A VibeAxis glossary for 2025 that skips the cosplay and gets to the point.
Core Ideas
- AI isn’t magic—it’s math with lipstick.
- Learn the words, spot the tricks, and don’t let anyone sell you neutral algorithms or moral fairy dust.
- If you want the teeth, see our field manual on ethics.
Artificial Intelligence
Umbrella term for software that imitates human-ish tasks: patterns, language, and decisions. Less “mind,” more “probability engine.” See also: What Is Conversational AI?
Machine Learning (ML)
How programs train on data instead of being hard-coded. Feed it examples; it learns the pattern—and your blind spots.
Deep Learning
ML with a lot of layers. Great at vision, speech, and language. Also great at hiding how it got there.
Neural Network
Math arranged like a brain diagram for vibes. Layers of “neurons” transform inputs to outputs. For a reality check, Neural Nets vs Old-School Code.
AGI (Artificial General Intelligence)
The hypothetical do-everything mind. Today’s systems are narrow specialists with swagger, not universal geniuses.
Supervised / Unsupervised / Reinforcement Learning
Supervised: labeled examples (“this is spam”).
Unsupervised: find patterns without labels (“these users cluster weirdly”).
Reinforcement: trial-and-reward loops (“win the game; don’t crash”).
Tools & Players (what you actually touch)
ChatGPT / Claude / Gemini
Chatbots that turn prompts into prose. Same sport, different coaches. For vibe checks and gaslight potential, Claude vs ChatGPT and teaching Gemini a personality.
GPT (Generative Pre-trained Transformer)
The engine class under many chatbots. Pretrained on oceans of text, then tuned to follow instructions without crying.
LLM (Large Language Model)
Text-in, text-out machines with long memories and short judgment. Great at drafts and options; pair with facts.
Midjourney / DALL·E
Describe a scene, get an image. Amazing for concepting; murky on training data. See Deepfakes & Digital Consent for the line between art and theft.
Sora (Text-to-Video)
Turns prompts into short video clips. Expect impressive demos, then asterisks the size of your GPU bill.
Text-to-Speech / Voice Cloning
Type words, get a voice. Type your boss’s words, get HR. Powerful, ripe for abuse.
Under the Hood (jargon you’ll hear in pitch decks)
Training Data
Everything the model learned from. Quality > quantity, unless your goal is scalable nonsense.
Model
The trained artifact (all those weights) that does the task. Think “frozen instincts.”
Fine-Tuning
Teaching an existing model a specialty with extra examples. Like sending your intern to night school.
Tokens
Chunks of text the model sees (pieces of words). Billing and context limits ride on these.
Inference
Using the trained model to produce an answer right now. The part you experience—and pay for.
Zero-Shot / Few-Shot Learning
Doing a task with zero or a few examples in the prompt. Magic trick powered by pretraining.
Embeddings
Vectors (coordinates) that represent meaning so machines can search/compare ideas. The backbone of “it found the right doc instantly.”
Parameters
The dials inside the model. More isn’t always smarter, but it sure is pricier.
Ethics & Impact (where the harm hides)
Bias
When model outputs mirror or magnify prejudice in data. If you felt that last loan denial in your bones, read The Algorithm Thinks You’re Poor, Dangerous, or Lying.
Hallucination
Confidently wrong output. The model isn’t lying; it’s guessing prettily. See Hallucination Nation for damage control.
Alignment
Trying to make models behave within human values and rules. Necessary, fragile, and endlessly argued.
Red-Teaming
Stress-testing models with adversarial prompts to find failures before the internet does.
Synthetic Identity / Digital Doppelgänger
AI-fabricated personas or clones of real people. Useful for testing; nightmare fuel for consent. Digital Doppelgängers goes deeper.
AI Governance
Policies, audits, and consequences. If there’s no kill switch or appeals process, it’s theater. Our take: Ethical AI, Without the Halo.
Explainability
Plain-language reasons for decisions. No explanation = no justice.
Turing Test / Singularity
“Can it fool you?” / “Will it outgrow us?” Old lenses. Modern risk is scale without recourse, not moustache-twirling robots. Try The Next 10 Years of AI for reality.
Culture & Media (where the weird gets legs)
Deepfake
Hyper-real synthetic media. Great for film labs; catastrophic for consent and elections. Again: Deepfakes & Digital Consent.
AI Girlfriend/Boyfriend
Chat companions with a script and a subscription. Emotional labor as a service. See AI Relationships & the New Age of Digital Love.
Prompt Injection
Smuggling instructions to hijack a model (often via tools or web content). The “eat at Arby’s” of security bugs.
Algorithmic Feed
Your discovery firehose, tuned for watch time over well-being. If music feels samey, Stuck on Repeat has the receipts.
Work & Biz (where AI meets payroll)
AI Copilot
An assistant inside your apps that drafts, summarizes, or auto-files. Great for options; verify before shipping.
Prompt Engineering
Writing instructions that get what you wanted the first time. Start here: Why Prompt Engineering Is the New Copywriting and Feel Like You Suck at Prompting? Read This First.
Generative AI
Models that produce new text/images/code/music. Amazing at scaffolding; keep humans in the loop for judgment.
LLMOps
Everything it takes to run big models in production—monitoring, guardrails, cost control, incident response. Translation: not “set and forget.”
Quick Filters
What data trained this? Named sources > vibes.
What happens when it’s wrong? Appeal path or purgatory?
Where does my data go? On-device beats “improve our models.”
Is this saving me time—or shaping my behavior? Turn off autoplay and find out.
