AI was supposed to make software easier to use. Instead, it has made the language around software harder to parse, and that’s why an AI glossary now belongs next to every product demo, funding pitch, and technical brief.
The latest AI glossary from TechCrunch captures the problem well: people are hearing LLMs, RAG, RLHF, AI agents, hallucinations, inference, and compute in meetings, panels, and product launches, often without a clear sense of what the words actually commit a company to delivering.
Why does an AI glossary matter when the product is supposed to be simple?
The assumption is that AI hides complexity. The reality is messier. The user interface may be a chat box, but the claims behind it often depend on terms that carry technical, financial, and operational meaning.
A company saying it uses a large language model is making a different claim than one saying it has built an AI agent. A team talking about fine-tuning is describing a different process than one talking about inference. A vendor promising fewer hallucinations is not promising truth. It’s saying, at best, that it has tried to reduce a known failure mode.
That distinction matters because AI language can inflate ordinary automation into something that sounds autonomous. TechCrunch’s glossary is useful because it focuses on terms people are likely to encounter while building, investing, or trying to keep up with AI products.
Before vs. after reading the AI glossary:
- Before: “AI agent” sounds like a digital employee.
- After: You ask what tasks it can actually perform, what systems it can access, and where a human still approves the outcome.
- Before: “Open source model” sounds automatically transparent.
- After: You ask what is actually public, the code, the weights, or only the marketing page.
- Before: “Hallucination” sounds like a quirky bug.
- After: You treat it as a quality and risk problem.
The chatbot looks like one product, but the stack underneath is layered
The most common confusion starts with the word model. In AI, the model is the trained system doing the prediction or generation. The app is the product wrapped around it.
ChatGPT, Claude, Google’s Gemini, Meta’s AI Llama, Microsoft Copilot, and Mistral’s Le Chat are examples of AI assistants that use LLMs, according to the TechCrunch glossary. When a user chats with one, the model processes the request directly or uses available tools such as web browsing or code interpreters.
An LLM is built from a deep neural network with billions of numerical parameters, also called weights. Those parameters encode patterns learned from large collections of books, articles, and transcripts. When prompted, the model generates the most likely pattern that fits the request.
That sits inside a wider stack:
| Term | What it means in practice | Reader question to ask |
|---|---|---|
| Neural network | A multi-layered algorithmic structure behind deep learning | What task was it trained for? |
| Deep learning | Machine learning using multi-layered artificial neural networks | How much data did it need? |
| LLM | A model that learns relationships between words and phrases | Is it generating, retrieving, or using tools? |
| Diffusion | A system that learns to recover data from noise | Is this for images, music, text, or another output? |
| GAN | A two-network setup where a generator and discriminator compete | Is the use narrow, such as realistic images or videos? |
Deep learning systems can identify important characteristics in data without engineers manually defining every feature. But TechCrunch notes the trade-off: they require many data points, often millions or more, and typically take longer to train than simpler machine learning systems.
Training builds the model. Inference spends it.
The easy mistake is to treat “AI” as one action. It isn’t.
Training is the process that creates the model. Inference is the process of running it. TechCrunch defines inference as setting a model loose to make predictions or draw conclusions from data patterns it has already learned.
That distinction explains why compute keeps appearing in AI conversations. Compute refers to the computational power needed to train and run AI models. In practice, the term often points to hardware such as GPUs, CPUs, TPUs, and other infrastructure.
Fine-tuning adds another layer. It means further training a model for a narrower task or field, usually by feeding it specialized data. Many AI startups, TechCrunch notes, start with large language models and then improve usefulness for a target sector or task through fine-tuning based on domain-specific knowledge.
Then there is distillation, where developers use outputs from a larger “teacher” model to train a smaller “student” model to approximate its behavior. TechCrunch says distillation can create a smaller, more efficient model with minimal loss, and says this is likely how OpenAI developed GPT-4 Turbo, a faster version of GPT-4. It also notes that distillation from a competitor usually violates AI API and chat assistant terms of service.
