开发者构建原生支持视频处理的决策模型Viveki-0.1
Jev can't process Images. I built a decisions model to process VIDEOS + text + images
提出了一种无需微调即可原生处理视频的决策新范式，实测延迟极低，为多模态推理提供了新的工程思路，值得跟进。
I built Viveki-0.1 - the first decisions model that can natively process VIDEOS along with image and text.
我构建了 Viveki-0.1——首个能够原生处理视频、图像和文本的决策模型。
Hugging Face - https://huggingface.co/not-ekalabya/viveki-0.1-20B
It has a latency of <150ms (1x A100-80GB) and can process videos in real-time.
其延迟低于 150ms（在单张 A100-80GB 上），并且可以实时处理视频。
Viveki was made without any training or fine-tuning based on a mechanistic insight that allows us to predict the models answer for any number of objective questions in its context before any tokens are generated (more details to be released later).
Viveki 是在没有任何训练或微调的情况下构建的，基于一种机制性洞察，使我们能够在生成任何 token 之前预测模型对上下文中任意数量客观问题的回答（更多细节稍后发布）。
Jev outperforms the model by <5% on BANKING77 and MuSR. However, I hypothesize that performance can improve significantly if the method is paired with fine-tuning like Perplexity Decider of Fastino GLiDE.
Jev 在 BANKING77 和 MuSR 上的表现比该模型低不到 5%。然而，我推测如果该方法与 Perplexity Decider of Fastino GLiDE 等微调技术结合使用，性能可能会有显著提升。
Further intelligence evaluations are in progress.
进一步的智能评估正在进行中。
I'm intentionally skipping details as this is an active research project. I'll add the method paper to the post when the time comes.
由于这是一个活跃的研究项目，我有意省略了细节。届时我会将方法论文添加到帖子中。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力