Expanding Brain Meme of the 5 Levels of AI Adoption.
Expanding Brain Meme of the 5 Levels of AI Adoption.

Here is how we climb up the ladder of success in using generative AI:

Level 1- Basic Prompting

  • Ask ChatGPT to write a professional email.
  • Create an image for a slideshow or marketing copy.
  • Write code, a web page, a Python function, a SQL statement, or CloudFormation YAML.
  • Use Perplexity, Consensus, one of the top custom GPTs for simple tasks like retrieval-augmented generation.
  • Find and use AI SaaS tools

Soon, you ask ChatGPT for something a bit more complex, and it doesn’t work very well. Now you are ready for…

Level 2 - Advanced Prompting

  • I wrote a blog post here. C-L-E-A-R is a good mnemonic acronym:
    • Contextualize - Specify a role or persona: “You are a copy editor with years of experience polishing articles and blog posts for web publication.”
    • Limits - Length; format like three bullet points; tone or style like concisely, or like a tech journalist, or only use facts from this text.
    • Elaborate and give Examples - Explain and provide as much detail and specifics as possible. Use chain of thought and other advanced prompting methods.
    • Audience - Identify the audience the response is addressed to, such as ‘explain like I’m 5’.
    • Reflect or Review - Ask ChatGPT to ask clarifying questions before answering, give itself space, such as “think step by step”, “make sure of x before answering”.
    • P-R-E-P-A-R-E-D is another. Or TCEPFT. Use what resonates with you.
  • Take a course on prompt engineering.
  • Explore other people’s prompts on the Web.
  • Make your own prompt repository on your PC, or use an app like AIPRM.
  • Use ChatGPT custom instructions.
  • Ask a custom GPT to write your prompts or improve them.
  • Start using integrations like Copilot; I use the Obsidian integration.

These integrations let you make templates, so you can create a keyboard shortcut to trigger a function like, “copy-edit this blog post”. Now you are moving on to Level 3!

Level 3 - Templates and Single-Prompt Apps

  • Develop your own custom GPT, with a single prompt and some data. Upload a document or knowledge base and set up your own retrieval-augmented generation to make a question-answering chatbot and talk to your knowledge base.
  • Develop an app that applies a prompt using a template using Jupyter or a Python script. For instance, here I ask ChatGPT to copy-edit a blog post and show markup of the before and after.
  • Use LangChain for different workflows. Try different LLM models like Google Gemini and Claude Sonnet through the API.
  • Create a one-turn agent that uses tools, picks the right tool and applies it, or have an agent write some SQL or Python code and run it, cautiously.
  • If you’ve been using an app for a while to perform sentiment analysis or other classification, and you have some good data, you can fine-tune a model on that data for better performance.
  • Try local models with Ollama or LM Studio, if you have a recent Nvidia GPU or Apple Silicon, and plenty of VRAM.
  • Custom GPTs and single-prompt agents that can use tools like web search, file search, even custom tools, are the simplest form of agent. They can be extremely effective, but eventually you run into limitations beyond a certain level of complexity. For instance the RAG in custom GPTs is great but you cannot customize it with metadata, complex chunking and reranking. If you want your agent to do more than a couple of steps based on user input, it quickly becomes unreliable. For relatively simple tasks like, make sure to include this but not that, you may need an editor agent in your pipeline. Then you move up to…

Level 4 - Multi-turn Agents, Pipelines, Chaining Prompts and APIs

  • An example pipeline chaining multiple tools:
    • Ask ChatGPT to write a poem about a current news event.
    • Ask Suno to take that poem and make a sad girl piano song.
    • Ask RunwayML to make some video clips from pictures of that news event.
    • Combine them all to make a music video, like the one at the bottom of this page.
    • If you have multiple APIs that integrate with e.g. Zapier or Workato, or if you can code with Python or ask Github Copilot to do it for you, you can write a program which takes the initial input description of the current event, and runs the entire pipeline end-to-end.
  • Or here is a notebook (video walk-through) that will:
    • Retrieve front pages of tech news.
    • Use ChatGPT to categorize headlines as about AI or not.
    • Use ChatGPT to extract topics from the headlines.
    • Identify the most popular topics today.
    • Use ChatGPT to summarize individual stories in the most popular topics.
    • Use ChatGPT to take the summaries and write a newsletter about today’s news in AI.
  • More complex multi-turn agents: Make a financial market question answering chatbot with a larger set of tools or a ReAct workflow.

These agentic workflows are like an aircraft autopilot that follows a set course, or a driverless train that follows a route along a track. AI performs a series of tasks, and possibly measures its progress and uses a feedback loop to stay on course toward a longer-term objective. But the decision process is mostly pre-determined and the degrees of freedom fall short of full autonomy.

We can of course, use coding tools like GitHub Copilot to build level 4 tools. Now we are approaching level 5.

Level 5 - Complex Agents

  • Try AutoGPT, a more complex ReAct workflow.
  • Try multi-agent workflows.
  • Create more complex ReAct agents that write and execute code.
  • Make an AI that creates a level 4 agent, writes code, and improves itself.
  • Make your own Devin, or Samantha, or Astra pie-in-the-sky AI.

Level 5 agents, that autonomously create solutions based on a general instructions about intent, are in their infancy. You can get advanced agents to do interesting things but they don’t currently always work consistently at a level you need in the enterprise. And there are years worth of Level 4 tasks that enterprises can automate with ‘sober’, as opposed to ‘pie-in-the-sky’ AI.

But low-code tools to create Level 4 agentic workflows using a combination of text-to-app AI, low-code configuration, and AI-assisted coding are well within reach. Once we have low-code tools, we’re only a short step away from having AI use that tool to create and improve workflows autonomously under loose human supervision.

Inspired by this post by Allie K. Miller, follow her for a ton of insight! (Ethan Mollick is good too.)