The Visual Encyclopaedia of AI for the Over-20s

What AI Actually Is

Artificial intelligence, in 2026, means large language models and related systems that process and generate text, images, code, and more. The core idea: train a neural network on massive data, and it learns to predict useful outputs. No magic, no consciousness — just patterns at scale.

• Modern AI = large language models trained on massive datasets
• The core mechanism is predicting the next token in a sequence
• No consciousness, no understanding in the human sense — just patterns

Did you know? GPT-4, the model behind ChatGPT, has an estimated 1.8 trillion parameters — numbers that encode the patterns it learned from the internet.

---

How LLMs Work

Large Language Models (LLMs) use an architecture called the Transformer. It processes all the words in your message at once, weighs how much each word relates to every other word, and uses that to predict the best next word in a reply. Repeat millions of times — that is a conversation.

• Transformer architecture processes whole sequences at once
• Attention mechanisms weigh word relationships across the text
• Training on trillions of tokens builds general language ability

Did you know? The Transformer architecture was invented by Google researchers in 2017. Their paper was titled 'Attention Is All You Need' — one of the most cited AI papers ever.

---

Training vs Inference

Training is when a model learns — processing vast data over weeks or months using thousands of GPUs. Inference is when a trained model answers your question — fast, cheap, and running in real time. You only ever experience inference. Training happens once, offline, by the AI company.

• Training: learning from data — slow, expensive, done once by the company
• Inference: using the trained model to answer questions — fast and cheap
• You only ever interact with the inference side of AI

Did you know? Training GPT-4 is estimated to have cost over $100 million. Running one conversation costs a fraction of a cent.

---

Context Windows

A context window is the maximum amount of text an AI can consider at once — its working memory. Models in 2026 have windows of 128K to 1M tokens (roughly 100K to 800K words). Longer contexts mean better multi-step reasoning and document analysis.

• Context window = how much the model can read at once
• Measured in tokens (about ¾ of an English word each)
• Larger context = can read whole books, codebases, or reports

Did you know? In 2023 context windows were 4K tokens (about 3,000 words). By 2026 some models reached 1 million tokens — a 250× jump in two years.

---

The Model Landscape

A handful of companies dominate the frontier: OpenAI (ChatGPT/GPT-4), Anthropic (Claude), Google (Gemini), Meta (Llama — open source), Mistral (open source, European). Each has a different philosophy: commercial, safety-focused, open, or efficient. Understanding the landscape helps you choose the right tool.

• OpenAI, Anthropic, Google, Meta, and Mistral lead the field
• Each model family has different strengths and pricing
• Open-source models (Llama, Mistral) can be self-hosted for free

Did you know? Meta released Llama 3 openly in 2024, and it quickly became the most downloaded AI model ever — used in millions of apps worldwide.

---

Multimodal AI

Modern AI handles not just text but also images, audio, code, and even video. You can upload a photo and ask questions about it, speak to your assistant and get spoken replies, or have AI analyse a chart. Multimodal means AI that works across all these modes at once.

• Multimodal AI handles text, images, audio, video, and code
• You can upload a photo and ask AI what it sees
• Models like GPT-4o and Claude 3.5 are strongly multimodal

Did you know? GPT-4o ('o' for omni) can see, hear, and respond in real time — making live spoken conversations with AI possible for the first time.

ChatGPT Deep Dive

ChatGPT is the world's most popular AI product with over 900M weekly users. The free tier uses GPT-4o Mini; paid ($20/mo) unlocks GPT-4o and o1 reasoning models. It has a huge plugin and tool ecosystem, code interpreter, image generation via DALL-E, and strong third-party integrations.

• Free: GPT-4o mini · Plus $20/mo: GPT-4o + o1 reasoning
• Has built-in code interpreter, image generation, and web search
• Largest user base and ecosystem of any AI product in 2026

Did you know? ChatGPT went from 0 to 100 million users in 2 months — faster than Instagram, TikTok, and every other app in history.

---

Claude Deep Dive

Anthropic's Claude is known for exceptional coding ability, long document analysis (200K context), and careful, nuanced reasoning. Claude 3.5 Sonnet and Opus are the top performers on many coding and writing benchmarks. Its Constitutional AI training makes it notably thoughtful on tricky ethical questions.

