Microsoft Azure OpenAI

Source: Azure OpenAI Service, OpenAI in Azure

Services

• GPT 3.5 - natural language
• Codex - code, development
• DALL-E 2 = text to image

Flow

• OpenAI studio > test > bring into code > call REST API
• Detects and mitigates against malicious use
• Protected by responsible AI and security access

1. Model Interactions

   • “Zero shot”: Text input given to AI for output. AI Model only uses inputs
   • “Few shot”: Text input with examples and AI, train on the fly, and AI output
     • Provide a bit of guidance to model
     • Seed information/words

Examples of Usage

• Copilot in Power Platform
• Microsoft Designer
  • Generate image from natural language to place into designs
• Minecraft
  • Codex to translate natural langauge commands into Minecraft commands to create objects
    • Built using “few shot” learning, by providing correct commands with natural language input
• Azure Cognitve search > search to retrieve relevant documents > documents parsed by GPT 3 > output sent to user of relevant results

Using Azure OpenAI Studio

• Use Azure OpenAI service, apply for access

Using the Playground

• Use AI for creative output like asking for branding ideas
  • Seed words
  • Provide prompt for what AI should do
  • Provide examples of what you want, like text example output you expect, adjust things like conciseness and format of answer for AI to learn using “few shot”
• Test prompts and results for accuracy
• Fine tune and custom models:
  • Choose base model with additional base data