AI - Understanding Basics
Generative AI
Generative AI has moved from research papers into daily business conversations faster than almost any technology in recent memory. When ChatGPT opened to the public in late 2022, it pulled an entire category of tools out of the lab and in front of everyday users. Within months, enterprise platforms, coding assistants, and AI-powered search had pushed the technology into mainstream IT. That is why so many IBM i teams are now hearing the same questions on repeat: What exactly is generative AI, how is it different from machine learning, and why should we care right now?
For IBM i professionals, the honest answer is that this topic matters more to you than it does to most people in IT. Generative AI is strongest in the areas where IBM i teams feel the most daily pressure: understanding complex systems, documenting legacy logic, speeding up research, assisting with code, and helping users access business information in plain language. It is not a replacement for RPG developers, architects, or support teams. It is a new class of tools that can help skilled people work faster, with better context, and with less friction. This chapter keeps the language simple and focuses on what is genuinely useful on the platform.
Start with the basics: AI, machine learning, and generative AI
Artificial intelligence, or AI, is the broad umbrella. It covers any machine-based system that can make predictions, recommendations, or decisions toward a human-defined goal. That definition is intentionally wide because AI includes many different approaches and has existed in various forms for decades.
Machine learning sits inside AI. Here the system learns patterns from data instead of relying only on fixed, hand-written rules. This is the branch of AI that powers recommendation engines, fraud detection, demand forecasting, and most modern language technology.
Generative AI is a newer category within that machine learning world. It refers to models that can learn the structure of their input data and then generate new content — text, images, audio, video, or code. The key distinction is simple: traditional models usually classify or predict, while generative models create. If a traditional model flags a transaction as suspicious, that is useful AI. If a generative model explains a block of RPG, summarizes a support ticket, or turns a plain-English question into working SQL against Db2 for i, that is generative AI in action.
A quick timeline: how we got here
AI did not start with ChatGPT. The earliest AI systems were rule-based expert systems that worked well inside narrow domains but depended on experts writing detailed logic by hand. Machine learning improved on this by letting systems learn from data, but language remained stubborn because human communication depends on context, sequence, tone, and ambiguity that rules and simple statistics cannot capture.
The big technical leap came in 2017 with the transformer architecture, introduced in the paper Attention Is All You Need. Transformers made it far easier for models to understand long-range relationships in text, and far easier to train at scale. Almost every modern large language model is built on that foundation.
The public breakthrough came on November 30, 2022, with the release of ChatGPT. It was not the beginning of AI, but it was the moment generative AI became visible, practical, and easy to experience. Expectations inside businesses changed almost overnight, because non-specialists could suddenly watch AI writing, explaining, summarizing, and answering questions in real time. From that point on, the question was no longer whether AI could do useful work. It was where and how to apply it safely.
How large language models actually work
A large language model, or LLM, is a deep learning model trained on enormous amounts of text so it can understand and generate natural language. The easiest useful explanation is that it predicts what comes next. When you give it a prompt, the model breaks the text into smaller units called tokens, then calculates which token is most likely to follow given everything before it. It repeats that process until the response is complete.
This is why people describe generative AI as “advanced word prediction.” The phrase is simplified, but it is still a good starting point. It explains why these systems are so fluent — and why they can sound confident while being wrong. The model is excellent at producing language that fits the pattern of the context. That does not mean it has verified the facts.
The same mechanism is why one model can handle many different tasks. The same LLM can summarize a document, explain code, draft an email, or translate a business request into SQL — not because it was programmed separately for each task, but because it learned broad patterns in language and applies them across contexts.
For IBM i professionals, the realistic framing is important. An LLM does not “understand” your business the way a veteran architect does. It does not carry twenty years of lived experience with your pricing rules, your product hierarchies, or the 1998 customization nobody has dared to touch. It works from patterns in training data and from whatever context you provide at the moment you ask. That is exactly why prompts, documentation, system design, and human review still matter.
Why 2022 to 2026 changed everything
Generative AI became mainstream because several things matured at once. Transformer-based models became dramatically more capable. Hardware made large-scale training and inference economically practical. User interfaces became so easy that a first-time user could get value in minutes. And enterprise vendors began packaging these capabilities into real workflows rather than treating them as research demos.
You can see that shift most clearly in development tooling. AI coding agents such as Claude Code, IBM Bob, Cursor, and Cline go well beyond next-line suggestions. They can read across a codebase, edit multiple files, run tests, coordinate multi-step tasks, and take direction in plain English. The move from chatbot to tool-using assistant is the real reason the conversation has become more practical. The technology is no longer just talking about work. It is starting to do the work.
