25 terms in plain English, a 15-question self-test, and free resources that cost you nothing but time. No hype, no paid courses, no guru talk.
Last verified: September 2026 · This page is re-checked every month — model names, free tiers, and prices change fast. If something below looks stale, treat the date as the warning it is.
AI literacy is becoming like English literacy: assumed, not impressive. Job ads already list “AI tools” as a requirement, and interviewers increasingly test whether you can direct AI, not just chat with it. The durable part isn’t memorizing this month’s tool names — it’s understanding the layers underneath: what a model is, what a token costs, why an agent needs guardrails. Tools change every quarter; the layers don’t.
AI literacy does not fix the degree-and-visa math on the rest of this site — a ₹60L loan is still a ₹60L loan. No certificate on this page guarantees a job, and “AI-powered” on a resume without a project to show is noise recruiters ignore. Never pay thousands for an “AI expert” course: everything worth learning at the start is free, and it’s listed below. If someone’s business model is selling you AI urgency, their advice about AI isn’t neutral.
Each term: what it actually means, then why a student should care. Read them in order — they build on each other.
Plain English
Software that does tasks which used to need human judgment — recognizing speech, writing text, predicting what comes next. It’s an umbrella word, not one technology.
Why you care: “AI” is also a marketing label slapped on everything from spell-check to robots. Employers pay for specific skills, not the word — learn to ask which AI, doing what.
Plain English
AI that creates new content — text, images, code, video — instead of just sorting or predicting (like spam filters or price forecasts).
Why you care: This is the part automating entry-level writing, coding, and design work. The career move isn’t competing with it; it’s learning to direct it.
Plain English
The engine behind ChatGPT, Claude, and Gemini. Trained on enormous amounts of text, it writes by predicting the most likely next word, over and over — which comes out looking like understanding.
Why you care: Nearly every “AI tool” you’ll meet at work is an LLM wearing a costume. Know the engine and you can see through the costume.
Plain English
The trained AI itself — a big file of numbers — before it’s wrapped in an app. ChatGPT is an app; the thing answering inside it is the model.
Why you care: News and job ads talk about models (new releases, benchmarks, “we use the latest model”). Knowing the layer the claim sits on keeps you from being impressed by packaging.
Plain English
The small chunk of text a model actually reads — roughly ¾ of a word in English. “Study abroad” is about three tokens. It’s also the unit AI companies bill by.
Why you care: Two practical reasons: long chats and big documents cost real money on paid APIs, and token limits are why an AI “forgets” the start of a long document.
Plain English
The instruction you give the model. “Write an essay” is a prompt; so is a three-paragraph brief with examples, audience, and format.
Why you care: The highest-leverage skill in the whole list. Vague prompt, vague output; specific prompt, useful output. This is the skill employers now put in job descriptions.
Plain English
How much text the model can “see” at once — its working memory. Measured in tokens; bigger windows fit whole documents.
Why you care: A big window lets you paste an entire SOP or contract in one go. But long contexts cost more, and models still miss details buried in the middle — size isn’t the same as attention.
Plain English
When the model states something false with complete confidence — invented citations, fake deadlines, wrong facts. It’s not lying; it’s predicting words without checking them.
Why you care: Never paste AI output into an SOP, application, or research summary without verifying. Professors and recruiters check — and a fake citation is worse than no citation.
Plain English
A way for software to talk to other software automatically. An AI API lets your code — or an app like n8n — send text to a model and get an answer back, with no chat window involved.
Why you care: Most student “AI projects” are really just calling someone else’s API, which is fine — but know what you actually built before you describe it in an interview.
Plain English
Models that handle more than text: images, audio, video, documents. You can show it a photo of a form, or talk to it instead of typing.
Why you care: Assignments and internships increasingly expect voice notes, screenshots, and PDFs as inputs — “chat” is only one mode now.
Plain English
The current most-capable models from the top labs — the ones every benchmark compares against. The frontier moves every few months.
