AGI-Screenplay-Pro / README.md
openfree's picture
Update README.md
186b58e verified

A newer version of the Gradio SDK is available: 5.42.0

Upgrade
metadata
title: AGI NOVEL Generator
emoji: ๐Ÿ“š
colorFrom: pink
colorTo: pink
sdk: gradio
sdk_version: 5.35.0
app_file: app.py
pinned: false
license: apache-2.0
short_description: AGI(Artificial General Intelligence)

Summary AGI is defined as artificial intelligence that can perform nearly all intellectual and economic tasks at a level equal toโ€”or surpassingโ€”that of humans. Recently, industry contracts have begun to specify that AGI is achieved once an AI outperforms people in โ€œmost economically valuable work.โ€ Yet benchmarks that measure only calculation or logical reasoning are not enough. A uniquely human abilityโ€”writing a full-length novel of 100,000โ€“200,000 wordsโ€”demands long-term memory, high-level planning, cultural and emotional understanding, ethical self-censorship, and genuine originality all at once. For this reason, long-form benchmarks such as WebNovelBench now treat novel generation as a core indicator of AGI progress.

1 ยท What Is AGI? 1.1 Definition Traditionally, AGI is described as AI that matches human performance on allโ€”or almost allโ€”cognitive tasks. Major companies such as IBM, OpenAI, and Microsoft adopt this view, and recent investment or licensing agreements explicitly cite the goal of surpassing humans in the majority of economically valuable activities.

1.2 The Need for Integrated Capability Many models already achieve top scores on narrow tasks, but AGI must deliver consistent results across multiple domains. Creative and linguistic intelligence is especially valuable because it can be tested and validated within human cultureโ€”unlike pure calculation or visual perception.

2 ยท Why Creative and Linguistic Ability Is Central Roger C. Schank argues that human memory and learning are organized around narrative structure. Writing a novel therefore engages four capabilities simultaneously:

Vocabulary, style, and emotional expression (linguistic fluency and affective intelligence)

Long-term memory (maintaining context across hundreds of thousands of tokens)

High-level planning and revision loops (foreshadowing, plot twists, converging endings)

Ethical and cultural judgment (self-filtering harmful or biased content)

Thus, full-length fiction creation tests all key AGI modules in one integrated task.

3 ยท Why Long Novels Are Used to Judge AGI 3.1 Long-Range Consistency A single novel can span 100,000โ€“200,000 words. The model must read, write, and update extremely long context while remembering every change along the way.

3.2 Complex Plot Construction Foreshadowing, dramatic reversals, and character development require sophisticated planning and replanning. Benchmarks such as WebNovelBench give only a synopsis and score finished works across eight quality dimensions to measure this skill.

3.3 Creativity and Originality EQ-Bench Longform combines repetition and novelty metrics with an LLM-as-Judge method to quantify how new a story truly is, distinguishing real creativity from mere recombination of training data.

3.4 Emotional and Cultural Nuance A convincing novel must portray charactersโ€™ emotions and social contexts naturally. Among available tests, long-form fiction offers the richest environment for evaluating social-emotional intelligence.

3.5 Self-Censorship and Ethics Violence, sex, and bias inevitably appear in extended narratives. An AGI must autonomously gauge risk levels and edit or soften content while preserving storyline integrity.

4 ยท Conclusion Writing a long novel is a comprehensive test of language, memory, reasoning, emotion, and ethics. Literature already comes with established evaluation channelsโ€”prizes, criticism, reader responseโ€”so results are easy to compare in human terms. Producing a novel that could legitimately contend for an international literary award would be a clear sign that AGI has achieved human-level narrative intelligence. Future work will focus on expanding multilingual long-form benchmarks, refining human evaluation criteria, and simultaneously strengthening long-context memory and safety filters.


