| Some more 'sloptuber' content for those who are enjoying it :) Model: unsloth glm 5.1 @ IQ2_XXS UD Prompt 1: I included the 'forward' direction here, because basically every attempt was using (-1,0,0) as forward, which is not correct for three.js world space After this first prompt, GLM proceeded to think for 32k tokens, output a full code block in its thinking, and continue overthinking. This overthinking behaviour seems way more common on heavily quantised models. I copied the code block, tested it, and gave some further feedback in a fresh session: Prompt 2: Now, I could have also asked it to fix the arm/leg swing direction here, but I thought the results were a good mix of amazing + ridiculous, and show what kind of initial results you're currently likely to get when asking any LLM for 3D outputs. I've been trying this prompt out with Minimax M2.7 too, but it really struggles with the relatively complex collision mechanics that are required for this. GLM will produce something workable most of the time. [link] [comments] |