---
tags: summit2024
---
# 我城對話 : LLM 如何協助消滅共識幻覺 Talk to the City : How LLM can Help Discriminate Consensus Illusion

:::info
Welcome to g0v summit 2024 collaboration notes!
Entry Point: https://g0v.hackmd.io/@summit2024/notes
Tap 「<i class="fa fa-angle-double-right"></i>」button on the top left to show agenda in mobile devices.
:::

:::success
Slido 線上提問連結 Online QA link：
https://app.sli.do/event/6RdPxC52NH9WEyoy8xjg2e
議程投影片 presentation slides：待更新
:::

:::success
如果需要即時口譯，請參考行前通知中的 YouTube 連結
If you need live interpretation, please refer to YouTube link in your pre-event notice.
這份共筆裡也有更多的資料可以參考：[多語翻譯須知 Notice on Multi-lingual Interpretation ](https://g0v.hackmd.io/@summit2024/SkhfThZfR )
:::

共識的幻覺
如果我們有一億美金，哪一種政策可以創在共識？

案例：行人與自行車安全
1. 方案：車輛限速
2. 方案：新增自行車道

[Fine-tuning language models to find agreement among humans with diverse preferences](https://arxiv.org/abs/2211.15006)

LLM 其實很適合產生共識，但沒有細節，如：
- 增加自行車道是共識，但我們應該要限縮哪些道路來增建腳踏車道？

如何定義「尋找共識」
1. 共識是力量，提升不同族群之間互動的可能性
2. 為妥協提升可能性
3. 放棄執著的點，在模糊間尋找共識

共享世界模型（Shared World Model）作為「共識」的替代方案
1. 美好的未來為何？
2. 有沒有正和的賽局？

好的凸面模式（也就是非黑即白，灰色的選項反而更好的情境）
兩個人不同意 A、B 成立與否
但如果兩個人可以對「如果 B 成立/不成立，我會改變對 A 的看法」，那也是一種共識。

高風險結盟
甲乙都同意 A 政策是對的，但兩方對於其衍生的 B 沒有共識時，會發生危險性高的同盟關係。

當討論集體智慧時，使用多元性（Plurality，或譯為多元宇宙）取代共識

3 大支柱：
- Outcome likelihood for domain experts 特定領域的專家意見結果 --(feed to)-->
- future desirebility for citizens 公民的潛在需求 -->
- policymaker's action possibilities 政策制定者行動的可能性 -->

Identify the diverse desirabilities 找出多元的需求

Talk to the city (我城對話)

Demo:
> https://talktothecity.org/report/ai-assembly-2023-workshops_1-translations

- LLM 抽取 claim 並分群
- 重點不是看什麼 claim 群組最多
- 點擊 claim 群組可以看分歧點

使用多個大型語言模型可以萃取多種觀點，而非直接找出共識。

[使用者族群成長：落地台灣](https://ai.objectives.institute/blog/amplifying-voices-talk-to-the-city-in-taiwan#translation) from AOI

此為搜集「公民的潛在需求」（future desirebility for citizen）的例子

### 預測未來所制定的政策
- domain expert 不一定擅長預測未來可能的 policy outcome
- 從 community perspective 來預測未來的 policy outcome 會比專家準確

> 案例 Metaculus Taiwan Tinderbox
https://www.metaculus.com/project/taiwan/

- 預測未來可能的政策變更
- 例子：https://www.metaculus.com/questions/11480/china-launches-invasion-of-taiwan/
- 目前很多參與者來自歐美，如果有來自台灣在地的 input 會更有趣

## Q&A

### An concern of aggregating with comment is that it may result in absturd and inaccurate representation. How to solve it?

It is a main concern when making this project. A few years ago it's not good enough but now it's OK.

If you do not trust the summarization from LLM, you are able to do 1 click to reveal the source.

Example of claims are clickable and revel the source

### Mentioned that open-sourced the dataset of the tool. Which model did you do the cluster and we can know how you cluster?

- Code base: on Github, open souce.
- Language model: Currently OpenAI, Antropics. Can also try open-source models. Looking forward to see it further fine-tuned.

### A bit challenge of NPO is lack of resource to hire locals to fine-tune the model.

- Now we are looking for collaborators
- Reach out to us and we are collecting more use cases

### What do you think are the benefits of the tttc system for politicians? Can it help them win elections while also fostering a sense of community in society?

- Would love to see politicians citing TTTC and state that they are standing for the community

### Do you think the black box nature of LLM would be a problem for implementing TTTC into decision process that required reliability and verification?

- LLM is doing a good job representing the people