02. From Computing Fundamentals to Generative AI: The Fall of the 70-Year Language Barrier

Published 2026-09-14

Core Theme: For seventy years, humans learned the computer's language — coding. Today the computer has learned the natural language of humans. That transition has broken the greatest technological barrier in the history of administration, and in human history itself.

What is actually happening inside a computer?

There is a widespread belief that understanding how a computer works requires a technical education. It does not. The mechanism is simpler than almost anything else people accept without explanation — simpler than a car engine, and considerably simpler than the human body.

It is worth starting here rather than with artificial intelligence, because almost every misconception about AI comes from not knowing what the machine underneath it does. An officer who believes a computer "thinks" finds AI mysterious and slightly frightening. An officer who knows it is an extremely fast arrangement of switches finds AI remarkable but comprehensible — and is far better placed to judge what to trust it with.

You already know more of this than you think. Walk into a shop to buy a phone or a laptop and you will interrogate the seller in the language of computer architecture without noticing. "How is the display? How is the camera?" — you are asking about output and input devices. "How much RAM? How big is the storage?" — temporary memory versus permanent memory. "Is it an Intel Core i7?" — the processor, the part that does the actual calculating. Figure 2.1 assembles those familiar questions into the design every computer on Earth still follows.

Figure 2.1: The basic architecture of a computer (Von Neumann Architecture) — input devices, the central processing unit (CPU), primary memory (RAM/ROM), permanent storage (SSD/Storage), and output devices.

The parts, and what each one is for

Information gets in through input devices. Keyboards, scanners, cameras and biometric readers all do the same single job: they convert something in the physical world — a finger pressing a key, light falling on a lens, the ridges of a thumbprint — into electrical signals, and push those signals into the machine along the input bus. Nothing reaches a computer except as electricity. Everything else is interpretation.

The Central Processing Unit does the work. The CPU is the engine: an arithmetic logic unit that calculates, and a control unit that decides what to calculate next. It runs one loop, endlessly — fetch the next instruction, decode what it means, execute it — billions of times a second. There is no cleverness in the loop itself. The astonishing capability of modern computing comes almost entirely from doing something very simple an unimaginable number of times.

Primary memory holds what is happening now. RAM, or Random Access Memory, is fast memory that holds programs while they run, and it is volatile: cut the power and its contents are gone, which is why unsaved work disappears when the electricity fails. ROM, Read-Only Memory, holds the permanent elementary instructions the machine needs to start itself up at all — the small set of things a computer must know before it knows anything.

Secondary storage holds what must survive. The SSD or hard disk is the vault: files, databases and the operating system itself, kept whether the power is on or not. The distinction between RAM and storage is the distinction between the desk you are working at and the filing cabinet behind you, and it explains a great deal of everyday computer behaviour.

Buses move the data between them. The data, address and control buses are the electrical highways carrying billions of bytes a second between processor, memory and storage. They are invisible in the specification sheet and frequently the reason one machine feels faster than another with identical numbers on paper.

Output devices return the result to a human. Screens and printers convert electrical state back into something a person can see. The loop closes where it began: a human on one side, electricity in the middle, a human on the other.

The limitation that made AI need different hardware

Notice one property of this design, because it explains the entire modern chip industry. The CPU executes instructions in sequence — one after another, however fast. That is exactly right for the work computers were built for: a payroll run, a land register, an accounts ledger, where step two genuinely depends on step one.

Artificial intelligence does not work that way. It requires hundreds of thousands of multiplications to happen at the same moment, over and over. A sequential processor, however fast, is the wrong shape for that job. This is why AI runs on GPUs — chips designed originally for drawing game graphics, which happen to be built for doing enormous numbers of simple calculations in parallel. It is also why a company that made graphics cards became one of the most valuable enterprises in the world. The hardware did not change because AI was invented; AI became possible because that hardware already existed.

How does an operating system bridge human and silicon?

The hardware in Figure 2.1 understands only two things: current and no current, one and zero. No person can usefully instruct a machine at that level. It is not merely difficult; it would take a lifetime to write what a phone does in a second.

