AI and Coding

What makes AI so fascinating is how flexible it is and how it can be applied successfully to an endless number of use cases. In the past, I’ve asked ChatGPT to create comprehensive meal plans with recipes for each meal and corresponding shopping lists. I also use it as a personal trainer and to answer general health questions. Recently, I’ve started using AI as a coding assistant and was surprised to see how effective it was at guiding me to real results. I’ve always been interested in coding, but the time and dedication required to learn the foundational skills and context have eluded me.

I recently had ChatGPT build a program that connects different API data sources into one file and then performs various functions in that file to clean up the data. ChatGPT patiently answered every one of my repetitive novice questions all the way from the idea stage to a functional prototype. At that point I could describe each feature I wanted to add, change or remove and it would update the code for me. I could always read the output to get a rough understanding, but I never had to write anything complicated or debug any syntax errors. From start to finish, this process was about 15-20 hours including learning where to store and run the code.

I have a few reflections on this experience:

• AI is the next generation of No-Code and Low-Code software: No-code software allows you to drag and drop different modules to essentially “build” an application or website without any coding skills. Squarespace does this for websites. Zapier does this for certain APIs. I’ve played around with No-Code applications but I’ve never built anything useful, mostly because the certain use cases they were designed for didn’t fit my specific interests or the integrations that I wanted were not available. AI-driven coding is completely different because it has no pre-defined limitations.
• Usage-Based Pricing: At just $20/mo, I don’t see how OpenAI is not losing money on my ChatGPT usage. I routinely get rate-limited, where they pause my access for 2-hour blocks. Yes, OpenAI does have usage-based pricing for its API, which I have used, but I don’t need it all that often. While companies compete to build, we all get access to heavily subsidized cutting-edge AI tools. That will have to go to usage-based pricing.
• Caution: I’m optimistic that Google’s Gemini model is one of the best available, but I haven’t found it particularly useful for my own regular projects. It either gives “lazy” short answers or tells me it can’t or won’t do the research I ask it to do. Especially in economics or health, it tends to tell me to talk to a professional. It’s understandable that they are being cautious and I’m sure I can get around their buffers, but I’d rather just have the answer I’m looking for, and I find that ChatGPT is more direct. I don’t do any work in AI video analysis but I hear Gemini is the leader in that category.
• Competition: When I don’t get a good answer, or I get rate-limited on one LLM, I can always jump to the next. I use Meta.AI, ChatGPT, Gemini and Perplexity knowing they give slightly different results. I only recently found Meta’s AI model. Claude is popular as well. I’m sure this list will continue to grow.
• Context Windows: Software programs offering APIs typically have developer pages with extensive documentation. That documentation outlines the available functions in plain English. Non-coders can typically make sense of it but cannot do anything with it if they are not functionally literate coders. When I drop all of the API documentation into ChatGPT, it gets the relevant context to build what I ask for. Context window, the closer you will get to a desired result.
• Operational efficiency opportunity: Over the past 10 years, tech-savvy people have had a perpetual opportunity to upgrade legacy business practices. Getting a good website, listings on Google and implementing specialized software packages have made a big difference even though they seem like basic upgrades. AI has reopened that door for a new set of potentially dramatic increases in operational efficiency. A friend was recently telling me about an acquisition where the seller could not make the asset profitable, but the buyer estimated they could add 20 pts of margin onto the business thanks to their engineering capabilities, which the seller did not have. There is an AI efficiency opportunity for every business.

The investment implications are pretty clear. AI is extremely useful, and the prices in mega-cap tech reflect that reality. AI could also end up in the standard private equity playbook: acquire a business, implement AI efficiencies and sell. I mention private equity because that asset class has been stalled for a number of years and is starting to get real pressure from investors to return capital on older funds and call capital on new funds. An operational efficiency boost thanks to AI could help private equity get unstuck. That’s an optimistic take, but not altogether unreasonable.