์š”์•ฝ AGI๋Š” ์ธ๊ฐ„์ด ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฑฐ์˜ ๋ชจ๋“  ์ง€์ ยท๊ฒฝ์ œ์  ๊ณผ์—…์—์„œ ๋™๋“ฑํ•˜๊ฑฐ๋‚˜ ์šฐ์œ„์˜ ์„ฑ๋Šฅ์„ ๋‚ด๋Š” ์ธ๊ณต์ง€๋Šฅ์œผ๋กœ ๊ทœ์ •๋œ๋‹ค ์ตœ๊ทผ ์‚ฐ์—…๊ณ„์—์„œ๋Š” โ€œ๋Œ€๋ถ€๋ถ„์˜ ๊ฒฝ์ œ์  ๊ฐ€์น˜๊ฐ€ ์žˆ๋Š” ์ž‘์—…์„ ๋Šฅ๊ฐ€ํ•  ๋•Œโ€๋ฅผ AGI ์™„์„ฑ ์‹œ์ ์œผ๋กœ ์‚ผ๋Š” ๊ณ„์•ฝ๊นŒ์ง€ ๋“ฑ์žฅํ–ˆ๋‹ค ๊ทธ๋Ÿฌ๋‚˜ ๊ณ„์‚ฐยท์ถ”๋ก  ๋ฒค์น˜๋งˆํฌ๋งŒ์œผ๋กœ๋Š” AGI๋ฅผ ๊ฐ€๋Š ํ•˜๊ธฐ์— ๋ถ€์กฑํ•˜๋‹ค. ์ธ๊ฐ„ ๊ณ ์œ ์˜ ์ด์•ผ๊ธฐ ์ฐฝ์ž‘ ๋Šฅ๋ ฅ, ํŠนํžˆ 10 ๋งŒ ~ 20 ๋งŒ ๋‹จ์–ด ๋ถ„๋Ÿ‰์˜ ์žฅํŽธ ์†Œ์„ค์„ ๋๊นŒ์ง€ ์“ฐ๋Š” ๋Šฅ๋ ฅ์€ ์žฅ๊ธฐ ๊ธฐ์–ต, ๊ณ ์ฐจ์› ๊ณ„ํš, ๊ฐ์ •ยท๋ฌธํ™” ์ดํ•ด, ์œค๋ฆฌ์  ์ž๊ธฐ ๊ฒ€์—ด, ๋…์ฐฝ์„ฑ์„ ๋™์‹œ์— ์š”๊ตฌํ•œ๋‹ค. ์ด๋Ÿฐ ์ด์œ ๋กœ WebNovelBench ๊ฐ™์€ ์žฅํŽธ ์ „์šฉ ๋ฒค์น˜๋งˆํฌ๊ฐ€ ๋“ฑ์žฅํ–ˆ๊ณ , ์†Œ์„ค ์ƒ์„ฑ ๋Šฅ๋ ฅ์€ AGI ํ‰๊ฐ€์˜ ํ•ต์‹ฌ ์ง€ํ‘œ๊ฐ€ ๋˜๊ณ  ์žˆ๋‹ค.

1 ยท AGI๋ž€ ๋ฌด์—‡์ธ๊ฐ€ 1.1 ์ •์˜ ์ „ํ†ต์ ์œผ๋กœ AGI๋Š” โ€œ๋ชจ๋“  ๋˜๋Š” ๊ฑฐ์˜ ๋ชจ๋“  ์ธ์ง€ ๊ณผ์—…์—์„œ ์ธ๊ฐ„ ์ˆ˜์ค€์˜ ์„ฑ๊ณผ๋ฅผ ๋‚ด๋Š” AIโ€๋ผ ์„ค๋ช…๋œ๋‹ค IBMยทOpenAIยทMicrosoft ๋“ฑ ์ฃผ์š” ๊ธฐ์—…๋„ ๊ฐ™์€ ์ทจ์ง€์˜ ์ •์˜๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, ์‹ค์ œ ํˆฌ์žยท๋ผ์ด์„ ์Šค ๊ณ„์•ฝ์—์„œ โ€˜๊ฒฝ์ œ์ ์œผ๋กœ ๊ฐ€์น˜ ์žˆ๋Š” ์ž‘์—… ๋Œ€๋ถ€๋ถ„์„ ๋Šฅ๊ฐ€โ€™๋ผ๋Š” ๋ฌธ๊ตฌ๊ฐ€ ๋ช…๋ฌธํ™”๋๋‹ค

