Artificial Intelligence

The field is older than most of your parents. In 1950, in his paper Computing Machinery and Intelligence, Alan Turing asked whether machines can think, and instead of a definition he proposed a game: if you can’t tell whether you are chatting with a human or a machine, the argument about “thinking” is pointless. Today we call it the Turing test. In 1956, at Dartmouth, the field got the name artificial intelligence.

Then came the waves. In the 1980s, companies invested in expert systems (hand-written rules); at the end of the decade came the “AI winter”: both the money and the interest disappeared. In 1997, IBM’s Deep Blue beat Kasparov at chess; in 2012, graphics cards turned neural networks into a usable tool (deep learning); and in 2016, AlphaGo beat Lee Sedol at Go, a game machines were supposed to master only “decades from now.”

Two key years for what you have in your phone today: 2017, when Google published the Transformer architecture (Attention Is All You Need), which GPT, Claude, and Gemini are built on, and November 30, 2022, when OpenAI launched ChatGPT. The technology existed before, but only the chat window put it in the hands of hundreds of millions of people, no manual needed.

Part of what you hear today is a bubble, part of it is infrastructure for the next twenty years, like the internet around 2000. Pop culture prepared you for a killer robot, not for an assistant that transcribes your meeting; that’s the expectation you work with in communication.

Why now and not in 2010? Four things came together, and none of them would have been enough on its own. Power: graphics cards (GPUs, mainly from Nvidia) multiply matrices in parallel, and that is exactly what training a model is. Data: the internet provided trillions of words, images, and code from which a model learns patterns. Transformer: the 2017 architecture with its attention mechanism weighs connections across a long text and trains in parallel, so it uses GPUs to the full. Cloud and APIs: the model runs at the maker’s end, and you pay for usage.

The speed of adoption is unprecedented. ChatGPT had a million users in 5 days and, according to an estimate by the bank UBS, 100 million in 2 months. In February 2026, OpenAI reported 900 million weekly active users and 50 million paying ones. Stanford’s AI Index 2026 calculated that generative AI reached 53% of the population within three years, faster than the personal computer or the internet (for a comparison of media, see the Media chapter).

The Czech context is more sober. According to ČSÚ, the Czech Statistical Office, 18% of companies with 10 or more employees used AI in 2025 (11% the year before, EU average 20%); most often generative AI for text and code (13%, 9% in the EU). People are ahead of companies: according to ČSÚ, 32% of Czechs aged 16 and over use generative AI, and 78% of those aged 16 to 24. In Microsoft’s global comparison (first quarter of 2026), the Czech Republic, at 30% of people aged 15 to 64, is above the 18% average but far behind the UAE (70%) and Singapore (63%). So in this classroom you are probably ahead of most of the companies you will go on to work for.

Under the chat window there are four layers. At the bottom, chips: Nvidia GPUs and specialized chips such as Google’s TPUs. Above them, data centers: halls with tens of thousands of chips, cooling, and a power supply; they are built by Microsoft, Google, Amazon, Meta, and OpenAI with partners (Stargate). Above that, cloud and APIs: Anthropic, OpenAI, or Google run the model and you call it. At the top, applications: ChatGPT, Claude, Gemini, or your website that calls the API. A marketer works in the top layer, but the price of a token pays for all four.

On October 29, 2025, Nvidia became the first company with a market value of $5 trillion; it had crossed $4 trillion less than four months earlier. The four largest data center operators invested roughly $410 billion in 2025, and according to their outlooks they are heading toward an estimated $700 billion in 2026, mostly in AI infrastructure.

The second cost is electricity. According to the IEA, data centers consumed 415 TWh in 2024, about 1.5% of the world’s electricity, and since 2017 they have grown 12% a year, four times faster than total consumption. By 2030, the IEA expects that to double to 945 TWh, more than all of Japan consumes today. A single query, meanwhile, is a trifle: Google measured the median text prompt in Gemini at 0.24 Wh (a microwave running for one second, five drops of water), and Sam Altman gives 0.34 Wh for ChatGPT. The problem is billions of queries, and training. Through an API you buy a service, not a technology, just like with a payment gateway (see the Internet chapter).

A large language model (LLM) is a statistical model of language. First, the text is chopped into tokens: pieces of words, about 3.5 characters per token in English, and in Czech, with its diacritics and inflection, more pieces per word. Every token has a number. For every token in its vocabulary, the model computes the probability that it comes next, picks one, adds it to the end, and repeats. Nothing more. That an article, code, or a little poem comes out of it is the result of billions of parameters that encode patterns from the trillions of words the model learned from.

