At least (as far as I can tell) they appear to be ranking the models by human evaluation rather than “benchmarks”, which is closer to measuring the real-world performance.

It would be interesting to consider the types of questions that users are posing. For example there is a difference between asking:

• A surface-level fact-based question such as “what is …”

• A creative question like “write a story/article about …” or “give me a list of possible talking points for a presentation on …”

• A question about reasoning/understanding like “why do you think the word … is more popular than … when referring to …” or “explain why … is considered socially acceptable while … is not”

• Anything coding-related

Also, some models seem to do well at things that can be answered after one or two replies, but struggle to follow an argument if you try to go more in-depth or continue a conversation about a topic.

As an autistic person I like to minimise change. I’m also good at learning how different appliances/equipment works and how to repair it. So I always try to repair things first before replacing them. I get the double satisfaction of not wasting an otherwise perfectly good appliance and also getting to keep the appliance that I am familiar with and like instead of having to try to find a different one.