• Claude 3.5 Sonnet tops many coding and writing benchmarks
• 200K token context window — can read entire codebases
• Trained with Constitutional AI for safe, thoughtful responses

Did you know? Anthropic was founded in 2021 by former OpenAI researchers — including the creators of the reinforcement learning from human feedback (RLHF) technique that made ChatGPT possible.

---

Gemini and Google

Google's Gemini Ultra is competitive with GPT-4 and Claude Opus on most benchmarks. Its killer advantage: deep integration with the Google ecosystem — Gmail, Drive, Docs, and Search. Gemini Advanced ($20/mo) includes a 1M token context and Workspace integration that no other model matches at scale.

• Gemini Ultra matches GPT-4 on most benchmarks
• Deep integration with Gmail, Docs, Drive, and Search
• 1M token context window in Gemini Advanced

Did you know? Google has more AI compute than any other company — its TPU (Tensor Processing Unit) chips power both Google Search and Gemini.

---

Open Source Models

Meta's Llama 3 and Mistral's models can be downloaded and run locally for free. With a mid-range GPU you can run a capable 8B or 13B parameter model entirely on your own hardware — no API costs, full privacy, and the ability to fine-tune for your specific needs.

• Llama 3 and Mistral can be downloaded and run locally for free
• No API costs and complete data privacy on your own machine
• Fine-tuning lets you specialise the model for specific tasks

Did you know? Llama 3 8B running on a MacBook M3 Pro can generate 50 tokens per second — fast enough for real-time conversation, entirely offline.

---

Choosing the Right Model

Different tasks call for different models. For complex coding: Claude. For fast, cheap, simple tasks: GPT-4o Mini or Haiku. For long document analysis: Gemini Advanced or Claude. For free and private: Llama. Build a mental model of the tradeoffs — speed, cost, intelligence, privacy — and match to your task.

• Match the model to the task rather than always using the biggest one
• Claude leads on coding; Gemini on Google Workspace integration
• Smaller models are dramatically cheaper and still very capable

Did you know? Running GPT-4 costs ~30× more per token than GPT-4o Mini — for most tasks, the cheaper model is good enough.

---

API vs Chat Interface

The chat interface (ChatGPT, Claude.ai) is the consumer product. The API is for developers: you send structured requests and get structured responses in code, at scale, with fine-grained control. If you build anything with AI, you work with the API. Learning the API is one of the highest-ROI skills of 2026.

• Chat interface = consumer product for everyday use
• API = programmatic access for building apps and automations
• API gives full control: temperature, system prompts, function calling

Did you know? Anthropic charges $3 per million input tokens for Claude 3.5 Sonnet. A 2000-word conversation costs less than $0.01.

What Is Prompting?

Prompting is the art of writing instructions for AI that produce the output you actually want. A good prompt specifies the role, task, format, constraints, and examples. Poor prompts get mediocre output. Great prompts get outputs that rival expert work. It is the most transferable AI skill you can learn.

• Prompting = writing instructions that get great AI output
• Specify: role, task, format, constraints, and examples
• Better prompting = better output, faster, every time

Did you know? Anthropic pays 'prompt engineers' six-figure salaries to write system prompts for its enterprise products. The craft is real and valued.

---

System Prompts

A system prompt is a hidden instruction that sets the AI's role, tone, and rules before the user speaks. In the API you set it explicitly. In chat products like Claude.ai or ChatGPT, you can add persistent instructions in settings. Mastering system prompts unlocks building custom AI tools.

• System prompts set role, tone, and rules before conversation starts
• Persistent custom instructions are available in all major chat UIs
• Via the API you have full control over system prompt content

Did you know? Most commercial AI products run on top of extensive system prompts — the 'personality' of a chatbot is usually just a well-crafted system prompt.

---

Chain of Thought

Telling AI to 'think step by step' dramatically improves performance on complex reasoning tasks. The model walks through its logic before giving an answer, catching errors it would have made in a single pass. This technique — chain of thought prompting — is one of the most powerful free optimisations available.

• Adding 'think step by step' dramatically improves reasoning quality
• The model catches its own errors during the thinking process
• Chain of thought is especially powerful for maths and logic tasks

Did you know? In a 2022 Google paper, 'let's think step by step' improved GPT-3's accuracy on maths problems from 18% to 79% — just from those five words.