So a category of software exists purely to stand between the two: the operating system. Its job is abstraction — hiding an overwhelming amount of physical detail behind a small number of things a human can reasonably ask for. When you drag a file into a folder, you are not addressing magnetic sectors on a disk. You are asking the operating system to do that on your behalf, and its entire purpose is that you never need to know it happened.

Figure 2.2: The fundamental layers of an operating system — physical hardware, the OS kernel, system libraries (APIs), command interfaces (shells), and user applications.

Reading the layers from the bottom up

The physical hardware sits at the bottom — motherboard, CPU, RAM, storage, keyboard, screen. Everything above it exists to make it usable.

The kernel is the layer that matters most, and it does five distinct jobs. Its process manager divides the CPU's time between the hundreds of programs that believe they are running at once, switching between them so quickly that each appears continuous. Its memory manager guarantees that two applications using RAM at the same time do not overwrite each other — which is why one badly written program no longer takes the whole machine down, as it routinely did thirty years ago. Its file system imposes the folders and filenames we recognise on what is physically an undifferentiated expanse of storage. Its device drivers are the specific translators that let the kernel speak to particular hardware chips, which is why a new printer needs one. And its system call interface is the protective wall through which an application must ask the kernel for anything it wants — the reason no ordinary program can reach past the operating system and damage the hardware directly.

System libraries and APIs sit above the kernel, offering developers standard, ready-built facilities — the POSIX and C runtime libraries — so that nobody has to reinvent the act of opening a file.

Shells and command interfaces are where a human gives instructions: a Bash or Zsh terminal for those who type commands, or the windows-and-icons desktop everyone else uses. Both are doing the same thing at different levels of convenience.

User applications are the top layer — Chrome, Word, Excel, and the enterprise and government systems people use all day. This is the only layer most users ever see, and it works only because every layer beneath it is honouring its contract.

This layered structure is why the operating system is the first question anyone asks about a new machine — Windows when buying a computer, Android or Apple's iOS when buying a phone. You are not asking about the hardware. You are asking which abstraction you will be living inside.

Figure 2.3: Windows NT Architecture and Android Architecture — how two operating systems, desktop and mobile, manage the abstraction barrier between hardware and software.

Two operating systems, the same idea

Figure 2.3 places a desktop operating system beside a mobile one, and the point is the resemblance rather than the difference.

Windows NT, running on desktops, stacks user applications — Win32, UWP, web apps — on a Win32 subsystem and its APIs (kernel32.dll, user32.dll, ntdll.dll), which sit on the Windows NT kernel and executive with its process manager, virtual memory and I/O manager, which in turn sits on a Hardware Abstraction Layer. That bottom layer is the interesting one: the HAL exists so the kernel never has to know the differences between one x86 or x64 motherboard and another.

Android, running on phones, stacks Kotlin and Java applications on a Java API framework — activity manager, window manager, content providers — above the Android Runtime and native C/C++ libraries such as Bionic libc, SurfaceFlinger and Vulkan, above its own hardware abstraction layer of camera, audio and sensor modules, above the Linux kernel with its Binder inter-process communication and power-management drivers, and finally the ARM system-on-chip itself.

Different vocabulary, different companies, different decades — and structurally the same solution: layer on layer, each hiding the one below, with an abstraction layer at the bottom absorbing the messiness of real hardware.

The lesson for government technical officers is practical and immediate. A phone and a computer are two genuinely different operating environments, and software written for one does not run on the other. This is precisely why a government digital service must be built platform-neutral — on the web and on open APIs — rather than tied to any one platform. Build for a single platform and you have chosen which half of the country's citizens to exclude, whether or not you meant to.

Why did we need programming languages — and compilers?

An operating system makes hardware usable, but it does not make it do the specific clerical work an office needs. To compute a government employee's monthly salary, register a land title, or calculate a pension, someone must specify the steps — the algorithm — and specify them with total precision.