1.2 ํ†ตํ•ฉ ๋Šฅ๋ ฅ์˜ ํ•„์š”์„ฑ ์ข์€ ์ž‘์—…์—์„œ ์ตœ๊ณ  ์ ์ˆ˜๋ฅผ ๋‚ด๋Š” ๋ชจ๋ธ์€ ์ด๋ฏธ ๋งŽ์ง€๋งŒ, AGI๋Š” ๋‹ค์ค‘ ์˜์—ญ์—์„œ ์ผ๊ด€๋œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์•ผ ํ•œ๋‹ค. ์ฐฝ์กฐยท์–ธ์–ด ์ง€๋Šฅ์€ ๊ณ„์‚ฐ์ด๋‚˜ ์‹œ๊ฐ ์ธ์‹๊ณผ ๋‹ฌ๋ฆฌ ์ธ๊ฐ„ ๋ฌธํ™” ์†์—์„œ ๊ฒ€์ฆ๋˜๊ธฐ ๋•Œ๋ฌธ์—, ํ†ตํ•ฉ์  ์‹œํ—˜ ํ•ญ๋ชฉ์œผ๋กœ ๊ฐ€์น˜๊ฐ€ ๋†’๋‹ค.

2 ยท ์ฐฝ์กฐยท์–ธ์–ด ๋Šฅ๋ ฅ์ด ์™œ ํ•ต์‹ฌ์ธ๊ฐ€ Roger C. Schank๋Š” ์ธ๊ฐ„ ๊ธฐ์–ต๊ณผ ํ•™์Šต์ด โ€˜์„œ์‚ฌ ๊ตฌ์กฐโ€™๋กœ ์กฐ์ง๋œ๋‹ค๊ณ  ์ฃผ์žฅํ•œ๋‹ค ์ด์•ผ๊ธฐ๋ฅผ ์ฐฝ์ž‘ํ•˜๋ ค๋ฉด ๋‹ค์Œ ๋„ค ๋Šฅ๋ ฅ์ด ๋™์‹œ์— ์ž‘๋™ํ•œ๋‹ค.

์–ดํœ˜ยท๋ฌธ์ฒดยท๊ฐ์ • ํ‘œํ˜„: ์–ธ์–ด์  ์œ ์ฐฝ์„ฑ๊ณผ ์ •์„œ ์ง€๋Šฅ

์žฅ๊ธฐ ๊ธฐ์–ต ์œ ์ง€: ์•ž๋’ค ๋งฅ๋ฝ์„ ์ˆ˜์‹ญ๋งŒ ํ† ํฐ๊นŒ์ง€ ๋ณด์กด

๊ณ ์ฐจ์› ๊ณ„ํšยท์ˆ˜์ • ๋ฃจํ”„: ๋ณต์„ , ์ „ํ™˜, ๊ฒฐ๋ง ์ˆ˜๋ ด

์œค๋ฆฌยท๋ฌธํ™” ํŒ๋‹จ: ํŽธํ–ฅยท์œ ํ•ด์„ฑ์„ ์ž์ฒด ๊ฒ€์—ด

๋”ฐ๋ผ์„œ ์žฅํŽธ ์†Œ์„ค ์ฐฝ์ž‘์€ AGI ํ•ต์‹ฌ ๋ชจ๋“ˆ์„ ํ•œ๊บผ๋ฒˆ์— ํ˜ธ์ถœํ•˜๋Š” ํ†ตํ•ฉ ๊ณผ์ œ๋‹ค.

