They basically achieve over 80% agreement with human evaluators [1]. This level of agreement is similar to the consensus rate between two human evaluators, making LLM-as-a-judge a scalable and reliable proxy for human judgment.
[1] https://arxiv.org/abs/2306.05685 (2023)
It sounds nice but it means at least 1 in 5 are bad. That's worse odds than rolling 1 on a d6. You'll be tripping over mistakes constantly.
> 80% is a pretty abysmal success rate and means it's very unreliable.
So if 80% is abysmal, you are getting the same abysmal level from human code reviewers.
The whole point of a tool is to reduce workload. If you have to frequently correct it, then it's not worthwhile to use.
Sure there's no bug with how the logic is defined in the CR or even in the context of the project, ti maybe won't throw an exception.
But the LLM won't know that the query is iterating over an unindexed field in the DB with the table in prod having 10s of millions of rows. The LLM won't know that even though the code says the button should be red and the comments say the button should be red, the corporate style guide says red should be a very specific hex code that it isn't.
I don't agree with this at all. Writing new code is trivially easy, to do a full in depth review takes significantly more brain power. You have to fully ascertain and insert yourself into someone elses thought process. Thats way more work than utilizing your own thought process.