---

Few-Shot Prompting

Give the model one or two examples of what you want before making your request. This dramatically improves output format and quality. 'Here are two examples of the style I want: [example 1] [example 2]. Now write: [your request].' Examples are worth far more than lengthy instructions.

• Including examples in your prompt dramatically improves output
• Show the format, style, and tone you want via examples
• Even one example can double output quality for formatting tasks

Did you know? Few-shot prompting was one of the first techniques shown to work with large language models, discovered accidentally at GPT-2 scale.

---

Prompt Templates

The best prompts are not one-offs — they are templates you reuse across tasks. A template for 'summarise a legal document', a template for 'generate a test suite from code', a template for 'draft a LinkedIn post'. Building a personal library of proven templates is a high-ROI practice.

• Build reusable prompt templates for your most common tasks
• Store them in a notes app or tool like Notion or Obsidian
• Good templates compound over time — each one saves you hours

Did you know? Claude Code (Anthropic's developer tool) stores reusable prompt knowledge in CLAUDE.md files that load automatically in every session.

---

Avoiding Hallucinations

To reduce AI hallucinations, add instructions like 'if you are not certain, say so clearly' and 'cite sources when possible'. For factual tasks, enable web search. For high-stakes work, ask AI to provide its reasoning so you can check it. Treating AI outputs as drafts, not facts, is essential.

• Add 'say if unsure' and 'show your reasoning' to reduce errors
• Enable web search for factual or recent-events queries
• Treat all AI outputs as drafts to verify, not final facts

Did you know? Studies show AI hallucination rates drop significantly when models are prompted to express uncertainty — they actually do know when they are less sure.

GitHub Copilot

GitHub Copilot is an AI coding assistant built directly into VS Code and JetBrains IDEs. It suggests whole lines or functions as you type, explains code on demand, generates tests, and has a chat interface for complex refactors. At $19/month it is one of the highest-ROI developer subscriptions available.

• Copilot integrates directly into VS Code and JetBrains IDEs
• Generates code completions, tests, and documentation inline
• Studies show Copilot users complete coding tasks 55% faster on average

Did you know? GitHub reports that 55% of code in new files is now written by Copilot in repos where it is enabled — more than half of new code is AI-generated.

---

Cursor Editor

Cursor is a VS Code fork built around AI from the ground up. Unlike Copilot, it can read your entire codebase and answer questions about it, make multi-file edits in one command, and generate complete features with context from your whole project. Many engineers in 2026 use Cursor as their primary IDE.

• Cursor reads your whole codebase for context-aware suggestions
• Multi-file edits in one command — change architecture, not just lines
• At $20/month it is many engineers' primary IDE in 2026

Did you know? Cursor reached $100M in annual revenue in under two years — the fastest-growing developer tool ever, built by a team of 20 people.

---

Claude for Code

Claude 3.5 Sonnet consistently tops coding benchmarks — SWE-bench (resolving real GitHub issues), HumanEval (code generation accuracy), and internal engineering evals at major companies. Its 200K context means it can read an entire large codebase and reason across all of it.

• Claude 3.5 Sonnet leads SWE-bench and HumanEval coding benchmarks
• 200K context window can hold an entire medium codebase
• Preferred by many senior engineers for complex architectural tasks

Did you know? Anthropic reported that Claude 3.5 Sonnet can resolve 49% of real GitHub issues automatically — without any human intervention.

---

AI-Assisted Testing

One of the highest-value AI coding use cases is generating tests. Give AI your function, ask for a comprehensive test suite, and get full coverage in seconds. This removes the friction that causes most engineers to skip testing — AI makes TDD dramatically easier to practise.

• AI can generate complete test suites from a function in seconds
• Dramatically reduces the friction that leads to skipping tests
• Enables real TDD: write the test prompt first, generate both test and code

Did you know? Studies show developers using AI for testing write 2× more test cases per hour — and higher-coverage tests — than those without it.

---

Code Review with AI

Paste a diff or function into Claude and ask for a code review: security issues, performance problems, edge cases, and style suggestions. AI catches 60-80% of the issues a human reviewer would find, in seconds. Use it before every human review to level up code quality.

• AI code review catches security issues, edge cases, and style problems
• Use before human review to catch obvious issues instantly
• Claude and GPT-4 find 60-80% of issues a human reviewer would find

Did you know? GitHub found that AI code review reduces the average pull request review cycle time by 40% — speeding up the whole engineering pipeline.

