在 AI 领域中,对齐(alignment)和涌现(emergence)是两个重要的概念。
对齐通常指的是在强化学习(RL)与大型语言模型(LLM)融合的情况下,如果 RL 下的自我博弈或相关技术能够带来新的洞察或泛化可能,那么对齐会遵循缩放定律。在真实物理世界中,分别在预训练与对齐过程中,体现了对世界泛化性的两种学习方式。而当前的 LLM 需要在后续环节完成 RLHF 或 RLAIF,是因为真实世界中人类历史沉淀的文本符号表征数据分布丰富,而需对齐方式获得的泛化能力的数据样本稀缺。
涌现则是指当训练的数据达到一定程度后,模型的能力会出现很多难以用逻辑解释的现象。比如在大模型中,尽管其学习是在降维的潜在空间中进行,只是计算相关性和概率,但却能表现出难以解释的表达。目前也有一些论文使用跟踪标注等方式尝试解释大模型内部的行为。
此外,在《促进创新的人工智能监管方法》中,提到了一些与监管相关的功能和活动,如支持前瞻性扫描活动、确保监管制度具有适应性、协作性和可信度,以及监测与国际方法的一致性等。
如果RL下的self-play亦或者HF/AIF能够尽可能多的带给我们一些历史真实世界数据所没有的、未见的、新的洞察或泛化的可能,那么alignment是会遵循scaling law的。尽管其中的数据也是AI合成的(也夹杂包含着模型本身的知识,习惯和泛化性),但新的环境和在去做alignment的过程中也许会将领域的泛化性进一步增强而出现领域内的“涌现”换句话说,在真实物理世界中,泛化性(或抽象概念)本身也有数量,特性,囊括度,在真实世界的分布性等特点的,因此分别在pre-training与alignment过程中,其实也是体现了两种对这个世界泛化性的学习(这种学习也是在token化中完成对齐或映射的),而之所以现在的LLM需要在后边的环节完成RLHF或RLAIF,表明在真实世界中这种人类历史所沉淀的文本符号表征即token的自然序列排列所体现的分布(训练后表现为常识性或领域性浅层表象知识)比某种需alignment方式才能获得的泛化能力的数据分布相比太丰富了,或者说需采用RLHF或RLAIF进行alignment的数据样本太稀缺了。比如生成内容的安全性,系统②的推理模式,复杂的数学推理或自然科学的探索洞察即AI4S。
rag,也就是Retrieval-augmented-Generation,检索增强技术,应该也是大家经常看到的一个名词了。那么他到底是个什么东西呢?首先我们先回顾llm的原理,其实就是把一堆互联网文献材料,降维到latentspace中,并通过transformer的方式用学习到了其中的“经验”。但文献资料是死的,如果仅通过固定的资料去回答问题,势必会出现一些无法回答的问题。那么rag的出现就解决了这个问题,rag允许模型可以到搜索引擎上去搜索问题相关的资料,并将获取到的信息,综合自己的知识体系内容,综合进行回复。当然rag不全是这些,rag的r,也就是搜索,其实不是一件简单的事情。所以这个环节还会有些传统的搜索的逻辑在,比如对于输入问题的纠错,补充,拆分,以及对于搜索内容的权重逻辑等等的事情。打个比方,比如问中国界世杯夺冠那年的啤酒销量如何,那就首先会做一次纠错,把“界世”纠错为“世界”,然后把问题拆分为两个问题,然后综合去搜索,把搜索到的资料在给到大模型,让其进行总结性输出[heading3]3.大模型的涌现[content]这个可能有些同学不怎么听过这个词,但这个是大模型里,我认为最核心的个概念。因为其实经上面的原理讲解,其实大家不难发现,这与目前大模型所表现出来的,仍然对不上啊。为什么只是在计算相关性和概率,就能让大模型表现出难以解释的表达?所以这就是涌现…也就是科学家们认为,当训练的数据到了一定程度后,模型的能力会涌现出很多难以用逻辑去解释的现象。说实话在我看来,因为本身模型的学习就是在一个降维的latentspace中进行的,那我们尝试去用三维世界中的if-else去解释其行为,似乎本身也不靠谱不是么。不过现在倒也有一些论文使用跟踪标注等的的方式来尝试解释大模型内部的行为,这个看看就好了
This function will support horizon-scanning activities in individual regulators but a central function is also necessary.As stakeholders have highlighted,an economy-wide view is required to anticipate opportunities that emerge across the landscape,particularly those that cut across regulatory remits or fall in the gaps between them.This function will help deliver a regulatory regime that is:•Adaptable:by identifying emerging trends to enable intelligent,coordinated adaptation of the regulatory framework.•Collaborative:by convening partners including frontier researchers,industry,civil society,international partners and regulators,to identify emerging trends.•Trustworthy:by ensuring that our regulatory framework is able to adapt in the face of emerging trendsEnsure interoperability with international regulatory frameworksActivities•Monitor alignment between UK principles and international approaches to regulation,assurance and/or risk management,and technical standards.•Support cross-border coordination and collaboration by identifying opportunities for regulatory interoperability.RationaleTo be effective,this function must be performed centrally.The whole regulatory landscape needs to be considered to understand how well the UK framework aligns with international jurisdictions.The impact of international alignment on innovation and adoption of AI in the UK is a key concern for businesses.The central oversight and monitoring of the global alignment of the framework will support UK engagement with like-minded international partners on AI regulation,building our influence in AI.This function will play a central part in delivering a regulatory regime that is:•Pro-innovation:by ensuring that UK innovators can trade internationally and UK companies can attract overseas investment.48A pro-innovation approach to AI regulation