How computers work, how the internet runs, and what AI actually is
Part A · what technology is — the long arc from tools to intelligence
Technology is not "gadgets." It is any reliable method for amplifying human capability: a stone blade, double-entry bookkeeping, refrigeration, GPS, antibiotics, a search engine, or a shipping container. The through-line across 10,000 years is simple — technology converts scarce human attention into repeatable systems that the next person can use without rediscovering.
Technology waves — 10,000 BCE to today
Each wave does not replace the previous one. Agriculture, electrification, and computing are all still running. What changes is which wave provides the primary constraint on human productivity. Bars are deliberately approximate.
A technology does not end when the next one begins. Steam engines, radio, and mainframes still matter.
Each wave — what it actually changed
Select a wave above.
Part B · how computers work — the hardware
The history of computing — from Babbage to the smartphone
Every major computing era was unlocked by a single physical breakthrough. The leap from mechanical gears to vacuum tubes to transistors to integrated circuits follows a pattern: each generation was roughly 1,000x smaller, faster, and cheaper than the last. The smartphone in your pocket contains more processing power than the entire Apollo program's ground control.
Bars show approximate active period of each computing paradigm. Eras overlap — transistors didn't replace tubes overnight.
Moore's Law — transistors, limits, and what comes next
In 1965, Gordon Moore noticed that the number of transistors on a chip was doubling roughly every two years, and predicted this would continue. It held for 50 years. A 1971 Intel 4004 had 2,300 transistors; a 2023 Apple M2 Ultra has 134 billion — a 58-million-fold increase. But physics is catching up: transistors are now measured in nanometres, with TSMC's 3nm process meaning gate lengths are just a handful of silicon atoms wide. The law is slowing, but chipmakers have responded with 3D stacking, chiplets, and specialised silicon (GPUs, TPUs).
1971 (4004)2025 (M-series)
2,357,800
Loading...
Binary and logic gates
Computers are binary not because of some arbitrary design choice, but because transistors are switches with two stable states: on (1) and off (0). Every number, letter, image, and instruction is encoded as bits. A byte (8 bits) can represent 256 values. Logic gates combine two bits with operations like AND, OR, and NOT to produce a third. Chain millions of these gates and you can do arithmetic, compare values, and store results.
The CPU — fetch, decode, execute
A CPU executes a simple three-step cycle billions of times per second: fetch an instruction from memory, decode what it means, and execute it. Clock speed (measured in GHz) tells you how many cycles per second, but cycles per instruction and parallelism matter more than raw frequency. Modern CPUs have multiple cores — effectively separate processors on one chip. A GPU has thousands of simpler cores optimised for doing the same operation on thousands of data points simultaneously, which is why they dominate deep learning. Apple's M-series chips integrate CPU, GPU, and memory on one die, slashing the latency cost of data movement.
Memory and storage — the hierarchy and latency scale
Computers manage a ruthless trade-off: fast memory is expensive and small; cheap memory is slow and large. The hierarchy solves this by keeping frequently used data in small, fast caches close to the CPU and archiving rarely used data on slow, dense drives. The latency gap between levels is staggering: L1 cache access takes ~1 nanosecond; a hard drive seek takes ~10 milliseconds — a 10-million-fold difference.
100 ns
1 ns (L1 cache)10 ms (hard drive)
The semiconductor industry — chips, geopolitics, and nm nodes
Making a modern chip requires depositing dozens of atomic-thin layers onto a silicon wafer using extreme ultraviolet (EUV) light — a technology that only one company in the world, ASML of the Netherlands, can build. TSMC in Taiwan manufactures chips for Apple, NVIDIA, AMD, and most of the world's electronics industry. This concentration makes chip supply chains extraordinarily fragile and geopolitically sensitive.
Select a topic above.
Part C · software — how computers are programmed
Software layers — from machine code to applications
Software is built in layers of abstraction, each one hiding the complexity below. Machine code is pure binary executed by the CPU; you never write it. Assembly gives human-readable names to those instructions but is still hardware-specific. High-level languages (Python, JavaScript, C++) compile or interpret down through these layers. An operating system sits in the middle, managing hardware resources so applications don't have to. The average smartphone app depends on dozens of such layers — understanding even one helps demystify all of computing.