Which Jobs AI Will Change Most

AI will transform — not eliminate — most knowledge work. The highest-risk tasks are repetitive writing, data entry, basic analysis, and first-pass customer service. The highest-value human skills become: complex problem framing, relationship management, creative direction, and knowing when AI is wrong.

• Repetitive writing and data tasks are being automated the fastest
• Complex problem framing and relationship work remains human
• The highest value is often knowing when to override AI output

Did you know? McKinsey estimates 60-70% of work activities could be partially automated with current AI — but very few entire jobs disappear overnight.

---

The AI Skills Premium

Knowing how to use AI tools — prompting, API integration, workflow design, evaluation — commands a meaningful salary premium in 2026. Roles explicitly requiring AI skills pay 20-30% more on average. The gap will widen. These skills are the new Excel.

• AI skills command a 20-30% salary premium in 2026
• Prompt engineering, API use, and evals are in high demand
• Employers increasingly filter for AI proficiency in early screening

Did you know? LinkedIn reported a 17× increase in job postings requiring AI skills between 2022 and 2025. The fastest-growing required skill in history.

---

Building an AI-Augmented Workflow

The highest-performing knowledge workers in 2026 use AI for: first drafts of any document, code and test generation, research synthesis, and email triage. The workflow is: AI generates first pass → human edits and decides → AI refines. This 10× throughput increase is the real competitive advantage.

• AI handles first drafts; humans handle decisions and direction
• Use AI for research synthesis, code, writing, and scheduling
• The 10× throughput increase is the real moat in knowledge work

Did you know? Some solo founders now run businesses with $1M+ in revenue that would have required 10 people in 2020 — by using AI for every automatable task.

---

Freelancing with AI

AI dramatically increases freelancer output. A solo content writer can produce 10× more work; a developer can build faster; a designer can explore 50 concepts where they used to show 5. The smart move: do not lower prices. Use AI to take on more work and deliver better quality faster.

• AI lets freelancers produce 5-10× more in the same time
• Do not lower prices — use AI to take on more clients
• Deliver higher quality by using AI to explore more options first

Did you know? Top-rated Upwork freelancers report earning 2-3× more since adopting AI tools — by delivering faster with the same or better quality.

---

Staying Relevant Long-Term

The skill that ages best is not knowing specific tools — those change every 6 months — but understanding HOW AI works, where it fails, and how to direct it well. Combine that with deep domain expertise in your field and you become the rare person who can both direct AI and evaluate its output critically.

• Tools change every 6 months — understanding beats tool-specific knowledge
• Combine AI fluency with deep domain expertise in your field
• Critical evaluation of AI output is a rare and valuable skill

Did you know? In 2026 the most sought-after AI hires are domain experts who can also use AI — not just AI experts without domain knowledge.

AI Writing

AI obliterates the blank page problem. For any writing task — report, email, blog post, script, cover letter — describe what you need and you get a solid first draft in seconds. The role of the writer shifts from typing to editing and directing. The skill becomes the taste, not the keystroke.

• AI removes the blank page — you get a draft in seconds
• Your job shifts from writing to editing, refining, and directing
• Taste and judgment become the differentiating skill

Did you know? The New York Times reported in 2025 that more than half of their writers now use AI for first drafts — including investigative journalists.

---

Image Generation

Text-to-image tools generate photorealistic or artistic images from written descriptions in seconds. DALL-E 3 is built into ChatGPT. Midjourney produces the most artistic results. Flux (via Replicate and Pollinations) is open-weight and free to run. All three are now professional-grade.

• DALL-E 3 is integrated into ChatGPT Plus — type a description and get art
• Midjourney produces the highest artistic quality for most styles
• Flux is open-weight and can be run free via Pollinations.ai

Did you know? Midjourney has no full-time employees dedicated to the AI model — a team of fewer than 15 people runs one of the world's most-used creative AI tools.

---

Video Generation

AI video generation reached a consumer inflection point in 2024-2026. Sora (OpenAI), Seedance (ByteDance), and Runway can generate 5-30 second video clips from text prompts. Quality is still improving rapidly. For short-form social content, AI video is already production-viable.

