As someone that trains LLMs, even if grok does have a "avoid ""woke""" fine-tuning, adding that needs a few thousand examples SFT and a system prompt. It costs nothing extra to do it to coding agents and is not the reason why grok sucks at coding. It's just not in the same league in general - it's 0.5T parameters only and not trained specifically for coding at all.

Even if the way they are doing it did damage coding performance, it is a simple matter of serving another model without that fine tuning in the enterprise API preferably only to the grok coding harness (or cursor, now). Coding performance for subscription plans don't move the needle in terms of revenue anyways and quality there doesn't matter as much.