I've spent 30min talking to candidates about their projects, experience, tech they used, choices made, etc. and been super excited and positive about them.
Then I ask them to find the intersection of two lists in better than O(n^2) and they completely fail to write functional code in their preferred language in the next 30 minutes. This isn't even "leetcode" level of challenge.
If I'm hiring you to be a professional software engineer for 6 figures income to spend 8-10 hours a day designing, writing, and debugging code, you need to be able to write code.
The disparity between the conversation and the coding performance can be significant, and both need to be evaluated.
If there's a better way to evaluate if someone can write code, I'm open to try it. But take home tests, portfolios, pair-programming, "a day in the life", etc. all have their problems and disadvantage different groups.