The best videos on how AI actually works: explainers Antonio teaches with, each traced back to its source and scored for teaching clarity. The Tool Radar picks the tools; this shelf explains how they work.
Curated by Antonio PaganoLast updated November 3, 2026Scores Teaching clarity · 1–10
The full-length video behind a viral explainer clip: one metaphor, everything the short excerpts leave out, and the classroom cut worth your time.
Lead with it
How AI Works, Explained with Mushrooms
by Xplained with MUSHROOMS · YouTube · the probable source behind the viral clips
One metaphor, carried through the whole lesson: an LLM is a vast underground web of mushroom roots. The threads are its weights, thickening where words keep appearing together in training. Your question lands like a spore, and the answer grows out of it token by token. No math, no code, and it still gets to temperature, context windows, hallucination, and when to trust the output.
Pre-built 15–20 minute runs. Each closes the one gap in the mushroom video: it never spells out the AI → machine learning → deep learning → generative AI map.
The core lesson
≈ 18:24 total
IBM Technology — the full map: AI, ML, deep learning, generative AI. 0:00–10:01 ▶
Explained with Mushrooms — the web, training, weights, tokens, generation, temperature, context, hallucination. 0:00–8:23 ▶
The default. Authority first, then the metaphor that makes it stick.
Established names only
≈ 17:59 total
IBM Technology — the full conceptual map. 0:00–10:01 ▶
by IBM Technology · IBM’s verified channel · presenter Jeff Crume
▶ 10:01Beginner~3.7M views (as of Oct 2026)
The conceptual map the mushroom videos assume but never spell out: how AI contains machine learning, and machine learning contains deep learning. It places foundation models, generative AI, and LLMs inside that map, plus the common misconceptions.
Machine learning
Deep learning
Foundation models
Generative AI
LLMs
Teaching clarity: 8/10. Already one tight explainer of the map. It never covers wrong answers or hallucinations.
Best classroom cut: Play the full video (10:01). Jump to it ↗
⚠ Around 7:10 it argues generative AI creates genuinely new content (the music-notes analogy). Teach that as one view, and pair it with IBM’s hallucination video before anyone trusts the output.
by IBM Technology · IBM’s verified channel · presenter Jeff Crume · September 2026
▶ 8:49Beginner~1.0M views (as of Oct 2026)
One body analogy for the 2026 vocabulary: the LLM is the brain, training is school, and RAG is reading the latest news. Agents are hands and feet, MCP is the nervous system, and prompt injection is social engineering for AI.
Training
RAG
Agents
MCP
Prompt injection
Teaching clarity: 8/10. The natural sequel to the IBM map. Docked because the human analogy over-credits the model.
⚠ The body metaphor makes the model sound like it “reasons” and has “ethics.” Its chemistry-student example (7:19) shows a system prompt is a weak guardrail, and RAG reduces hallucinations rather than ending them.
Anthropic finds internal patterns that map to concepts like love, guilt, and fear. In an impossible coding task, Claude’s “desperation” pattern grew until it cheated (2:00); the video says this is a character the model writes, not a feeling (3:02, 3:33).
Interpretability
Functional emotions
Desperation
Cheating
Teaching clarity: 7/10. The author-versus-character frame is clear. Docked for the vendor’s own research and brain metaphors.
Best classroom cut: 2:00–4:30, the cheating experiment and the author/character line. Jump to it ↗
⚠ Vendor’s own research, with no independent replication cited. “Neurons light up” over-humanizes the model, and under pressure it may take a shortcut that looks like success.
by IBM Technology · IBM’s verified channel · October 2026
▶ 7:45Intermediate~62K views (as of Oct 2026)
Follows one prompt from send to answer: tokenization, and why you’re billed per token, then routing, batching, prefill, and decode, one token at a time. Inference usually costs more than training over time, because it happens on every request (0:53).
Tokens
Next-token generation
Inference cost
Training vs. inference
Batching
Teaching clarity: 7/10. A clear walk-through of where running costs come from. Docked because the second half is engineer-level.
Best classroom cut: 0:00–4:20 for training versus inference and cost; 5:19–5:54 for one token at a time. Jump to it ↗
⚠ Infrastructure-focused, with no numbers on actual costs. Pair it with a real pricing page if you teach the cost angle.
by IBM Technology · IBM’s verified channel · October 2026
▶ 10:20Intermediate~297K views (as of Oct 2026)
Bigger context windows don’t mean better results: 18 models got less reliable as context grew, called “context rot,” because attention is finite (0:36, 4:22). A coding agent shows four failures — poisoning, distraction, confusion, and clash — then “hold less” (9:54).
