Glossary Category: Ai Machine Learning

A large language model (LLM) is an AI system trained on enormous amounts of text — books, articles, code, conversations — so it can predict, word by word, what should come next in a piece of writing. That simple prediction task, done at massive scale with billions of parameters, turns out to be powerful enough to draft emails, answer questions, summarize documents, write code, and hold a conversation that reads as genuinely coherent.

For a founder, the useful mental model is: an LLM is not a database and does not “look things up” by default — it generates the most statistically plausible continuation of your prompt based on patterns learned during training. This is why LLMs are excellent at fluent, flexible language tasks but can still state incorrect information confidently (see Hallucination) unless grounded with real data through techniques like RAG.

Most Philippine startups never train an LLM from scratch — that requires enormous compute budgets few companies anywhere can justify. Instead, they build on top of an existing foundation model via an API (Claude, GPT, Gemini, and others), paying per use, and customize behavior through prompt engineering, fine-tuning, or retrieval rather than retraining the whole model.

🇵🇭 Philippine Example

In June 2026, the Department of Information and Communications Technology (DICT) and Google Cloud announced a multi-year partnership to deploy Google's Gemini LLM to more than 50,000 Philippine government workers through the Gemini Enterprise platform, with plans to expand to over 200,000 officers — one of the largest verified real-world LLM rollouts in the country to date.

Related Terms

Prompt engineering is the practice of designing the instructions, examples, and context you give an AI model so it produces the output you actually want. In practice this means being specific about the task, format, and constraints; providing a few examples of good output when the task is unusual; and iterating — testing a prompt, seeing where the model gets it wrong, and adjusting the wording or structure until the results are reliable.

The nuance beginners miss is that prompt engineering isn’t about finding secret “magic words.” It’s closer to writing a clear brief for a very capable but literal-minded new hire: the more precisely you specify the goal, audience, tone, and edge cases, the better the output, and the less you’ll need to fix afterward. As models have gotten more capable, some of the finicky tricks from a few years ago matter less — but clear, specific instructions still consistently outperform vague ones.

For teams shipping AI features rather than just chatting casually, prompt design directly affects cost and reliability at scale: a poorly structured prompt can produce inconsistent output across thousands of requests, while a well-tested one behaves predictably — which is why it has become a real, in-demand skill rather than a novelty.

🇵🇭 Philippine Example

KDCI, a Philippine outsourcing company founded in 2011, has built an AI-native division (KDCI.ai) that places Filipino professionals trained specifically in prompt engineering and large-language-model workflows with international clients — a real, verifiable example of prompt engineering becoming a distinct, paid skill set inside the Philippine outsourcing industry.

Related Terms

Vibe coding is a style of building software where a developer (or non-developer) describes what they want in natural language, and an AI model generates the working code, with the person guiding further changes through follow-up prompts rather than by editing syntax directly. The term was coined by AI researcher Andrej Karpathy in February 2025 and spread quickly enough to become a widely recognized word within a year.

The practice has matured noticeably since it first appeared. Early “vibe coding” often meant accepting AI-generated code with little or no review — fine for a quick prototype, risky for anything handling real users or real money. By 2026, the more disciplined version of the same idea involves structured specs, automated testing, and layered human review before code ships, which is also where the related, more rigorous term Agentic Engineering comes in.

For a first-time founder, the honest nuance is: vibe coding is genuinely useful for speeding up early prototypes and MVPs, but skipping code review entirely creates real risk once a product has paying customers or handles sensitive data — security flaws and technical debt introduced this way don’t announce themselves until something breaks.

🇵🇭 Philippine Example

No single verified Philippine startup built entirely through vibe coding was found through research for this entry, so rather than invent one: Filipino developer communities and coding bootcamps have visibly picked up mainstream vibe-coding tools (AI coding assistants built into editors and terminals) at the same pace as the rest of the world, since these tools are usage-based and globally available rather than region-restricted.

Related Terms

Agentic engineering describes a more structured evolution of vibe coding: instead of a developer prompting an AI for a snippet and copy-pasting the result, one or more AI agents are given a goal and largely handle the loop of writing code, running tests, and even deploying it — while a human sets direction and reviews what actually ships, rather than typing most of the code by hand.

The distinction from vibe coding is discipline and oversight, not just autonomy. As the practice has been described by its own popularizers (including Karpathy, who used the term to describe the “mature” version of vibe coding by mid-2026), the developer’s job shifts from writing code line-by-line to orchestrating agents, defining guardrails, and being accountable for whatever the agents produce — closer to being a technical lead for a team of very fast, very literal-minded junior engineers than to being a solo coder.

