Module 1: Foundations of AI · BoK IV.A
Model face-offs the exam loves
Four head-to-head comparisons straight from the performance indicator: classic vs generative, proprietary vs open source, small vs large LMs, and language vs multimodal. Complex tasks may require multiple types of models working together.
Four head-to-heads, straight from the performance indicator. Each row is a potential question.
| Classic | Generative |
|---|---|
| Specific tasks, deterministic outputs | Creates new data instances resembling training data |
| Structured algorithms, fixed rules → decision trees, linear regression | Learns the underlying distribution of input data → generates novel, diverse text and images (GPTs) |
| Proprietary | Open source |
|---|---|
| Developed by specific organisations, restricted access and use | Publicly available to use, modify and distribute |
| Commercial focus → can limit transparency and independent auditing | Promotes collaboration, innovation, transparency → but risks in quality control and security |
| Small LMs | Large LMs |
|---|---|
| Millions to several billion parameters | Billions to trillions of parameters (GPT-4) |
| Efficient, often specialised for specific tasks | Highly complex → human-like text across a wide range of topics |
| Lighter resource consumption, easier to fine-tune narrowly | Huge training data and compute appetite, broader bias surface to manage |
| Language models | Multimodal models |
|---|---|
| Process and generate text only (GPT-4 in text mode) | Handle text, images, audio and video |
| Shine at translation, summarisation, conversational AI | Understand and generate across modalities → CLIP, DALL-E |
Exam flash
Expert quote worth remembering → complex tasks may require multiple types of models working together to complete the mission.
Key terms - quick answers
What is “Proprietary models”?
Models developed by specific organisations with restricted access; can limit transparency and auditing.
What is “Open source models”?
Models publicly available to use, modify and distribute; risk lies in quality control and security.
What is “Small language models”?
LMs with millions to several billion parameters, efficient and often specialised.
What is “Large language models”?
LMs with billions to trillions of parameters (GPT-4), complex generalists.