Why you care: Benchmark headlines (“beats X at Y”) are about these. They’re usually paid or rate-limited; the free chatbot you use is often a smaller, cheaper sibling.
Plain English
A model whose trained weights — the actual “brain file” — are downloadable, so anyone can run, study, or modify it. Llama, Kimi, and NVIDIA’s Nemotron families work this way.
Why you care: No per-use cost, and it can run on your own machine (private — your data never leaves). The catch: the big ones need serious GPU hardware you probably don’t own.
Plain English
The opposite: usable only through the company’s app or API — ChatGPT, Claude, Gemini. You never see the weights.
Why you care: Usually the strongest and the most convenient, but your inputs go to their servers and the pricing can change under you. Don’t build your whole workflow on a free tier you don’t control.
Plain English
The GitHub of AI: a public hub where people publish, share, and download models and datasets — most of them free.
Why you care: The fastest free way to try real models, and a portfolio of Hugging Face projects is something employers can actually click on. Its free courses are in the resources below.
Plain English
NVIDIA’s software layer that lets AI code run fast on NVIDIA graphics chips (GPUs). It’s the reason “I need a GPU for AI” almost always means “I need an NVIDIA GPU.”
Why you care: This one piece of software is a big part of why NVIDIA dominates AI hardware — and why AI computing is expensive. You don’t need to learn it; you need to know why it’s mentioned everywhere.
Plain English
Anthropic’s agentic coding tool: it reads your project, edits files, runs tests, and carries out multi-step coding tasks from a plain-English request — not just autocomplete.
Why you care: This is what “AI coding” looks like in real jobs now: supervising an agent that does the work, reviewing what it changed. “Can you code?” is becoming “can you direct coding?”
Plain English
Folders of instructions, scripts, and resources that Claude loads on demand — so it follows your workflow without you re-explaining it every time. Anthropic ships ready-made ones for Word, PDF, PowerPoint, and Excel documents, and you can write your own custom skills.
Why you care: This is how “AI that knows your process” actually works — reusable playbooks instead of repeated prompting. Honest caveat: Skills need a paid Claude plan (Pro, Max, Team, or Enterprise) — not free, as of September 2026. That’s why it’s in the glossary, not the free list below.
Plain English
Google’s free playground for trying Gemini models in a browser — including a free API key so you can build small AI projects without paying.
Why you care: The cheapest way to put a real AI project on your resume. Honest caveat: don’t paste private data into it — free-tier inputs may be used to improve Google’s models.
Plain English
AI that doesn’t just answer — it works toward a goal, choosing its own steps and using tools (web search, code, apps) along the way. A chatbot answers questions; an agent books the tickets.
Why you care: This is the shift employers are hiring around: from “chatbot that replies” to “system that does the task.” New roles are about supervising these systems, not typing into them.
Plain English
An agent that runs with less human supervision — longer leashes, fewer checkpoints. Note the word “less,” not “none.”
Why you care: “Autonomous” describes the supervision level, not reliability or consciousness. Fewer checkpoints means mistakes travel further — never give one unsupervised access to money, email, or accounts.
Plain English
The limits placed on an AI system: what it’s allowed to do, say, or touch. Content filters, spending caps, and “ask a human before sending” rules are all guardrails.
Why you care: Every serious company deployment has them, and designing them is a real, hireable skill — part engineering, part judgment. If someone demos an agent with no guardrails, that’s the part to ask about.
Plain English
Connecting apps so work happens without you: a new form response appears in a spreadsheet, which triggers a confirmation message — no human in the middle.
Why you care: This is the non-programmer’s superpower. Internships, campus clubs, and small businesses all have repetitive work that automation deletes — and “I automated X” is a resume line that gets interviews.
Plain English
The two automation tools students actually use. Both are visual — you drag boxes, no code. n8n is fair-code licensed and can be self-hosted free; Make.com is cloud-based with a friendlier free tier (about 1,000 operations/month as of 2026).
