尽管大语言模型（LLMs）非常惊艳，但当下提升它们的知识能力，过程远比人们普遍认为的要零散和碎片化。我以前写过：AI 很了不起……但也没那么了不起。同样，LLM 的确是通用的……但并没有那么通用。 我们不应该被一种不准确的炒作所误导，认为 LLM 在短短几年内就能通向 AGI；但我们同样也不该陷入另一种同样不准确的叙事，认为它们只是“演示级技术（demoware）”。相反，我认为，更有帮助的是对当前构建更智能模型的真实路径形成一种更精确的理解。 首先，LLM 确实比以往几代技术更具通用智能特征。 这正是为什么同一个 LLM 可以被用于广泛不同的任务。 第一波 LLM 技术之所以能做到这一点，是因为它们在公共互联网数据上进行训练，而互联网包含了大量、覆盖面极广的信息。这使得它们的知识结构，比那些只被训练来完成单一任务的早期算法（比如预测房价，或只下国际象棋、围棋）要通用得多。 但即便如此，它们距离人类的通用能力仍然差得很远。 例如，在预训练时“读遍”整个公共互联网之后，一个 LLM 依然很难适应某些写作风格（而许多编辑却可以轻松做到），或者稳定、可靠地使用一些简单的网站。 当几乎所有开放的网络信息都被利用之后，进步就变得更困难了。 今天，如果一家前沿实验室希望一个 LLM 在某个特定任务上表现出色——比如使用某一种特定编程语言写代码，或者在某个高度细分的领域（如医疗或金融的某个子领域）中给出合理判断——研究人员往往需要经历一个非常繁琐的过程： * 寻找或生成大量该领域的数据 * 对这些数据进行处理（清洗低质量文本、去重、改写等） * 最终构造出可用于向 LLM“灌输”该领域知识的训练数据 又或者，为了让模型学会某些操作性任务（比如使用网页浏览器），开发者可能需要经历更为复杂的流程： 构建大量强化学习“训练场（RL gyms）”——即模拟环境——让算法反复练习一小部分、非常狭窄的任务集合。 相比之下，一个普通人，尽管接触过的文本数量远远少于当今最前沿模型，也几乎没有在“计算机使用训练环境”中进行过系统练习，却依然能够泛化到远多于前沿模型的任务范围。 人类之所以能做到这一点，可能依赖于：从反馈中进行持续学习。对非文本输入拥有更优越的内部表征（坦率说，把图像“分词”给 LLM 的方式，在我看来仍然很像一种权宜之计），以及许多我们尚未理解的机制。因此，当下推动前沿模型进步，往往需要做出大量人工决策， 并采用一种以数据为中心的 AI 方法，对训练数据进行精细工程化设计。 未来的突破，或许能让我们以一种不那么碎片化的方式来提升 LLM；但即便做不到这一点，这种持续的、零散的改进，加上模型有限但真实存在的泛化能力，以及偶尔出现的“涌现行为”，仍将推动技术继续快速前进。 无论如何，我们都应该做好心理准备：前方还有很多年的艰苦工作要做。构建更智能模型的道路，将是一段漫长、艰难——但也充满乐趣的——跋涉之旅。 As amazing as LLMs are, improving their knowledge today involves a more piecemeal process than is widely appreciated. I’ve written before about how AI is amazing... but not that amazing. Well, it is also true that LLMs are general... but not that general. We shouldn’t buy into the inaccurate hype that LLMs are a path to AGI in just a few years, but we also shouldn’t buy into the opposite, also inaccurate hype that they are only demoware. Instead, I find it helpful to have a more precise understanding of the current path to building more intelligent models. First, LLMs are indeed a more general form of intelligence than earlier generations of technology. This is why a single LLM can be applied to a wide range of tasks. The first wave of LLM technology accomplished this by training on the public web, which contains a lot of information about a wide range of topics. This made their knowledge far more general than earlier algorithms that were trained to carry out a single task such as predicting housing prices or playing a single game like chess or Go. However, they’re far less general than human abilities. For instance, after pretraining on the entire content of the public web, an LLM still struggles to adapt to write in certain styles that many editors would be able to, or use simple websites reliably. After leveraging pretty much all the open information on the web, progress got harder. Today, if a frontier lab wants an LLM to do well on a specific task — such as code using a specific programming language, or say sensible things about a specific niche in, say, healthcare or finance — researchers might go through a laborious process of finding or generating lots of data for that domain and then preparing that data (cleaning low-quality text, deduplicating, paraphrasing, etc.) to create data to give an LLM that knowledge. Or, to get a model to perform certain tasks, such as use a web browser, developers might go through an even more laborious process of creating many RL gyms (simulated environments) to let an algorithm repeatedly practice a narrow set of tasks. A typical human, despite having seen vastly less text or practiced far less in computer-use training environments than today's frontier models, nonetheless can generalize to a far wider range of tasks than a frontier model. Humans might do this by taking advantage of continuous learning from feedback, or by having superior representations of non-text input (the way LLMs tokenize images still seems like a hack to me), and many other mechanisms that we do not yet understand. Advancing frontier models today requires making a lot of manual decisions and taking a data-centric AI approach to engineering the data we use to train our models. Future breakthroughs might allow us to advance LLMs in a less piecemeal fashion than I describe here. But even if they don’t, the ongoing piecemeal improvements, coupled with the limited degree to which these models do generalize and exhibit “emergent behaviors,” will continue to drive rapid progress. Either way, we should plan for many more years of hard work. A long, hard — and fun! — slog remains ahead to build more intelligent models.