Artificial intelligence is a complex and multifaceted field. Researchers often use specialized language to describe their work, which we incorporate into our articles on AI. To assist readers, we’ve compiled a glossary explaining key terms we frequently use.
We’ll update this glossary regularly as new AI techniques emerge and as safety concerns are identified.
Artificial General Intelligence (AGI) is a term often discussed but loosely defined. Generally, it refers to AI that can outperform humans in most tasks. OpenAI CEO Sam Altman describes AGI as comparable to an average human co-worker, while OpenAI’s charter defines it as highly autonomous systems excelling at economically valuable work. Google DeepMind sees AGI as being as capable as humans in cognitive tasks. Even leading AI researchers find the definition challenging.
An AI agent is an advanced tool using AI to perform tasks like expense filing, booking tickets, or coding, surpassing basic chatbots. The term “AI agent” can vary in meaning as the field evolves. Infrastructure is developing to support their capabilities, implying autonomous systems performing multi-step tasks using various AI systems.
For simple questions, humans can quickly respond, e.g., whether a giraffe or cat is taller. Complex problems might require pen and paper. In AI, chain-of-thought reasoning in large language models breaks down problems into smaller steps for better results, especially in logical or coding contexts. These models are enhanced for chain-of-thought via reinforcement learning.
Compute generally refers to the crucial computational power for AI models. It fuels AI’s model training and deployment, often used to denote the hardware like GPUs, CPUs, and TPUs essential to AI infrastructure.
Deep learning involves a multi-layered artificial neural network structure, enabling complex correlations. These AI models identify data features without human input, learning from errors to improve outputs. However, they need vast data and time-consuming training, raising development costs.
Diffusion is central to many art-, music-, and text-generating AI models. Inspired by physics, it dismantles data structure (e.g., photos, songs) with noise until nothing remains, learning reverse diffusion to restore data from noise.
Distillation extracts knowledge from large AI models through a “teacher-student” model approach, enabling the creation of smaller, efficient models from larger ones. Distillation is internally used by AI firms but may breach AI service terms if applied to competitor models.
Fine-tuning further trains AI models for specific tasks, using new, specialized data. Many startups enhance large language models for targeted sector-specific applications.
A Generative Adversarial Network (GAN) is a machine learning framework for generating realistic data, not limited to deepfake tools. GANs feature two neural networks; one generates output, and the other evaluates it. This adversarial setup refines data realism without extra human input.
Hallucination in AI refers to models generating incorrect information, a significant quality issue. Hallucinations can mislead, with risks like harmful medical advice. Training data gaps often cause this, especially for general-purpose models, prompting a shift to specialized AI models to minimize risks.
Inference is the process of using an AI model to predict or conclude from trained data. Diverse hardware can perform inference, but efficiency varies. Large models may run slowly on laptops compared to cloud servers with advanced AI chips.
Large language models (LLMs) drive popular AI assistants like ChatGPT, Claude, and Google’s Gemini. Interacting with them involves processing requests directly or using tools like web browsing. LLMs are vast neural networks with billions of parameters, learning word relationships and generating likely response patterns.
Memory cache enhances inference by optimizing calculations, reducing algorithmic workload and power use. KV caching in transformer models accelerates response generation by saving calculations for future queries.
A neural network is an algorithmic structure behind deep learning and generative AI tools. Inspired by the brain, it was unlocked by the gaming industry’s graphical processing hardware, enabling better performance across domains like voice recognition and autonomous navigation.
RAMageddon highlights the tech industry’s RAM chip shortage, affecting daily tech product availability. As AI grows, major firms absorb RAM for data centers, driving up prices in sectors like gaming and consumer electronics.
Training in AI is feeding data to a model to identify patterns and create useful outputs. Pre-training, the AI structure is merely layers and random numbers; training shapes the model to meet goals like image recognition or poetry creation. Not all AI needs training — rules-based systems follow predefined instructions. Training is resource-intensive, requiring considerable inputs.
Human-AI communication involves tokens, data segments for LLM interaction. Tokenization refines data into units an AI can process. Tokens vary, impacting enterprise AI costs, as pricing often relates to token usage.
Transfer learning uses a trained AI model for related tasks, leveraging past learning to boost efficiency, especially with limited data. Limitations exist; models may need extra training for refined focus.
Weights are crucial in AI training, determining the importance of data features, shaping model output. They are parameters adjusting during training