That is what programming languages are for, and the reason there are so many of them: Fortran, COBOL, C, C++, Java, Python and the rest, each designed for a different balance of speed, safety and human convenience.

Figure 2.4: The programming language — the abstraction barrier between human logic and the computer's binary signals (0 and 1).

The gap a programming language crosses is worth stating plainly. A policy in an administrator's head, a concept, a line of reasoning — none of it means anything whatsoever to a silicon chip. To get from one to the other, humanity settled on languages with rigid grammar, called syntax, and a small set of structures for expressing logic: conditions, loops, functions. The rigidity is not carelessness on the designers' part. It is the price of removing ambiguity, because the machine cannot ask what you meant.

Figure 2.5: The step-by-step pipeline of a compiler — from source code to machine code (binaries).

What a compiler does, stage by stage

A compiler is the translator that turns human-written code into machine instructions, and it works in stages because translation is not a single act.

It starts with source code: the text a person writes in C++, Java or Python. The first stage, lexical analysis, reads that text character by character and breaks it into tokens — separating words, punctuation and commands, much as a reader separates a line of text into words before understanding it. Parsing then checks those tokens against the language's grammatical rules, and this is the stage that produces the syntax errors every programmer knows: the compiler has found a sentence that is not a sentence.

Next the compiler builds an abstract syntax tree and intermediate code, restructuring the program's logic into branches so its actual flow becomes explicit rather than implied by the order of the text. Finally, code generation produces executable binary machine code — the ones and zeros the hardware can run directly.

Worth noticing: the compiler is itself a translation layer, doing for code what the operating system does for hardware. Abstraction on abstraction is not an accident of computing. It is the only known way humans manage systems too complex to hold in one mind.

What changed in 2022: from code to conversation

When ChatGPT appeared in late 2022, what broke was precisely that seventy-year barrier. It is easy to describe the change as a better chatbot, and easy to miss what actually moved.

Figure 2.6: From Code to Conversation — natural human language replacing complex coding languages as the computer's new operating interface.

For seven decades, the burden of translation sat on the human side. If you wanted a computer to do something, either you learned its language or you paid someone who had. That single fact determined who held computing power, and it is the reason the digital economy concentrated in the hands of people with a particular and fairly narrow training.

Generative AI moved the burden of translation to the machine. You no longer need C++ or Python; you describe what you want in your own language — in Sinhala, in Tamil, in English — and the system works out the rest. Natural language has become, in effect, the new operating interface.

The consequence is a redistribution rather than an improvement. Computing power that was once confined to programmers has passed into the hands of the management assistant at the Divisional Secretariat and the Grama Niladhari officer in the village. That is the shape of this revolution — and whether it amounts to anything depends on something the technology cannot supply: whether those officers know the capability exists, and know how to ask well. That is what the rest of this book is about.

What is the difference between traditional AI and generative AI?

A distinction that gets lost constantly, and costs organisations real money: "AI" and "generative AI" are not the same thing. They are two different technologies, built on shared foundations, doing different jobs. Buying one when you needed the other is one of the more common expensive mistakes in institutional technology.

Figure 2.7: Traditional AI versus Generative AI — classifying and predicting from existing data (Analytical AI) versus generating novel content and reasoning (Generative AI).

The shared foundation

Both rest on the same stack: machine learning, deep neural networks, and large volumes of training data. Neither is programmed with rules in the traditional sense; both learn patterns from examples. The divergence is in what they are asked to produce.

Traditional, or analytical, AI reads what exists

Analytical AI takes data that already exists, examines it, and returns a classification or a prediction. It forecasts — next month's tax revenue, the coming harvest, the likely number of dengue patients. It classifies and matches patterns — detecting a fraudulent document, filtering spam out of an inbox, matching a photograph against a driving-licence record. It tags content, applying stable classification labels across files and reports.

Its output is characteristically narrow and precise: a value, a label, or a probability. "The probability that this transaction is fraudulent is 92 per cent." That narrowness is a strength, not a weakness. Such an output can be measured against reality, audited, and trusted in proportion to its measured accuracy — which is exactly what you want for anything touching money, entitlement or safety.

