Glossary Category: Ai Machine Learning

A foundation model is a large AI model trained on a broad, massive dataset — text, code, images, and more — using self-supervised learning at scale, producing a general-purpose base that can then be adapted to many different downstream uses rather than being built for just one task. Models like Claude, GPT, Gemini, and Llama are all foundation models: instead of training a separate model for translation, one for summarizing, and another for coding, one large model learns broadly enough to do reasonable versions of all of them.

The adaptation step matters as much as the base model itself. A foundation model becomes useful for a specific business through prompting (giving it clear instructions), retrieval (RAG — grounding it in your own documents), or fine-tuning (further training it on a narrow dataset) — three different ways of specializing the same underlying model for a specific job.

The nuance most founders should internalize early: training a foundation model from scratch costs an amount of compute and data that only a handful of companies worldwide can justify. Virtually every startup, in the Philippines or anywhere else, builds on top of an existing foundation model through an API rather than attempting to train its own — the real competitive work happens in what you build around the model, not in building the model itself.

🇵🇭 Philippine Example

The Philippine government's DICT-Google Cloud partnership, announced June 2026, builds its public-sector AI agents and services directly on top of Google's Gemini foundation model via the Gemini Enterprise platform — a real, verified example of a large Philippine institution adapting an existing foundation model rather than training one of its own.

Related Terms

Inference is the process of running a trained AI model to produce an actual output — answering a question, generating an image, writing code — as opposed to training, which is the earlier, one-time (or occasional) process of teaching the model in the first place. Every single time someone sends a message to an AI chatbot or an app calls an AI API, that’s an inference request.

The distinction matters commercially more than it might first appear. Training a model is a large but largely one-off cost. Inference, by contrast, happens continuously, every time the model is used — so for a company building a product on top of someone else’s foundation model via API, inference cost is an ongoing, per-use expense that scales directly with how many people use the product and how much text they send and receive each time.

For a founder doing basic unit economics, inference cost is a genuine line item to plan for, not an afterthought — a feature that calls an AI model on every user action can look free to build and then turn out to have real per-user cost at scale, the same way cloud hosting costs do, just billed per token instead of per server-hour.

🇵🇭 Philippine Example

A general, honest note rather than a specific invented company: Philippine startups building AI-powered features typically pay foundation-model providers directly for inference by usage (per token processed), which means inference cost shows up as a real, variable line item in their unit economics rather than a fixed cost — a pattern true of AI-using startups broadly, not something unique to a single named Philippine company that could be verified here.

Related Terms

Training data is the raw material an AI model learns from before it’s ever used to answer a question or generate an output. For a large language model, this typically means enormous amounts of text pulled from sources like websites, books, code repositories, and licensed datasets; for a more specialized model, it might mean a much smaller, purpose-built set of examples relevant to one specific task, like graded student assignments or labeled customer support tickets.

The quality, diversity, and accuracy of training data directly shapes what a model can and can’t do well. A model trained mostly on English text will be weaker in other languages; a model trained on outdated information can’t know about anything more recent; and biases present in the training data tend to show up in the model’s outputs, since the model is fundamentally learning statistical patterns from what it was shown, not independently reasoning about the world from scratch.

The nuance beginners often miss is that “garbage in, garbage out” applies at real scale here — a huge but low-quality or narrow dataset can produce a worse model than a smaller, carefully curated one, which is why serious AI teams spend significant effort on data cleaning, filtering, and documentation, not just data collection volume.

🇵🇭 Philippine Example

GradeChum, built by Philippine company CodeChum Software Solutions Inc. and featured among Filipino startups at GITEX AI Asia 2026, uses AI to automatically evaluate handwritten and digital student assessments — a product category that inherently depends on training data made up of large volumes of real, previously graded student work to learn what an accurate evaluation looks like, even though the company's specific internal training data pipeline is not itself publicly documented.

Related Terms