Architectures and the buzzwords that matter
Governance pros must hold a credible conversation about architectures: transformer models (process inputs in parallel), multimodal models/LMMs (WHO 2024 ethics guidance), generative and specialised networks (CNN/RNN/GNN), and RAG for pulling in external information.
You will not build these, but you must hold a credible conversation about them to assess risk.
Transformer models (foundation models):
- Deep learning that learns context and meaning by tracking relationships in sequential data (words in a sentence)
- Find patterns mathematically → no need for large labelled datasets
- Process inputs in parallel → efficient training and inference
- Enable modern NLP and multimodal models
- Bonus uses → protein sequencing for medications, DNA sequencing
Multimodal models (LMMs):
- Inputs and outputs across image, video, audio and text (unimodal = one modality)
- NLP is a key component
- Use cases → weather forecasting, medical diagnoses, code generation
- WHO released AI ethics guidance for LMMs in 2024 → concerns about inaccurate or biased output affecting health decisions, poor training data, patient privacy
- Tools → Gemini, ChatGPT, ImageBind (Meta), Inworld AI
Other architectures: generative architectures create new text, images, audio or code from learned patterns (GPT, LLaMA, DALL-E 2). Specialised networks → know the acronyms: CNN (convolutional, images), RNN (recurrent, sequences), GNN (graph). RAG → retrieval-augmented generation lets a GenAI system pull in external information when answering, boosting LLM accuracy and relevance.