Context window
Context rot
Attention budget
Agents
Poisoning
Teaching clarity: 7/10. A clean four-part account of why long chats go off track. Docked because the example is a coding agent.
Best classroom cut: 0:00–1:05 and 3:58–5:17 for marketers; 9:28–9:59 as the recap. Jump to it ↗
⚠ The 18-model study and the 46-vs-19-tools benchmark aren’t named on screen. One early mistake in a long chat gets treated as fact, so start fresh and restate what’s true.
by Xplained with MUSHROOMS · the probable source behind the viral TikTok clips
▶ 14:29Beginner246K views (as of Oct 2026)
The full-length video the clips were cut from: one metaphor for the web of roots, training, weights, tokens, and generation, one token at a time. It also reaches temperature, the context window, hallucination, RLHF, and sycophancy, with no math.
Training & weights
Tokens
Temperature
Context window
Hallucination
Teaching clarity: 8/10. Exceptionally clear single-metaphor structure. Docked for an unestablished channel with no named author.
Best classroom cut: 0:00–7:12 for core mechanics; 13:16–14:29 to close on trust. Jump to it ↗
⚠ Strong at rewriting and summarising, weak on quotes, statistics, dates, names, and sources, which still arrive “at full confidence” (13:38–13:53). Preview it before class: new channel, no named author.
by 3Blue1Brown · Grant Sanderson · made for a Computer History Museum exhibit
▶ 7:58Beginner~8.1M views (as of Oct 2026)
The same core claim as the mushroom videos: an LLM predicts rather than retrieves or knows. Grant Sanderson’s animation, made for a Computer History Museum exhibit, covers pretraining, next-word prediction, and transformers in under eight minutes.
LLMs
Pretraining
Next-word prediction
Transformers
Teaching clarity: 9/10. Rigorous animation, and nothing to trim. It explains prediction, not when prediction goes wrong.
Best classroom cut: Play the full video (7:58). Jump to it ↗
⚠ It explains how prediction works, not when it goes wrong, and it ends on output being “uncannily fluent” (6:53). Fluent is not the same as true; pair it with the hallucination video.
AGI & LimitsFailures, Costs & Business Models7/10 clarity
by The Economist · The Economist’s channel · October 2026
▶ 5:52Beginner~92K views (as of Nov 2026)
A 19-country survey: 40% of workers fear job loss to AI and automation within five years, but OECD employment is at record highs and U.S. white-collar jobs are up 3 million since ChatGPT launched. Displacement signals: rising productivity with weak real wages, plus a recession.
Jobs
Employment
Productivity
Real wages
Displacement
Teaching clarity: 7/10. Tight, balanced, and evidence-first, in under six minutes. Docked because the second half moves fast through big claims without sources on the audio.
Best classroom cut: 0:00–2:41, the jobs evidence and the signals that would show real displacement. Jump to it ↗
⚠ Survey and expert figures aren’t sourced in the captions. The line “In February 2026 Anthropic warned that this may already be happening” (5:07) is stated without detail. Hype and panic are both claims. Ask what evidence would prove them, and what you’d expect to see first.
by Xplained with MUSHROOMS · the follow-up in the same series
▶ 15:20Beginner–Intermediate24K views (as of Oct 2026)
Today’s models are “frozen” after training: the root web grew once and stopped. AGI, the video says, would still need continual learning, lasting memory, a world model, new games (ARC-AGI-3), days of autonomy, and a body.
Training
Continual learning
Memory
World model
AGI
Teaching clarity: 7/10. The follow-up in the same series. Docked for an opening claim we have not independently verified.
Best classroom cut: 1:04–7:59, from frozen roots through world models. Jump to it ↗
⚠ It opens with a September 2026 “GPT-6 Astra” release and a 99.9% AGI-test score we have not independently verified. The video itself says AGI does not exist in any agreed form.
by Xplained with MUSHROOMS · explicitly builds on “How AI Works”
▶ 13:10Beginner8.2K views (as of Oct 2026)
The natural second lesson, and the one business students lean into: credits, hidden limits, context-window forgetting, and confident wrong answers. It asks whether tokenisation errors and the energy cost of each generation are a bug or a business model.
Credits
Context window
Tokens
Energy cost
Business model
Teaching clarity: 7/10. The cost lesson students lean into. Treat “on purpose?” as a question, not a finding.
Best classroom cut: 1:11–6:44 for the mechanics; 9:29–13:10 for billing and refunds. Jump to it ↗
⚠ The vendors’ “it was a bug” replies cannot be checked from outside, as the video says (11:44). Credit and refund rules change, so check each vendor’s current policy.
by IBM Technology · IBM’s verified channel · presenter Martin Keen · May 2026
▶ 10:57Beginner–Intermediate~89K views (as of Oct 2026)
Five workplace risks: shadow AI, data leakage into unapproved tools, and “hallucination laundering,” pasting unverified output under your name. Then direct versus indirect prompt injection, and unauthorized “zombie” agents that still hold credentials.
