At first, I thought bigger parameters were the best, so I mainly looked for and used large models. If you think about it that way, the Chinese model should be the best.
But when you look at Claude, GPT, etc., there are Chinese models that surpass them in benchmarks, but they can't keep up in actual use.
I think this is because when all the services are combined, such as the harnesses and assistant services (MCP) attached to services like Claude App and Codex App, even with similar parameter sizes, a big difference can occur.
At first, I thought, "Is AI a genius? Do we need experts anymore?" But now it seems that these things are getting dumber than I thought.
They can work in an instant and then evaporate it while managing memory, but sometimes they act like they're hearing what I asked 30 minutes ago or said for the first time. And if the context capacity is high and compact, I have to tell them again.
Sometimes it goes crazy on its own, receives all the feedback again, restarts what it finished, fixes what it set up...
Claude and GPT have better assistant systems for managing this kind of thing, so they are better.
