The problem with LLM written code is that you can't trust it. It will hallucinate dependencies (with security risks) and is prone to writing unsecure code. So you need a human to review it. And code review is difficult, and not very interesting. It's like having a self-driving car but having to stay alert and ready to take control because it can randomly fail at any moment. You'd probably rather drive.
Ok. That sounds like what a decent coder would say.
Now, in the other hand, if you'd be a pointy haired CEO ... why not trust the LLM service your code, your database and your devops ? That's innovation !
I guess you all know my feelings about service as a software substitute and vibe coding, so the read was fun, and se non è vero è ben trovato.
SaaStr CEO first week of vibe coding for Replit's profit.
You actually want it the other way around: you drive and the car warns you if you miss something. And my limited experience with LLM-assisted code review wasn't very conclusive.
Well, some people with deep enough tongues and deep pocketed contacts are probably right now in a city near you planning to start a startup hiring 3D printed robotaxis from a LLM generated blueprint and paperwork run by the same chatbot running the customer support.
The control software in the robotaxi will likely be generated by itself from training in a virtual reality footage generated by a browser agent playing grand theft auto for many hours. And the only question the investor will probably have is :
- But did Rockstar agree to the plan, or are we doing it without them noticing ?
- Your second guess is right. As a contingency plan, we plan on adding a lawyer persona to our compliance/customer support chatbot.
- I see that's all well thought out, how much money are you asking for this round ?
There are some tasks where LLM could help (what TeDaDeS calls "Advanced Find&Replace" and extended autocomplete) but in my experience you'd be better off with a dedicated script/program/IDE extension for this. I actually had this in my last job: some of my colleagues were using LLM to write boilerplate code but I wrote a script that did the job in a far more reliable way.
Well, I guess that depends on how we look at it. Replit's bots have spared some suffering from the staff SasStr would have hired. Not sure any salary is worth having such a boss...