What generative AI can do well today
Generative AI is strongest when the task involves language, explanation, summarization, pattern recognition, or first-draft creation. In IBM i environments, that translates into concrete everyday wins. It can summarize long design documents, support tickets, or change requests into something a busy lead can read in a minute. It can explain old RPG III, RPGLE, CL, or embedded SQL — invaluable when the original author left a decade ago. It can draft first-pass technical documentation instead of asking a developer to start from a blank page. It can turn natural-language questions into SQL against Db2 for i, opening data access to people who know the business but not the schema. And it can act as a conversational layer on top of green-screen or browser-based applications, shortening training time for new users.
Consider a realistic example. A new developer joins a team supporting a fifteen-year-old manufacturing application with dozens of interdependent RPG programs, scattered business rules, and only partial documentation. A generative AI assistant can walk the developer through program flow, summarize what each module appears to do, identify related files and objects, and draft a first pass of technical notes. It does not remove the need for an experienced developer. It simply shortens the ramp from months to weeks.
What generative AI cannot do reliably
For all the excitement, generative AI still has real limits that IBM i teams need to respect.
The most famous is hallucination. A model can produce an answer that sounds polished and authoritative but is still false, incomplete, or misleading. This is not theoretical. It is an everyday reality whenever AI outputs are used without validation. An LLM can happily invent a system API that does not exist, cite a manual that was never written, or describe program behavior that contradicts what the code actually does.
The second limitation is business context. A general-purpose model understands patterns in code and language, but not your company’s pricing tiers, inventory exceptions, approval workflows, tax treatments, or audit requirements. IBM i environments are full of domain-specific logic built up over many years, and that logic often matters more than the syntax.
The third is governance and security. Enterprise AI is not only about model capability. It is about where data goes, who can see outputs, how systems are monitored, and how risks are controlled. A good AI answer is not enough on its own — the path that produced it has to be defensible too. In an IBM i environment where a single application might touch finance, HR, and customer records, that is not optional.
Why IBM i professionals should pay attention
IBM i systems often sit at the operational core of the business. They run manufacturing lines, ledgers, claims, logistics, and distribution processes that companies depend on every day. Their strength is reliability and the depth of business logic they carry. Their challenge is that this knowledge is often trapped inside code, job logs, scribbled notes, and the heads of a shrinking group of experts who know why a particular flag in a particular file changes the behavior of a particular nightly job.
Generative AI matters because it can help surface and translate that knowledge. It can make old logic easier to explain, support faster onboarding, improve documentation, and reduce the amount of institutional memory required for routine understanding tasks. In modernization programs, it can support analysis, planning, and knowledge transfer long before any actual code changes begin — and that groundwork is often the hardest part of the project.
None of this means IBM i professionals should fear being replaced. If anything, the opposite is true. The more complex and business-critical the environment, the more valuable experienced people become. AI is most useful when it amplifies expert teams by removing friction, not when it tries to bypass expertise. On IBM i, expertise is the whole point.
Final thoughts
Generative AI is real, but it is not magic. It works especially well with language, explanation, summarization, and first-draft creation — exactly the everyday frictions that slow IBM i teams down. Approach it with curiosity and discipline at the same time. Learn the basic ideas. Test the tools in low-risk workflows first. Stay honest about limitations. Keep humans in the loop for anything that touches production data, customer records, or financial logic.
IBM i has survived every wave of IT change for decades because the people running it have known when to adopt, when to wait, and when to adapt. Generative AI is simply the next wave. Approached with the same judgment, it is likely to become one of the most useful tools the platform has seen in years.
AI Hype Vs Reality
Generative AI has moved from research papers into daily business conversations faster than almost any technology in recent memory. When ChatGPT opened to the public in late 2022, it pulled an entire category of tools out of the lab and in front of everyday users. Within months, enterprise platforms, coding assistants, and AI-powered search had pushed the technology into mainstream IT. That is why so many IBM i teams are now hearing the same questions on repeat: What exactly is generative AI, how is it different from machine learning, and why should we care right now?
For IBM i professionals, the honest answer is that this topic matters more to you than it does to most people in IT. Generative AI is strongest in the areas where IBM i teams feel the most daily pressure: understanding complex systems, documenting legacy logic, speeding up research, assisting with code, and helping users access business information in plain language. It is not a replacement for RPG developers, architects, or support teams. It is a new class of tools that can help skilled people work faster, with better context, and with less friction. This chapter keeps the language simple and focuses on what is genuinely useful on the platform.