• Sora, Seedance, and Runway generate video from text prompts
• Quality sufficient for social media content is now achievable
• Cost and generation time are falling fast with each new version

Did you know? Sora generated a one-minute, photorealistic video from a single sentence of text — something thought impossible just 18 months earlier.

---

AI Music

Tools like Suno and Udio generate full songs with vocals, instruments, and production from a text prompt in under a minute. Quality is rapidly reaching professional levels for certain genres. AI music is already used in YouTube videos, advertisements, and game soundtracks.

• Suno and Udio generate full songs from a text prompt in under a minute
• AI music is production-ready for social media, ads, and game audio
• Generation costs have dropped from dollars to fractions of a cent per track

Did you know? Suno generated its one-millionth song before it reached its second birthday — an output no human music catalogue could rival.

Solo Businesses with AI

In 2026 it is realistic for a single person to build and run a business that used to require a team of 10. AI handles: content creation, code, customer support, data analysis, marketing copy, and basic design. The bottleneck shifts entirely to judgment, direction, and customer relationships.

• One person can do the output of a 10-person team using AI
• AI handles: content, code, support, analysis, and design
• The bottleneck becomes judgment and customer relationships

Did you know? Several hundred solo founders crossed $1M ARR in 2025 running businesses entirely with AI assistance and no employees.

---

Building with the API

Every major AI company offers an API that lets you build products on top of their models. The API returns structured outputs you can embed in any application. You can build: custom chatbots, document processing tools, content generators, data analysers, and much more — without training any model yourself.

• Build AI-powered apps without training any model yourself
• Anthropic, OpenAI, and Google APIs are pay-per-use with free tiers
• You can ship a working product in a weekend using the API

Did you know? Anthropic's API pricing in 2026: Claude 3 Haiku (fast and cheap) costs $0.25 per million tokens — cheap enough for millions of requests per dollar.

---

No-Code AI Tools

Tools like Make (formerly Integromat), Zapier, and Voiceflow let you build AI workflows without writing code. Connect Claude or ChatGPT to Gmail, Notion, Sheets, Slack, and hundreds of other apps. You can automate entire business processes in an afternoon using drag-and-drop.

• Make, Zapier, and Voiceflow connect AI to your existing tools
• No code required — drag-and-drop AI workflow building
• You can automate entire business processes in an afternoon

Did you know? Make.com has over 5 million users and connects to 1,400+ apps — most workflows that used to require a developer can now be built in under an hour.

---

AI for Marketing

AI accelerates every part of marketing: writing ad copy, generating images for social posts, A/B testing headlines, building email sequences, and analysing campaign data. The marketer who uses AI well produces 10× more content in the same time — and can test 10× more variations.

• AI writes ad copy, emails, headlines, and social posts instantly
• Generate 10× more creative variations to A/B test
• AI analytics tools surface campaign insights from raw data automatically

Did you know? HubSpot found that AI-using marketing teams produced 3× more content and saw 2× better click-through rates compared to teams that did not use AI.

---

Pricing AI Work

The biggest mistake AI-era freelancers make: discounting because AI made the work faster. Price on value delivered, not hours spent. If AI lets you produce in 2 hours what took 20, your margin improves — your rate should not drop.

• Value-based pricing: charge for the outcome, not the method
• Clients pay for expertise in directing AI, not raw output volume
• Publish an AI-enhanced tier: same result, faster turnaround, same price

Did you know? Top AI-native freelancers on Toptal and Upwork earn 40-60% more than peers by positioning AI as a speed multiplier, not a cost-cutter.

---

Vibe Coding — Ship Without a CS Degree

Vibe coding is the practice of describing what you want to build in plain English and letting AI write the code. Claude Code, Cursor, and Replit Agent can build working prototypes from a conversation. You direct; AI implements.

• Non-technical founders are now shipping MVPs in days using vibe coding
• The skill is product thinking and clear specifications, not syntax
• Best for: web apps, bots, automation scripts, internal tools

Did you know? Pieter Levels (Nomad List, Remote OK) shipped 12 profitable products in one year before AI existed. With AI, the same output now takes weeks, not years.

How AI Uses Your Data

When you use a chat AI product, your conversations may be used to improve the model. OpenAI and others allow you to opt out in settings. API users' data is not used for training by default. Enterprise tiers guarantee no data retention. Know which tier you are on before sharing sensitive information.

