Pakistan turns 79 today, and one of the more serious things the country built this year is a four-level artificial intelligence curriculum for its civil servants. The Civil Services Academy in Lahore, working with atomcamp and two federal ministries, now runs AI training as a required part of officer education. I read the manuals. This post explains what is in them, and where I think they are strong and weak.
Happy Independence Day
To every Pakistani reading this, at home or abroad, happy Independence Day. 79 years is not a long time for a country. It is long enough to see what we are good at, and we are good at building things under constraint.
I want to mark the day by writing about something concrete rather than something sentimental. National pride is easier to feel when you can point at work. So let me point at some work.
What the four AI manuals actually are
The Government of Pakistan, through the Ministry of Information Technology and Telecommunication and the Ministry of Planning, Development and Special Initiatives, partnered with the Civil Services Academy and the training company atomcamp to produce a structured AI course for public servants. It is not a single workshop. It is a 4 level series.
The levels are AI 101, AI 201, AI 301, and AI 401. Each is a two-day course with its own manual. The first volume covers 101 and 201. The second volume covers 301 and 401, and was reviewed and edited by Dr Momina Moetesum.
The progression is deliberate. AI 101 teaches what AI is. AI 201 teaches how to govern an AI system you did not build. AI 301 teaches how to actually deploy one inside a ministry. AI 401 teaches how a state controls the hardware and data infrastructure underneath all of it.
According to reporting from the launch, AI 101 is now mandatory at the Civil Services Academy. Around 150 probationary officers went through a two-day programme, and 30 faculty members from civil service training institutions were trained as master trainers so the course can run without the original vendor.
That last detail matters more than the headline number. Training the trainers is the difference between a programme and an event.
AI 101 and AI 201, the foundation
AI 101 is a beginner course and it reads like one, in a good way. It defines AI as a system that learns from data and uses that learning to predict, recommend, or act. It walks through the three layers of an AI system, which are computing infrastructure, data and data management, and processing and modelling.
It also makes a point I wish more corporate AI decks made. AI is not new. The manual traces it through mathematics, statistics, programming, and logic, notes John McCarthy naming the field in 1955, and says plainly that what changed recently is accessibility, not the underlying idea.
The strongest line in AI 101 is the one it repeats. AI does not think and does not exercise judgement. It produces outputs from data, design assumptions, and the objectives it is given. Responsibility for the decision stays with the human official.
AI 201 turns that principle into a job description. It tells civil servants they are AI product owners and system architects, not developers. They define what the system should achieve. They ensure it matches law and public interest. They oversee vendors. Somebody else writes the code.
The module gives officers five questions to ask about any proposed AI system. Where is data stored. Is sensitive data sent to external cloud services. How is privacy protected. What happens if the AI makes an error. Is there a human review step.
Those five questions are better procurement guidance than most enterprise buyers in the private sector operate with.
AI 301, where the engineering starts
AI 301 is the point where I stopped skimming. It is a genuine technical manual, split into two lanes. Lane A is for governance and policy officers. Lane B is for technical delivery teams. Both lanes share the modules, but Lane B sections go into transformer internals, retrieval engineering, and deployment mechanics.
The course sets six control objectives that every implementation must satisfy before going live. Data sovereignty by design. Grounding outputs on approved sources. Role-based access with full auditability. Human approval gates on anything affecting rights, penalties, or entitlements. Protection against prompt injection and insecure output handling. Operational limits on tokens and cost.
Why smaller models, not bigger ones
The core argument of AI 301 is that model size is not a proxy for value, and governance is. It says frontier models operated by a handful of foreign firms are the wrong default answer for most Pakistani government workloads. Not because they lack capability, but because sending sensitive government data to them creates sovereignty risk, uncontrollable cost, and accountability gaps.
So the manual pushes small language models, which are compact models that can run on hardware a ministry can actually own. It gives a model selection matrix covering Llama 3, Phi-3, Gemma 2, and Qwen 2.5, with minimum VRAM figures, licence types, and Urdu capability ratings for each. VRAM is the dedicated memory on a graphics card, and a model has to fit inside it to run well.
The manual also insists that open weights do not mean unrestricted use, and that legal and procurement teams must review every licence before deployment. That is a small line with large consequences, and it is correct.
The recommended techniques are RAG and LoRA. RAG, short for retrieval augmented generation, connects a model to an approved document store so answers are grounded in real sources and can cite them. LoRA is a fine-tuning method that adapts a model's tone and format without retraining all of its weights. QLoRA adds compression so the work fits on a single government server.
This is a realistic stack. It is what a competent team with a limited budget would actually choose.
The Urdu problem nobody can skip
The section I found most valuable is the one on Urdu and multilingual AI, because it is the part no imported curriculum would contain.
Pakistan's government runs in Urdu and English. Citizens write in Urdu, in Roman Urdu, and in Sindhi, Punjabi, Pashto, and Balochi. The manual states directly that no existing open-source model handles that full range well.
