The AI Jargon Glossary You Actually Need: From AGI to Diffusion, Explained Plainly

Key points

• AGI, or artificial general intelligence, has no single agreed definition, even among the labs racing to build it.
• OpenAI defines AGI as systems that "outperform humans at most economically valuable work", while Google DeepMind sets the bar at matching human capability across most cognitive tasks.
• AI agents, software tools that carry out multi-step tasks on your behalf, are reshaping how developers build products in 2025.
• Techniques such as chain-of-thought reasoning and distillation are now standard building blocks inside most commercial AI products.
• Knowing a handful of these terms makes it far easier to judge AI headlines, product claims and policy debates on their merits.

Every week brings a new batch of terms that smart people feel embarrassed to admit they don't know. TechCrunch AI published a living glossary to help; below is our own plain-English guide to the ones that matter most right now.

What do all these AI terms actually mean?

Most AI jargon names a specific technique or a specific goal. Once you know what problem each term is solving, the word usually sticks.

AGI (artificial general intelligence) is the idea of an AI that can do most things a person can do, at least as well. OpenAI's founding charter describes it as "highly autonomous systems that outperform humans at most economically valuable work." Google DeepMind frames it as AI that matches human ability across most cognitive tasks. Neither lab agrees with the other, and neither has built it.

An AI agent is software that carries out a sequence of tasks on your behalf, without you doing each step manually. Think: booking a flight or filing an expense report. It's different from a chatbot, which just answers questions. The infrastructure to make agents reliable is still being built.

API endpoints are, in plain terms, hidden buttons inside software that other programs can press. A developer building an app might press the button on a mapping service to get directions, or on a payment provider to charge a card. As AI agents grow more capable, they can find and press these buttons without a human involved, which is powerful and occasionally surprising.

Chain-of-thought reasoning means asking an AI model to show its working, the way a maths teacher asks a student to write out every step rather than just circling an answer. Breaking a problem into smaller pieces produces more accurate results, especially for logic tasks. It takes longer, but the answer is more likely to be right.

A coding agent is an AI agent built for software development. Rather than suggesting code for a human to paste in, it writes, runs and debugs code autonomously, cycling through trial and error the way a developer would. A human should still review the output.

Compute is shorthand for the raw processing power that AI models need, both to learn and to run. The hardware doing that work includes GPUs (specialised chips built for heavy number-crunching), CPUs (the general-purpose chips inside most computers) and TPUs (chips Google designed specifically for AI workloads).

Deep learning is a method of building AI where the system is structured in many layers, loosely inspired by the way neurons connect in a human brain. Each layer spots patterns in data, and the whole system learns by correcting its own mistakes over millions of examples. We covered how deep learning relates to the broader field of machine learning in our plain-English guide from 23 July.

Diffusion is the technique behind most AI image and music generators. The system learns to destroy an image by adding random noise until nothing recognisable remains, then learns to reverse that process. Once it can reverse the noise, it can generate new images from scratch.

Distillation is how labs compress a large, expensive AI model into a smaller, cheaper one. A big "teacher" model answers thousands of questions; a smaller "student" model studies those answers and learns to copy the teacher's behaviour. The result is a faster, lighter model that keeps most of the quality. Using a competitor's model as the teacher almost always breaks that company's terms of service. It's worth the context: as we reported on 28 July, distillation sits at the centre of a live US-China technology dispute, with Washington arguing China used it to extract American know-how and Silicon Valley calling it standard engineering.

Fine-tuning means taking a general-purpose AI model and training it further on a specific subject, such as medical records or legal contracts, so it performs better in that narrow area. Many AI startups build their products this way rather than training a model from scratch.

What should ordinary people take from all this?

You don't need to build with these tools to benefit from understanding them. If a product claims to use an "AI agent" to manage your finances, you now know that means software acting on your behalf, often without asking permission at each step. Worth knowing before you hand over access.

When a news story mentions that a lab has "fine-tuned" a model on health data, you can picture what that process involved. When someone says a company distilled a rival's model, you know that likely violated a contract.

The terms keep multiplying. But the underlying questions are always the same: what data did this learn from, what's it doing on its own, and who checked the work?

Common questions

Is AGI actually close?

No lab has built AGI, and researchers disagree on what crossing that line would even look like. The definitions from OpenAI, Google DeepMind and independent researchers all differ in meaningful ways, so treat any claim of imminent AGI with healthy scepticism.

Are AI agents safe to use for personal tasks?

Agents are still maturing. They're most reliable for contained tasks with clear goals and easy-to-check results, such as drafting a document or writing a code snippet. For anything involving money, sensitive data or irreversible actions, review every step before it goes through.

Do I need to understand these terms to use AI tools?

Not to get started. But knowing what an agent, a fine-tuned model or an API endpoint actually does helps you ask better questions, spot marketing exaggeration and make smarter choices about which tools to trust.