For a Philippine founding team without a large engineering budget, the appeal is obvious: a small team can credibly attempt what used to require many more engineers. The real nuance is that “agentic” does not mean “unsupervised” — the human review, testing discipline, and security judgment still have to exist somewhere, or the speed gained up front turns into cleanup cost later.

🇵🇭 Philippine Example

A general, honest note rather than a specific invented example: the Philippines' large business-process-outsourcing sector, which employs a significant share of the country's tech-adjacent workforce, is already publicly grappling with what more autonomous "agentic AI" tools mean for jobs and workflows, as reported by outlets like the Philippine Daily Inquirer in 2026 — a sign the concept is landing in the local market even where a specific agentic-engineering software team hasn't been individually verified.

Related Terms

Retrieval-Augmented Generation (RAG) is a technique that pairs a language model with a search step. Instead of relying purely on what the model memorized during training, the system first retrieves relevant passages from an external source — a company’s internal documents, a product manual, government regulations — and feeds those passages into the model alongside the user’s question, so the answer is generated with real, current source material in front of it rather than from the model’s general training alone.

This matters for two practical reasons. First, it substantially reduces hallucination, since the model has actual text to reference rather than having to guess. Second, it means a business doesn’t need to retrain or fine-tune an entire model every time its internal information changes — updating the searchable document set is enough, which is far cheaper and faster.

The nuance many beginners miss is that RAG only helps as much as the retrieval step is good — if the system retrieves the wrong or outdated document, the model will still confidently answer based on bad information, just now with the appearance of a cited source. Good RAG systems require real attention to how documents are organized, updated, and matched to queries, not just plugging in an LLM.

🇵🇭 Philippine Example

No specific Philippine company's internal RAG architecture could be verified through research for this entry, so rather than assign the label to a system whose technical details aren't public: RAG is the standard approach any Philippine business would reach for if it wants an AI assistant to answer accurately from its own documents — HR policies, compliance manuals, product catalogs — instead of the model's general internet-trained knowledge.

Related Terms

Fine-tuning takes an already-trained foundation model and continues training it on a smaller, focused dataset — customer support transcripts, legal documents in a specific format, a particular tone of voice — so the model’s behavior shifts toward that narrower task without starting from scratch. It’s a middle ground between using a general-purpose model as-is (prompting) and building an entirely new model from zero.

The nuance a beginner often misses is that fine-tuning is not always the right tool. It requires real data preparation, some machine learning expertise, and ongoing compute cost, and it doesn’t teach a model new factual knowledge as reliably as people assume — for grounding a model in current, specific facts, retrieval (RAG) usually works better and is far cheaper to update. Fine-tuning tends to be worth it for narrow, high-volume tasks where consistent style or format matters more than fresh factual lookup — for example, classifying support tickets in a company’s exact categories, or writing in a very specific brand voice repeatedly.

For a startup deciding between prompting, RAG, and fine-tuning, the practical order is usually: try clear prompting first (cheapest), add retrieval if the task needs current or private facts, and reach for fine-tuning only once a specific, repetitive task has proven itself worth the extra investment.

🇵🇭 Philippine Example

Philippine AI-recruitment and salary-benchmarking sources list fine-tuning as one of the named, in-demand skills (alongside LLM and RAG work) that local AI engineers are expected to know, reflecting real market demand for this skill in the Philippines even without a single named company's fine-tuning project to point to.

Related Terms

Every AI model has a limit on how much text — measured in units called tokens, roughly pieces of words — it can hold in view at one time while generating a response. That limit is the context window. It includes everything: your instructions, any documents you paste in, and the ongoing conversation history. Once a conversation or document exceeds that limit, older content has to be dropped or summarized for the model to keep working.

Larger context windows are genuinely useful — they let a model read an entire contract, codebase, or long meeting transcript in one pass instead of needing it chopped into pieces. But the nuance beginners often miss is that a bigger context window doesn’t guarantee a model will use all of it equally well; models can still lose track of details buried in the middle of a very long input, a known effect sometimes called “lost in the middle.” Cost and response speed also generally scale with how much context is fed in, so bigger isn’t automatically better for every task.

For founders building on top of AI APIs, context window size is a real, practical constraint to plan around — not just a marketing spec — since it determines whether a use case, like reviewing a full legal document at once, is even possible with a given model.