• Chat product conversations may be used to train models by default
• You can opt out of training data use in settings on most products
• API and enterprise tiers typically guarantee no training data use

Did you know? In 2023 Samsung banned employees from using ChatGPT after a developer accidentally pasted proprietary source code into a conversation — and it was sent to OpenAI's servers.

---

Privacy Best Practices

Never paste into an AI chat: passwords, private keys, personal identification numbers, sensitive client data, unreleased product details, or medical records unless you are on an enterprise plan with a data processing agreement. Treat any AI chat like an email you are CCing to a stranger.

• Never paste passwords, API keys, or client data into AI chats
• Treat AI conversations like emails CCed to an unknown third party
• Enterprise plans with DPAs offer contractual data protection

Did you know? GDPR compliance requires companies using AI with EU citizen data to have a data processing agreement — without one, using AI for customer data may be illegal.

---

RAG — Retrieval-Augmented Generation

RAG is the technique of feeding relevant documents to an LLM at query time, so it can answer questions about private or recent data without fine-tuning. Your company knowledge base, PDFs, and emails can power a custom AI assistant via RAG.

• RAG injects up to 200K+ tokens of your documents into the model context
• No model retraining needed — knowledge is injected fresh each query
• Use cases: legal doc Q&A, internal wikis, customer support bots

Did you know? Most enterprise AI products marketed as custom AI are RAG wrappers around a standard LLM — the differentiation is in which data you feed in.

---

PII and AI Privacy Risks

Sending personal data to an AI API means it may be stored, logged, or used in training. GDPR classifies AI outputs about individuals as personal data processing. Know what you share and with which providers.

• OpenAI prompts may be used for training unless API calls opt out via data controls
• Anthropic API calls are not used for training by default
• GDPR Article 22 restricts fully automated decisions about individuals

Did you know? A Samsung engineer accidentally leaked proprietary chip designs by pasting source code into ChatGPT in 2023, prompting Samsung to ban the tool internally.

AI Bias

AI models learn from human-generated data — which is full of historical biases. Models can perpetuate racial, gender, and socioeconomic biases in hiring, lending, and criminal justice systems. Understanding this is essential for anyone building AI products that affect real people's lives.

• AI learns from biased historical data and can perpetuate those biases
• Bias has been documented in hiring, lending, and law enforcement AI
• Model evaluation for bias is now a standard part of responsible AI development

Did you know? Amazon scrapped an AI recruiting tool in 2018 after discovering it had learned to downgrade resumes from women — trained on 10 years of mostly male hires.

---

Deepfakes and Misinformation

Generative AI makes it trivially easy to create realistic fake images, audio, and video of real people saying or doing things they never did. The political, legal, and social consequences are still being worked out. Critical media literacy — asking 'is this real?' — is now a fundamental life skill.

• AI deepfakes of real people are now trivially easy to create
• Political and legal frameworks for deepfakes lag behind the technology
• Critical media literacy is now essential for everyone — not just journalists

Did you know? In 2024 an AI-generated fake audio of a politician was used in an election — and was shared millions of times before it was debunked.

---

AI and Copyright

Whether AI-generated content can be copyrighted is being contested in courts worldwide. In the US, courts have ruled that pure AI output without human authorship is not copyrightable. In the EU the situation is more nuanced. If you build products on AI output, understand the legal uncertainty in your jurisdiction.

• Pure AI output is not copyrightable in the US (current case law)
• EU law is more nuanced — creative AI-assisted work may qualify
• Using copyrighted data to train models is itself being litigated in multiple countries

Did you know? The New York Times sued OpenAI in 2023, arguing that training on its articles without a licence constituted copyright infringement — setting up a landmark case.

---

Using AI Responsibly

Responsible AI use comes down to four questions: Is the output accurate? Is it fair? Is it transparent (does the reader know AI was involved)? And does it harm anyone? Applying these four questions before deploying AI-generated content or decisions protects you legally and ethically.

• Before deploying AI output ask: accurate, fair, transparent, harmless?
• Disclosing AI involvement is both ethical and increasingly required by law
• Building in human oversight at decision points reduces liability

Did you know? The EU AI Act, effective 2026, requires disclosure and risk management for AI systems used in hiring, lending, and public services — with fines up to 3% of global revenue.