Start with the basics: AI, machine learning, and generative AI
Artificial intelligence, or AI, is the broad umbrella. It covers any machine-based system that can make predictions, recommendations, or decisions toward a human-defined goal. That definition is intentionally wide because AI includes many different approaches and has existed in various forms for decades.
Machine learning sits inside AI. Here the system learns patterns from data instead of relying only on fixed, hand-written rules. This is the branch of AI that powers recommendation engines, fraud detection, demand forecasting, and most modern language technology.
Generative AI is a newer category within that machine learning world. It refers to models that can learn the structure of their input data and then generate new content — text, images, audio, video, or code. The key distinction is simple: traditional models usually classify or predict, while generative models create. If a traditional model flags a transaction as suspicious, that is useful AI. If a generative model explains a block of RPG, summarizes a support ticket, or turns a plain-English question into working SQL against Db2 for i, that is generative AI in action.
A quick timeline: how we got here
AI did not start with ChatGPT. The earliest AI systems were rule-based expert systems that worked well inside narrow domains but depended on experts writing detailed logic by hand. Machine learning improved on this by letting systems learn from data, but language remained stubborn because human communication depends on context, sequence, tone, and ambiguity that rules and simple statistics cannot capture.
The big technical leap came in 2017 with the transformer architecture, introduced in the paper Attention Is All You Need. Transformers made it far easier for models to understand long-range relationships in text, and far easier to train at scale. Almost every modern large language model is built on that foundation.
The public breakthrough came on November 30, 2022, with the release of ChatGPT. It was not the beginning of AI, but it was the moment generative AI became visible, practical, and easy to experience. Expectations inside businesses changed almost overnight, because non-specialists could suddenly watch AI writing, explaining, summarizing, and answering questions in real time. From that point on, the question was no longer whether AI could do useful work. It was where and how to apply it safely.
How large language models actually work
A large language model, or LLM, is a deep learning model trained on enormous amounts of text so it can understand and generate natural language. The easiest useful explanation is that it predicts what comes next. When you give it a prompt, the model breaks the text into smaller units called tokens, then calculates which token is most likely to follow given everything before it. It repeats that process until the response is complete.
This is why people describe generative AI as “advanced word prediction.” The phrase is simplified, but it is still a good starting point. It explains why these systems are so fluent — and why they can sound confident while being wrong. The model is excellent at producing language that fits the pattern of the context. That does not mean it has verified the facts.
The same mechanism is why one model can handle many different tasks. The same LLM can summarize a document, explain code, draft an email, or translate a business request into SQL — not because it was programmed separately for each task, but because it learned broad patterns in language and applies them across contexts.
For IBM i professionals, the realistic framing is important. An LLM does not “understand” your business the way a veteran architect does. It does not carry twenty years of lived experience with your pricing rules, your product hierarchies, or the 1998 customization nobody has dared to touch. It works from patterns in training data and from whatever context you provide at the moment you ask. That is exactly why prompts, documentation, system design, and human review still matter.
Why 2022 to 2026 changed everything
Generative AI became mainstream because several things matured at once. Transformer-based models became dramatically more capable. Hardware made large-scale training and inference economically practical. User interfaces became so easy that a first-time user could get value in minutes. And enterprise vendors began packaging these capabilities into real workflows rather than treating them as research demos.
You can see that shift most clearly in development tooling. AI coding agents such as Claude Code, IBM Bob, Cursor, and Cline go well beyond next-line suggestions. They can read across a codebase, edit multiple files, run tests, coordinate multi-step tasks, and take direction in plain English. The move from chatbot to tool-using assistant is the real reason the conversation has become more practical. The technology is no longer just talking about work. It is starting to do the work.
What generative AI can do well today
Generative AI is strongest when the task involves language, explanation, summarization, pattern recognition, or first-draft creation. In IBM i environments, that translates into concrete everyday wins. It can summarize long design documents, support tickets, or change requests into something a busy lead can read in a minute. It can explain old RPG III, RPGLE, CL, or embedded SQL — invaluable when the original author left a decade ago. It can draft first-pass technical documentation instead of asking a developer to start from a blank page. It can turn natural-language questions into SQL against Db2 for i, opening data access to people who know the business but not the schema. And it can act as a conversational layer on top of green-screen or browser-based applications, shortening training time for new users.
