That was GPT-2 which was what originally powered Chat GPT when it came out in the fall of 2019. GPT-4 was trained on 2023 data, and was what was driving Chat GPT since for the past few years (though there was a 4.5 version). GPT-5, which was released by OpenAI 3 days ago, and is now what you are using if you use Chat GPT, was trained on data up to 2025.
Also, Chat GPT is only one LLM and first to market. They all have the same problems though.
This statement makes it sound like LLMs think and reason. They do not. They are just fantastic pattern matchers. Often the new ideas they generate are hallucinations. One of the most common hallucinations that LLMs suffer from when asked about Godot is making up functions that do not exist and explaining in great detail - with code examples - how to use them.
Also since LLMs scrape the web, they often cannot tell the difference from a correct and incorrect answer - and most of the data comes from these forums. (You can tell in the format of the answers.)
I personally would not describe them as an abstraction layer. I would describe them as an obfuscation layer. Especially considering the rise in posts here that start with something like: “I used an LLM to make this code. I don’t understand how it works or how to fix it. The AI can’t fix it either. Someone please help me.”
It would certainly be an interesting experiment. Though not one I would take without precautions. Most service contracts for web services have a 99.99% uptime clause. If your AI takes your server out completely for a few days over the course of a year you’re going to owe lots of money. Services like finance and medical usually have three nines, i.e. 99.999% uptime, which means if your service is out for more than 8 hours in a year you’re in breach of contract.
Having said that, I do think it’s going to become more common in the future - but it’s going to probably going to need to be specialized models doing that kind of maintenance and some people will be bit by downtimes. still I’d love it if AI managed my infrastructure because I hate dealing with AWS.