AI & LLMs

Foundation Models: The Beginning of the LLM Era

Stanford researchers have given a name to the models reshaping AI: foundation models, trained once on broad data and adapted to countless tasks. We explain what the term means, why a single model underpinning many applications is both powerful and risky, and what it means for organisations building with AI.

By EthicalHat Research · · 3 min read
Infographic: One foundation model, many applications

Last month, more than a hundred researchers at Stanford's new Center for Research on Foundation Models published a 200-page report: On the Opportunities and Risks of Foundation Models. Its most immediate impact was a name. The large pre-trained models behind so much recent progress (BERT, GPT-3, CLIP, Codex) now have a collective term: foundation models.

The name is deliberate. These models are not finished applications; they are foundations on which many applications are built. And like any foundation, their strengths and cracks show up in everything built on top.

What makes a model a "foundation model"

The report defines a foundation model as one trained on broad data at scale, usually with self-supervision, that can be adapted to a wide range of downstream tasks.

Two properties stand out:

Emergence. Capabilities appear that were not explicitly trained for. GPT-3 was trained only to predict the next word, yet it can translate, answer questions and write code when prompted. These abilities emerge from scale rather than being designed in.

Homogenisation. Increasingly, many different applications are built on the same few models. Sentiment analysis, search, chatbots and summarisers that once used separate bespoke models now share one pre-trained backbone.

A year of foundation models

2021 has made the pattern obvious:

  • CLIP and DALL·E (January) connect images and text. CLIP can classify images into categories it was never explicitly trained on, just from text descriptions.
  • GPT-3 has been adapted, through its API, into hundreds of applications since its launch in 2020.
  • Codex (August), a descendant of GPT-3 fine-tuned on public code, powers GitHub Copilot and translates natural language into code.
  • Open alternatives are appearing: EleutherAI released GPT-J in June, and several labs outside the US have announced very large models.

How adaptation replaces training from scratch

The typical AI project used to start with data collection and model training. With foundation models, the workflow looks different:

flowchart LR
  subgraph Before["Task-specific ML"]
    D1[(Task data)] --> T1[Train model A] --> A1[App A]
    D2[(Task data)] --> T2[Train model B] --> A2[App B]
  end
  subgraph After["Foundation model era"]
    Broad[(Broad data at scale)] --> FM[Foundation model]
    FM -->|prompting| B1[App A]
    FM -->|fine-tuning| B2[App B]
    FM -->|embeddings| B3[App C]
  end

Adaptation methods include:

  • Prompting: describing the task, possibly with a few examples, as with GPT-3.
  • Fine-tuning: further training on a smaller, task-specific dataset.
  • Using embeddings: taking the model's internal representations as features for search, clustering or lightweight classifiers.

For many organisations, the practical consequence is that the hard part shifts from building models to adapting and evaluating them.

The opportunities

  • Lower barriers. Teams without large labelled datasets can build useful NLP and vision features.
  • Better performance on many tasks, especially with limited data.
  • Faster iteration. Prototypes take days instead of months.
  • New capabilities, such as code generation and cross-modal search, that were impractical before.

The risks the report highlights

The Stanford report is as much about risk as opportunity:

  • Inherited flaws. Biases, errors and security weaknesses in a foundation model are inherited by every application built on it. Homogenisation turns one model's flaw into a systemic one.
  • Opacity. Emergent behaviour is hard to predict and hard to test comprehensively. Nobody fully knows what these models can and cannot do.
  • Concentration. Training frontier models requires resources only a few organisations have, which concentrates power and limits independent scrutiny.
  • Misuse. The same capabilities that help also enable disinformation, spam and fraud at scale.
  • Data and legal questions about training on web-scale data without explicit consent.

What organisations should do

  1. Evaluate for your use case. General benchmarks say little about performance on your data. Build your own test sets.
  2. Assume inherited risk. Test for bias, harmful outputs and failure modes relevant to your domain.
  3. Keep humans in the loop for consequential decisions.
  4. Think about dependency. Relying on a single external model provider has implications for cost, availability, data handling and control.
  5. Track the field closely. The pace of change is fast; today's best approach may be superseded within a year.

The bottom line

"Foundation model" is more than a label. It describes a structural shift in AI: from many narrow models built from scratch to a few broad models adapted everywhere. That shift brings real gains in capability and speed, and it concentrates both value and risk in the foundations. Building on them well means understanding both.

Our agentic AI security team helps organisations evaluate foundation models and adapt them to real business problems.

EthicalHat Research

The EthicalHat research team: penetration testers, SOC analysts and engineers writing about security, software engineering and AI.

Want an expert opinion on your own systems?

Email info@ethicalhat.com or call +1-844-838-4422. Our security engineers are happy to talk it through.