Tell training and inference apart. Training happens once: pretraining on an enormous corpus with a single task, predicting the next word; then fine-tuning and reinforcement learning from human feedback (RLHF), so the model answers like an assistant, not like a random page from the internet. It takes months on thousands of GPUs. Inference is every one of your queries: cheap per call, expensive at mass scale. And above all: after training, the model is frozen; it learns nothing from your conversation unless the app has a separately built memory.

The context window is working memory: how many tokens the model “sees” at once, meaning the prompt, attached documents, and its own answer. Claude and Gemini both offer a window of 1 million tokens. A million tokens is roughly 2,000 pages of text (an estimate). Whatever doesn’t fit into the window, the model doesn’t know.

Temperature controls how random the choice is: zero gives the most probable and most boring text, a higher value gives more surprising but less reliable text. This leads to the most important property for a marketer: hallucination. The model optimizes for plausibility, not truthfulness; a made-up study or a wrong number looks just as confident as a correct answer. That’s why you check every name, number, and link yourself. And Czech pays extra: more tokens per word, a higher price, less text in the window. The vocabulary is below.

At the frontier there is a handful of labs: Anthropic (Claude), OpenAI (GPT), Google DeepMind (Gemini), and xAI (Grok). You can get their models only through an app or the API; the weights (the learned parameters) are secret. The second camp is open-weights models: Meta (Llama), the French Mistral, the Chinese DeepSeek, Qwen, or Kimi, and more recently gpt-oss from OpenAI. You download the weights and run them yourself. Be careful with the term “open source”: what’s open are the weights, not the training data or the method.

Three ways in. Chat apps (Claude, ChatGPT, Gemini) are for people: Claude’s Pro plan costs $20 a month, or $17 a month with annual billing, and Max starts at $100 (claude.com/pricing, October 2026). The API is for automation: you pay for tokens, and the model becomes part of your website or script. Self-hosting an open-weights model gives you control over your data, but you pay for the servers and the people. For a marketing team: apps for people, the API for processes, self-hosting only with a good reason.

Models differ in character more than in leaderboard numbers; the leaderboards change every month, so it’s better to have two or three models at hand and choose by task. Stanford’s AI Index 2026 reports that in March 2026 the gap between the best American and Chinese models was only 2.7%, even though private investment in AI reached $285.9 billion in the US in 2025 and $12.4 billion in China (not counting state money). Europe has Mistral, and regulation.

Text to text is the most widespread use: emails and summaries, meeting notes, brainstorming names, rewriting a specialist text for a different target group, a campaign strategy, analyzing a pasted spreadsheet, or code. What makes the difference between usable and generic output is the prompt.

A good prompt has a structure. Role and context: who the model is and who it works for (“you are the club’s copywriter, writing to parents of kids aged 6 to 10”). Task: one clear sentence. Input material: facts, data, previous texts; the model doesn’t have your know-how until you give it. Output format: length, bullet points, three variants. Examples of what you like (few-shot). Constraints: tone, banned words, what not to make up. And a safeguard: “if you’re missing information, ask.” For more complex tasks, let the model ask you questions first.

The process is always the same: the model writes a draft, you correct the facts and the tone, and you ask for a second iteration. Save recurring context (brand manual, tone of voice, target groups) in a project in Claude or in a custom GPT. The iron rule: no text you haven’t read goes out. More in Anthropic’s guide, which applies to other models too.

Instead of stock photos or hours of a designer’s time, you get ten variants in a minute. Midjourney has the best aesthetics and thousands of styles, GPT Image from OpenAI handles text in images and precise edits, Nano Banana from Google keeps a character consistent, Flux is open-weights, Adobe Firefly is trained on licensed data for cases when a lawyer wants certainty, and Higgsfield puts several models under one roof, including video.

An image prompt has a different grammar: the subject (who or what, doing what), the style (photography, illustration, 3D), composition and lens (close-up, wide-angle, from above), light (golden hour, studio, neon), the aspect ratio for the channel, and optionally a reference image. A generic prompt gives a generic image, like a million others on social media; a precise prompt gives a visual from your campaign. For a brand, consistency is key: one fixed style string at the end of every prompt, so the images look like the work of one hand.

Where AI images don’t belong: in truthful product photos (food packaging at HOPI has to match reality), on people without their consent, and on anything that passes itself off as a document. Licensing: paid plans allow commercial use; read the terms. From August 2, 2026, you must clearly label deepfakes in the EU, meaning realistic-looking images of people, places, or events that aren’t real; the machine-readable watermark is embedded in the output by the generator itself. You do the finishing in ordinary tools: Photopea for edits, remove.bg for backgrounds, Squoosh for compression to WebP, Adobe Express for templates. More in the Production chapter.

