Foundation Model
Foundation Model
What is a foundation model?
- A general-purpose model pretrained once at scale
- Versatile across translation, summarization, coding, and more
- The foundation of modern AI
What foundation models changed is not performance but the starting point of an AI project. It used to take six months and a dedicated team just to build one inquiry classifier, with data collection, labeling, training, and validation, and any change in business criteria meant starting over. Now that a pretrained model can be called through an API, the question a company asks shifts from “how do we train it?” to which model do we choose, and how do we connect it to our data? The nature of the cost changes too, from a one-time development expense to monthly usage fees, and the skills a team needs move from building models to choosing, connecting, and validating them.
Not all foundation models are alike. Closed models keep their weights private and are available only through an API, while open-weight models can be downloaded and hosted on your own servers. Models also differ in whether they accept text only or multimodal input such as images and audio. There are three layers for putting a model to work, and cost and difficulty climb from prompting to RAG to fine-tuning. Contrary to popular belief, bringing in company knowledge is a job for RAG, not fine-tuning. Fine-tuning is for locking in how a task is done: tone, output format, classification criteria. Miss this distinction and you can end up with a model that cost tens of millions of won to fine-tune and still cannot read yesterday’s updated price list.
The most common failure in the field is the belief that you connect it once and you are done. Models get new versions every few months and old versions are announced for retirement. When the version changes, a prompt that ran fine can quietly alter its output format and bring a downstream system to a halt. Only teams that pin the model version and keep a regression test set get through these swaps without incident. Another is the expectation that a general-purpose model will somehow handle your work too. Internal policies, product codes, and yesterday’s revised price list are not in its training data, so you get plausible but wrong answers. A foundation model you have not connected to your data is just a smart outsider.
John KimInstructor John Kim, author of "B2B Marketing Design" and CEO of Nanum Consulting
Sungil KimGlobal pitching expert from KAIST and Samsung Electronics who coached the winning teams at the UN Tourism and Select USA IR pitch competitionsIn 2021, Stanford’s Institute for Human-Centered AI (HAI) named these large general-purpose models “foundation models” and established the concept.
Consider a hypothetical case: Company A, an industrial valve manufacturer with 520 employees and KRW 180 billion (≈ US$134,670,100) in annual revenue. It receives 700 technical inquiries a month from overseas clients, and answering each one meant digging through drawings and spec sheets, so the first reply took 2.5 days on average. Someone proposed building a dedicated AI model, but the quote was KRW 800 million (≈ US$598,500) with a 12-month development period.
- Kim, the CTO and an executive director, rejected the quote and asked, "Can you guarantee this will beat commercial models a year from now?" Instead, Kim told the team to run a two-week model comparison test first.
- A technical sales team lead pulled 200 overseas inquiries from the past year and ran the same prompt through three models. Scored by the practitioners, the correct-answer rates came out at 71%, 64%, and 58%.
- The information strategy team connected the top-ranked model to 14,000 drawing specs and quote records through RAG and put off fine-tuning. Their reasoning was: “Feed in knowledge through search, and set the tone through prompts.”
- The legal team wrote a no-training-on-input-data clause and the data processing region into the contract, while the development team decided to pin the model version and build a 120-question regression test to run at every update.
- Three months later, the average lead time for a first reply fell from 2.5 days to 4 hours. At a monthly token cost of KRW 2.4 (≈ US$1) million, less than one-twentieth of the initial quote, the backlog of unanswered inquiries dropped 60%.