3 ms·
GPT 5.6 has similar language quirks that makes its comments nearly unusable. I wonder if this is a side effect of MoE models — they can write excellent prose,
by jiggawatts 2mo ago
GPT 5.6 has similar language quirks that makes its comments nearly unusable.
I wonder if this is a side effect of MoE models — they can write excellent prose, but not simultaneously with writing code.
- notfromhere 2mo agoI think it’s just where they focused RLHF resources. The models have generally only gotten worse at writing. And writing doesn’t have validators like code so you can’t really scale it in the same way