It then lists the failure modes with their governance consequences. Code-switching inconsistency, where a model mixes Urdu and English inside one reply, produces informal official communication. Register mismatch, where the model writes casual Urdu in a formal notice, damages institutional authority. Transliteration inconsistency breaks record searchability, because the same word gets spelled differently across responses. Hallucination rates are higher in low resource languages than in English. Script confusion can garble Nastaliq against Arabic, Farsi, or Shahmukhi.
There is a cost dimension too. Under standard tokenization, Urdu script produces roughly 40 to 80 percent more tokens per word than English. Tokens are the units a model is billed and limited by. That means the same document costs meaningfully more to process in Urdu, and government budgeting has to account for it.
The manual's conclusion on regional languages is the honest one. For Sindhi, Pashto, and Balochi, current systems should be limited to classification and routing, not direct citizen communication, until better models exist and are tested.
I have not seen a clearer public statement of this constraint from any Pakistani institution.
AI 401, sovereignty as an infrastructure question
AI 401 moves up a level. Its guiding principle is that sovereignty in AI is not determined by which model you pick. It is determined by where computation happens, how data moves, and who controls the stack.
It splits sovereignty into three layers. Hardware sovereignty is where computation physically occurs. Data sovereignty is where data is stored, processed, and replicated. Operational sovereignty is who controls deployment, access, and model behaviour.
The manual then audits what Pakistan actually has. The National Data Centre serves over 100 federal agencies but has no GPU infrastructure yet. NADRA runs the most operationally advanced AI in the country through biometric verification, but it is not connected to any national AI platform. Raast, the instant payment system, processes millions of daily transactions and is described as the highest return use case for real-time fraud detection, yet it has no AI inference layer. There are more than 22 commercial data centres, concentrated in Karachi, Lahore, and Islamabad, but they are judged suitable only for lower sensitivity workloads without a sovereign cloud certification framework.
That is a fair assessment. It is also a more candid inventory than most official documents publish.
The phased plan, and what it would cost
AI 401 proposes a three-phase build.
Phase 1 runs from 2026 to 2027 at an estimated 60 to 150 million US dollars. It deploys 200 to 500 AI-grade GPUs at the National Data Centre, commissions an Urdu foundation model fine-tuned on Pakistani legal and government text, connects NADRA to that cluster, and puts fraud detection on Raast.
Phase 2 runs from 2027 to 2030 at 300 to 600 million US dollars. It expands to 1,000 to 2,000 GPUs, builds an inter-agency data exchange layer modelled on Estonia's X-Road, and launches a single government AI portal billed on usage.
Phase 3 starts in 2030. It imagines a dedicated AI factory of 5,000 to 10,000 GPUs drawing 50 to 100 megawatts, plus a shared sovereign compute pool for the wider OIC bloc.
The immediate actions listed for 2025 and 2026 need no new hardware at all. Enact the Personal Data Protection Bill with AI localisation clauses. Issue a sovereign compute policy. Commission an infrastructure audit. Build an AI procurement framework under existing public procurement rules. Run civil service AI literacy training.
What the manuals get right
Three things stand out.
First, the governance-first posture is consistent across all four levels. AI is advisory. Humans decide. That is stated in AI 101 and still stated in AI 401, which suggests one editorial mind held the line.
Second, the content is written for Pakistan's real conditions. It accounts for unreliable power, uneven broadband, scanned documents with poor OCR quality, and mixed-script corpora. It recommends hybrid keyword and vector retrieval specifically because government archives are messy. That is not generic material with local place names dropped in.
Third, the use cases are specific and appropriately constrained. Policy drafting with enforced citations. Tax audit risk flagging where no AI output may ever be the sole basis for enforcement action. Beneficiary eligibility screening where the officer decides. Each one names the deployment zone, the human gate, and the failure risk.
What still worries me
The plan assumes a legal foundation that does not exist yet. As of 2026, Pakistan still has no enacted comprehensive data protection law. The Personal Data Protection Bill remains a draft. PECA 2016 is a criminal statute, not a privacy framework. Every sovereignty control in AI 401 is easier to write than to enforce without that law.
Funding is the second gap. Phase 1 alone needs 60 to 150 million US dollars. Nothing in the manual explains where that comes from, and infrastructure plans in Pakistan have a long history of stalling at exactly this step.
Third, capacity is thin. The manuals ask civil servants to evaluate LoRA configurations, VRAM budgets, and retrieval architectures. Two days of training gives you vocabulary, not judgement. Without a permanent technical cadre inside government, ministries will keep depending on the vendors they are supposed to be supervising.
None of these make the curriculum less worthwhile. They are the reason it exists. But a training programme cannot substitute for a law, a budget line, and a hiring plan.
What you can do with this
If you build software in Pakistan, read AI 301. It is the clearest public statement of what the government will and will not accept in an AI system, which makes it the closest thing we have to a requirements document for public sector work.
If you are a civil servant, the five questions in AI 201 are enough to start. Ask them of every vendor.
If you work on Urdu or regional language models, the failure mode table in AI 301 is a research agenda. Somebody has to fix transliteration consistency and formal register in Urdu generation. There is no reason it should not be a Pakistani team.
79 years in, the most useful patriotic act available to most of us is to be very good at something and do it here. Happy Independence Day.
