Foundation Models

Work in Progress

This section is a draft: It will probably drastically change in the upcoming days.

Common Architectures

Architecture	Notes	Example
Direct Model Access	User Directly Exposed to Open Prompt	GPT3 API
Indirect Model Access	User Presented with Preset Options	Notion Continue Writing
User Prompt + Predefined Prompt		Replit Code Generation
"Stateful" Foundation Model Apps	By Putting History in Context	ChatGPT
"Retrieval Augmented QA"	Embedding Store + FM Default architecture for search/synthesis/QA use cases. Use K-Nearest Neighbors Search to find top N documents to query (retrieval), then use LLM to summarize (Synthesis)	Elicit (AI Research Assistant)
Model Chaining & Cascading	Recursively/Iteratively apply LLMs to create index data structures which can then be traversed with an LLM

Training

Useful Resources

Link	Description
A Foundation Models Primer - March 2023	Quick Technology Primer Early Products & Use Cases Prompting - Basics Common Architectures & Approaches to Building Foundation Model Apps Training, Fine Tuning & Alignment Cost Structure & Economics Productization of FMs Images, Videos, and Other Modalities MultiModal Models Market Structure and Landscape