AI Notes for Developers

Practical AI for Developers

Command Line

• Warp command line tool w/ AI capabilities

Coding

• Co-pilot and code augmentation
  • Initially from OpenAI’s fine-tuned GPT-3 version called Codex
  • AI-powered Code Generation, a smarter text expansion
    • Trained on open source code from Github
    • Reinforcement learning means it gets better with time, but maybe just incrementally
  • Prompt Engineering
  • Surprising use cases
    • Giving enough in the comments for generate decent code. Still needs some analysis
  • Giving a template for unit tests.
    • A lot of times, developers skip this step. By having an outline for it, the time to write tests go down
  • AI-powered tools as paired programming
    • Great alternative to the question and answer sites like StakeOverflow
• Limitations
  • Answers are generative, meaning what you use in the prompt and how the model was training will determine the output
  • If you are looking for structured data, Google and Wolfram|Alpha do a much better job
    • 
    • Code that’s complexity is still hard to write
      • Can get very buggy
      • Training models in a rut w/ a Glitch Token
      • SMEs need to be the adult in the room
  • Legality

    A class-action lawsuit against GitHub, OpenAI, and Microsoft claims that the training of Codex violated open source licensing agreements. The outcome could have legal implications for models that generate text, images, and other media as well.

    • We Ask a Lawyer about Github Copilot
  • NeoVim plugin
• Other tools
  • AlphaCode
  • Ebay’s low-code AI
• Low-code scripting
  • Alternative to writing your own code for general scripting
  • For example, Google App Scripts

Researching tools

• Feedback from Chuks about learning new topics
• Notemaking - https://fireflies.ai/
  • Summarizing meetings, take action items, etc.
• https://www.producthunt.com/topics/artificial-intelligence
• PR Reviewer powered with GPT models

Reading

• Ghostreader from Readwise Reader that helps augment reading
  • Like Shortform or Blinkest, as a pair next to you
  • Doesn’t replace reading, but good at summarizing what you are reading and some questions to help with critical thinking
• Reading resumes
  • Could it give you resume quality back?
• Summarizer
  • Brave Search summarizer
  • Bing search summaries
  • Readwise Reader ghostreader

Writing

• Emails
• Resumes
• Cover Letters
• Slack messages?
• Brainstorming
• Wiki Pages, blog posts, meeting agenda, social media, PRs, Job descriptions, sales emails, poems, pros and cons lists, outlines
  • Notion AI can handle a first draft
• Technology matricies
• Tools
  • Notion AI
  • Lex
• Obsidian forum has [users ideating what AI could look like](AI Assisted Q&A / Chatbot / Text Generation) for the app
  • Learning to summarize with human feedback
    • 

Reflecting on different applications for AGI

• Deep Learning’s newsletter, “The Batch”, reflected on the general uses and stories for 2022.
  • Synthetic Images
    • New technologies
      • [[OpenAI]]‘s DALL·E 2
      • Stable Diffusion
    • Prompt Engineering arises
      • Places like PromptBase are an open marketplace for text strings that generate interesting output
    • Limitations
      • Training set based off online images, which have inherit biases (see bad algorithms)
        • For example, sexualizing women unfairly
    • Concerns
      • Concern with legality
      • Concern with artists

        ArtStation, an online community for visual artists, launched its own text-to-image features. Many artists, feeling threatened by computer programs that can reproduce an artist’s hard-won personal style in seconds, boycotted the website.

    • Things I still don’t understand
      • GAN - Generative Adversarial Networks
  • Prompt Engineering is Probably More Important Than You Think
    • Current LLMs provide windows into their process through prompts
    • In the future, we may have something more precise
    • In the interim, we have prompts to guide us through the alien silicon
• Knowledge Base Q&A