I’ve written software since 1971. Since the mid-1980’s I’ve heard that a new computer program is going to write the code from now on instead of programmers. That’s left me with skepticism about what AI might do. After spending time with them, here are a few early impressions.
AI can write working code
I started in March of 2025 coding with the chatbots. The CLI and VS Code based tools were only beginning to appear. But given a fairly simple problem in writing SQL the chatbots were able to produce working code. ChatGPT and Gemini were able to write some code but Grok went past that and found a bug in the code it was give to to adapt and fixed it. Now all the foundational models are good at code. The harnesses, like Claude Code, Codex, Cursor, and JetBrains, have taken that further and are very good at it. I documented that experience here:
Testing still matters
The AI tools are confident even when wrong, so testing matters more than ever, not less. The good news is that debugging skills are still useful. However, I’m working on some projects were I’ve not looked at the code and have interacted with the AI by doing my own testing and then going back and requesting changes like a typical user might.
Image Generation Isn’t so Great
Given the logos of Gemini, Grok, Claude, and ChatGPT, I asked each:
Using the four logos provided make an image for the theme of AI-Assisted coding. Include a few lines of code somewhere in the image, maybe in the backround. Make sure to produce a square image.
The results were pretty bad. They all butchered at least two of the logos. Most of the output wasn’t square. The image at the top of this post is from ChatGPT and here’s a try from Grok.
If you’re exploring the same territory, I’d love to compare experiences.