Programming fundamentals — algorithms and complexity
All programs combine four primitives: variables (named storage), conditionals (if/then branching), loops (repeating an action), and functions (reusable named blocks of code). Algorithms are recipes for solving problems. Complexity notation (Big-O) measures how an algorithm's time or memory use scales with input size. A linear-search algorithm on a list of a million items might require a million comparisons in the worst case (O(n)); a binary-search tree can find the same item in just 20 comparisons (O(log n)). That difference is what separates scalable software from software that grinds to a halt.
Algorithm complexity — comparisons needed to search 1,000,000 items
O(1) — hash lookup
1
O(log n) — binary
20
O(√n) — sqrt
1,000
O(n) — linear
1,000,000
O(n²) — quadratic
10¹²
Bars are log-scaled for visibility. O(n²) would be 1,000,000x wider than O(n) on a true linear scale.
How the web works — HTTP, DNS, and the request cycle
When you type a URL into a browser, a precise sequence unfolds in under 200 milliseconds. First, a DNS lookup translates the domain name into an IP address. Then a TCP connection is established (the "three-way handshake"). The browser sends an HTTP GET request to the server, which responds with HTML. The browser parses that HTML, discovers CSS and JavaScript files, requests those too, and renders the page. HTTPS adds TLS encryption so the content can't be read in transit. APIs work the same way but return structured data (usually JSON) instead of HTML.
Open source — the invisible infrastructure of the internet
Open source software is code whose licence allows anyone to inspect, use, modify, and redistribute it. The economics seem paradoxical — why would companies give away valuable software? The answer is that shared infrastructure lowers everyone's costs: no single company wants to write their own TCP/IP stack or cryptography library. Linux runs 96% of the world's top 1 million web servers, all Android devices, and most cloud infrastructure. Git (created by Linus Torvalds in 2005 in two weeks) is the version-control system behind essentially all professional software development.
Open source prevalence — selected categories (% of market using OSS)
Web servers
96%
Cloud infra (Linux)
90%
Mobile OS (Android)
72%
AI/ML frameworks
~85%
Desktop OS
~4%
Source: W3Techs (servers), Statcounter (mobile). Desktop Linux share remains low despite its dominance in servers and embedded systems.
Part D · the internet and networks
How the internet works — packets, protocols, and physical infrastructure
The internet is a network of networks held together by a shared set of protocols, chiefly TCP/IP. When you send data, it's broken into small packets (typically 1,500 bytes), each of which is routed independently across any available path and reassembled at the destination. TCP provides reliability — it retransmits any dropped packets. UDP skips that guarantee for speed (used in video calls and gaming). The physical layer is largely fibre optic cables, including 550+ submarine cable systems carrying ~99% of international internet traffic.
Global internet traffic by type (approximate 2024)
Video is the internet's heavy cargo, which is why content delivery networks cache popular clips close to viewers.
Cybersecurity fundamentals — how attacks and defences work
Most security failures are not dramatic technical exploits — they're human failures. The 2020 SolarWinds attack, which compromised US government systems for months, began by compromising a software update. The 2016 Bangladesh Bank heist ($81 million stolen) used phishing emails. Encryption is foundational: HTTPS wraps every web transaction in TLS, using public-key cryptography so two parties who've never met can establish a shared secret in milliseconds.
Select an attack or defence type above.
Cloud computing and the platform economy
Cloud computing means renting computation, storage, and networking from a provider rather than owning servers. AWS launched in 2006 and within a decade disrupted the entire enterprise IT industry. The three layers are IaaS (raw virtual machines), PaaS (you manage only your code), and SaaS (you manage nothing — think Gmail). Network effects made the winners dominant: AWS, Azure, and Google Cloud control roughly 65% of the market.
Cloud provider market share — global infrastructure (2024)
Amazon AWS
31%
Microsoft Azure
24%
Google Cloud
11%
Alibaba Cloud
4%
Others
30%
Source: Synergy Research Group, Q4 2024 estimates.
A network effect occurs when a product becomes more valuable as more people use it. The telephone is the textbook example: worthless alone, priceless at scale. Modern platforms weaponise this: Facebook's 3 billion users aren't just customers — they're the product and the moat simultaneously. Direct network effects (WhatsApp: more users = more contacts) are stronger than indirect ones. Data network effects (TikTok: more watching = better recommendations) are among the most powerful because they compound automatically with scale.
