From LLMs to Physical AI — an AI Glossary for Logistics and Manufacturing Teams

LLM, RAG, fine-tuning — the terms pile up in meetings. Everyone nods, and then nobody can define them out loud. You do not need to memorize them like flashcards. AI evolved in a sequence: it learned language, then became an expert in specific domains, then stepped into the physical world. Follow that arc and the vocabulary connects.

Act 1: AI Learns Language

The starting point is the LLM (Large Language Model) — a model trained on vast amounts of human text. It does not assemble words at random; it understands the context of a question and predicts the most natural continuation. ChatGPT and Claude both sit on this foundation.

Two concepts determine capability.

• Parameter: roughly the model's "brain size," the number of connections in the network. More parameters generally means handling more complex context.
• Token: the smallest unit in which the model perceives language — a character or word fragment.

Prompts Decide the Output

Even a large model returns nonsense when the instruction is vague. How precisely you write the prompt changes output quality dramatically on the same model.

Act 2: AI Gets Domain Expertise

There are two main ways to adapt a general model to your operation.

Fine-Tuning — Training It

Fine-tuning adds your internal data — equipment manuals, internal policies — to an existing model through further training. A manufacturer can build a Q&A assistant specific to its own plant.

RAG — Letting It Look Things Up

RAG (Retrieval-Augmented Generation) works like an open-book exam. Just before answering, the system retrieves external documents or current database records and composes the answer from them. In a distribution center where data changes hourly, the model checks today's live inventory before responding.

Fine-tuning trains; RAG consults. Frequently changing data favors RAG, while stable specialist knowledge favors fine-tuning.

Why Accuracy Gets So Much Attention

Because of hallucination — the model producing convincing but fabricated information. In casual conversation it is a joke. On an industrial site, one wrong answer can halt a line or create a safety incident. Fine-tuning and RAG are the control mechanisms for that risk.

Once accuracy is under control, the model moves beyond answering into AI agents that set goals, plan, and execute. Purchase order checks, dispatch requests, and settlement work are increasingly handled this way.

Act 3: AI Enters the Physical World

Multimodal AI is where the model leaves the text box, understanding images, video, and audio alongside text. Inspection systems that analyze factory camera feeds in real time to catch micro-scratches and defects are the common example.

Field deployment adds two constraints: security and latency. On-device AI runs the model directly on the chipset inside a phone, robot, or machine rather than routing through cloud servers. Data never leaves the site, and network delay disappears.

The destination is physical AI — a stage where the model judges and acts in the physical world through 3D vision as its eyes and a robot as its body. Older robots repeated pick-and-place from fixed positions. A physical AI robot recognizes randomly stacked boxes, sorts them itself, and handles parts of varying shapes safely.

The Question That Matters More Than the Terms

The center of gravity in AI is moving from chat windows to logistics and manufacturing floors. But the real question in an adoption review is not a definition. It is "which problem on our floor does this actually solve?" Answer that, and the relevant vocabulary narrows on its own.

For a view on turning AI tools into an actual working system, see AX Design Starts in a Note App.