智谱或推原生多模态模型 GLM-Omni
Zhipu AI might drop a native multimodal AI called GLM-Omni.
AI 导读
X.PIN 称智谱可能推出原生多模态模型 GLM-Omni，该系统将图像生成与文本训练结合，让模型在作答前先画出视觉推理步骤，使文本问答准确率从 12.57% 提升至 41.14%。作者还调侃这就像自己考试时涂鸦。
正文
Zhipu AI might drop a native multimodal AI called GLM-Omni.
By pairing image generation and text training, the system allows the AI to draw out visual reasoning steps to look at before answering. This drove text QA accuracy up from 12.57% to 41.14%.
Basically me doodling during exams.