稀疏MOA:为何我构建了集体智能而不是选择单一模型

📄 中文摘要

AI模型市场正在趋同,前沿模型之间的差距每个季度都在缩小,边际成本趋向于零,形成了商品市场。因此,选择构建一个受管控的集体智能,而不是单纯出售模型的访问权限。该架构的核心是Grok,它将每个查询分类为认知需求类型,如检索、推理、创造、编码、社交和反应,而不是直接回答问题,而是进行路由。这一过程类似于生物体的前额叶皮层,根据查询类型选择合适的处理模型。此外,并非每个模型在每个查询上都激活,这避免了计算上的低效,类似于全脑癫痫的情况。

📄 English Summary

Sparse MOA: Why I Built a Collective Intelligence Instead of Picking a Model

The AI model market is converging, with the gap between frontier models narrowing each quarter and marginal costs approaching zero, leading to a commodity market. Instead of selling access to a model, a governed collective intelligence was built. At the core of this architecture is Grok, which classifies every query into cognitive demand types such as retrieval, reasoning, creation, code, social, and reflex. Rather than providing direct answers, Grok routes the queries appropriately. This process resembles the prefrontal cortex of an organism, selecting the right models based on the type of thinking required. Furthermore, not every model is activated for every query, preventing computational inefficiencies akin to a whole-brain seizure.

Powered by Cloudflare Workers + Payload CMS + Claude 3.5

数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等