Why you care: Pick one, build one real automation (say, a form-to-spreadsheet-to-email flow), and you’ve learned the concept employers mean by “AI automation” — most of which is this, not robots.
Plain English
A way to design agents: each step is a node, the routes between steps are edges, and loops repeat a reason→act→observe cycle until the job is done. LangGraph is one popular toolkit for this.
Why you care: This is the engineering pattern behind agents that actually work reliably instead of wandering off. Note: a graph here is an orchestration pattern, not an AI model — don’t let the jargon impress you.
Plain English
An open-source personal AI agent (created by Peter Steinberger) that runs on your own machine and can act through your files, apps, messaging, email, and browser — your own assistant rather than a chatbot.
Why you care: It shows where agents are heading: personal, local, acting on your behalf. The same power is the risk — 2026 coverage (Reuters, Scientific American) stressed real security trade-offs if it’s misconfigured. Power without guardrails (see #21) is the whole story.
Kimi K3 — Moonshot AI’s open-weight model; TechCrunch (Sept 11, 2026) noted its breakout success. What it proves: the frontier isn’t only American, and strong open weights mean students anywhere can run serious models.
Meta Muse Spark — Meta’s model family; Reuters covered Spark 1.1 (July 9, 2026) and Muse Code running on Spark 1.2 (Aug 5, 2026). A 1.3 release was reported in early September 2026 — treat version claims as “reportedly” until confirmed. The lesson: models now ship inside apps you already use; AI literacy includes tools you didn’t choose.
Nemotron — NVIDIA’s family of open models. And a warning: there is no credible model called “Memtron.” If you hear that name, it’s almost certainly a misheard “Nemotron.” Verify a model name before you cite it — this field invents names weekly.
These entries age fast — that’s exactly why this page carries a monthly re-verification date.
One question at a time, with the explanation the moment you answer. Your answers never leave this page — no account, no tracking, works offline.
Every link below was checked in September 2026 and costs nothing. Each carries a realistic effort note — “free” still costs your evenings.
106 short courses from Andrew Ng’s team (Sept 2026), most 1–2 hours; course pages show “Enroll for Free.” Start with ChatGPT Prompt Engineering for Developers.
Effort: 1–2 hours each. Some courses assume basic Python — check prerequisites.
Free courses from the team behind the model hub — including an AI Agents course (smolagents, LangGraph) and a Model Context Protocol course built with Anthropic. Described on their own pages as “completely free.”
Effort: a few evenings. The hands-on courses assume Python.
Beginner module, no prerequisites: basic AI terms and prompting best practices, straight from Microsoft’s free training library.
Effort: about an hour.
Google’s free Gemini playground — try models in the browser and grab a free API key to build small projects for your resume.
Don’t paste private data into it: free-tier inputs may be used to improve Google’s models.
The official docs and templates for the workflow-automation tool — the fastest free route to putting “I automated X” on your resume.
n8n’s free path is self-hosting. If you don’t want servers, Make.com’s free cloud tier (about 1,000 operations/month as of 2026) is friendlier.
Anthropic Skills (term #17 above): reusable instruction folders — including ready-made Word, PDF, PowerPoint, and Excel skills — so Claude follows your workflow without re-prompting. Needs a paid Claude plan (Pro, Max, Team, or Enterprise), so it’s not in the free list. Official reference: github.com/anthropics/skills.
Google AI Essentials (Coursera): several 2026 guides list it as free-to-audit, but 2024 reporting said the audit option wasn’t offered — genuinely disputed. The certificate costs about $49/month. Check the enrollment page yourself before counting it as free; everything it teaches is also covered free above.
Microsoft AI Skills Fest: ran June 8–12, 2026 — already over. The free Microsoft Learn modules above stay available year-round.
