Lateral thinking
Prior to 2026 100% of all models in the world couldn't solve any of those questions, now some (frontier only) on ocasion can.
There is one exception - Qwen3.5. It solved all lateral thinking questions i threw at it. Car wash, inverted cup, this one about limbs too. Even the 4B version of Qwen3.5 mentioned the man not having hands, while giving a generic answer, and 27B/35B-a3b answered correctly without doubts. It's so uniquely size-effective that it could as well be called a frontier model.
Qwen cooked well π€
With how good Qwen3.5 27B, it makes me wish for a 70B π₯²
Is there a plan to use newer model wth 4chan training? Im very interested in this project and wonder how would be be if the base model is already smart. Or is there actually a specific reason why it has to be llama3?
There's a plan for a ~32B version, a prototype was already made, but I'm not too happy with it, so it's possible that I'll retrain it.
I considered Gemma-4, but so far training it is a bit of a pain.