Thoughts on Software and AI
So much has already been said about AI (LLMs), but here's my take.
1 Power tools, anyone?
There's a woodworker named Paul Sellers (look him up). He's well-known as a longtime woodworker, famous for using hand tools.
When you're working with wood, one of the first things you have to do is square it. That is, make it flat on all surfaces. Squaring it reduces your margin of error when making measurements and such.
Now, you can square a piece of wood by hand. However, most (all?) modern woodworkers use power tools (you can use a jointer for this). Even Paul Sellers!
2 AI as a tool
Like power tools for the woodworker, AI can be thought of as another tool. In the day-to-day tasks of the average software dev, there are many rather mundane, mechanical tasks. Below are some examples:
Fixing simple merge conflicts. These can be rather annoying, but it's something that can (and should) be automated away.
Creating (even just half-decent) documentation. No one likes writing docs after already having written the code, because it's oftentimes a very mechanical, dry, mundane task.1
Doing some research on some technical area you're not familiar with.
The last one hits hard --- in the pre-LLM era, I used to open up 10, 15 tabs on topics while searching the internet, just to get a quick glimpse into what <INSERT UNFAMILIAR TECHNICAL TOPIC> was about. This was really hard and such a huge drain --- you had to read blog posts and StackOverflow answers and random internet comments sometimes. I save so much time by asking LLMs, because about 99.99% of the time, I am only interested in generic principles and well-lit paths.
There is no doubt in my mind that AI usage, as heretical as it may seem to a lot of purists, is here to stay. Just like power tools though, AI has to be used very carefully.
3 Where do we draw the line?
I don't think we should only rely on LLMs when writing software. There needs to be some human input, somewhere, in the process. You just can't ask 1 prompt, let the LLM go crazy on something, and walk away from it.2 As a developer we have to direct the tool to do what we want it to do.
You may think of the port of Bun from Zig to Rust by AI agents as a counter-example, but even in that case they had to be given very specific instructions by a human.
4 Inevitable doom for the job market?
Do I think AI will render all software devs useless, in the long term? No. I don't think this will ever happen.
Whhy? The simple reason is trust. Imagine if an airline said "hey, your plane is flown by our new AI model, there is no human pilot aboard the aircraft". I think they will lose many customers overnight.
I think the software world will react similarly. If Microsoft said "yeah, we fired all of our engineers, we only have AI agents running the show from a technical standpoint" --- I think they will lose a lot of trust.
You must understand how the software (written by AI or otherwise) works. If you don't, you're sort of rolling the dice.
I will never use AI to generate blog post content on this site --- it crosses a line I am not comfortable with. However, I have no issues with writers using LLMs to proofread their drafts, especially where English is not their first language.
5 Conclusion
If you use AI, just make sure you fully understand what it's doing.
6 Footnotes
1. OTOH, I personally enjoy writing docs at the same time I am writing code (or even before I write code) --- all as part of literate programming. ↑ ¶ (Cell 10)
2. Unless you're Anthropic and you literally have unlimited tokens to spend. ↑ ¶ (Cell 17)