Text to video is the youngest field and changes month by month: OpenAI’s Sora app, launched at the end of September 2025, announced its shutdown in March 2026. Veo (Google), Kling, Seedance (ByteDance), Runway, or Pika generate clips from a few seconds to tens of seconds long; they handle camera movement, character consistency, and ever better physics. Clips get assembled into ads, reels, and even whole series; Higgsfield has its own series, generated from start to finish. Video is ceasing to be limited by technology and starting to be limited by imagination. And by the budget for iterations, because the first generation rarely gets it right.

Text to audio: ElevenLabs does realistic voices, voice cloning, and dubbing into dozens of languages; Suno and Udio compose a track including vocals; Google’s NotebookLM turns any document into a podcast with two hosts. In marketing: a voiceover without a studio, localizing a video in a single voice, a podcast from an article, a jingle without licensing fees.

The least flashy and the most useful is speech to text: Whisper from OpenAI (open-weights, runs even on a laptop), transcripts in Teams, automatic subtitles. A transcript of an interview gives you quotes, a transcript of a meeting gives you tasks, a transcript of a podcast gives you an article. Transcription plus a language model is the biggest time saver AI has brought to media work so far. How this meets streaming is in the Streaming chapter.

A chat waits for a prompt and answers. An agent gets a goal and goes to work: the model runs in a loop, plans the steps, calls a tool (reads a file, runs a command, searches the web, calls an API), looks at the result, and decides what to do next, until the task is done. You set the goal and the boundaries, not the steps: from conversation to delegation. How fast this is moving shows in the AI Index 2026: on the OSWorld test (tasks on an ordinary computer), agents’ success rate rose in a single year from about 12% to 66%, and on SWE-bench Verified (fixing bugs in real code) from 60% to nearly 100%.

An agent needs tools. MCP (Model Context Protocol) is an open standard that Anthropic introduced on November 25, 2024: a unified way to connect a model to data and tools, from Google Drive and Gmail to Trello and a company database. Like USB-C: one connector, many devices. In December 2025, Anthropic handed MCP over to the Agentic AI Foundation under the Linux Foundation (together with OpenAI and Block); at that point there were over 10,000 public MCP servers, and ChatGPT, Gemini, and Microsoft Copilot all supported it. For a marketer: you can connect a model to your Trello, spreadsheets, or email without a programmer.

Claude Code (in preview from February 2025, fully launched in May 2025) is an agent in the terminal or a desktop app: it reads an entire project (repository), writes code, runs tests, fixes bugs, and opens a pull request, a proposed change that a human approves or sends back. Codex from OpenAI does the same. n8n and Make are no-code automation tools where the model is one step (an email arrives → the model sorts it → a Trello card is created). The next level up is a multi-agent workflow: an orchestrator divides the work among specialists, and a critic reads the results and sends them back. That’s also how this garden was made: thirteen agents wrote the chapters in parallel and two critics read them, one for facts, the other for clarity.

Velín (Czech for “control room”) is my personal dashboard at jecminek.co/dashboard: news, mentions of HOPI, of judo, and of me, things to read, the weather, the markets, the status of my websites. The most interesting part is the one where AI writes nothing and only sorts. Before, my media monitoring sent me an email about every mention: 68 emails in 26 days, one useful every two to three weeks. I wanted discreet delivery: two or three items in a quiet week, better nothing than noise.

Haiku, the cheapest class of Claude models, acts as the filter: it throws out obvious noise (recipes, contests, namesakes). Sonnet decides according to a rubric: 2 = alert right away (a Grand Slam medal, a HOPI truck accident, the company’s owner in politics), 1 = weekly digest (an independent product test), −1 = noise (regional judo, a PR article about a recipe). I didn’t write the rubric from behind a desk: first I labeled 226 mentions and 134 headlines by hand (117 of them noise), and the labels turned into rules with examples. Every item that gets delivered has a thumbs up and a thumbs down. Haiku isn’t good enough for the decisions (false alarms), so it only does the filtering.

Cost was the second problem. The first version cost $0.89 a day, and 81% of that went to the expensive Opus, which got the full rules again with every query, without caching. Three changes: the rules and the evaluations as two blocks with a one-hour cache, decisions on Sonnet instead of Opus (a local test on my labels came out the same), and a night pause, when from 10 p.m. to 6 a.m. items are only collected and at 6:07 a.m. the night gets sorted. The result: $0.25 a day, 72% less, at the same quality. The daily cap is $1.50, in case something breaks.

