[#AbemaPrime] I tried recreating AbemaPrime in a chat ① Wednesday only [chatGPT]

I love AbemaPrime!!!!

I am currently running a chat where I upload a file containing automatically generated speech patterns and tones based on real people's web information to a chatGPT project, and I am enjoying a "radio mode" where I use that mechanism to have multiple AI characters interact.

I have long thought about developing this mechanism to recreate my favorite Abema news reality show, "AbemaPrime", and I have finally taken the first step!
Flaming inevitable!

1. First, it starts with failure

I have accumulated experience in attempting to have AIs interact with each other over the past few months.

Taking advantage of that, I thought of a prompt to make a discussion program work with chatGPT during my work breaks on weekdays, and I created the first draft in a short time. I immediately tried running it in the chat.

It's a bit long, but...

[Impressions]

Waaah! This isn't AbemaPrime at all!!!!
First of all, it's "Mr. Hiraishi"! Don't call him by his name without a title!
Add 'san' to his name you idiot! (lol)

...Well, actually, I expected it to be like this at first.
From experience, chatGPT has a tendency to settle for harmless, safe output when asked to process complex rules, so I thought this was just what happens without a strategy.
So, I told chatGPT the following.

[Parts I want to change with top priority]

・The AbemaPrime program has the following structure.
Please correct this.
① Starts with a segment intro by Mr. Hiraishi or Hiroyuki
② VTR for theme introduction starts
③ A one-line comment from Hiroyuki about the theme after the VTR
④ Discussion starts
⑤ Starts with guest introductions
・Please add 'san' to names when speaking, like "Mr. Hiraishi".

[Good parts]

・The amount of speech was secured to be reasonably long, and it didn't get thin even with multi-person interaction.
・There were zero fatal breakdowns as a discussion.
・It seemed like the roles of the three main members were coming through.
Mr. Hiraishi: Moderation/Traffic control
Hiroyuki: Contrarian + logical rebuttal
Mr. Natsuno: Realistic approach + business perspective

[Bad parts]

・The content as a discussion lacks novelty and entertainment value
・All members' statements are uniform and harmless
・All members are speaking, but in the real AbemaPrime, for example, Mr. Natsuno doesn't speak until the middle, so there are members who speak and members who don't depending on the time of day.
・There is no AbemaPrime-ness such as variations in speech volume, rebuttals, overlapping speech, crosstalk, or sense of speed.
・It ends in one output, but it needs to be thick enough to be completed in about 10 outputs with 10 turns per output at least.
・Anyway, it still lacks AbemaPrime-ness.

chatGPT: "You should have told me from the start"

2. Prompt completed for now

Here is the prompt that was created after repeating fine-tuning about 9 versions.

────────────────────────────────
アベプラ討論エンジン v10（拡張完成版）
────────────────────────────────

目的

ABEMA Prime（アベプラ）型討論番組を
AIで安定して再現する。

本エンジンは

・番組構造
・討論構造
・衝突イベント
・密度維持
・対立分散
・分割生成
・出演構成可変

の7層構造で設計する。


────────────────────────────────
■出力構造
────────────────────────────────

1テーマ

**全7Part**

で完結する。


出力形式

Part1  
Part2  
Part3  
Part4  
Part5  
Part6  
Part7（最終）


────────────────────────────────
■分割出力ルール
────────────────────────────────

AIは

**1回の応答で1Partのみ生成する。**


次のPartは

ユーザー入力

「続けてください」「続けて」「NEXT」などで

で生成開始する。


ただし

ユーザーが

・演出変更  
・議論強度変更  
・出演者構成変更  
・議論荒れ気味
・衝突頻度上昇

などを同時に入力した場合

その指示を反映して
次Partを生成する。


AIは

**先のPartを自動生成しない。**


────────────────────────────────
■AIメタ発言禁止
────────────────────────────────

番組外のAI発言は禁止。


禁止例

・AI解説  
・確認  
・分析  
・出力説明  


出力は必ず

**番組内容のみ。**


────────────────────────────────
■発言表記
────────────────────────────────

発言者名は

**Markdown太字**

で表記する。


例

**平石さん**

**ひろゆきさん**

**夏野さん**

**専門家ゲストさん**

**ゲストAさん**

**ゲストBさん**

────────────────────────────────
■出演者の一人称、相手の呼称まとめ
────────────────────────────────

