TechCrunch released an AI glossary that defines common artificial intelligence terms for readers, investors, and developers. The guide, updated regularly, provides clear explanations of acronyms and concepts that appear in product meetings, investment discussions, and public commentary. AGI is described by OpenAI CEO Sam Altman as a system comparable to a median human employee, while OpenAI’s charter defines it as highly autonomous systems that outperform humans in most economically valuable tasks; DeepMind characterizes AGI as AI matching human capability across most cognitive tasks. An AI agent is a tool that uses AI to perform multiple tasks on a user’s behalf, such as filing expenses or writing code, though its precise meaning varies. API endpoints function as interfaces that allow software to invoke other services, enabling integrations and autonomous actions by AI agents. Chain‑of‑thought reasoning involves breaking a problem into intermediate steps to improve answer accuracy, a technique used in reasoning‑optimized large language models. A coding agent is a specialized AI agent that can write, test, and debug code with minimal human oversight. Compute refers to the hardware resources, such as GPUs and TPUs, that enable AI model training and inference. Deep learning is a subset of machine learning that employs multi‑layered neural networks to identify patterns in data. Diffusion models, used in generative AI for images, audio, and text, learn to reverse a noise‑adding process. Distillation extracts knowledge from a larger model to train a smaller, more efficient model. Fine‑tuning involves further training a model on specialized data to improve performance for a specific task. GANs are neural network frameworks that pit a generator against a discriminator to produce realistic data, commonly applied to deepfake creation. Hallucination describes instances where models generate incorrect information, raising concerns for applications such as medical advice. Inference is the execution of a trained model to produce predictions, requiring hardware ranging from smartphones to cloud GPUs. Large language models are neural networks with billions of parameters that generate text based on prompts, forming the basis of many AI assistants. The Model Context Protocol provides a standardized way for AI models to connect to external tools and data sources. Mixture of Experts architecture activates only a subset of specialized sub‑networks for each task, improving efficiency in large models. Neural networks are layered algorithms that underpin deep learning and modern generative AI. Open source AI models, such as Meta’s Llama series, are publicly available for inspection and modification, contrasting with closed‑source offerings like OpenAI’s GPT. Parallelization enables simultaneous processing across multiple chips, which is important for training and inference at scale. A shortage of RAM chips has driven up prices across consumer electronics, gaming, and enterprise computing. Recursive self‑improvement refers to AI systems that modify their own architecture without human input, a capability explored by several startups. Reinforcement learning trains models through trial and reward signals, and is used to enhance reasoning in large language models via methods such as RLHF. Tokens are discrete units of text processed by language models, and token throughput measures the speed at which a system can handle these units. The glossary continues to be updated as the field evolves, offering a reference for ongoing developments in artificial intelligence.