Consider a realistic example. A new developer joins a team supporting a fifteen-year-old manufacturing application with dozens of interdependent RPG programs, scattered business rules, and only partial documentation. A generative AI assistant can walk the developer through program flow, summarize what each module appears to do, identify related files and objects, and draft a first pass of technical notes. It does not remove the need for an experienced developer. It simply shortens the ramp from months to weeks.
What generative AI cannot do reliably
For all the excitement, generative AI still has real limits that IBM i teams need to respect.
The most famous is hallucination. A model can produce an answer that sounds polished and authoritative but is still false, incomplete, or misleading. This is not theoretical. It is an everyday reality whenever AI outputs are used without validation. An LLM can happily invent a system API that does not exist, cite a manual that was never written, or describe program behavior that contradicts what the code actually does.
The second limitation is business context. A general-purpose model understands patterns in code and language, but not your company’s pricing tiers, inventory exceptions, approval workflows, tax treatments, or audit requirements. IBM i environments are full of domain-specific logic built up over many years, and that logic often matters more than the syntax.
The third is governance and security. Enterprise AI is not only about model capability. It is about where data goes, who can see outputs, how systems are monitored, and how risks are controlled. A good AI answer is not enough on its own — the path that produced it has to be defensible too. In an IBM i environment where a single application might touch finance, HR, and customer records, that is not optional.
Why IBM i professionals should pay attention
IBM i systems often sit at the operational core of the business. They run manufacturing lines, ledgers, claims, logistics, and distribution processes that companies depend on every day. Their strength is reliability and the depth of business logic they carry. Their challenge is that this knowledge is often trapped inside code, job logs, scribbled notes, and the heads of a shrinking group of experts who know why a particular flag in a particular file changes the behavior of a particular nightly job.
Generative AI matters because it can help surface and translate that knowledge. It can make old logic easier to explain, support faster onboarding, improve documentation, and reduce the amount of institutional memory required for routine understanding tasks. In modernization programs, it can support analysis, planning, and knowledge transfer long before any actual code changes begin — and that groundwork is often the hardest part of the project.
None of this means IBM i professionals should fear being replaced. If anything, the opposite is true. The more complex and business-critical the environment, the more valuable experienced people become. AI is most useful when it amplifies expert teams by removing friction, not when it tries to bypass expertise. On IBM i, expertise is the whole point.
Final thoughts
Generative AI is real, but it is not magic. It works especially well with language, explanation, summarization, and first-draft creation — exactly the everyday frictions that slow IBM i teams down. Approach it with curiosity and discipline at the same time. Learn the basic ideas. Test the tools in low-risk workflows first. Stay honest about limitations. Keep humans in the loop for anything that touches production data, customer records, or financial logic.
IBM i has survived every wave of IT change for decades because the people running it have known when to adopt, when to wait, and when to adapt. Generative AI is simply the next wave. Approached with the same judgment, it is likely to become one of the most useful tools the platform has seen in years.
Glossary – 30 Key Terms
AI conversations move fast, and the vocabulary moves even faster. When people start talking about LLMs, context windows, embeddings, RAG, agents, copilots, and coding assistants, it can feel like the discussion is happening in a different language. This glossary fixes that.
The goal is simple: give IBM i professionals a plain-English reference they can actually use. Each definition is short and practical, and every entry is paired with an IBM i analogy to make the concept stick.
1. Artificial Intelligence (AI)
A broad term for systems that perform tasks normally requiring human intelligence — pattern recognition, prediction, or decision-making.
IBM i analogy: AI is the umbrella category, in the same way “the IBM i ecosystem” is an umbrella over RPG, CL, Db2, and every tool around them.
2. Machine Learning (ML)
A branch of AI where systems learn patterns from data instead of relying on hand-written rules.
IBM i analogy: Instead of coding every rule in RPG or CL, the system learns the rules from historical records.
3. Generative AI
AI that creates new content — text, code, summaries, answers — rather than only classifying or predicting.
IBM i analogy: Not an SQL query that returns rows, but a tool that writes a new paragraph of explanation about them.
4. Large Language Model (LLM)
A deep learning model trained on massive amounts of text that can understand and generate language.
IBM i analogy: Imagine a model trained on every IBM Redbook, every RPG manual, and every business document ever written.
5. Transformer
The neural network architecture behind most modern LLMs, designed to handle long sequences of text efficiently.
IBM i analogy: The engine architecture inside today’s AI — similar to how Power is the processor architecture inside IBM i.
6. Token
A small piece of text the model processes — roughly a word, part of a word, or a punctuation mark.