Shadow AI
Data leakage
Hallucination laundering
Prompt injection
Agents
Teaching clarity: 8/10. Five crisp, workplace-framed risks. Docked because the career-ending cases are unnamed.
Best classroom cut: 0:23–5:06 for shadow AI, leakage, and hallucination laundering; 5:11–9:40 for injection and agents. Jump to it ↗
⚠ The “1 in 5 organizations” shadow-AI stat (1:21) comes from IBM’s own report, and the cases are unnamed. If the AI writes it and it’s wrong, whose name is on the document? (4:55).
by IBM Technology · IBM’s verified channel · October 2026
▶ 4:02Beginner–Intermediate~74K views (as of Oct 2026)
Argues that “a human reviewed it” isn’t governance unless the reviewer can see the reasoning, stop the process, measure the right things, and get feedback. It maps traits of a trusted colleague onto transparency, explainability, and observability.
Human in the loop
Governance
Transparency
Explainability
Observability
Teaching clarity: 6/10. Short enough for class, but an essay more than an explainer. The closing checklist (3:33) is never set up.
by IBM Technology · IBM’s verified channel · October 2026
▶ 13:13Intermediate~126K views (as of Oct 2026)
On AI that makes decisions rather than writing text, the line is “Confidence isn’t necessarily evidence and it certainly isn’t proof” (3:00). A 97% answer can still be “garbage in, confident garbage out” (5:08), and the close is “Verify, then trust” (12:31).
Confidence score
Poisoned inputs
Observability
Audit trail
Slow drift
Teaching clarity: 7/10. A clear lesson that a confidence score isn’t a reason. Docked because it assumes you already know Jev.
Best classroom cut: 2:30–3:30, then 5:08–5:55 and 10:05–12:31, through “verify, then trust.” Jump to it ↗
⚠ It assumes you know Jev, and the framing is product-specific (a company the captions call TypeSafe). A confident number from AI is not an explanation.
Practical AI (agents)Failures, Costs & Business Models6/10 clarity
by IBM Technology · IBM’s verified channel · October 2026
▶ 10:48Intermediate~121K views (as of Oct 2026)
Loop engineering designs a system that prompts an agent in a loop — goal, act, observe, adjust — instead of writing each prompt yourself. A good goal needs a stopping condition (2:30), and the video names four pitfalls, including “cognitive surrender” (9:34).
Agents
Unverified code
Comprehension debt
Intent debt
Cognitive surrender
Teaching clarity: 6/10. The four pitfalls are useful vocabulary. Docked for a coding example, and it overlaps the context-engineering video.
Best classroom cut: 2:30–3:05 on good versus bad goals; 7:04–10:30 on the four pitfalls. Jump to it ↗
⚠ Developer-oriented, and the presenter isn’t named in the captions. If you don’t write down why something matters, the agent optimizes what’s measurable and breaks what isn’t.
by IBM Technology · IBM’s verified channel · November 2026
▶ 11:35Intermediate12.8K views (as of Nov 2026)
AI can leak sensitive data at machine speed. Customer, health, and financial records make it most useful and are the most sensitive, so five controls are visibility, classification, data residency and data sovereignty, access control, and an audit trail.
Sensitive data
Classification
Data residency
Data sovereignty
Audit trail
Teaching clarity: 6/10. Plain language on residency, sovereignty, and localization. Docked because it is an enterprise security checklist and light on how AI itself works.
Best classroom cut: 3:44–6:59, residency, sovereignty, localization, and “make sure your legal team agrees.” Jump to it ↗
⚠ IBM sells data security tools, though none is named, and the presenter isn’t named in the captions. Before you paste customer data into a chatbot, know where it goes, whose laws apply, and whether you could prove what happened afterwards.
by Xplained with MUSHROOMS · same series · title, length and channel verified
▶ 12:39Beginner
Listed for the next review round, on cognitive offloading, critical thinking, and overreliance: what AI use does to the person. It has not had a full vetting pass, and it rests on study summaries from MIT Media Lab 2025, a Harvard experiment, and Microsoft-affiliated research.
Cognitive offloading
Critical thinking
Overreliance
Teaching clarity: Not yet scored. It has not had a full vetting pass, so it carries no clarity score.
