AI Fundamentals Fluency

AI is any technique that lets a machine do something we associate with human intelligence: perceiving, reasoning, deciding, communicating or creating. That definition is deliberately broad. It covers a 1980s rule-based tax program and a 2026 agentic assistant that books your travel. What separates them is not whether they count as AI, but how they achieve the behaviour, which is what the subfields describe.

The subsets, in detail

Picture nested circles, then go one level deeper than the usual overview.

• Artificial intelligence is the umbrella. It holds both classical rule-based approaches and modern learning-based ones.
• Machine learning is systems that learn patterns from data instead of following hand-written rules. It contains three paradigms, supervised, unsupervised and reinforcement, covered properly in Week 3.
• Deep learning is machine learning using layered neural networks, able to learn directly from raw unstructured data (images, audio, raw text) with no manual feature engineering.
• Natural language processing is the branch aimed specifically at understanding and generating human language. Today's large language models are the state of the art within NLP.
• Computer vision is the branch aimed at interpreting images and video: crop-disease detection from a phone photo (the Week 4 case scenario), medical imaging, facial recognition, quality control on a production line.
• Robotics and embodied AI act in the physical world through sensors and actuators: warehouse robots, autonomous vehicles, inspection drones. The 2026 framing is AI learning through motion rather than through code, moving out of purely virtual environments.
• Generative AI is deep learning trained to produce new content, text, images, audio, video, code, rather than only to classify or predict. LLMs are its text branch.
• Agentic AI is the fastest-moving subfield as of 2026: systems that plan and execute multi-step tasks with tools, searching, booking, writing and running code, coordinating with other agents. The shift is from what AI can say to what AI can do.

These are not separate boxes. Nearly every tool below spans NLP, generative AI and increasingly agentic AI at the same time.

The timeline, with the turns that mattered

• 1956, the Dartmouth Workshop. The term artificial intelligence is coined, with confident predictions of human-level machines within a generation.
• 1960s and 70s, symbolic AI and expert systems. Hand-coded logic dominates. The successes are narrow and brittle.
• 1970s and late 1980s, the AI winters. Funding and interest collapse twice as hand-coded systems fail to scale to real-world complexity.
• 1997, Deep Blue beats Kasparov. A landmark for brute-force search, and still not learning in the modern sense.
• 2012, AlexNet. A deep network decisively outperforms older computer-vision methods, opening the current era. It was possible because big data and GPU compute arrived together.
• 2017, Attention Is All You Need. The transformer architecture is published: the technical foundation under every major LLM since.
• 2022, ChatGPT launches publicly. Generative AI reaches non-technical users at scale, compressing years of gradual enterprise adoption into months of public familiarity.
• 2023 to 2025, the multi-model era. Claude, Gemini and a wave of open-weight models (DeepSeek, Kimi, Qwen, GLM) establish real competition. Reasoning-focused models emerge and costs fall sharply.
• 2026, the agentic and evaluation era. The mood shifts from AI evangelism to AI evaluation: enterprises want proven return rather than experiments, agents move into production, and organisational adoption reaches roughly 88 percent of surveyed companies per the 2026 Stanford AI Index.

Where AI is actually used

• Healthcare: diagnostic imaging support, triage assistants, drug discovery.
• Finance: fraud detection, credit scoring, algorithmic trading, and closer to home, mobile-money fraud pattern detection in Kenya's M-Pesa economy.
• Agriculture: crop-disease detection from phone photos, yield prediction, precision irrigation.
• Education: personalised tutoring, grading support, adaptive learning.
• Retail and e-commerce: recommendation engines, demand forecasting, generated product copy.
• Logistics and manufacturing: route optimisation, predictive maintenance, autonomous sorting and inspection moving from pilot to deployment.
• Creative industries: assisted writing, design, video and music, with measured productivity gains in marketing content specifically.

What happens next, honestly

Credible experts disagree sharply, and fluency means holding that disagreement rather than picking a side to repeat.