IBM i analogy: Tokens are to an LLM what bytes are to a program: the smallest unit it actually sees.
7. Prompt
The instruction or question you give the model.
IBM i analogy: Like a CL command, but written in plain English.
8. Prompt Engineering
Crafting prompts so the model produces clearer, more accurate outputs.
IBM i analogy: Tuning command parameters until you get the result you actually want.
9. Context Window
The maximum amount of text the model can consider at once.
IBM i analogy: The model’s working memory for the current job — anything outside the window is effectively forgotten.
10. Next-Token Prediction
The core mechanism of an LLM: predicting the next piece of text one step at a time.
IBM i analogy: Building an output record one field at a time, based on everything written so far.
11. Training
The process of teaching a model by exposing it to huge datasets.
12. Inference
The stage where a trained model generates answers in response to input.
IBM i analogy: Running the program after it has been compiled.
13. Parameters
The internal values a model learns during training, often numbering in the billions.
IBM i analogy: Internal weights baked into the model — not settings you read or edit directly.
14. Hallucination
When a model produces confident but incorrect information.
IBM i analogy: A junior developer who confidently explains a program in a way that doesn’t match what the code actually does.
15. Fine-Tuning
Adapting a general-purpose model using your own organization’s data or examples.
IBM i analogy: Customizing a generic template to match your shop’s coding standards and business rules.
16. Embedding
A numerical representation of text that captures its meaning, not just its letters.
IBM i analogy: A semantic index of a record — two different sentences end up “close” if they mean the same thing.
17. Vector Database
A database built to store and search embeddings by meaning rather than exact text.
IBM i analogy: A search engine that finds records with similar meaning, not just similar wording.
18. RAG (Retrieval-Augmented Generation)
A pattern where the system retrieves relevant documents first, then feeds them to the model before it answers.
IBM i analogy: Handing the developer the right specs before asking them to write the program.
19. Grounding
Tying AI answers to approved, trusted enterprise data so they reflect reality.
IBM i analogy: Verifying a result against the system of record before trusting it.
20. Agent
An AI system that can plan steps, use tools, and complete multi-stage tasks — not just answer a single question.
IBM i analogy: Less like F4 help on a single field, more like a supervised assistant that can actually run the next step of the job.
21. AI Coding Agent
An agent specialized in reading code, making edits, running tests, and supporting multiple stages of development. Examples include Claude Code, IBM Bob, and Cline.
IBM i analogy: A junior developer who can explore a codebase, draft changes, and run jobs — always under review.
22. Copilot
A general-purpose AI assistant that supports a user’s work rather than taking full control.
IBM i analogy: An experienced colleague looking over your shoulder — quick to offer help, but never pressing Enter for you.
23. Agentic Workflow
A workflow where AI can take actions, call tools, and progress through steps with check-ins along the way.
IBM i analogy: A controlled job stream with approval gates at each stage.
24. Human in the Loop
A system design where people review and approve AI actions before they take effect.
IBM i analogy: Change management with required review gates — AI proposes, humans approve.
25. Multimodal
Models that can work with multiple input types — text, images, audio, and code — in the same conversation.
IBM i analogy: A system that understands data records, screenshots of green screens, and PDF specs all at once.
26. Structured Data
Data stored in well-defined formats such as database tables.
IBM i analogy: Db2 for i files — rows, columns, and known data types.
27. Unstructured Data
Text, PDFs, notes, logs, or documents with no strict schema.
IBM i analogy: Free-form comments in a program, spool files, emails, and support tickets.
28. Model Governance
Processes for controlling, reviewing, and auditing how AI models behave in an organization.
IBM i analogy: Change control, auditing, and authority structures — applied to AI instead of to programs.
29. Prompt Injection
A security risk where malicious or unexpected input hijacks how an AI behaves.
IBM i analogy: A crafted input that forces an unexpected branch in a program — but through language instead of code.
30. AI Modernization
Using AI to help understand, document, analyze, and refactor legacy applications.
IBM i analogy: An acceleration layer for modernization — not a replacement for architecture, testing, or experienced people.
Summary
You do not need to be an AI researcher to speak confidently about AI. You need a clear vocabulary and a few strong mental models. For IBM i teams, that is exactly what this glossary is for: a way to join the conversation, evaluate new tools, and identify which AI capabilities are genuinely relevant to your systems.
With these 30 terms, you have the foundation to follow AI discussions, push back on hype, and spot realistic opportunities for productivity, modernization, and knowledge capture inside your IBM i environment.