少点错误 2024年11月12日
(WIP) Exploring memetics -- hub
index_new5.html
../../../zaker_core/zaker_tpl_static/wap/tpl_guoji1.html

 

本文探讨了如何通过提升文明的认知水平来实现 AI 对齐。作者认为,需要更精确的术语来描述诸如模因、协调和概念清晰化等现象,并将其与实际应用结合起来。文章以 TikTok 为例,分析了缺乏自主性导致的模因传播现象,并对比了对话中人类意志主导的交流方式,指出 TikTok 的推荐算法和模因传播过程难以理解和控制。最终,作者希望通过研究如何提升 AI 与人类之间的带宽,促进协调和哲学进步,从而实现 AI 对齐的目标。

🤔 **目标:拯救世界**,文章旨在通过持续编辑,探索如何通过提升文明认知水平来实现 AI 对齐,最终目标是拯救世界。

💡 **术语澄清**:文章强调需要更精准的术语来描述“模因”、“协调”和“概念清晰化”等现象,并指出协调描述的是系统中个体间的关系,而“概念清晰化”和“模因”则是效果的名称。

🗣️ **案例研究**:文章以 TikTok 为例,指出其缺乏用户自主性,导致模因传播过程类似于一种难以理解和控制的“外星”过程,并将其与人类对话中意志主导的交流方式进行对比。

🚀 **AI 与人类带宽**:作者认为,研究如何提升 AI 与人类之间的带宽,可以促进协调和哲学进步,有助于实现 AI 对齐。

🌐 **社区分化**:文章指出社区分化会带来两种效应:增加总能量和提高迭代速度,这可能导致“思想进化/模因/协调/其他”的加速发展。

Published on November 11, 2024 6:18 PM GMT

This post (AI alignment via civilizational cognitive updates) is sort of the precursor to this one, but not in terms of the a thread of reasoning, but more for me personally: I posted it, received comments, someone DMed me and we talked a bit, I started thinking more about it and writing down my thoughts, and here I am.

 

General goals and expectations

The goal is to save the world

This post is intended to be continuously edited, I don't expect it to ever be frozen or act as a coherent thing with a beginning and end and clear message and such.

I want to find more precise terminology, for example these 3 are all different even though  they feel similar/it hasn't fully sank in that they are different:
- memetics
- coordination
- increasingly clear definitions and concepts

For one, coordination is a description of how agents in a system relate to one another (and a slightly narrower version is that this coordination is good), whereas "clarity of concepts" and "memetics" are more, names of effects.

 

Initial thoughts on this post

- separating terminology: - expecting exploration and many posts, versus clear beginning and end
- requires factoring out sub-topics into self-contained posts
- abstract discussion lame without practical application (actually trying to implement ideas while working on answering actual concrete questions or solving real problems)
- what makes tiktok bad (hint: lack of agency)
- start with simple case studies like: a conversation, iterating on a post, the game of telephone <-- see if you can build models and concepts that apply to these
- hilariously ambitious project
- concrete direction: how and why does tooling that increases AI-human bandwidth help with coordination and with philosophical progress?
- direct tradeoff between resolution and broadness of discussion
- things I want to define: broadness, resolution, conversation, mental context, information, information sharing, what is "an AI"? (simple definition is "the simulacrum in any given conversation")

 

Collection of effects and phenomena.

Bifurcation of communities

1) increases total energy, because some people will enjoy talking more in the niche community

extreme case: imagine a party of only introverts where everyone is in 1 big group, versus everyone talking exclusively in pairs <-- massive difference in total conversation-energy expended (leading to more "idea-evolution/memetics/coordination/whatever tf im supposed to call it")

2) fewer agents == individual outputs reach more of the community with less loss of resolution, and higher iteration speed (or "applies bigger relative update to the global state") --> faster evolution of ideas
(and the ideas that are good in a broader way, can still be shipped to the main community, which is cool)

 

TikTok case study: Less agency --> more alien memetic processes?

Tiktok scrolling basically has 1 action: "how long until you scroll?", it's a single integer value with a pretty small range (video length) and often it's actually a boolean: "did you immediately scroll further?".

(There are a bunch of other actions too, but they're negligable in terms of how important they are for the algorithm, and how many of decisions in terms of quantity are formed by these: like, comment, input search term, click channel, share, close app)

Compare this to a conversation, where you have word choice, intonation, topic choice, social cues about what you find interesting, it's a set of very high-dimensional pieces of information that you share repeatedly and they all have a big effect, crazy amounts of agency.

You can make a similar analysis on the "supersets" of both of these.

the entirety of Tiktok and how viral videos and hashtags spread and how they influence humans, compared to, when people start getting annoyed by how their friendship is going, they can explicitly talk about it and steer it or abort mission.

Conversations very much do not feel like "runaway memetic phenomena", they feel like 2 people converging on their values (finding out they dislike each other is included, advancing a dialogue while subtly learning about each other is included).

A dialogue is ruled by human will, Tiktok scrolling is ruled by the recommendation algorithm and more broadly, some alien process of meme spread that we don't understand or control, it's not even purely explained by recommendation algorithm. I doubt that even the recommendation algorithm plus the decisions of the engineers of the algorithm and the entire TikTok codebase would explain it very well.



Discuss

Fish AI Reader

Fish AI Reader

AI辅助创作,多种专业模板,深度分析,高质量内容生成。从观点提取到深度思考,FishAI为您提供全方位的创作支持。新版本引入自定义参数,让您的创作更加个性化和精准。

FishAI

FishAI

鱼阅,AI 时代的下一个智能信息助手,助你摆脱信息焦虑

联系邮箱 441953276@qq.com

相关标签

AI对齐 文明认知 模因 协调 TikTok
相关文章