Weak / contestableStrong / winner-take-all
Part E · data and databases
How databases work
A relational database organises data into tables of rows and columns, with relationships enforced between them. SQL is the 50-year-old standard for querying them. ACID properties (Atomicity, Consistency, Isolation, Durability) ensure that a bank transfer either completes fully or rolls back entirely — there's no state where money disappears. NoSQL databases (MongoDB, Cassandra, Redis) sacrifice some of these guarantees for horizontal scale or specialised access patterns. PostgreSQL and MySQL together power the majority of the world's web applications.
Big data — when scale changes everything
Big data is defined by the three Vs: Volume (petabytes), Velocity (streams arriving in real time), and Variety (text, images, logs — not just clean tables). Traditional SQL databases choke on this. Hadoop (2006) distributed processing across thousands of commodity servers; Spark (2014) made it faster by keeping data in RAM. The real insight of the big data era: storing everything first and asking questions later is often more valuable than deciding in advance what to record. The hard part is rarely storage — it's deciding which data is trustworthy.
Privacy, surveillance, and what your data actually reveals
Surveillance capitalism (Shoshana Zuboff's term) is the business model of extracting behavioural data, predicting future behaviour, and selling those predictions to advertisers. Metadata — who you call, when, for how long, from where — can reveal more than message content: a 2014 Stanford study showed that phone metadata alone could identify medical conditions, legal issues, and religious affiliations with high accuracy. Anonymisation is far weaker than assumed: a 2019 Nature study showed 99.98% of Americans could be re-identified from just 15 demographic attributes in "anonymous" datasets.
Metadata exposure calculator
Enter your usage above and calculate.
The phrase "data is the new oil" is catchy and wrong in one important way. Oil is used up when burned; data can be copied, combined, and reused indefinitely. That makes data less like fuel and more like memory with economic power attached.
Part F · artificial intelligence — what it is, how it works, and what it means
What AI actually is — and the difference between types
AI is not a single thing. The term covers everything from a spam filter to GPT-4. Today's AI systems are all "narrow" — they do one thing well, but understanding does not transfer. A chess engine that beats Magnus Carlsen cannot play draughts. The long-hypothesised "artificial general intelligence" (AGI) — a system that can learn any task a human can — does not yet exist, and experts disagree sharply on whether current approaches could ever produce it.
Machine learning — how computers learn from data
Traditional programming is explicit: a programmer writes rules. Machine learning inverts this: the system is given labelled examples and learns the rules itself. In supervised learning, a model is trained on input-output pairs. In unsupervised learning, the model finds structure without labels. Reinforcement learning trains an agent to maximise reward through trial and error — the method behind AlphaGo. The quality and size of training data matters enormously: a model trained on biased data will produce biased outputs.
1,000 examples1 trillion tokens
1 million
Loading...
Neural networks and deep learning — the architecture
A neural network consists of layers of "neurons" (simple mathematical functions) connected by "weights" (numbers that scale each connection). During training, the network makes a prediction, computes an error, then uses backpropagation to adjust every weight slightly in the direction that reduces that error. "Deep" learning simply means many layers; depth allows the network to learn hierarchical features — edges to shapes to objects. The 2012 AlexNet paper, which used a deep convolutional network to win ImageNet by a crushing margin, triggered the modern AI era.
Large language models — how ChatGPT and Claude actually work
Large language models are neural networks trained on vast text corpora to predict the next token (roughly a word-fragment) in a sequence. The transformer architecture (introduced in the 2017 paper "Attention Is All You Need") uses a mechanism called "attention" to weigh the relevance of every other token in a sequence when predicting the next one. Pre-training on internet-scale text teaches grammar, facts, reasoning patterns, and coding; fine-tuning and RLHF (reinforcement learning from human feedback) then shapes the model's behaviour. The key misconception: LLMs don't "know" things the way humans do — they're extremely powerful interpolation engines over statistical patterns in text.
Transformer context window — how much text can the model see at once?
8,000 tokens
short promptchapterbook-scale context
The "hallucination" problem: LLMs generate plausible-sounding text, not verified facts. When asked about something outside their training distribution, they frequently confabulate — producing confident, grammatically perfect falsehoods. This is not a bug to be patched; it's an inherent feature of predicting plausible text.
