How to Dive Into AI

I know a surprising number of people who don't take advantage of their AI tools at work. They usually thumb their nose at the hype, doubt the quality of the output, or worry about its impact on human authenticity. These are all valid concerns. But I am here to tell you that staying sidelined on AI is a mistake. Despite its shortcomings and potential for dystopian outcomes, it remains a truly transformative technology that can't be put back in the bottle. In order to put limits and guardrails on AI, we need to understand it. Here are some concepts to help you dive in.

Be a Skeptic

Set preconceptions aside. Expect neither miracle nor disaster. Be skeptical of the machine and yourself. Skepticism is warranted because AI is prone to hallucinating. It will sometimes share false information and confidently present it as fact. Identifying wrong information and pushing back will keep the conversation on track. However, if the conversation significantly derails, it is often better to start over. This doesn't happen often but it's an important to be aware of.

Avoid this pitfall by adopting the mindset of a truth seeking philosopher. Ask it a question and think about the answer. If you didn't understand it, ask more questions about it. Ask the AI why it gave the answer that it did. Think about the assumptions that the machine is making. Also consider the assumptions baked into your own questions. Turn ideas over in your mind. Examine points from every angle.

Here are some ways to practice a skeptical approach:

1. You want to architect code a certain way. Explain your approach to the AI in great detail, upload files, and ask it to critique it for positives and drawbacks. Tools like Claude Code and Windsurf utilize agentic modes for enhanced context awareness. If it provides code to use, ask about the code.
2. Suppose you see a social media post about a new study that draws a dramatic conclusion. If you want more context, try uploading the study itself to AI and ask it to provide context about the findings. Ask about its strengths and weaknesses and how it fits into the broader research.
3. Your app crashed and you don’t know where to start debugging. Upload the crashlog to AI and ask it to provide analysis. When it eventually provides a theory, probe it.

Use the Socratic Method

The Socratic method, named after Socrates, is a popular technique used in law school and scholarly debate. It dissects ideas through questioning, forcing the student to examine every assumption. Rather than lecturing, the instructor asks questions to probe the subject, deeply engaging the learner. Being able to interact with a machine in this manner is extraordinary.

Open any chatbot and try this prompt: "I want you to teach me using the Socratic method." Then ask it a question about anything. The AI will then ask you questions to clarify your meaning. Think, answer, repeat. After some back-and-forth you will be start thinking about your subject from another perspective.

Example

How do combustion engines work?

Keep an Open Mind

Make it fun. If you don't know where to start with AI at work, start with low stakes tasks: summarize a document or revise an email. Interrogate it. Ask it why it did what it did. Try to find the bounds and push the limits.

The first thing I ever asked ChatGPT to do was to flip a coin, tell me the result, and count it. Flip it 5 times. 10 times. 50 times. For every time it lands on heads, multiply the count by 1000.

Why?

No good reason; I was playing. This kind of play-exploration manifested itself in further lines of questioning.

Some examples of this:

1. Do you have a favorite color? → What is color? → Why does mixing blue and yellow make green?
2. Why was Napoleon such a historic commander? → How did map making work in the 18th / 19th centuries? → How did triangulation surveys work?
3. How much time does a batter have to react to a fastball? → When did the first pitcher throw 95 mph? → When was the curveball invented?

Through hundreds of playful interrogations, I develop a sense for the model. Some questions I knew the answer to already; I used those to occasionally test the machine's accuracy. Others I did not know; I asked more questions for clarification. Then I used Google search to verify the answers. I repeated this for hours every day. It was like learning guitar and getting familiar with the fret board.

Different Models, Different Results

Some models are designed for short, fast answers. Others have longer thinking cycles. Some are better for coding. My recommendation for newcomers is to try three: OpenAI's ChatGPT, Anthropic's Claude, and Google Gemini. These products are actually a series of Large Language Models: each product has models optimized for speed or thinking. For example, in the Anthropic product suite use Claude Haiku for fast answers, Claude Opus for intense research, and Claude Sonnet for a balance. Try them all. Test their boundaries with play-exploration.

Explore with Your Hobbies

Try integrating AI into your hobbies in ways that are genuinely useful. For example, I use it in my photography. I use it for things that are vital to my photography – like storage and backup management – but not for photography itself. (For a deep dive on that, see my earlier post here.)

I'm not using AI to do my hobby. I'm still composing, shooting, cataloging, critiquing, editing, and printing. Rather it is the elimination of the tedium that affords me more time to actually do the thing.

Start building practical applications after getting your footing. Coding is another great example. I introduced Agentic AI into my personal iOS quote-keeping app, InkLeaf. I use Claude Code to go from ideation to prototype in minutes, not hours. The prototype leads to further ideation and prototyping. This is an iterative cycle until I bring the feature to production, which is when I write the code myself. Of course, at this point, I am still using Claude for debugging and questioning, the code is my own.

Examples:

1. Quickly build a "Current Book" feature, usable within 15 minutes. The implementation and testing would probably have taken me a couple of hours.
2. Build a linguistic analysis system to analyze the structure, meaning, and use of the words in a given quote. The implementation was done in 1-2 hours. This would have taken me weeks to build.

Coding is a great playground for AI. When bringing AI generated or influenced code to production, remember to slow down. Understand the underlying fundamentals of the system. Build it again without assistance. Let other people read it and contribute to your code.

Get After It

In sum, AI is a transformative technology that we should all be using. Play with these tools and get a feel for the models. Try it out with low stakes tasks, like summarizing emails or searching Slack. Bring your own hobbies into the mix. Challenge yourself to see where AI can eliminate the mundane, but be smart about it. Practice healthy skepticism at all times. Always keep learning and devise opinions through experience. Develop a strong voice in this new world.