🇵🇭 Philippine Example

A general point rather than an invented specific: Philippine legal, compliance, and policy teams reviewing lengthy government circulars, contracts, or regulatory filings are a natural beneficiary of larger context windows, since these documents can run to dozens of pages that previously had to be broken into smaller chunks for AI review.

Related Terms

Hallucination is the term for when an AI model generates information that sounds plausible and is stated confidently, but is factually wrong, fabricated, or unsupported — a made-up statistic, a citation that doesn’t exist, a wrong date, or a confidently incorrect answer to a factual question. It happens because a language model’s core job is predicting plausible-sounding text based on patterns, not verifying truth against a database, unless it’s specifically given a way to check real sources.

The nuance beginners frequently miss is that hallucination isn’t a rare bug that gets fixed as models improve — it’s an inherent property of how these models generate text, and it shows up even in very capable, current-generation models, especially on niche facts, exact numbers, or recent events outside their training data. The main practical defenses are grounding the model in real source material through retrieval (RAG), keeping a human in the loop for anything consequential, and being skeptical of specific factual claims an AI produces without a checkable source.

The stakes scale with what the AI’s output is used for — a hallucinated fact in a casual brainstorm is harmless, but the same failure mode inside a system giving official guidance to the public is a real governance concern, which is exactly why large-scale public-facing AI deployments put real effort into grounding and review before launch.

🇵🇭 Philippine Example

As the Philippine government moves to deploy AI agents through DICT's Gemini Enterprise partnership to answer citizen questions on things like business registration and disaster relief guidance for hundreds of thousands of officers, hallucination is precisely the risk such a rollout has to guard against — wrong information in those specific contexts carries real consequences, which is why grounding and review matter, not because any confirmed hallucination incident has been reported in that program.

Related Terms

An AI agent is a system built on top of a language model that can do more than respond to a single prompt: it can break a goal into steps, decide which tools or data sources to use, take an action, observe the result, and re-plan — repeating that loop until the goal is met or it needs to stop and ask a human. The core distinction from a plain chatbot is behavioral: a chatbot answers when asked; an agent pursues an objective.

The nuance beginners miss is that in 2026, “agent” is used loosely in a lot of marketing — a simple chatbot with a slightly expanded feature set is sometimes labeled an “agent” even though it doesn’t actually plan or act autonomously. A genuine agent typically needs guardrails: clear boundaries on what actions it can take without a human approving first, since letting a system act independently is a meaningfully bigger risk than letting it draft a suggested reply for a person to review.

For founders, agents are attractive because they can automate multi-step workflows, not just single questions, but the honest tradeoff is that more autonomy also means more that can go wrong unsupervised — which is why serious agent deployments, including government ones, are typically rolled out with human-in-the-loop review for anything consequential.

🇵🇭 Philippine Example

In June 2026, the Department of Information and Communications Technology (DICT) announced it will build and manage AI agents on Google's Gemini Enterprise Agent Platform to handle citizen-facing e-government tasks — answering questions on business registration, community health center scheduling, and disaster relief guidance in local languages — a real, verified, large-scale Philippine deployment of AI agents rather than a hypothetical one.

Related Terms

The Model Context Protocol (MCP) is an open standard, introduced by Anthropic in November 2024, that defines a common way for AI applications to connect to external tools, files, and data sources — things like a company’s database, a search engine, or a local codebase — without a developer having to write a brand-new, one-off integration for every single combination of AI app and data source.

The commonly used analogy is that MCP is like a USB-C port for AI applications: before a shared standard, connecting any given device to any given AI system meant custom wiring; with a shared protocol, any MCP-compatible tool can plug into any MCP-compatible AI application. An MCP “server” exposes a particular tool or data source, and an MCP “client” inside the AI application talks to it in a standard way, translating requests and responses between the two.

The nuance worth knowing is that MCP itself doesn’t make an AI smarter — it solves a plumbing problem, not a reasoning problem. Its real significance is that it was adopted quickly beyond Anthropic itself, including by OpenAI and Google DeepMind, which is what makes it genuinely useful: a tool built to speak MCP can work across multiple AI platforms instead of being locked to just one.

🇵🇭 Philippine Example

No specific Philippine company publicly building or adopting MCP integrations could be verified through research for this entry, so instead of naming one speculatively: Philippine developers and outsourcing teams building custom integrations for AI-assisted tools for overseas clients are a natural audience for MCP, since it lets a connector they build once work across multiple AI platforms rather than being rebuilt per client.

Related Terms