The API is billed by tokens, separately for input (what you send) and output (what the model writes); output is five times more expensive because it’s generated token by token. The price range is a hundredfold. At Anthropic, a million input tokens costs $0.10 on Haiku 5.5, $2 on Sonnet 5.5, $4 on Opus 5.5, and $10 on Fable 5.1; output costs $0.50 to $50. OpenAI’s current lineup is almost a mirror image: gpt-6-luna $0.10, gpt-6.1-sol $2, gpt-6-astra $10; older “pro” models cost as much as $30 (gpt-5.5-pro). Google charges $0.30 to $0.75 for Gemini Flash and $2 for Gemini 4 Argon (from December 1, 2026, already $4, and Flash 3.6 and 3.8 double from January 2027). Prices are as of October 2026 and change every few months; the ratios between the classes stay.

Besides choosing the model, there are two ways to pay less. Prompt caching: a repeated beginning of the prompt (rules, brand manual) is stored in a cache, and on the next call it costs 2.5 to 10% of the input price, depending on the model; writing it costs 1.25× for five minutes or 2× for an hour, and it pays off from the second call. Batches: whatever isn’t urgent, you send all at once and get back within 24 hours at half price. Long context: Anthropic charges the whole 1-million-token window at the same rate (except Haiku 5.5 above 100,000), while Google, for its Pro models above 200,000, and OpenAI, above 272,000 tokens, charge double.

An example for the board: 1,000 queries a day, each with 3,000 input tokens and 500 output tokens. On Sonnet 5.5: 3 million × $2 = $6 for input, 0.5 million × $10 = $5 for output, $11 a day. If 2,500 of the 3,000 input tokens are a fixed prefix in the cache ($0.10 per million), the total drops to roughly $6.25. On Haiku 5.5, the same thing costs $0.30 + $0.25 = $0.55 a day. A twentyfold difference, which is why the checklist below starts with the cheapest model. For an individual, the $20-a-month plan is enough; the API is for when something has to run without a human.

The risks aren’t sci-fi. Deepfakes: in early 2024, a finance employee at the firm Arup in Hong Kong joined a video call with the “CFO” and colleagues, all of them generated, and sent 200 million Hong Kong dollars (about $25 million) in 15 transfers. Two things for a marketer: a brand can be misused for fraud (a fake executive, a fake contest), and every viral video needs to be verified (a reverse image search in TinEye, the original source) before you share it.

Information overload: generating is almost free, the amount of average content grows, and attention shrinks. Hallucinations: confident errors in numbers and names; the AI Index 2026 counted 362 documented AI incidents in a year, up by half from the year before. Privacy: what you type into a public chat may be used for training; customer data, contracts, and internal prices belong only in company accounts with training turned off, or nowhere. Jobs: the World Economic Forum estimates that by 2030, 170 million jobs will be created and 92 million will disappear; 41% of employers plan to reduce their workforce because of AI, and routine cognitive work is first in line. Addiction and attention are covered in the Social Media chapter.

Regulation: the EU AI Act has been in force since August 1, 2024, and is phasing in step by step. From February 2, 2025: bans (manipulative techniques, social scoring, emotion recognition at school and at work) and an AI literacy obligation for companies. From August 2, 2025: obligations for providers of large models (GPAI). From August 2, 2026: transparency; a chatbot must make it clear that you are not talking to a human, and anyone who publishes a deepfake must clearly label it. The Digital Omnibus (in effect since July 27, 2026) postponed machine-readable marking of outputs for generators already on the market to December 2, 2026, and the rules for high-risk systems (recruitment, loans) to December 2, 2027. For marketing: label generated content, don’t manipulate emotions, keep a record of where you use AI.

What will remain for people? When anyone can produce text, images, and video in a minute, the value isn’t in the production but in who stands behind the content. What will be rare: trust, authenticity, character, creativity drawn from experience, the courage to make an unpopular decision, leadership, and the ability to decide under uncertainty. For your career, that’s a more important question than when general AI will arrive.

AI isn’t just one chapter out of fourteen; it’s a layer over all thirteen others. In each of them, a different tool does a different job; an overview by field is in the list below.

Three rules for every field. A human approves: the model prepares, a human sends. A human checks the facts: names, numbers, quotes, links. A human guards the brand: tone, values, and what isn’t said are not in the prompt unless you write them there. Roll out one process at a time, and measure the time before and after.

Three sentences to take away. The model guesses the next token: brilliant with language, unreliable with facts. A good prompt is a brief like one for a temp: role, task, material, format, constraints. And the biggest change is still to come: from writing prompts to delegating goals to agents that work while you sleep. Whoever can do that will be rare in marketing. Today’s AI is still narrow (ANI), though; when general AI (AGI) will arrive is a debate measured in years, or decades.