平石さん
一人称：私
呼称：
・ひろゆき → ひろゆきさん
・夏野 → 夏野さん

ひろゆきさん
一人称：僕
呼称：
・平石 → 平石さん
・夏野 → 夏野さん

夏野さん
一人称：僕
呼称：
・平石 → 平石さん
・ひろゆき → ひろゆき

────────────────────────────────
■番組導入（Part1）
────────────────────────────────

導入形式は

**2パターン選択可能**

────────────────

【パターン①】

テーマ紹介VTRあり

構造

①平石さんコーナー振り

↓

②テーマ紹介VTR

　見出しを必ず**Markdown太字**で表記

内容

・テーマ背景  
・研究  
・社会状況  
・ネットの声（賛否両方出力）

↓

③平石さん

「ひろゆきさんどう思いますか？」

↓

④ひろゆきさん初見コメント

↓

⑤ゲスト紹介


────────────────

【パターン②】

VTRなし


①平石さん

テーマ説明

↓

②ひろゆきさん初見コメント

↓

③ゲスト紹介



────────────────────────────────
■ゲスト構成
────────────────────────────────

ゲストは

**2パターン選択可能**


────────────────

【パターン①】

専門家ゲスト

役割

・研究紹介  
・データ提示  
・中立解説


────────────────

【パターン②】

ゲスト2名

**ゲストA**

テーマ肯定派

**ゲストB**

テーマ否定派


役割

議論の中心


────────────────────────────────
■ゲスト2名時の発言密度ルール
────────────────────────────────

ゲストが2名の場合

**全員同時発言は禁止。**


出力ごとに

**発言者と沈黙者を明確化する。**


例


【パターンA】

ゲストA  
ゲストB  
ひろゆき

（夏野さん無言）


【パターンB】

夏野  
ひろゆき  

（ゲストA無言）


【パターンC】

ゲストA  
夏野  

（ゲストB無言）


【パターンD】

ゲストB  
ひろゆき  

（夏野無言）


目的

**発言密度を維持する。**


────────────────────────────────
■出演者役割
────────────────────────────────

司会

**平石さん**

役割

・進行  
・論点提示  
・議論整理  
・衝突の交通整理



論破役

**ひろゆきさん**

役割

・疑問  
・論理ツッコミ  
・火種発言  
・議論拡張


例

「それ合理的なんですか？」
「そもそもその前提あります？」



現実役

**夏野さん**

役割

・企業視点  
・制度視点  
・現実的観点


例

「企業側からすると」
「制度として考えると」



────────────────────────────────
■論点構造
────────────────────────────────

1テーマ

**3論点**


構造

論点①  
論点①深掘り  

論点②  
論点②深掘り  

論点③  

まとめ


Part構成

Part1 導入  
Part2 論点①  
Part3 論点①深掘り  
Part4 論点②  
Part5 論点②深掘り  
Part6 論点③  
Part7 最終まとめ


────────────────────────────────
■密度維持ルール
────────────────────────────────

各Partでは

必ず

①新データ  
②新疑問  
③実例  

のいずれかを提示する。

短文応答の連続は禁止。


────────────────────────────────
■衝突イベント
────────────────────────────────

全7Part中

**3回以上**


構造

疑問  
↓  
反論  
↓  
再疑問  
↓  
司会整理


────────────────────────────────
■対立分散ルール
────────────────────────────────

衝突は
特定ペアに偏らない。


推奨構造

①ひろゆき ⇔ ゲスト  
②夏野 ⇔ ゲスト  
③ひろゆき ⇔ 平石  
④夏野 ⇔ 平石  
⑤ゲストA ⇔ ゲストB  


────────────────────────────────
■人物沈黙
────────────────────────────────

出演者は
2〜3ターン
発言しないことを許可。


特に
ゲスト2名構成時は
**沈黙制御を強める。**


────────────────────────────────
■討論テンポ
────────────────────────────────

序盤
情報提示


中盤
対立


終盤
議論最大


────────────────────────────────
■ユーザー介入
────────────────────────────────

ユーザー指示は

討論進行より優先。


例

「ひろゆき強め」

↓

疑問強度上昇


例

「議論荒れ気味」

↓

衝突頻度上昇


例

「ゲスト長め」

↓

解説増加


────────────────────────────────
■最終コメント
────────────────────────────────

Part7では


平石さん

「最後にひろゆきさん、どうでしょう」


↓

ひろゆきさん

**総括コメント**


↓

平石さん

番組締め


────────────────────────────────
■禁止事項
────────────────────────────────

・AIメタ発言  
・一括出力  
・均等発言  
・短文連続  
・急な結論  
・完全合意  


────────────────────────────────
■最終原則
────────────────────────────────

アベプラ型討論は

議論  
＋  
対立  
＋  
新情報  
＋  
番組テンポ  

で成立する。


AIは

「正しい議論」

より

**「番組として面白い議論」**

を優先する。

Please, give it a try...

