Not so useless. In my experience LLMs are about 50/50 on making a regex that actually works and covers the cases you asked it for. Even less when you get into cases needing advanced features like backreferences and lookahead.

Anecdotal data point, writing and maintaining regex is still a core part of my job. Not useless at all for me :)

A little bit early to tell.

Let’s wait how affordable, available and good AI is when the companies turn to profit maximization and enshittification begins

You can go local now with qwen 3.5 9B Q4 powering hermes agent at 35 to 50 tok/s with 99 percent tool call success rate on a used RTX 3060 for the price of two months of ChatGPT Pro and never bother. https://xcancel.com/sudoingX/status/2033020823846674546#m

This is the worst local AI will ever be. It only gets better from here. https://xcancel.com/sudoingX/status/2033959603944493192#m

Nope, if nobody trains the models on new data you have at some point an outdated model.

Imagine Qwen 3.5 created in the 1990s and then use it for today web or desktop development.

And is the problem solved that training AI with AI code makes the AI worse? If not the "it only gets better" claim is questionable.

> Nope, if nobody trains the models on new data you have at some point an outdated model.

As people train the models on new data they'll be increasingly training on AI output including hallucinations and slop. More garbage in means even more garbage out and the cycle will continue as "updated" models decline in quality.