and relevant responses or actions.
Hallucination refers to a situation wherein an AI system, especially one dealing with natural language processing, generates outputs that may be irrelevant, nonsensical, or incorrect based on the input provided. This often occurs when the AI system is unsure of the context, relies too much on its training data, or lacks a proper understanding of the subject matter.
A type of deep learning model trained on a large dataset to perform natural language understanding and generation tasks. There are many famous LLMs like BERT, PaLM, GPT-2, GPT-3, GPT-3.5, and the groundbreaking GPT-4. All of these models vary in size (number of parameters that can be tuned), in the breadth of tasks (coding, chat, scientific, etc.), and in what they're trained on.
A subfield of AI that involves the development of algorithms and statistical models that enable machines to improve their performance with experience. Example: A machine learning algorithm that can predict which customers are most likely to churn based on their past behavior.
Multi-hop is a term often used in natural language processing and, more specifically, machine reading comprehension tasks. It refers to the process by which an AI model retrieves answers to questions by connecting multiple pieces of information present in a given text or across various sources and systems, rather than directly extracting the information from a single passage.
Zero/Single/Few shot learning are variations of the same concept – providing a model with little or no training data to classify new data and guide predictions. A “shot” represents a single training example. Fun fact: Within the GPT prompt, you can ask for “N” examples to improve the accuracy of the response.
The process of adjusting the parameters of a model to minimize a loss function that measures the difference between the model's predictions and the true values. Example: Optimizing a neural network's parameters using a gradient descent algorithm to minimize the error between the model's predictions and the true values.
Parameter-Efficient Fine-Tuning, also known as PEFT, is an approach that helps you improve the performance of large AI models while optimizing for resources like time, energy, and computational power. To do this, PEFT focuses on adjusting a small number of key parameters while preserving most of the pretrained model's structure.
Identifying inputs — prompts — that result in meaningful outputs. As of now, prompt engineering is essential for LLMs. LLMs are a fusion of layers of algorithms and, consequently, have limited controllability with few opportunities to control and override behavior. An example of prompt engineering is providing a collection of templates and wizards to direct a copywriting application.
Recursive prompting is a strategy for guiding AI models like OpenAI's GPT-4 to produce higher-quality output. It involves providing the model with a series of prompts or questions that build upon previous responses, refining both the context and the AI's understanding to achieve the desired result.
A type of machine learning in which a model learns to make decisions by interacting with its environment and receiving feedback through rewards or penalties. GPT uses reinforcement learning from human feedback. When tuning GPT-3, human annotators provided examples of the desired model behavior and ranked outputs from the model.
Responsible AI refers to the approach of creating, implementing, and utilizing AI systems with a focus on positively impacting employees, businesses, customers, and society as a whole, ensuring ethical intentions and fostering trust, which in turn enables companies to confidently scale their AI solutions.
Stacking is a technique in AI that combines multiple algorithms to enhance overall performance. By blending the strengths of various AI models, stacking compensates for each model's weaknesses and achieves a more accurate and robust output in diverse applications, such as image recognition and natural language processing.
AI steerability refers to the ability to guide or control an AI system's behavior and output according to human intentions or specific objectives. This involves designing AI models with mechanisms that understand and adhere to the preferences provided by users, while avoiding unintended or undesirable outcomes. Improving steerability requires ongoing research and refinement, including techniques like fine-tuning, rule-based systems, and implementing additional human feedback loops during AI development.