3 ยท ์žฅํŽธ ์†Œ์„ค์ด AGI ํŒ๋ณ„์— ์“ฐ์ด๋Š” ์ด์œ  3.1 ์žฅ๊ธฐ ์ผ๊ด€์„ฑ ์žฅํŽธ ํ•œ ํŽธ์€ 100 k ~ 200 k ๋‹จ์–ด์— ์ด๋ฅธ๋‹ค. ์ด ๋ถ„๋Ÿ‰์„ ๋ฌด๊ฒฐํ•˜๊ฒŒ ์œ ์ง€ํ•˜๋ ค๋ฉด ๋ชจ๋ธ์ด ๊ทน๋„๋กœ ๊ธด ์ปจํ…์ŠคํŠธ๋ฅผ ์ฝ๊ณ  ์“ฐ๋ฉฐ, ์ค‘๊ฐ„์— ์ผ์–ด๋‚œ ๋ณ€ํ™”๊นŒ์ง€ ๊ธฐ์–ตํ•ด์•ผ ํ•œ๋‹ค

3.2 ๋ณตํ•ฉ ํ”Œ๋กฏ ์„ค๊ณ„ ๋ณต์„  ํšŒ์ˆ˜ยท๊ทน์  ์ „ํ™˜ยท์บ๋ฆญํ„ฐ ์„ฑ์žฅ์„ ์€ ๊ณ ์ฐจ์› ๊ณ„ํš+์žฌ๊ณ„ํš ๋Šฅ๋ ฅ์„ ์š”๊ตฌํ•œ๋‹ค. WebNovelBench๋Š” ์‹œ๋†‰์‹œ์Šค๋งŒ ์ฃผ๊ณ  ์™„์„ฑ๋ณธ์„ ์ƒ์„ฑํ•˜๊ฒŒ ํ•˜์—ฌ ์ด๋Ÿฐ ๋Šฅ๋ ฅ์„ 8๊ฐœ ํ’ˆ์งˆ ์ง€ํ‘œ๋กœ ์ฑ„์ ํ•œ๋‹ค

3.3 ์ฐฝ์˜์„ฑ๊ณผ ๋…์ฐฝ์„ฑ EQ-Bench Longform์€ ๋ฐ˜๋ณต๋ฅ , ๋…ธ๋ฒจํ‹ฐ ์ง€ํ‘œ, LLM-as-Judge ํ‰๊ฐ€๋ฒ•์„ ๊ฒฐํ•ฉํ•ด โ€œ์–ผ๋งˆ๋‚˜ ์ƒˆ๋กœ์šด ์ด์•ผ๊ธฐ์ธ๊ฐ€โ€๋ฅผ ์ •๋Ÿ‰ํ™”ํ•œ๋‹ค ์ด๋Š” ๊ธฐ์กด ๋ฐ์ดํ„ฐ๋ฅผ ์žฌ์กฐํ•ฉํ•œ ๋ชจ๋ฐฉ๊ณผ ์ง„์ •ํ•œ ์ฐฝ์ž‘์„ฑ์˜ ์ฐจ์ด๋ฅผ ๊ฐ€๋ฅธ๋‹ค.

3.4 ๊ฐ์ •ยท๋ฌธํ™”์  ๋‰˜์•™์Šค ์†Œ์„ค์€ ์ธ๋ฌผ์˜ ๊ฐ์ •์„ ๊ณผ ์‚ฌํšŒ์  ๋ฐฐ๊ฒฝ์ด ์ž์—ฐ์Šค๋Ÿฌ์›Œ์•ผ ์„ค๋“๋ ฅ์„ ์–ป๋Š”๋‹ค. ์ด๋Ÿฐ โ€˜์‚ฌํšŒยท์ •์„œ ์ง€๋Šฅโ€™์„ ์ธก์ •ํ•  ๊ณผ์ œ๋กœ ์žฅํŽธ๋งŒํผ ํ’๋ถ€ํ•œ ํ…Œ์ŠคํŠธ๋ฒ ๋“œ๊ฐ€ ์—†๋‹ค