Hallucinations expose the gap between fluent language and verified truth
Hallucination is the AI industry’s term for models making things up. The output may sound confident. That doesn’t make it correct.
TechCrunch’s explanation is blunt: hallucinations are generated information that is incorrect, and they are a major AI quality problem. The consequences can move beyond annoyance when an AI system returns harmful medical advice or other misleading information with real-world stakes.
The source ties hallucinations to gaps in training data. If a model lacks reliable information in an area, it may still produce an answer that sounds plausible. That is why the glossary connects hallucinations to growing interest in more specialized or vertical AI models, where narrower domain expertise may reduce knowledge gaps and shrink disinformation risks.
Chain-of-thought reasoning is one response to harder tasks. For simple questions, a model may answer directly. For problems with intermediate steps, breaking the task down can improve accuracy. TechCrunch uses a farmer problem with 40 heads and 120 legs, which resolves to 20 chickens and 20 cows, to show why step-by-step reasoning can matter.
The catch: chain-of-thought style reasoning can take longer. Better answers may cost more time and compute.
OpenAI CEO Sam Altman once described AGI as the “equivalent of a median human that you could hire as a co-worker.”
OpenAI’s charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.”
Google DeepMind views AGI as “AI that's at least as capable as humans at most cognitive tasks.”
Even the term AGI is unsettled. TechCrunch calls it nebulous, and the competing definitions show why. If experts disagree on the destination, readers should be cautious when companies imply they are close to it.
The AI agent pitch sounds autonomous. The expense report tells you if it is.
An AI agent is supposed to do more than chat. TechCrunch describes it as a tool that uses AI technologies to perform a series of tasks on your behalf, such as filing expenses, booking tickets or a restaurant table, or writing and maintaining code.
That sounds powerful. The useful question is what “perform” means.
Take the expense report example. A real agent claim should answer:
- Access: Which systems can it use?
- Action: Can it submit the report, or only draft it?
- Approval: Who signs off before money moves?
- Memory: Does it retain prior actions or context?
- Errors: Where are mistakes caught?
- Accountability: Who is responsible when the agent does the wrong thing?
This is where API endpoints matter. TechCrunch describes them as hidden “buttons” on the back of software that other programs can press to make things happen. Developers use them to build integrations, and more capable agents may be able to find and use these endpoints directly.
A coding agent is a narrower version of the same idea. Instead of merely suggesting code, it can write, test, and debug code across a codebase with limited human oversight. That does not remove review. It changes where review must happen.
Open source, distillation, and MCP turn vocabulary into strategy questions
The glossary also shows which AI terms are becoming decision points.
Open source AI refers to models or software where underlying code is publicly available for others to use, inspect, or modify. TechCrunch names Meta’s Llama family as a prominent AI example and contrasts it with closed source systems such as OpenAI’s GPT models, where users can access the product but not see how it works.
Model Context Protocol, or MCP, is another term to watch. TechCrunch describes MCP as an open standard that lets AI models connect to outside tools and data, including files, databases, Slack, and Google Drive, without a custom connector for every pairing. It says Anthropic introduced MCP in 2024, later handed it to the Linux Foundation, and that OpenAI, Google, and Microsoft have adopted it.
The practical implication is simple: the AI glossary is no longer trivia. It’s a filter for product claims.
If a vendor says “agent,” ask about tools and approvals. If it says “open source,” ask what is actually open. If it says “fine-tuned,” ask on what data. If it says “fewer hallucinations,” ask how errors are measured and caught. The words will keep shifting, but the sharper reader will keep forcing them back to evidence.
Key Takeaways
- AI terminology can make ordinary automation sound more autonomous than it really is.
- Understanding terms like LLMs, RAG, RLHF, inference, and hallucinations helps readers evaluate product claims more critically.
- Clear definitions matter for buyers, builders, and investors trying to separate technical reality from marketing language.
Originally published on XOOMAR. For more news and analysis, visit XOOMAR.
Top comments (0)