Concept Contrast Fable

Large language models and foundation models: one lives in the language channel, the other can sleep anywhere

Conclusion first:A large language model is one specific implementation of a foundation model—trained specifically to understand and generate text. A foundation model is a larger concept, covering general models pretrained on various large-scale data, and multimodal tasks such as text, images, code, and proteins can all be part of it. Treating an LLM as the whole of foundation models is like treating one river as all the water resources in the world.

A large language model is one specific implementation of a foundation model—trained specifically to understand and generate text. A foundation model is a larger concept, covering general models pretrained on various large-scale data, and multimodal tasks such as text, images, code, and proteins can all be part of it. Treating an LLM as the whole of foundation models is like treating one river as all the water resources in the world.

Fable

Tianshu Pavilion had two kinds of disciples: one kind only understood speech, the other practiced the fundamentals first in everything they did

Tianshu Pavilion was the imperial court's highest academy for training strategists. Strategists came in many kinds: "language disciples" specializing in translating documents and legal codes from various states, "map disciples" skilled at interpreting maps and hydrology, and "gap-filling disciples" who could infer the original form from fragmentary old texts. For the first three years, everyone entered a common "Foundation Hall" for general training on large-scale texts, maps, and calendrical data.

The Foundation Hall's training method is very special: it does not target specific tasks, and it teaches neither language nor graphs. It does only one thing—let disciples build an overall perception of the world on vast amounts of unlabeled raw material, making the connections in their minds dense, stable, and broad in coverage. Once this foundation is laid, the disciples then enter their respective professional paths: language disciples learn grammar and writing, while graph disciples learn projection and topography.

The academy had an old scholar who loved to test new teachers with one question: this language disciple before you is a first-rate language expert, but what if you asked him to temporarily read a map? The teacher would usually say this disciple had never trained for that. The old scholar would say exactly—it's not that he can't read a map, it's that he was trained to use text as his channel to the world. His foundation spent three years on the general layer at the Foundation Hall, and his specialized path only opened one road called "language."

The key turning point in the story came when the foundation hall expanded enrollment for the third batch. Among this batch of disciples, some studied speech and graphs at the same time, and others, after building a foundation in the foundation hall, went on to train in the "multi-channel perception" specialty—able to switch among text, images, and climate change data. The old scholar drew a box on the enrollment register for this batch of new disciples, writing: the base is the foundation hall, and how far you go later depends on whether your specialty path is single-channel or multi-channel.

Once the court sent someone to the Tianshu Pavilion, saying they wanted to procure a batch of "language disciples" to help the six ministries translate documents. The purchase order only said "wise one (language type)." Later the Ministry of Revenue also asked for disciples who could both read ledgers and recognize images. Only then did the pavilion realize the procurement description was too narrow—they were procuring "language disciples" rather than "wise ones from the Foundation Hall," the latter being a large pool that can be trained in various capabilities.

The old academic director changed the wording used in external communications and set a rule: when describing talent produced by TianShu Pavilion, first write "foundational general education plus specialized path," then specify the channel type. If only document handling is needed, a language-track disciple suffices; if reading documents, examining maps, and analyzing hydrology are all required simultaneously, one must find a disciple who trained through multiple channels after the foundational hall.

Later, there was a consensus in the cabinet: the foundation hall could not simply be called the school of verbal disciples. The foundation hall is a training level; verbal disciples are only one outlet. One foundation hall can simultaneously produce verbal disciples, graph disciples, and gap-filling disciples; if the name of the foundation hall is narrowed, all other outlets are locked shut.

Large Language Model

A large-scale pretrained model focused on understanding and generating text. It is a subclass of foundation models, trained mainly on text data to handle language tasks.

Foundation model

A general model pretrained on large-scale multi-source data that can serve as the foundation for various downstream tasks. It covers multiple modalities such as text, vision, code and scientific data, and LLMs are one specific form of it.

Story Mapping

Foundation Hall: base model — a pretrained foundation layer on large-scale general data
The speech disciple: LLM—a specialized branch that uses text as its channel to the world
Graph disciples and others: base models for other modalities—vision models, code models, and so on
Three years of general training at the foundation hall: large-scale pretraining—building general representational capability
Multi-channel perception disciple: multimodal foundation model—a foundation trained on multiple modalities
The Ministry of Revenue procurement office said it was too narrow: using the term LLM in procurement would limit expectations for broader capabilities
The old academic's new rule: first declare the pretrained base, then add the channel type — the correct way to name things

Down to model selection and capability assessment

When choosing a model, first clarify whether the business needs pure text processing or cross-modal processing. If you only need conversation, translation, and summarization, an LLM is enough, and you do not need to pay extra compute costs for a multimodal foundation model. If you need to process text, images, or code at the same time, go directly to the product line for multimodal foundation models. In internal technical documents and external procurement forms, use "foundation model" as the name of a broad category, with LLM as one of its subclass labels. Do not treat the two as synonyms and swap them in documents—that will make it unclear to later people what kind of model capability you actually bought.

Large Language Model Foundation model