The AI Glossary Every Mid-Career Professional Needs to Survive Board Meetings (Without Pretending to Understand Prompt Engineering)

The Problem You're Actually Facing

You're in a meeting. Someone mentions their model's "hallucination problem." Someone else talks about "fine-tuning vs. RAG." A director casually references "transformer architecture." You nod. You have no idea what they said. And you're about to walk into an interview where getting this wrong costs you the offer.

You're not dumb. The jargon is just noise masquerading as expertise. Let's fix that.

Why This Matters Right Now

Hiring loops at mid-senior levels aren't just assessing your technical chops anymore. They're assessing whether you understand the business problem AI is solving. That gap—between panic and actual knowledge—is where you lose credibility. You don't need to be an ML engineer. You need to sound like someone who actually works in a world where AI is already here.

The Essential 12 Terms (In Plain English)

1. LLM (Large Language Model)

The thing doing the talking. GPT-4, Claude, Gemini—these are all LLMs. They're AI systems trained on massive amounts of text data to predict the next word, then the next, then the next. That's it. Not magic. Pattern matching at scale.

Where you'd use this: "Our team evaluated three LLMs for the customer support automation project." Boom. Credible.

2. Hallucination

When an AI confidently makes stuff up. It generates plausible-sounding but completely false information. A model telling you that your company was founded in 1842 when it was founded in 2015. This is the #1 reason AI isn't trusted in regulated industries yet.

Red flag if you ignore it: Hiring managers know hallucinations exist. If you pretend they don't, you sound out of touch.

3. Prompt Engineering

Fancier term for "how you ask the question." Changing "Write a summary" to "Write a one-paragraph executive summary for a C-suite audience familiar with SaaS metrics" gives you different (usually better) results. It's not witchcraft. It's specificity.

You don't need to master this. You need to know it exists and why it matters for consistency.

4. Fine-tuning

Taking a pre-trained model and training it further on your specific data. Think of it as "teaching" the model your company's way of doing things. Expensive but powerful. RAG (below) is usually the smarter choice for most companies.

5. RAG (Retrieval-Augmented Generation)

This is the workhorse of modern AI implementations. The system retrieves relevant documents from your knowledge base, then generates answers based on that retrieval. It's how ChatGPT can talk about your internal wiki without being trained on it. Most competent AI deployments use RAG.

6. Token

A unit of text the model processes. Roughly 1 token ≈ 4 characters. When you hear "token limit," it means "how much text can we feed this model at once." GPT-4 has a 128K token context window (roughly a novel). Matters for cost and accuracy with large documents.

7. Embedding

Converting text into numbers (vectors) that a machine can understand and compare. "Embedding" your product documentation lets an AI search system find the most similar help article without using traditional keyword matching. It's semantic matching. Way better than Ctrl+F.

8. Transformer

The neural network architecture underlying most modern AI. Released in 2017. Changed everything. You don't need to know how it works internally (that's what PhD students are for). Know that "transformer-based" = modern, reliable, probably good.

9. Training Data

What you fed the model to make it smart. GPT-4 was trained on publicly available internet text (books, articles, websites) up to a cutoff date. Your proprietary data isn't in there. This is why fine-tuning or RAG exists.

Critical context for interviews: When someone says "our model is trained on..." they're telling you their knowledge ceiling. Older training data = staler knowledge.

10. Inference

When the model actually generates output. "Training" is learning. "Inference" is doing. Inference is slower and cheaper per-unit than training, which is why you can use these models via API affordably.

11. Temperature

A setting that controls how "creative" vs. "consistent" the AI is. Low temperature (0.1-0.3) = predictable, safe, good for customer service. High temperature (0.7-1.0) = creative, risky, good for brainstorming. Think of it as a dial.

12. Latency

How long it takes the AI to generate an answer. Critical for real-time applications (chatbots, autocomplete). High latency = slow AI = frustrated users. Why some companies use smaller, faster models instead of bigger ones.

What You Actually Need to Demonstrate

You don't need to pass an ML certification. You need to show that you:

1. Understand the tradeoff. Bigger model = smarter but slower. Smaller model = faster but dumber. This applies to nearly every AI decision.
2. Know hallucinations exist and matter. It's not a cute quirk. It's a business risk.
3. Can spot RAG vs. fine-tuning conversations. RAG = safer, faster to implement. Fine-tuning = more expensive, more powerful. Context matters.
4. Won't pretend when you don't know. The moment you BS about transformer attention mechanisms is the moment you lose the room.

The Move Before Your Next Interview

Pull up the job description. Spot every AI/ML term they mention. Spend 20 minutes on each one. Not to memorize it. To understand why it's in the JD. What problem are they solving?

Then in the interview, when they ask about your experience with AI, don't say "I haven't worked directly with models, but I understand the landscape." Say: "I've evaluated RAG approaches for document retrieval on three projects. Fine-tuning was overkill for our use case and cost. RAG got us 88% relevance within two weeks."

Specific. Credible. You.

That's not pretending. That's the whole game.