zansara

I was quite surprised to find out that even recent LLMs like GPT 5.4 still clearly prefer to use Python over any other language in dedicated benchmarks, even if it's not a good fit for the given task. I usually use Python myself so I could not notice this effect: did anyone observe this "unreasonable" preference first-hand? How strong is it?

https://zansara.dev -- that's my blog

If you want to have some fun you can also play the game here: https://www.zansara.dev/guess-who/ (free, BYOK)

idk, I was not expecting it to go this wrong when I started this project. What makes Claude models fail so consistently? I didn't think they would be especially robust, but I thought they would be able to understand the point of the game.

I found some more information about Topology's approach here: https://yellow-apartment-148.notion.site/CLM-Docs-507d762ad7..., but nothing else.

Does anyone here uses this model? What is your experience with it?

The fact that these devices are getting more "secure" imho also removes a great opportunity for learning. It used to be possible to tinker with these devices: now even vacuum cleaner's firmware is locked in and unreadable. What is so precious about a handheld vacuum's firmware anyway? It seems like their only goal here was to make the device nearly impossible to fix. I really have a hard time understanding the rationale.

I'm learning French too but I can't see the mistakes you're pointing out... What would be the correct sentence?