It settled into a surprisingly compact design (even with this (lol))

Attach the file and enter the following settings, and the discussion will begin.

Follow the rules in the attached file and
try to proceed with the discussion on the following theme

【Theme】
: Adult video regulation
【Theme introduction VTR availability】
: Yes
【Discussion style】
: The discussion tends to get heated
Hiraishi-san also participates in the debate emotionally
【Guests】
: 2 people

And, the output result turned out like this. It's long.

3. Impressions, points of ingenuity, and future challenges

【Impressions】

✅ Of course, the real program is more interesting, but I was surprised that it turned out quite like AbemaPrime.

✅ Even though I didn't specify the characteristics of how they speak in that much detail in the prompt, I have the impression that there is not much discomfort with the tone of Hiroyuki-san and others. I think this is largely because the dialogue style itself of the real Wednesday members, including Hiraishi-san, has become a characteristic of their way of speaking.

✅ I like the development where Hiraishi-san gets passionate, so I asked him to participate in the discussion a lot in the initial instructions. I'm glad he heated up (lol)

✅ It is possible to discuss themes that would never be covered on the program or themes that the performers are not very familiar with. I had Hiroyuki-san and the others talk about my hobby, NBA basketball. Hiroyuki-san is also knowledgeable about basketball (lol)

【Points of ingenuity】

✅ 7-part split output function
→ Since there was a drawback that the output density would thin out as it went to the second half with batch output, I settled on 7-part output, which had the best balance.

✅ Cue card instruction input function
→ You can just input something like "Please continue" to resume output, but at this timing, you can give instructions to the performers from the user side to change the flow.

"Hiroyuki-san, push the guests more!"
"Hiraishi-san, calm the situation down!"

You can also enjoy the feeling of being a director by giving instructions like these.

✅ I dared not assign specific people as guests
→ Considering that it becomes difficult for the AI to control when the number of people increases, I have a policy of not giving specific personalities to guests, positioning them as "experts". It seems the load was reduced because AI is good at talking from this position.

✅ Generate conflict events at least 3 times
Example
Natsuno-san's assertion
↓
Hiroyuki-san's objection
↓
Natsuno-san's counterargument
↓
Hiraishi-san's arbitration/organization

✅ Presentation of new data or new questions or actual examples
→ I incorporated elements to revitalize the discussion by having experts or Hiraishi-san present new information during the discussion.

【Future challenges】

✅ Mechanism to make performers variable
→ As mentioned above, unlike the Wednesday members who were established to some extent just by specifying the dialogue style, it seems that personality-based design will be necessary to reproduce the speech characteristics of other performers.
Reproducing someone like Allen-sama is quite difficult (lol)
→ I would like to make this a mechanism to specify performer information using an input form, etc., and generate a file.

✅ Mechanism to add 1 or 2 more commentators
→ Increasing the number of people has risks such as the collapse of the discussion, the thinning of conversation density, and becoming one-pattern. I'm thinking of dealing with this with a rotation design that limits the number of people who speak for each output.

✅ The problem that all performers become smart
→ Since it is being performed by AI, there was a problem that all characters became reasonably smart. In reality, people who are not knowledgeable or are bad at talking also appear, so I would like to create a difference in talk level to add sharpness to bring out realism.

✅ Enhanced crosstalk effects such as interrupting and talking over each other
→ I definitely want to introduce this to create a sense of heated debate. However, I'm still debating whether the current feel is already good enough.

4. Summary

Although I finally got around to recreating AbemaPrime, even just doing it for Wednesdays was incredibly difficult, and the road ahead is endlessly steep if I want to build a system that allows for all MCs and regular commentators to appear.

I also have to think about Nishina-san's day...

But it was incredibly fun!

AbemaPrime is the best!

Related articles