Generative AI creates what did not exist

Generative AI synthesises the patterns in its training data to produce genuinely new content — text, images, audio, code, plans. These are synthesis engines, not lookup engines, and the distinction matters enormously in practice. It drafts circulars, legal opinions and working software from scratch. It generates voiceovers, illustrations and video. And it composes logic: multi-step plans and reasoning chains aimed at solving a problem it has not been shown before.

The corresponding weakness follows directly from the strength. Because a generative system produces new content rather than retrieving stored facts, it can produce new content that is entirely wrong while remaining perfectly fluent. Analytical AI gives you a number you can check. Generative AI gives you a paragraph you must check. Chapter 08 deals with what that means in practice, and it is the single most important operational fact about this technology.

The practical rule: if the question is "what does this data tell us?", you need analytical AI. If it is "produce something for me", you need generative AI. Institutions that reach for a language model to do forecasting, or for a forecasting model to draft communications, are using an excellent tool for a job it was never built to do.

What can generative AI actually do?

Generative AI is routinely introduced as a tool that drafts office letters, which undersells it roughly as much as introducing the internet as a way to send messages. Figure 2.8 maps the capability across twelve domains now being reshaped in global industry and public governance.

Read this as a map rather than a syllabus. Nobody needs all twelve. The value in scanning them is recognising which one touches the work in front of you, and noticing how far the frontier has moved beyond text.

Figure 2.8: The full scope of generative AI — a comprehensive capability map spanning 12 domains across global industry and public governance.

Text, knowledge and translation

The domain everyone knows, and still underestimated. Autonomous multi-document research means reading thousands of circulars, reports and cabinet decisions in mixed formats at once and producing an accurate synthesis — work that was previously measured in officer-months. Cross-cultural translation transmits meaning while preserving local and cultural nuance rather than substituting word for word, which is the difference between a usable Sinhala notice and an embarrassing one. And style transfer recasts the same fact into the legal register of a gazette or the plain, warm language a villager actually understands, which is a capability with real democratic consequences.

Visual and creative media

Image generation now composes scenes with coherent lighting and perspective, making public-awareness visuals producible without a studio. Long-form video generation maintains character and environment continuity across a sequence, which is what separates a usable educational video from a novelty clip. Three-dimensional and virtual-reality environments can be generated from a text description — city plans, archaeological sites — and interactive presentation experiences adapt lecture visuals in real time to an audience's responses.

Agentic and executive operations

This is the frontier, and the domain with the largest institutional implications. Systems now take end-to-end ownership of a task, completing procurement or payment workflows through software interfaces without a human in each step. They coordinate across platforms — website, database and email at once — to finish one piece of work. They improve themselves, analysing a failure and correcting their own instructions. They hold persistent memory and plan over long horizons, steering a months-long project while remembering its history. They verify their own output, auditing it before handing it to a human. And they collaborate with each other: a finance agent, a legal agent and an administrative agent negotiating among themselves to resolve a single file. Chapter 09 returns to this properly, because it changes what an institution is, not just what it can produce.

Code and technical artefacts

Full-stack development agents write software, test it, and release it to the cloud. AI-native architecture design shapes systems so computing resources are used at maximum efficiency. Bug detection finds security holes and patches them immediately. Legacy code revitalisation restructures decades-old COBOL or Java systems into modern cloud systems without breaking them — a capability of enormous relevance to any state whose critical records still live in software written before its officers were born. And intent-driven infrastructure builds entire cloud server networks from a plain sentence in English or Sinhala.

Business, policy and strategy

Hyper-personalised sequences send each citizen official notifications matched to their actual circumstances rather than a single generic template. Risk simulation stress-tests a decision against hundreds of thousands of alternative scenarios — economic crises, currency shocks, port strikes — before it is taken rather than after. Proactive compliance monitoring watches new gazettes and legal amendments as they appear and updates institutional regulations accordingly. Supply chain optimisation reroutes fuel, medicine and food distribution in real time as a disaster unfolds, which for an island economy is not a theoretical benefit.

