Curiosity is all you need: How GenAI will Transform Creativity

AI empowers everyone to build their own software

“Everyone is a programmer now - you just have to say something to the computer. Human language is the new programming language.” - Jensen Huang, CEO of NVIDIA

GenAI breaks the barriers to creativity. I’m not referring to the creative professional fields like design and art, although it will undoubtedly have a substantial impact there too. What I mean is creativity in its broadest sense — the act of envisioning and bringing an idea to life, your own idea, independently.

Curiosity is all you need

For me personally, LLMs have sparked a new enthusiasm and energy for learning and creation. Any topic, any thread, any thought, any idea has the potential to be realized if you are curious enough to see it through.

In high school, I was given a science award for being the most inquisitive. I would often annoy teachers with my questions - many of which were ignorant, sometimes repetitive, and often distracting - but they were an essential part of the way I learned. My questions were my brain trying to form and validate my mental model of a concept. LLMs like GPT are teachers with limitless energy, infinite patience. It doesn’t judge, it comes along with you on your learning journey and it never tires. To get the most value out of an LLM like GPT, you really have to probe it. You have to experiment with the technology itself and see what’s possible.

Here are a few ways I have prompted GPT to help me understand an idea:

1. Ask it to provide an example(s) (a simple tactic that can you can easily continue to modify to simpler and simpler examples)
2. Ask it to create an analogy or metaphor (GPT helped me understand the difference between policy driven and reward driven reinforcement learning algorithms through the lens of cooking)
3. Ask it to explain the context or framework in which this concept is situated. (a strategy I often use when I’m trying to assess the scope and depth of the concept)
4. Ask it to validate or build on a partially-formed hypothesis (a tactic that I use when the picture is beginning to clear but I need help bringing it home; sometimes I use the tactic “complete this sentence” where I leave an ellipsis for GPT to fill in)
5. Ask it to translate between the language you know and the language you’re trying to learn (whenever you’re learning a new concept, it is often the case you don’t intuitively have the language to describe it, GPT is incredible at helping you translate ideas between domains)
6. Give more context about yourself (this one’s self-explanatory — tell GPT more about why you’re learning and any relevant information about yourself, so it can tailor its response).

From three weeks to three days

As an example of how different it is for a non-technical person to create something now compared to the pre-GenAI era: In 2019, I built my personal website using Gatsby, following a step-by-step tutorial and browsing Stack Overflow every time I encountered an issue. It took me about three weeks with lots of head banging to reach a point where I felt it was presentable.

Contrast that experience to this website which I built using Hugo. It took me three days with no tutorial, no instructions, just persistent prompting with Github Copilot. I customized an existing theme, streamlined the deployment process with Github actions and also added fun little elements like the Typed.JS animation on the homepage.

Over the past six weeks, I’ve developed eight prototypes, starting with basic scripts and culminating in a full-stack news summarization app that utilized reinforcement learning with human feedback, a React frontend, a Python backend, an SQL database, and integrations with several external APIs. I’ve described some of my prototypes in the following posts: Three ways I played with OpenAI’s APIs and Experiments in Supervised and Reinforcement Learning.

My tools were curiosity, patience and skepticism. I’ve already written about curiosity above. Patience is required to iterate through a prompt and its generated response until you achieve what you’re looking for (and it’s far faster than sifting through Stack Overflow). A healthy skepticism is required to ensure you don’t blindly follow GPT’s instructions because it does make mistakes and can lead you astray.

Prompting a process: A DIY approach to building software

For each of my projects, I took a similar process. Each of these steps builds critical context for it to help you move forward into the next step.

• Brainstorm: Begin by generating a basic idea for your project, such as “I want to create X.” During this phase, engage with GPT to ask you follow-up questions one at a time to help you develop your project goals and ideas. This process establishes the context necessary for GPT to offer more relevant recommendations in subsequent discussions.
• Define: Collaborate with GPT to create a product specification or a requirements document that outlines the project’s features. GPT can assist you in distilling the most critical aspects based on the context established in Step 1. You can also prompt GPT to help you prioritize these requirements by providing additional context that you believe may affect the prioritization.
• Spec: Use GPT to assist you in generating a technical specification. In this phase, provide GPT with information about your technical expertise, time commitment, and whether the project serves as a proof of concept or needs to be more scalable. During this step, you may iterate with GPT by asking it to simplify certain aspects of its recommendations, reminding it of your learning goals, or requesting multiple implementation options with their respective pros and cons to aid in your decision-making.
• Plan: Seek GPT’s guidance in determining where to begin and formulating a plan. For example, when I was working on my Reinforcement Learning app, GPT suggested starting by learning the basics of Q-learning (a reinforcement learning algorithm) with dummy data. This initial step was essential for me to learn the foundations.
• Implement: Through purely natural language prompting, you can start to build your app by copying and pasting the code that GPT generates into your IDE. Initially I used Replit (and copy/pasted between GPT for help writing and debugging code). I’ve since moved to using Visual Studio Code to access Github Copilot directly in the interface. Expect a significant amount of de-bugging and trial and error but in all cases, I was able to complete each step with Copilot’s assistance. From a learning perspective, I recommend that you ask GPT to help you understand each piece of code it generates so that you can still hold all of the moving parts of the system you are building. This is particularly important as you build systems that have begin to have multiple modules.
• Remind: Continuously provide context from the previous steps as you proceed to each new step in your project. This ensures that GPT remains informed and can offer relevant assistance at each stage.

A few additional observations

• GPT rarely gets it right the first time. There is still a significant amount of debugging, but GPT guides you through this process.
• GPT will hallucinate. For example, it once hallucinated about an API endpoint that didn’t exist. I didn’t realize this until lots of troubleshooting. Increasingly, I will pass documentation from external websites to make sure it provides me valid answers.
• GPT has to be constantly reminded of its context. As my prototypes got more complex involving multiple files and modules, I would have to remind GPT about the existence of these components.

Read more

I also found Geoffrey Litt’s article on Malleable Software to be particularly insightful if you want to read more about end-user programming.