Nevkontakte
@me@m.nevkontakte.com
Lawful neutral. Definitely not a cat in a hat. Opinions are of Cthulhu. what/why.
For a long time I was curious if Claude Code works so well because of Claude (the model) or Code (the CLI tool / agent). This weekend, I tried to find out. Turns out that both matter, but more than anything post-training fine tuning of the model makes a big difference. If the model has been tuned for planning and tool usage in a certain way, it would provide much more reliable results.
Details as in https://nevkontakte.com/2025/swap-ai-brains.html
Likes
@me Love the experiment.
Video clip is my mood when I saw Gemini2.5 was the secondary. "That model is nasty in general."
Leaderboards aren't everything but Gemini's position on https://gorilla.cs.berkeley.edu/leaderboard.html are a cry for help. "You okay over there, google?"