Advanced science and material discovery

Generative molecular synthesis designs new drugs and proteins. Material lattice design creates new materials for more efficient solar cells and batteries. Automated hypothesis generation reads millions of research papers and proposes discoveries a human reader would never have connected. Hyper-local climate simulation predicts floods and cyclonic conditions at a precision that matters when the question is which village to evacuate.

Education and personalised mentorship

Socratic tutoring guides each student individually, at their own pace and along their own line of reasoning — the thing a hall of five hundred students structurally cannot provide. Automated curriculum creation composes textbooks and exercises adapted to the local environment. AI grading with early intervention identifies a learning weakness and acts on it before a student ever fails an examination, rather than four months later on a report card. Immersive learning lets a student experience a historical event or outer space rather than read about it.

Healthcare and precision medicine

Medical imaging diagnosis recognises disease indicators in X-ray and MRI scans within seconds. Personalised treatment planning sets drug dosage from a patient's genetic data and history. Clinical scribe automation writes the medical record while the doctor talks to the patient, returning the doctor's attention to the person in front of them. Patient triage manages emergency-room queues by severity at speed.

Embodied AI and physical robotics

Zero-shot motor control lets robots navigate and handle objects in environments they have never seen. Visuomotor policy generation converts camera data directly into motor movement for delicate surgery or assembly. Sim-to-real training raises robots in virtual worlds before releasing them into the physical one. Humanoid coordination puts human-form robots to work on heavy lifting in factories and warehouses.

Cybersecurity and synthetic defence

Generative red teaming has AI attack state systems deliberately and safely, to find weaknesses before criminals do. Polymorphic defence mutates system code and keys continuously so an attacker never faces the same target twice. Deepfake detection and provenance tracking identify forged voice recordings, videos and documents scientifically rather than by intuition. Threat intelligence synthesis combs the dark web and warns of attacks on critical infrastructure in advance.

Finance and economic simulation

Black swan stress-testing measures a national economy's resilience against severe global financial collapse or sanctions. Hyper-personalised wealth architecture shapes pensions and social safety nets to each citizen's actual income. Automated regulatory synthesis updates internal rules the moment international financial regulations — FATF, IMF — change.

Entertainment, culture and spatial reality

Three-dimensional preservation techniques such as NeRF reconstruct historic sites as complete navigable environments from a handful of photographs. Interactive narrative engines let public communication adapt live to a citizen's questions. Voice and persona cloning preserves the emotional depth of historical voices for future generations — a capability that is simultaneously a gift to cultural memory and, as Chapter 08 discusses, a serious hazard.

Where is AI heading: generative, agentic, embodied

The twelve domains above describe capability at one moment. The more useful question for anyone planning more than a year ahead is the direction of travel, and Figure 2.9 sets out the shift now underway: artificial intelligence moving from creating content to operating in the world.

Figure 2.9: The 3 phases of the AI transition — Generative AI (Creating), Agentic AI (Doing), and Embodied AI (Moving).

Phase one: generative — creating

The first phase is the one most people have experienced. Focus: generating text, images, music, video and basic code, shaping ideas, acting as an assistant. The defining property is where control sits — entirely with the human. One prompt, one answer, and a person decides what happens next. Nothing occurs that a human did not just ask for, which makes this phase both limited and comfortably safe.

Phase two: agentic — doing

The current frontier, and a genuine change in kind rather than degree. Focus: making decisions, composing multi-step plans, and completing tasks independently. An officer says, "check this permit application and prepare it for approval," and the agent queries the database on its own, computes the fees, produces an audit note, and presents the completed file for signature.

Read that sequence carefully. The human gave one instruction and received finished work, not a draft to work from. The role has moved from operator to approver — which is a substantial gain in throughput and a substantial shift in where responsibility sits. An approver who signs without reading has not saved time; they have transferred their judgment to a system that cannot be held accountable for it.