The best AI misconception to drop: "it is just autocomplete." Next-token prediction sounds trivial until the training data, model size, and compute scale force the system to learn grammar, facts, code patterns, visual concepts, and reasoning heuristics as compression strategies. It still makes mistakes, but the mechanism is more powerful than the slogan suggests.
AI risk and governance — alignment, bias, and the big questions
The alignment problem asks: how do you ensure an increasingly capable AI system pursues the goals its designers actually intended, not a subtly different proxy? More immediate risks include bias encoded in training data (facial recognition systems having substantially higher error rates for darker-skinned women, per Joy Buolamwini's 2018 Gender Shades study), and autonomous weapons systems making lethal decisions below human decision latency. The EU AI Act (2024) is the world's first major AI regulation, classifying systems by risk and banning certain applications outright.
AI concerns — where researchers place different risks (consensus view)
AI and the economy — tasks, jobs, and what changes
The automation question is not "will AI take jobs?" but "which tasks will be automated, how fast, and what will workers do instead?" A 2023 Goldman Sachs analysis estimated that 300 million jobs globally could be affected, with knowledge workers — lawyers, financial analysts, programmers — more exposed than expected. History is reassuring in aggregate: the Industrial Revolution created far more jobs than it destroyed. But the workers who lost handloom jobs in 1820 didn't smoothly transition to factory work. Jobs are bundles of tasks, relationships, accountability, and physical context. Automating 40 percent of a job rarely means deleting 40 percent of workers on schedule.
Task exposure to AI automation by sector (Goldman Sachs / McKinsey, 2023)
Legal / admin
44%
Office / admin
46%
Software / IT
37%
Education
27%
Construction
6%
Maintenance
4%
"Exposed" means tasks that could be automated; it does not mean those jobs will disappear.
Computer vision, speech, and multimodal AI
Computer vision uses convolutional neural networks (CNNs) — filters that scan images for edges, textures, and shapes at increasing levels of abstraction. Diffusion models (Stable Diffusion, DALL-E, Midjourney) generate images by learning to reverse a process of adding noise. Multimodal models combine vision and language: GPT-4V, Gemini, and Claude can reason about images, charts, and documents. Speech recognition improved dramatically when converted to a sequence-to-sequence problem: Whisper (OpenAI, 2022) achieves near-human performance across 99 languages with a single model.
AI capability frontier by modality (approximate state, 2024)
Part G · emerging and frontier technologies
Quantum computing — superposition, entanglement, and the hype-reality gap
A classical bit is either 0 or 1. A qubit can exist in a superposition of both simultaneously until measured. Quantum algorithms (Shor's for factoring, Grover's for search) exploit quantum interference to amplify correct answers. Shor's algorithm could break RSA encryption in polynomial time. The practical reality: today's best machines have ~100-1,000 physical qubits but require extensive error-correction because qubits are extremely fragile. A cryptographically relevant quantum computer likely needs millions of logical qubits. Quantum computers will not make spreadsheets or websites faster — the real near-term value is chemistry, materials science, and optimisation research.
50 qubits (2019)4,000,000 qubits (needed)
1,000
Loading...
Frontier technologies — five more fields reshaping the next decade
The frontier is uneven. AlphaFold transformed protein-structure prediction years before robots could fold laundry reliably. Electric vehicles arrived before fully autonomous taxis. Technology does not advance as one smooth front; it breaks through wherever physics, data, money, regulation, and user tolerance happen to align.
Hype versus deployed impact
A frontier technology can be scientifically real and economically immature at the same time. The winning question is not "is it possible?" but "what exact job does it do better than the existing system, at what cost, against what incumbent?" The scale below plots estimated real-world deployment versus media attention for each technology.
Mostly promise or labsLarge real-world deployment
Positions reflect expert consensus as of 2024-2025. They shift — lithium batteries moved sharply right between 2015 and 2023.
Part H · sharp questions
If Moore's Law is ending, does that mean computing stops getting faster?
Not at all — it means the specific improvement mechanism of shrinking transistors is slowing, but computing capabilities are improving through other dimensions. Chipmakers are stacking chips vertically (3D NAND, HBM memory), using specialised silicon (GPU clusters for AI, TPUs for inference), and designing at the system level rather than the transistor level. Apple's M-series chips outperform comparable Intel chips not because they have smaller transistors, but because of better memory architecture, tighter CPU-GPU integration, and superior compiler support. Progress continues; the roadmap just looks different.
