LLM is not Gen AI.

Aref Karimi — Wed, 03 Dec 2025 17:55:07 +0000

People often say things like I used an LLM to generate an image, I used an LLM to clone my voice or I used an LLM to make a video. In most cases, this is not technically correct. What they actually used was Generative AI. The confusion comes from the fact that most users interact with AI through chat style tools, so it feels like one single model is doing everything. In reality, different models play very different roles.

Generative AI is the broad field that includes any system that can create new content. That includes text, images, audio, video, code and even scientific structures. Image generators, voice synthesis tools, video creators and music generators all fall under Generative AI. These models are trained on very different types of data such as pixels, sound waves or motion patterns.

An LLM, or large language model, is just one type of Generative AI. It is trained only on text. Its job is to understand language, generate language, reason, summarise and follow written instructions. It does not generate images, sound or video by itself. It only works with words.

In tools like ChatGPT, the LLM acts as the language interface. It reads your prompt, understands your intent and decides what should happen next. If you ask for an explanation, it responds directly. If you ask for an image or audio, it sends structured instructions to a specialised Generative AI model that actually creates the content. The LLM then explains the result back to you in plain language.

This is why saying I used an LLM to generate an image is usually giving credit to the wrong part of the system. The LLM helped translate your idea into instructions. The Generative AI model did the actual creative work.

Understanding this difference leads to better expectations, better tool choices and smarter AI system design. LLMs are powerful, but they are not all purpose generators. They are the language brain sitting on top of a much larger Generative AI ecosystem.

Read the full article here:
LLMs and Generative AI are not the same!

Making Sense of MCP and ACP, two wings of high flying AI Agents!

Aref Karimi — Wed, 03 Dec 2025 07:32:35 +0000

There is a lot of excitement around AI agents right now, yet most engineers still feel unsure about how these systems actually gather information or communicate with each other.

There are several ways to provide the large language models with external data or context, and one of them is through Model Context Protocols (MCP). MCP is pretty popular right now, however, it usually goes hand-in-hand with ACP, or Agent Communication Protocol that is a concept cooked by IBM! ACP is there to provide AI Agents with a unified interface to streamline the Agent-to-Agent communicaiton and Agent discovery.

If you are building or experimenting with AI agents, understanding these two ideas makes everything far less mysterious and gives you a clearer path to building reliable workflows.

What MCP really handles

MCP stands for Model Context Protocol. Its purpose is simple. It gives your AI agent a structured way to reach out to external sources and bring the correct context into the model at the right time. Instead of hardcoding integrations or relying on unpredictable plugins, MCP acts like a contract that tells the agent how to retrieve data from tools, services or internal systems. They can also do stuff. For example, they can update the customers' records in a CRM system.

This becomes important the moment you try to use AI in a real workplace. Models are only as good as the context they have. With MCP, your agent can fetch live information, query APIs, inspect databases or read documents in a consistent and predictable way. It removes guesswork and stops the usual mess of one off integrations that break every time your environment changes.

Why ACP matters as the next step

ACP means Agent Communication Protocol. Once you have multiple agents working together, you need a clear and reliable way for them to talk to one another. Without standards, each agent ends up with its own method of sending messages, sharing state or coordinating tasks, and it quickly becomes unmanageable.

ACP gives you a shared interface for communication. Agents know how to exchange information, request actions and hand off work without custom glue code. This leads to cleaner orchestration, easier debugging and faster development because the rules stay consistent across your entire system.

The important part is the interaction between the two

MCP brings the right context into the agent. ACP lets agents communicate that context and act on it together. When both are in place, you get something far more powerful than a single model responding to a prompt. You get a small ecosystem of agents that can gather information, collaborate, adapt to new inputs and maintain reliability across complex tasks.

Why this matters for everyday engineers

If you are building AI features into existing software, experimenting with agent workflows or planning internal automation, MCP and ACP give you the foundation you need. They remove the frustration of unpredictable behaviour, reduce the risk of incorrect assumptions and give you patterns that scale beyond toy examples.

If you want to learn how to implement MCP and ACP in a practical environment with real tools and working examples, I have put together a complete course that walks through everything step by step. It covers the concepts, the architecture and the hands on side that most tutorials skip. If you want to build agents that actually work in real systems and avoid the usual early mistakes, this is the best place to start.

Happy MCPing!

Forem: Aref Karimi

LLM is not Gen AI.

Making Sense of MCP and ACP, two wings of high flying AI Agents!