You don’t need your own graphics card to use AI seriously. These companies run the models on their hardware and hand you a free API key. The key drops into any OpenAI-compatible tool — the OpenAI Python library, LangChain, n8n’s HTTP node, or a one-page script.
Free tiers below are as of September 2026. They are rate-limited and providers change them often — the number in your provider’s console outranks this page.
Google’s free Gemini playground — generous per-model quotas, no card, and a key you can use in code, not just the browser.
How to use: sign in at aistudio.google.com/apikey → “Create API key” → paste the key wherever an OpenAI-compatible tool asks.
Free-tier inputs may be used to train Google’s models. Don’t paste private data, coursework you can’t share, or secrets.
Specialized AI chips built for one thing: answering fast. The free tier (no card) gives roughly 30 requests a minute and about 14,400 a day — plenty for learning and small projects.
How to use: sign up at console.groq.com → API Keys → create a key. OpenAI-compatible at https://api.groq.com/openai/v1.
One key that reaches 500+ models from many providers. Models whose names end in :free cost nothing: about 50 requests a day and 20 a minute on a free account, no card.
How to use: sign up at openrouter.ai → Keys → create a key, then call any model with the :free suffix. A one-time $10 top-up (paid) permanently raises the free-model limit to 1,000/day.
Another speed-focused chip company. The free tier (no card) is about 1 million tokens a day — the most generous daily budget on this list for bigger tasks.
How to use: sign up at cloud.cerebras.ai → API Keys → create a key. OpenAI-compatible endpoint.
Europe’s leading AI lab. The free “Experiment” plan gives API access to all their models at about 1 request a second — fine for learning, not for apps.
How to use: sign up at console.mistral.ai → choose the Experiment plan → API Keys → create a key.
Requires phone verification, and the Experiment tier uses your data for training by default — you can opt out in the admin console before pasting anything real.
NVIDIA’s catalog of ready-to-run open models (Llama, DeepSeek, Nemotron and more). Free developer signup includes about 1,000 inference credits and roughly 40 requests a minute, no card.
How to use: sign up at build.nvidia.com → generate an API key (starts with nvapi-). OpenAI-compatible at https://integrate.api.nvidia.com/v1.
Try GPT, Llama, Mistral and more side by side with nothing but the GitHub account you probably already have. Free, but tightly rate-limited (roughly 50–150 requests a day depending on the model).
How to use: visit github.com/marketplace/models → generate a personal access token with the models scope → use it as your bearer token against https://models.github.ai/inference.
One Hugging Face token that routes to hundreds of open models hosted by partner providers (Groq, Cerebras, Together and more) — a single key instead of eight.
How to use: create a free account at huggingface.co/join → generate a fine-grained token at hf.co/settings/tokens with “Make calls to Inference Providers” permission.
The free tier covers trying things out, not heavy use. Docs: huggingface.co/docs/inference-providers.
Free tiers are for learning, not for running a business on — rate limits (a few dozen requests a day on some providers) will interrupt anything ambitious, and quotas change without much warning. Free also isn’t private: some free tiers train on your inputs, so treat them like a public whiteboard — no passwords, no API keys of your own, no personal data, no unpublished work you care about.
And the durable skill here isn’t memorizing which provider is free this month — it’s the pattern: one key, one OpenAI-compatible call, and you can swap providers when the deals move. That’s what the glossary terms above (API, token, open-weight) were preparing you for.
If you’re drowning in coursework, don’t add “learn AI” as guilt — the glossary plus one finished project beats ten half-finished courses. Free resources still cost time, so pick one and finish it before opening the next tab.
AI skills are leverage, not a lottery ticket: they multiply a strong degree and real projects, they don’t replace them. And anything on this page with a month attached — model names, free tiers, prices — can be wrong by next quarter. That’s why the date at the top exists. When in doubt, the official docs linked above outrank this page.
Not sure AI is even your lane? Take the 2-minute country filter or the brutal self-assessment first — career direction before career tools.