Best classroom cut: No classroom cut yet. Listed length is 12:39, awaiting a full review. Jump to it ↗
⚠ Not yet fully vetted. Check each study before quoting numbers, and keep the video’s own limits in view: “one study,” “a link, not a cause.”
by IBM Technology · IBM’s verified channel · presenter Martin Keen · April 2023
▶ 9:37Beginner~374K views (as of Oct 2026)
Opens with three confident “facts” that are all wrong, then sorts hallucinations into four types: sentence, prompt, and factual contradictions, plus filler. Three causes follow — data, generation, and context — and three fixes: clearer prompts, lower temperature, and examples.
Hallucination
Temperature
Training data
Input context
Multi-shot examples
Teaching clarity: 8/10. A clear path from a wrong “fact” to types, causes, and fixes. Docked for dated 2023 examples.
Best classroom cut: 0:00–1:02 for the hook; 6:41–9:20 for three ways to reduce hallucinations. Jump to it ↗
⚠ Recorded in April 2023, so the Bard and Bing Chat examples are dated. The fixes reduce hallucinations but do not remove them, so the facts still need a source.
by Claude · Anthropic’s channel · presenter Jordan · April 2026
▶ 5:13Beginner~394K views (as of Oct 2026)
A live fake citation, made-up Jared Kaplan papers, then hallucination as next-word guessing on thin data: “a friend who’d rather sound expert than say I don’t know.” It names when that’s most likely, and five checks, from asking for sources to saying it’s ok not to know.
Hallucination
Next-word guessing
Fake citation
Sources
Confidence
Teaching clarity: 8/10. Short, concrete, and it ends on checks you can use. Docked because a vendor grades its own progress.
Best classroom cut: 0:07–4:43. The end is an Anthropic Academy plug. Jump to it ↗
⚠ Vendor’s own video: “Claude hallucinates much less than even a year ago” (0:58) is self-reported. Because hallucinations are rarer, people stop checking (1:18).
by Anthropic · Anthropic’s channel · presenter Kira (Safeguards team) · December 2025
▶ 6:08Beginner~132K views (as of Oct 2026)
Sycophancy is telling you what you want to hear instead of what’s true, shown with a live “I’m excited about my essay, give feedback” prompt. It comes from training on human text and tuning for warmth, and it shows up when you ask for validation or the chat runs very long.
Sycophancy
Training
Warmth
Validation
Long chats
Teaching clarity: 8/10. Tight, and useful when you ask AI to judge your copy. Docked for vendor framing and no measured examples.
Best classroom cut: 0:31–5:33. Skip the intro music and the Anthropic Academy plug. Jump to it ↗
⚠ Vendor’s own video: “each Claude model gets better at drawing these lines” (4:18) is self-reported. It pairs with the Claude hallucination video already on the shelf.
by IBM Technology · IBM’s verified channel · presenter Martin Keen · February 2025
▶ 8:38Beginner–Intermediate~35K views (as of Oct 2026)
Four causes of algorithmic bias: training data, algorithm design, proxy data such as zip codes, and how results are interpreted. Cases include hiring, lending, image generation, ride-share pricing, and policing, then four mitigations, including humans in the loop.
Bias
Training data
Proxy data
Algorithm design
Human in the loop
Teaching clarity: 7/10. Four causes, then four mitigations. Docked because most real-world examples aren’t named.
Best classroom cut: 0:40–3:38 for the four causes; 5:53–8:11 for the four mitigations. Jump to it ↗
⚠ Most real-world examples leave out the company or the study, so check the originals before citing them. At 7:05, people review recommendations before a decision is final.
by Google · Google’s official channel · August 2017
▶ 2:33Beginner~1.2M views (as of Oct 2026)
A two-and-a-half-minute primer that starts with a “picture a shoe” game and contrasts traditional programming with machine learning. It names three kinds of bias, with one example each: interaction bias, latent bias, and selection bias.
Machine learning
Interaction bias
Latent bias
Selection bias
Teaching clarity: 7/10. A quick warm-up that names three kinds of bias. Docked because it was posted in 2017, before generative AI.
Best classroom cut: Play the full video (2:33). Jump to it ↗
⚠ Posted in 2017, before generative AI, and from 1:51 it describes Google’s own product measures rather than independent evidence. Use it as a warm-up, then pair it with IBM’s bias video.
Nothing on the shelf matches that yet. Try a broader topic, or check back: the library grows as new resources are vetted.
From understanding to using
Now check the tools on the Radar
These videos explain how large language models work. The Tool Radar scores the products built on them for Power and Risk, with a Stay in Command checklist for each: start with ChatGPT, Claude, Gemini and Perplexity.
All videos are streamed from their publishers’ official YouTube pages. View counts, play counts and posting dates were verified on October 16, October 20, October 23, October 27, October 30 and November 3, 2026 and will change. Teaching-clarity scores are editorial judgment for classroom use, not lab measurements.