38 min read
Three ways to see it
You have completed seven lessons on the LLM landscape. You know how a model works, how the major makers differ, where open and closed split, what multimodal adds, when a small local model wins, how to pick well, and where the field is moving. The last step in any serious adult education is application. This lesson is a capstone: by the end of it, you will have produced a one-page model selection report for your own department or business, written in the voice of someone who has thought about the trade-offs, ready to defend in front of a CFO who is allergic to vendor jargon.
The Anand Kumar opener. Imagine you are sitting across from the chief financial officer of your organisation. They have one question for you: why should we sign this AI contract instead of that one. You have three minutes. Nothing else matters about the meeting. If you can answer in plain language, with three reasons rooted in cost, privacy, and capability, you keep the contract. If you cannot, the CFO sends you back to think harder. This exercise is the same conversation, run on paper, before the meeting. Doing it well today is what makes you the person the CFO actually calls when the moment arrives.
Step one: pick three real use cases. Not hypothetical ones. Three pieces of actual work your department already does that AI could plausibly help with in the next ninety days. Examples to think with, but pick your own. A federal ministry: drafting English notes for the cabinet division, translating internal policy circulars into Urdu, summarising long inter-ministerial committee meeting transcripts. A commercial bank: classifying customer complaints by category and language, drafting first-response letters in English and Urdu, summarising regulator circulars for compliance teams. A textile exporter: translating European buyer emails, drafting compliance documents for EU import rules, summarising weekly production reports for the chief executive. Write each of your three use cases in two sentences each. No fluff. State the input, state the output, state how often the task happens.
Quick check
Quick check: what makes modern AI different from a rule-based program?
The why-tree
Three frames for your final pick, listed in order of seriousness. Cost-first: pick the cheapest model that clears the quality bar, accept the privacy compromise where the data is not sensitive. Capability-first: pick the model that does the best work, accept the cost where the task warrants it. Sovereignty-first: pick an open-weight self-hosted model wherever the data is sensitive, even when it costs more in engineering effort. Different rows of your table will lean on different frames. State the frame in the report so the CFO knows you chose, not accepted.
Sources
Sources for further reading, gathered from the eight lessons. Anthropic Claude model card and responsible scaling policy on the Anthropic site. OpenAI GPT-5 system card and preparedness framework on the OpenAI research site. Google Gemini technical report and AI Studio documentation on ai dot google dot dev. Meta Llama 3 and Llama 4 model cards on llama dot meta dot com. Mistral, Qwen, and DeepSeek documentation on their respective sites. Microsoft Phi 3 and Phi 4 technical reports on the Microsoft research site. Google Gemma model card on ai dot google dot dev. LMSys Chatbot Arena leaderboard at chat dot lmsys dot org. Hugging Face open LLM leaderboard at huggingface dot co. Stanford HAI AI Index at hai dot stanford dot edu. Artificial Analysis at artificialanalysis dot ai. For Pakistan, the National AI Policy draft from MoITT, P@SHA AI position papers, NUST Atlas announcements, and the Tabadlab and PIDE notes on AI and Pakistani labour. Read three of these before your next AI procurement meeting and you will outclass the room.