---

Fairness and Bias in AI

AI models inherit biases from their training data. A hiring model trained on historical data will replicate historical discrimination. Facial recognition is less accurate on darker skin. These are not bugs to patch — they require structural change in how models are trained and evaluated.

• Amazon scrapped an AI hiring tool in 2018 after it penalised CVs from women
• Facial recognition error rates: 0.8% for light-skinned men, 34.7% for dark-skinned women
• The solution: diverse training data, bias audits, human oversight on high-stakes decisions

Did you know? The EU AI Act classifies AI used in hiring, credit scoring, and education as high risk, requiring bias testing and human review before deployment.

---

AI and Copyright

When AI generates an image, song, or paragraph, the legal status is contested. US courts have ruled AI-generated art cannot be copyrighted with no human author. The training data question — did models train on copyrighted work without permission? — is in active litigation.

• US Copyright Office: works with no human authorship are not copyrightable
• Getty Images sued Stability AI for training on 12M copyrighted photos
• The EU AI Act requires disclosure if training data includes copyrighted material

Did you know? Some studios now require AI-generated assets to be disclosed — and some publishers have blanket bans on AI-written content submissions, with contracts being rewritten accordingly.

AI Agents

The next wave of AI is not chatbots — it is agents. AI agents can use tools (browsers, code runners, APIs, databases), plan multi-step tasks, and execute them autonomously. Give an agent a goal like 'research competitors and draft a report' and it works through the steps without you prompting each one.

• Agents can use tools, plan tasks, and execute multi-step workflows
• They use browsers, code runners, APIs, and databases autonomously
• Claude, GPT-4, and Gemini all support tool use / function calling

Did you know? Devin, a software engineering AI agent, autonomously resolved real GitHub issues end-to-end — writing code, running tests, and submitting PRs without human intervention.

---

Multi-Agent Systems

Rather than one AI trying to do everything, multi-agent systems assign specialised agents to different parts of a task: a planner, a researcher, a coder, a reviewer. The agents pass work between each other. This mirrors how human teams work and produces dramatically better results on complex tasks.

• Multi-agent = different AIs with different roles working together
• A planner breaks down the task; a researcher gathers info; a coder builds
• Reduces hallucinations by having one agent verify another

Did you know? Claude Code, Anthropic's engineering tool, can spawn sub-agents for research, writing, and building — all orchestrated by a master agent in a single conversation.

---

AGI — What It Means

Artificial General Intelligence (AGI) means AI that can do any intellectual task a human can do. We are not there yet. The debate is over how far away it is: months? years? decades? Most researchers agree that current AI is narrow and domain-specific, not truly general — but the pace of progress makes confident predictions hard.

• AGI = AI that matches human ability across ALL intellectual tasks
• We do not have AGI yet — current AI is narrow and specialised
• Expert estimates range from 5 years to never — the honest answer is: uncertain

Did you know? Sam Altman (OpenAI CEO) has said AGI could arrive 'in a few years'. Many researchers strongly disagree. The range of expert predictions has never been wider.

---

AI Infrastructure

AI runs on specialised chips — GPUs (Nvidia) and TPUs (Google). The Nvidia A100 and H100 GPUs power most of the frontier model training and inference. Demand for these chips is so high that delivery times stretched to a year at peak. Whoever controls the chips controls the pace of AI.

• Nvidia's H100 GPUs are the primary hardware for AI training
• TPUs are Google's custom AI chips used for both training and Gemini inference
• Chip access is the primary bottleneck in AI development in 2026

Did you know? Nvidia's market capitalisation passed $3 trillion in 2024 — more than any company except Apple and Microsoft — driven almost entirely by AI chip demand.

---

Emerging Model Capabilities

The next wave of model capabilities includes: better long-horizon planning, real-time audio/video understanding, memory across sessions, and reliably following complex multi-step instructions. Each of these was a research problem two years ago; now they are shipping into products.

• Long-horizon planning and reliable multi-step execution are arriving now
• Real-time audio and video understanding is shipping in frontier models
• Persistent memory across sessions is now in paid plans of major products

Did you know? GPT-4o's live voice mode processes speech in 300ms — fast enough that it can interrupt, laugh, and react mid-sentence in real time.

