文章探讨了语言模型未来应用的多种可能性，指出其潜力远不止于简单的助手或工具。未来，语言模型可能取代我们处理信息的方式，例如，通过接收JSON对象而非网页，由本地设备上的语言模型将其转换为人类可读的语言。此外，语言模型还可能在终端中执行自然语言命令、提供个性化信息流，甚至根据个人喜好生成电影等媒体内容。文章强调，当前语言模型的采用率较低，主要原因是可靠性和基础设施问题，但这些都是暂时的。真正的挑战在于，现有系统并非为利用强大的机器智能而设计，需要从根本上进行重新设计，以充分发挥语言模型的潜力。最后，文章提出了“软性权力剥夺”的概念，即人们可能因AI提供的更佳体验而自愿放弃对生活的控制。

📖 未来语言模型可能取代传统的信息获取方式，例如，用户可以向语言模型提问，而非阅读长篇文本，或让语言模型将要点整理成文章。

💻 语言模型可能直接在终端中执行自然语言命令，进行即时编码，或者在浏览器中提供个性化的信息流，甚至根据用户偏好生成电影、书籍等媒体内容。

💡 未来产品设计将以AI为中心，信息可能以JSON对象或要点形式传递，由本地设备上的语言模型将其转换为人类可读的语言，从而更高效地处理海量信息。

🤔 现有系统未能充分利用语言模型的潜力，需要重新设计，以适应机器智能时代。语言模型不应仅被视为助手，而可能成为我们感知现实的“基础层”。

🪞 人们可能会因AI提供的更佳体验而自愿放弃对生活的控制，这是一种“软性权力剥夺”，而非传统意义上的“人类权力丧失”。

Published on January 24, 2025 11:16 PM GMT

Language models are really powerful and will continue to get more powerful. So what does the future of language model usage look like? 

Imagining the future

Here are some things people likely already do. 

Instead of reading a long piece of text, you might give a book to an LLM and ask it questions about things you want to know.Instead of writing an essay, you might draft some points and ask an LLM to format it into prose. 

Here's some other things LLMs might be able to do. 

Instead of you manually deciding what to do, you might give an LLM access to a large collection of tasks you've written at some point and ask it for a small subset to do today.Instead of writing a script, an LLM might live in your terminal and ingest natural language commands, then do "just-in-time coding" to execute them.   Instead of browsing the Internet, an LLM might live in your browser and deliver you a personalized 'feed' of relevant information aggregated from everywhere in the world.Instead of you going on Netflix, an LLM might just write a new movie for you based on your preferences, or according to a specification. This generalises to other kinds of media like books, plays, music, etc.

AI-first product development

In the future it's likely that we'll design products explicitly for AI rather than for human consumption. Instead of websites and human-digestible reading we might just send information as bullet points, or as JSON objects, and trust the language model running on our local device to 'decompile' this into human-readable language.  

The modern world contains volumes of information orders of magnitude higher than what we can process. Making sense of this involves efficiently aggregating, distilling, and presenting this information in digestible chunks. Language models are likely to be able to do this way better than our existing systems can. 

Current systems restrict adoption 

It must first be acknowledged that current adoption is low because of mundane reasons like reliability and lack of infrastructure. However, these are transient issues. Furthermore, I think they are overblown - most people could get way more useful work out of language models than they currently do, if they really tried. 

The main problem is just inertia. Existing systems are just poorly designed to take advantage of this. They were designed in a world that didn't have extremely cheap and powerful machine intelligence, and must be re-designed from the ground up accordingly. 

Language models have so much potential. They won't just be assistants or tools. Consider the five senses we use to experience the world. There's no reason why all of those can't be replaced by language models generating the  equivalent. Language models will form a 'base layer' over reality through which you perceive everything. C.f. Plato's cave.

"Soft" disempowerment

We often express concern that AI is likely to take over the world leading to "human disempowerment", and this phrase conjures up something like 1984. However, a functionally equivalent outcome is "soft disempowerment" of the kind seen in Brave New World, where we very willingly cede more and more control over their lives to AI simply because this is an objectively better experience.  

Discuss