The fast-timeline camp, including some frontier-lab leaders, has publicly discussed AI matching or exceeding human performance on most cognitive tasks within a few years, with serious economic disruption starting as early as 2026 or 2027. The cautious camp, including economists and senior researchers, argues general intelligence remains decades away, that measured productivity gains outside programming and customer support are still limited, and that most enterprise AI pilots do not show measurable return.

What is well supported either way: adoption is high and rising, inference spending is growing fast, agents are moving into production, and regulation is catching up to deployed capability rather than running ahead of it.

Treat this as a spectrum, not a debate to win. The useful question is what evidence would move you toward either end.

The tools, and why each is strong where it is strong

• ChatGPT (OpenAI). The most widely used general assistant, largest ecosystem, fastest feature cadence. Strength: all-round versatility, and usually the right first tool for a new user.
• Claude (Anthropic). Often the stronger choice for document-heavy business work, long proposals, contracts, compliance review, policy writing, and for coding, with large context windows and the ability to ground answers in a company's own knowledge base securely. That last point is why Claude comes up specifically in enterprise buying conversations.
• Microsoft Copilot. Built on OpenAI models and differentiated almost entirely by where it lives: inside Word, Excel, Outlook and Teams, under an organisation's existing Microsoft 365 security and compliance. Strength: near-zero adoption friction for a Microsoft shop, not superior raw intelligence.
• Gemini (Google). Strong multimodal handling of text, image, video and audio together, very large context windows, tight Google Workspace and Cloud integration. Strength: large mixed-media inputs in a Google-centric organisation.
• Kimi (Moonshot AI). Known for very long context and strong multi-step reasoning with heavy tool use, effective on large documents, spreadsheets and log files. For regulated or sensitive data, ask the data-residency question explicitly.
• DeepSeek. The model most associated with proving near-frontier reasoning and coding could be reached at dramatically lower training and API cost, with open weights. Strength: budget-conscious, high-volume work where cost per query matters. Same data-residency caveat.

This landscape moves in months, not years. Verify current versions, pricing and capability claims before you rely on them.

What AI does well, and badly

Today's assistants are genuinely strong at drafting, summarising, rewriting for a different audience, translating, brainstorming, explaining a concept several ways, and turning messy notes into structure. They are weak, often confidently so, at exact arithmetic, anything that happened after their training cutoff, niche local facts, precise citations, and knowing what they do not know. That last one is the dangerous one: a model rarely says "I am not sure", so the checking has to come from you. The rest of this course keeps returning to that gap, and the Responsible Use policy you agreed to starts from it: treat output as a draft, not a fact.

Case scenario

Neurobyte Logistics Co. is now choosing a tool, not just a category. Its three founders each favour a different assistant for the same task, drafting customer-facing delivery-status replies, based purely on what each has used before, with no evaluation criteria at all. Design a fair bake-off: what exactly would you test, with which prompts, judged against what criteria, before recommending anything?

Practical bit

• In groups, run one identical prompt through at least two different tools. For example: draft a professional but warm reply to a customer whose delivery is two days late.
• Compare tone, length, accuracy and any invented details. The invented details are the interesting part.
• Hits and misses: run five everyday tasks through an AI assistant (a summary of a long email, a translation, a sum with large numbers, a question about this week's news, and a fact about your own town). Score each one hit or miss, and note which of the weaknesses above explains each miss.
• Map today's milestones onto a shared timeline as a closing activity.

Resources and homework

• The 2026 AI Index Report from Stanford HAI, the anchor for the adoption and investment figures used here: https://hai.stanford.edu/ai-index/2026-ai-index-report
• Stanford HAI's expert-predictions roundup, as a pre-read for the range-of-views framing: https://hai.stanford.edu/news/stanford-ai-experts-predict-what-will-happen-in-2026
• A plain-language explainer on the transformer architecture. Read a beginner summary, not the original 2017 paper.
• A current side-by-side enterprise AI tool comparison. Check its date: anything more than a few months old is describing a different market.