Leetcode was never about being similar to real work. Companies aren't stupid - they know it's different and still think that they get useful signal from testing skills that don't match the day-to-day of the job. Whether that's appropriate or not is where the actual debate around the value of Leetcode is; the debate isn't whether leetcode is like day-to-day work, nobody is claiming that it is.

And LLMs don't make the difference you think they make to the viability of LC interviews. Before LLMs, you could solve an interview problem by looking it up in a database with solutions. Or (in real work) you solve these algorithmic problems by importing a library. LLMs are just the (n+1)th way to do the same thing.

Where LLMs play some role is that they make cheating in remote leetcode interviews easier.