I've found that capabilities are only "peaky" in GPT models (including 6 Astra). On Claude models (particularly Fable 5 and Opus 5.5), capabilities are fairly well-rounded. I'm astounded at how good Claude is at writing, storytelling, sociology, psychology etc.

I wonder if Karpathy has also let his opinion about frontier LLMs be informed a bit too much by OpenAI's models instead of exploring the other ones. It's not Coke and Pepsi at all.

At my own little sociology benchmark, if I'd rate Fable's performance at 7/10 (subjective scoring), Astra would be 2/10. They're that different.