---

Open Source vs Closed

The frontier AI race is split: OpenAI, Anthropic, and Google keep their best models closed. Meta, Mistral, and others release weights openly. Open models are now within 10-20% of closed models on benchmarks and are rapidly closing the gap. The outcome of this race determines who controls AI infrastructure.

• Closed models (OpenAI, Anthropic, Google) keep weights proprietary
• Open models (Meta Llama, Mistral) release weights anyone can download
• Open models are closing the capability gap — now within ~10-20% on benchmarks

Did you know? Llama 3 70B, released free by Meta in 2024, scored higher than GPT-3.5 on most benchmarks — and GPT-3.5 cost OpenAI billions to develop.

---

AI Regulation

Governments worldwide are rushing to regulate AI. The EU AI Act is the most comprehensive law — banning high-risk AI applications and requiring transparency and testing. The US approach is lighter: executive orders and voluntary commitments. China mandates registration of AI systems. The regulatory landscape will shape your career.

• EU AI Act (2024-2026) is the world's most comprehensive AI law
• Bans certain high-risk AI uses and requires transparency and risk assessments
• US approach is lighter: voluntary commitments and sector-specific guidance

Did you know? The EU AI Act defines AI systems used in hiring, credit scoring, and law enforcement as 'high risk' — requiring human oversight and bias audits.

---

Your Decade in AI

The best preparation for an AI-shaped decade: learn to prompt well (now), understand APIs and how to build with them (2025-26), follow key model releases and benchmark updates, and maintain deep domain expertise. Generalists who can direct AI will outperform pure AI specialists without domain knowledge.

• Learn prompting now — it is free and pays immediately
• Learn the API when you build your first side project
• Maintain deep domain expertise — AI fluency × domain = rare and valuable

Did you know? The 10 skills most in demand in 2030 do not exist yet in their current form — the ability to learn and adapt is more valuable than any fixed skill set.

---

Multimodal AI

Multimodal models process and generate text, images, audio, and video together. GPT-4o can see your screen and speak. Gemini 1.5 can watch a 1-hour video and answer questions about it. The interface is becoming the conversation.

• GPT-4o processes text, image, and audio in a single model
• Google Gemini 1.5 Pro has a 1M token context window — roughly 1 hour of video
• Multimodal opens AI to accessibility use cases: live transcription, image description, real-time translation

Did you know? Within 2 years, the chat with your documents use case will expand to chat with your video library — the context window just needs to grow another order of magnitude.

---

AI Agents

An AI agent is a model that takes a goal, breaks it into steps, uses tools (browser, code runner, file system), and iterates until done. Claude Code writes and runs code. AutoGPT browses the web. Devin (Cognition) can open a pull request autonomously.

• Agents fail at long-horizon tasks — reliability degrades with each step
• The key primitive: tool use (letting the model call APIs and execute code)
• Multi-agent systems: one model plans, another executes, a third reviews

Did you know? Claude Code, Cursor Agent, and GitHub Copilot Workspace are all agent-mode products — the shift from autocomplete to autonomous implementation is already in your IDE.

---

The AGI Debate

AGI — AI that matches or exceeds human capability across all cognitive tasks — is the stated goal of OpenAI, Anthropic, and DeepMind. Timeline estimates range from 2027 to never. What matters to you: the productivity step-change is already here, regardless of whether AGI arrives.

• OpenAI CEO Sam Altman: AGI may arrive within a few thousand days
• Most AI researchers define AGI as a moving target — we keep raising the bar after each milestone
• The practical impact: plan for AI that is 10x better than today within 5 years, whatever we call it

Did you know? Every time AI beats a benchmark — chess, Go, protein folding, bar exam — the definition of real intelligence shifts. This is known as the AI effect, or Teslers theorem.

---

Your Path in the AI Era

The over-20s generation will live their entire careers in an AI-shaped economy. The winning strategy: develop deep domain expertise plus strong AI collaboration skills. Neither alone is enough. Be the person who knows the domain AND knows how to direct AI to work in it.

• Prompt engineering fluency: achievable in 30 hours of deliberate practice
• Build one AI-powered side project in 2025 — the learning compounds
• Follow 3-5 AI researchers or practitioners on X/LinkedIn — the field moves weekly

Did you know? The roles that will grow most by 2030 are not AI jobs — they are every existing job, amplified. The doctor, lawyer, and designer who use AI well will outperform those who do not.