Phase three: embodied — moving

The horizon. Focus: sensing, moving and manipulating objects in the physical world. Humanoid robots, disaster-response drones, autonomous city-cleaning systems — intelligence joining physical labour rather than only cognitive labour. This phase is further off and less certain in its timing than the confident predictions suggest, but the direction is not in doubt, and its arrival reshapes categories of work that the first two phases leave untouched.

What are the five layers of the AI ecosystem — and where does Sri Lanka fit?

As NVIDIA's chief executive Jensen Huang points out, an AI system stands on five foundational layers. This is the most strategically useful diagram in this module, because it converts an overwhelming question — what should our country, or our organisation, do about AI? — into a specific one: which layer are we actually able to compete in?

Figure 2.10: The 5 layers of an AI system — Energy, Infrastructure, silicon Chips (GPUs), Foundation Models, and Applications.

Layer 1: energy and the power grid

The foundation, and the binding constraint on everything above it. Processing intelligence at scale demands enormous electrical current and equally serious heat dissipation. Without continuous, stable and increasingly renewable power, an AI data centre simply cannot operate. It is worth registering how unglamorous the primary bottleneck is: the limiting factor in the most advanced technology on Earth is electricity supply.

Layer 2: data centres and infrastructure

The AI factories — land, liquid-immersion cooling systems, and high-speed optical interconnects moving terabits of data between machines. This layer is civil engineering and plumbing as much as computing, and it is where a great deal of current global capital expenditure is going.

Layer 3: microchips and GPUs

The engine: tensor processors performing massively parallel computation, such as NVIDIA's H100 and B200 Blackwell chips. This is the layer that connects directly back to Figure 2.1 — the reason AI needed hardware of a different shape from the sequential CPU, and the reason a handful of firms now sit at the narrowest point of the global supply chain.

Layer 4: foundation models

The cognitive layer: large language models pre-trained on trillions of tokens — ChatGPT, Gemini, Claude, DeepSeek. Building one of these from nothing costs hundreds of millions of dollars and requires the three layers beneath it at industrial scale. Using one costs a few rupees a query.

Layer 5: applications

The only layer where technology becomes economic value. Health, agriculture, transport, manufacturing, citizen services. Nothing in layers one to four produces a single rupee of benefit for a single citizen until somebody builds something at layer five. This is worth stating as plainly as possible, because public discussion of AI concentrates almost entirely on the layers where the value is not created.

🇱🇰 A Reality Check for Sri Lanka

As public officers, we must recognise where our country's real capability lies — and read honestly, the five layers answer the national strategy question themselves.

Layer 3 is not available to us. Building a state-of-the-art chip fabrication plant costs on the order of twenty billion dollars, and two advanced chips alone can cost a million. Sri Lanka will not be fabricating chips at this moment, and pretending otherwise wastes attention.

Layer 4 is not our priority either. Pre-training a foundation model from zero costs hundreds of millions — billions, at the frontier — to reproduce something already available for a subscription fee. There is no return in it for us.

Layer 5 is where our opportunity actually lies. Taking the world's most powerful models and using them to eliminate file delays in our public service, accelerate land registration, and deliver fair service to the citizen — that is entirely within reach, today, at a cost measured in training and effort rather than billions. Strengthening green energy at layer 1 and local edge data centres at layer 2 matters too, as a question of digital sovereignty: it determines whether our data and our critical services depend on infrastructure we do not control.

There is a further point, and it is the one to hold on to. The world's powers will build the chips and the models. But the application that settles the land dispute at our own Divisional Secretariat, distributes Aswesuma benefits fairly, speeds up services in the Grama Niladhari division, or audits our own accounts — no foreign company will build those, because nobody outside understands the problem well enough to. Layer 5 is not the leftover. It is the layer where domain knowledge is the scarce input, and domain knowledge of Sri Lanka is something only Sri Lankans hold. Taking the finest models in existence and building applications that solve this country's problems with a product engineering mindset is, on that reading, less an opportunity than a national duty.

In the next chapter, we explore the Transformer brain inside these models — tokenization, and how A/L Combined Mathematics connects directly to AI.