In a comprehensive concept guide published by TechCrunch, the publication's writers present a series of key terms in artificial intelligence to provide plain-language definitions for developers, investors, and readers following the field. The glossary is updated regularly alongside the evolution of the systems it describes, spanning fundamental concepts and new computational techniques.
Artificial General Intelligence, AI Agents, and Coding Agents
The term artificial general intelligence (AGI) is defined in the article as a nebulous concept, but it generally refers to an AI system that is more capable than the average human at many, if not most, tasks. OpenAI CEO Sam Altman previously described AGI as the equivalent of a median worker you could hire as a co-worker. OpenAI's charter defines AGI as highly autonomous systems that outperform humans at most economically valuable work. Conversely, Google DeepMind views AGI as a system with capabilities at least equal to those of humans in most cognitive tasks. Experts at the forefront of AI research are also confused.
An AI agent is a tool that uses AI technologies to perform a series of tasks on a user's behalf — beyond the capabilities of a basic chatbot — such as filing expense reports, booking tickets or a table at a restaurant, and even writing and maintaining software code. There are many moving pieces in this emergent space, and the infrastructure is still being built out to deliver on its envisaged capabilities. The basic concept describes an autonomous system that may rely on multiple AI systems to carry out multistep tasks.
Coding agents represent a dedicated version of AI agents applied to software development. Rather than simply suggesting code for a human to review and paste, a coding agent can write, test, and debug code autonomously, handling iterative and trial-and-error work. These agents can operate across entire codebases, spot bugs, run tests, and deploy fixes with minimal human oversight, similar to a fast intern, although human review of outputs is still required.
API endpoints are described as interfaces that allow software programs to trigger actions in other software. Developers use them to build integrations, such as pulling data or allowing an AI agent to control third-party services without manual operation. Most smart home devices and connected platforms include such interfaces. As AI agents evolve, they are increasingly able to find and use these endpoints independently.
Reasoning Methods, Training, and Model Architectures
Chain of thought in the context of large language models describes breaking down a problem into small intermediate steps to improve the quality of the final result. It usually takes longer to receive an answer, but the likelihood of correctness increases, especially in logic and coding contexts. Reasoning models were developed from traditional language models and optimized for step-by-step thinking using reinforcement learning.
Deep learning is a subset of machine learning in which algorithms are structured as multi-layered artificial neural networks, enabling the formation of complex relationships compared to linear models or decision trees. These models identify characteristics in data on their own and learn from errors, but require many data points (millions or more), their training typically takes longer compared to simpler machine learning algorithms, and development costs tend to be higher.
Diffusion is the technology underpinning image-, music-, and text-generation models. These systems gradually add noise to the data structure until none remains, and learn a reverse process designed to reconstruct the original data from the noise.
Distillation is a technique for extracting knowledge from a large model using a "teacher-student" model, where the outputs of the teacher model are used to train the student model to approximate its behavior. This process enables the creation of a smaller, more efficient model with minimal loss, which is likely how OpenAI developed GPT-4 Turbo. Distillation from a competitor's model typically violates API terms of service.
Fine-tuning refers to the additional training of an existing model to optimize it for a specific task or domain, typically by feeding in dedicated data. Many startups build upon large language models and add fine-tuning based on domain expertise.
Generative Adversarial Networks (GANs) involve a pair of neural networks operating against each other: one generates output and the other evaluates whether it is authentic or synthetic. This structured competition can optimize AI outputs to be more realistic without requiring additional human intervention, making it suitable for narrow applications such as generating realistic photos, videos, and deepfake tools.
Mixture of Experts (MoE) is an architecture that splits the neural network into many specialized sub-networks and activates only a fraction of them for any given task using a built-in router. Mistral AI's Mixtral model is a well-known example, and OpenAI's newer GPT models are widely believed to use some version of this approach, though the company has not officially confirmed it.
Compute, Memory, Protocols, and Hardware Constraints
Compute refers to the computational processing power essential for running and training AI models, serving as shorthand for hardware infrastructure such as GPUs, CPUs, and TPUs.
Inference is the process in which an AI model is actively executed to make predictions or draw conclusions based on data learned during training. Inference requires prior training and is supported across a range of hardware, from smartphone processors to accelerated GPUs in cloud servers.
Memory caching, and specifically KV caching, helps streamline inference in transformer-based models by saving previous calculations for reuse, thereby reducing response times and compute consumption.
Model Context Protocol (MCP) is an open standard that allows AI models to connect to external tools and data repositories (such as files, databases, Slack, and Google Drive) without requiring a custom connector for each pairing. The standard was introduced by Anthropic in 2024, handed over to the Linux Foundation, and subsequently adopted by OpenAI, Google, and Microsoft.
Another new term in the hardware market is RAMageddon, describing an increasing shortage of random access memory (RAM) chips caused by massive data center procurement by AI labs and tech companies. The bottleneck increases memory component costs and affects the gaming, consumer electronics, and enterprise computing industries.
Safety Terms, Opaque Recurrence, and Other Key Concepts
Opaque recurrence, or recurrent depth, describes a technique where an AI model loops a query repeatedly through its internal layers rather than reasoning step-by-step in human-readable language. This approach improves compute efficiency but leaves fewer readable traces compared to a standard chain of thought, raising concerns among safety researchers due to the difficulty of monitoring model behavior.
Neuralese describes a hypothetical scenario where a model reasons entirely in internal numerical representations without human language. Safety researchers noted that the use of opaque recurrence in OpenAI's Astra model (launched in September 2026) represents a first step in this direction, though OpenAI stated that the model maintains a readable chain of thought.
Hallucination refers to the generation of incorrect information by models, a problem thought to arise from gaps in training data, and it contributes to the push to develop domain-specific or vertical models.
Additional concepts covered include reinforcement learning from human feedback (RLHF), recursive self-improvement (RSI), tokens and token throughput, parallelization, open source versus closed source, transfer learning, weights, and the validation loss metric used to detect overfitting.