Two technical lessons from 3 projects with ChatGPT
During 2023 I built 3 projects utilizing ChatGPT API:
- CoockItEasy (completely disabled) — platform to help people find dish receipts.
- CarComparisonAI (landing page is here, but application was disabled) — platform to help people compare used cars and choose the best to buy. Inspired by utilizing ChatGPT to bought used car in Toronto, which I did in last year.
- Indie Hacker News — platform which generates business ideas from last tech news. Inspired by my old project for sending hacker news updates in Telegram — https://github.com/vitalii-honchar/hacker-bot
First two projects had one big problem — using ChatGPT as a core in software system.
Let’s see on examples: CookItEasy had UI for user and requested recipes from ChatGPT.
CarComparisonAI worked in the same way.
In another works ChatGPT was rock star for my software and did all job to solve a problem.
In fact CoockItEasy and CarComparisonAI were just front-end with some predefined prompts for ChatGPT.
As a conclusion ChatGPT generated primary value for software which made my services at least very easy to copy by another developers and at most useless, because ChatGPT alone provided full value for them.
Also as far as everybody can use ChatGPT to solve their problems, ChatGPT can’t be used as a core feature of the system, because it’s too easy for competitors to copy a project with utilizing ChatGPT. Need to create differentiaton for a project, which makes project unique and this should be done with another tools, not with ChatGPT.
First lesson learned:
ChatGPT shouldn’t be core feature of the software. Differentiate project by some unique feature which is hard to copy.
During autumn 2023 I found a very interesting course — ChatGPT Prompt Engineering for Developers which provides an idea that ChatGPT is not silver bullet which should solve everything, instead it’s machine learning model which can be very quickly adjusted for a problem by plain English.
This provided me an idea that ChatGPT can be used not only for generate content, but also for analyze existing content and provide some result for an output.
Second lesson learned:
ChatGPT is a framework for rapid development of machine learning models and solve tasks: analysis, making decision and etc.
With this two lessons I decided to build PoC project “Indie Hacker News” which should send business ideas in Telegram every day by analysing latest news.
For this project I used ChatGPT as an analysis tool and prepared data for it before analysing it.
From an architecture picture above we can see that ChatGPT is doing analysis only by provided data from “Scrapping Stage” and core feature moved from ChatGPT to News Sites set, so quality of content primarly depends on the quality of scrapped news. Of course ChatGPT still plays critical role in the software system, because it provides analysis of data.
But here ChatGPT is like just machine learning model which provides analysis for input data and in theory I can change ChatGPT with my own model in the future for similiar system. Of course not for “Indie Hacker News”, because using own machine learning model is very expensive in terms of it’s learning. Purpose of “Indie Hacker News” project was to test two approaches:
- Use ChatGPT as additional tool in the software system and not make it core feature.
- Use ChatGPT as a machine learning model which can be trained by just plain English.
Conclusions
During 2023 I built 3 non-profitable projects with ChatGPT which provided me very useful insights:
- Project should have differentiation point which makes it hard to copy by competitors.
- ChatGPT should be used at the place of machine learning model for rapid development with a possibility to replace ChatGPT with in-house trained machine learning model.