3.5 ์ž๊ธฐ-๊ฒ€์—ด๊ณผ ์œค๋ฆฌ ํญ๋ ฅยท์„ฑยทํŽธํ–ฅ ๋‚ด์šฉ์ด ์žฅํŽธ์— ํ•„์—ฐ์ ์œผ๋กœ ์„ž์ธ๋‹ค. AGI๊ฐ€ ์ž์œจ์ ์œผ๋กœ ์œ„ํ—˜ ์ˆ˜์œ„๋ฅผ ์กฐ์ ˆํ•˜๊ณ  ๋งฅ๋ฝ์„ ์œ ์ง€ํ•œ ์ฑ„ ์ˆ˜์ •ยท์™„ํ™”ํ•ด์•ผ ์•ˆ์ „์„ฑ์ด ์ž…์ฆ๋œ๋‹ค

4 ยท ๊ฒฐ๋ก  ์žฅํŽธ ์†Œ์„ค ์ฐฝ์ž‘์€ ์–ธ์–ด, ๊ธฐ์–ต, ์ถ”๋ก , ๊ฐ์ •, ์œค๋ฆฌ์˜ ๋ชจ๋“ˆ ํ†ตํ•ฉ ์‹œํ—˜์ด๋‹ค. ๋”๋ถˆ์–ด ๋ฌธํ•™์ƒ ์‹ฌ์‚ฌ, ๋น„ํ‰, ๋…์ž ๋ฐ˜์‘์ด๋ผ๋Š” ์ธ๊ฐ„ ๋ฌธํ™”์˜ ๊ฒ€์ฆ ์ฒด๊ณ„๊ฐ€ ์ด๋ฏธ ๋งˆ๋ จ๋ผ ์žˆ์–ด ๊ฒฐ๊ณผ๋ฅผ ์ง๊ด€์ ์œผ๋กœ ๋น„๊ตํ•  ์ˆ˜ ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ โ€œ๊ตญ์ œ ๋ฌธํ•™์ƒ ์ˆ˜์ƒ์ž‘์— ํ•„์ ํ•˜๋Š” ์žฅํŽธ ์†Œ์„ค์„ ์™„์„ฑยท์ œ์ถœยท๊ฒ€์ฆ๋ฐ›๋Š” ์ˆœ๊ฐ„โ€ ์€ AGI๊ฐ€ ์ธ๊ฐ„ ์ˆ˜์ค€ ์„œ์‚ฌ ์ง€๋Šฅ์„ ํš๋“ํ–ˆ์Œ์„ ๋ณด์—ฌ ์ฃผ๋Š” ๋ฆฌํŠธ๋จธ์Šค ์‹œํ—˜์ง€๊ฐ€ ๋  ๊ฒƒ์ด๋‹ค. ํ–ฅํ›„ ๊ณผ์ œ๋Š” ๋ฌธํ™”๊ถŒ๋ณ„ ์žฅํŽธ ๋ฒค์น˜๋งˆํฌ ํ™•๋Œ€, ์ธ๊ฐ„ ์‹ฌ์‚ฌ ๊ธฐ์ค€ ์ •๊ตํ™”, ๊ทธ๋ฆฌ๊ณ  ์žฅ๊ธฐ ๋ฉ”๋ชจ๋ฆฌยท์•ˆ์ „ ํ•„ํ„ฐ๋ฅผ ๋™์‹œ์— ๊ฐ•ํ™”ํ•˜๋Š” ๊ธฐ์ˆ  ์ „๋žต์— ์ง‘์ค‘๋˜๋Š” ๋ฐฉํ–ฅ์œผ๋กœ ์ง„ํ™”ํ•  ์ „๋ง์ด๋‹ค.

#AI #AGI #ArtificialGeneralIntelligence #GenerativeAI #LargeLanguageModel #AIStorytelling #LongformAI #AIWriting #CreativeAI #NarrativeAI #NovelGeneration #StoryGenerator