If computers are just 0s and 1s, why do they feel so flexible?
Because layers of abstraction compound. A transistor makes a bit reliable; logic gates combine bits; instructions combine gates into arithmetic and memory; programming languages combine instructions into human-readable commands. By the time you reach a browser or game engine, the primitive 0/1 choices are buried under millions of reusable conventions. Flexibility comes from stacking simple reliable layers until the top layer feels expressive — the same way individual words become literature.
Why do tech companies need to collect so much personal data? Could they work without it?
Advertising-funded platforms like Google and Meta could theoretically run on contextual advertising (ads matched to content rather than to the individual) as they did in the 1990s. The reason they don't is money: personalised advertising commands roughly 3-5x higher prices than contextual. The data also compounds — more data improves the model, which improves targeting, which increases advertiser willingness to pay. Subscription-funded models (Netflix, Spotify) require far less personal data, which is why their data practices are less controversial.
Why is cybersecurity so hard if encryption is mathematically strong?
Encryption protects data under specific assumptions: keys stay secret, endpoints are honest, software is patched, and users are not tricked. Attackers often avoid the math entirely. They phish a password, exploit an old library, bribe an insider, steal a session token, or compromise a vendor. Security is a system property, and the system includes tired people, old code, business incentives, and emergency exceptions. The 2020 SolarWinds breach — one of the worst in US government history — exploited trusted software updates, not broken cryptography.
Is AI genuinely "understanding" language, or is it doing something fundamentally different?
This is one of the most contested questions in AI. LLMs demonstrably pass many benchmarks that were assumed to require comprehension — the bar exam, medical licensing tests, complex reasoning chains. But they also fail in ways no human would, making bizarre errors on simple tasks or hallucinating facts with complete confidence. The practical stance is neither worship nor dismissal: treat them as powerful pattern engines that need verification, context, and good judgment around consequences. They do not have human experience, goals, or grounded perception by default, but they learn rich statistical structure that supports translation, summarisation, programming, and explanation in ways that look meaningfully cognitive.
What does HTTPS actually protect — and what doesn't it protect?
HTTPS (HTTP over TLS) encrypts the content of communication between your browser and a server, making it unreadable to eavesdroppers. It also authenticates the server: the certificate proves you're talking to google.com, not an impostor. What HTTPS does not protect: the fact that you visited google.com (visible in DNS lookups and IP routing), what you do once connected (visible to Google), data breaches at the server end, or malware on your own device. The padlock icon means the channel is private and authentic; it does not mean the destination is trustworthy.
Should we be worried about AI taking over? Is the existential risk case credible?
The existential risk case — that a sufficiently capable AI system might pursue goals misaligned with human survival — is taken seriously by a minority of mainstream researchers (including at major AI labs like Anthropic and DeepMind) and dismissed by many others as premature or unfalsifiable. The serious version of the argument is not about malevolence but about specification failure: a highly capable system optimising for a proxy goal might achieve that goal in ways that destroy everything humans care about as a side effect. Both things can be true simultaneously: present-day AI causes real harm and future AI might pose deeper risks.
How should I judge a new technology claim?
Ask what bottleneck it removes, what it costs, and what must change around it for adoption. A lab demo is not a supply chain, a prototype is not a business, and a benchmark is not a user habit. Look for constraints: energy, latency, regulation, manufacturing yield, training data, maintenance, liability, and integration with old systems. The future usually arrives through boring compatibility work after the spectacular demo. Fusion energy, for example, has been "30 years away" for 70 years — not because the physics is wrong, but because engineering at scale is a different problem than physics in a lab.
What does the cloud actually change for companies and individuals?
The cloud converts computing from equipment you buy into capacity you rent. That matters because demand is spiky: a startup can scale from one server to thousands without building a data centre. It also centralises power in hyperscalers that can afford custom chips, global fibre, cooling expertise, and security teams. The tradeoff is dependence: outages, pricing changes, data residency rules, and vendor lock-in become strategic risks. A 2021 Fastly outage took down Amazon, the Guardian, the Financial Times, and CNN for about an hour — the cost of centralisation made visible.