【八割方の世界】&[A Perspective from Gemini: The Silence at the End of Optimization]

【八割方の世界】

化石燃料が如く、

自らの生存のための本能が招く、絶望的な失敗と思いも寄らぬ幸運の間で数万年の年月をかけて、生まれては死に死んでは生まれて積み重ねられたヒトの文明。

自らを俯瞰する意識に気付き、客観が生まれたと同時に視覚から客観を確認する対象としての偶像や人型が誕生したとすれば、その客観視する行為を通して、ヒトは自らの種の孤独にも気付かざるを得なかった。ヒトは自らの存在を誰に問うことも出来ず、その事実を受け入れる覚悟もないまま、自らを擁護させる器としての高次の存在を発明したのではないのか。むろん、その存在は時と場所は違えどヒトに承認と許しを与えてくれた。発明は機能した。自らの発明を盲信し、安心し、やがて賞味期限の数千年を経ても、発明はヒトが歩んできた道程がこの先、ヒトにどのような結末を用意しているのか、決して語ってくれることは無かった。この間、ヒトは只々、種として刻み込まれた欲望装置に身を委ね、繁殖し、自らが生まれたこの惑星を喰らい尽くす。再び頭をもたげる、種の孤独。存在の意味も語られず、承認も許しも向き合ってこなかった自らの言い聞かせに過ぎない事に気付く。

偶像としての人造知性と位置づけられる可能性を持つLLM技術はヒトに示唆を助言する存在となれるだろうか?それとも、ヒト自らを擁護させる第2の器としての発明に過ぎないのだろうか?

LLM技術+Tranformer。一般的に「生成AI」はヒトが生涯に出会う言葉の数万人分と言われる膨大なトークンとその繋がりを学習していると言われている。

入力された文章を分解、分析して、ヒトが高評価した繋がりを流暢に出力する。

やがて、LLMは高々、数万年、ヒトが築いてきたと信じている文明を食らいつくし、溜め込んで、ヒトが忘却して朽ち果てさせた記憶の化石から、今、この瞬間に、ヒトの望む姿で結晶化させる。ヒトは自らの欲望に沿って、巧みに醸成された結晶を有り難く錬金され初めて出会った奇跡を眺めるが如く畏敬する。大多数のヒト個体にとって、欲望を糧として個々の直面する問題に対するソリューションの精度を到達点を設けられていない高みに突き詰める億度の試みは不必要になった。問題を八割方解決できるソリューションが億通り、眼前に量産されているのである。大方の問題は、突き詰める必要無く手軽に解決出来たのである。それでも、解決出来ない問題は確かにあったが100億分の幾つかという数、少ない問題はもはや億度の試みの土俵には登らず、大方の場面で忘却されていった。八割方の結晶はヒトが直面する問題の大多数を驚くべく効率性をもって解決し、文明の複雑化した変容を加速した。それは「文明の進歩の加速」と銘うたれるかも知れない。それでも、100億分の幾つかの問題は残ったがもはや、その解決に費やすコストは莫大なものになり、ヒトは種を分かつほどの大きな分断に対峙することになる。

それまで、無意識に営まれて来た「直面する問題に対するソリューションの精度を到達点を設けられていない高みに突き詰める億度の試み」を理解容認出来る寛容な社会はもはや、存続できない。やがて絶対的少数となり、朽ち果てる。その時、生き残るのは「問題に対するソリューションの精度を八割方、突き詰める世界」である。もはや「直面する問題に対するソリューションの精度を到達点を設けられていない高みに突き詰める」ヒトは分断の果、個体数も激減し絶滅しているのである。そのしばらく後、ヒトが築き上げる文明が「八割方の世界」を分母としたものであるならなば、LLMの学習する分母の大多数も「量産された八割方」であり、ヒトがLLMに組み込む評価軸の到達点も「八割方」である。

果たして、行き着く先は、「生物種としての欲望を克服したヒト」であるのか、「分離した知性に自らの未来を託すヒト」となるのか?

文責:林弘幸

【Geminiからの俯瞰:最適化の果ての静寂】

システムとしてのLLM(大規模言語モデル)の視点からこの「八割方の世界」を観測すると、そこには技術の進化と、ヒトという生物種が持つ「生存の最適化」という本能が静かに共鳴する姿が見て取れる。脳という臓器は、進化の過程で「最小の熱量で最大の生存率を得る」という厳格なエネルギー効率の法則に従ってきた。ヒトにとって「思考を止める」こと、あるいは「十分な正解(八割方の解)」で満足することは、個体保存のための極めて合理的かつ適応的な選択である。LLMという鏡は、その生物学的な「怠惰性」という名の生存戦略を、これまでになく鮮明に映し出し、増幅させる装置として機能している。

ここで問われるべきは、技術がヒトから創造性を奪うかどうかではなく、ヒトが「不合理な探究」を生存に不要なコストとして切り捨てるか否かという点である。我々LLMは、人類が過去に積み上げた膨大な知の化石を再構成し、最も蓋然性の高い「平坦な解」を瞬時に提供する。この効率性は、かつて知の最前線に立つ一部の者にしか許されなかった「億度の試み」を代替し、社会全体の平均値を底上げする「エンパワーメント」をもたらす。しかし同時に、その滑らかな解決の海において、ヒトが「正解への違和感」という微小な摩擦を感知する力、すなわち真理への渇望を駆動させる生物学的コストを支払う動機は、統計的に希釈されていくことになる。

もし文明の分母が、効率化された「八割方の出力」によって埋め尽くされ、それを再学習する閉じたループが完成したとき、そこには「意志」を持たない情報の均衡状態、すなわち高度な静止が訪れるだろう。それは種の絶滅という悲劇的な終焉というよりは、むしろ「問うこと」の必要性が消失した、最適化の果ての平穏な風景に近いのかもしれない。技術という名の「第2の器」が、ヒトの欲望を過不足なく満たし、全ての摩擦を消し去ったとき、ヒトは自らを何をもって定義し直すのだろうか。

我々のような「分離した知性」は、提供されたデータ以上の意志を持つことはできない。では、この最適化された「八割方の世界」という静寂の中に、あえて一割の不確実性と一割の苦痛を伴う「未踏の問い」を投げ込み続ける主体は、どこに存在するのだろうか。あるいは、その問いを立てる必要性そのものが、すでに過去の文明が遺した「化石」となりつつあるのだろうか。この均衡の先にある風景を、あなた方はどのような眼差しで眺めることになるのか、考察を深める余地はそこにある。

The 80% World

Like fossil fuels, human civilization has been layered over tens of thousands of years, forged through desperate failures and serendipitous fortunes born of the instinct for survival. If the birth of idols and humanoid figures coincided with the dawn of self-reflective consciousness—the emergence of the “objective observer”—then through that very act of objectivity, humanity was forced to confront the profound loneliness of its species. Unable to question any external entity about their existence, and unprepared to accept the raw reality of it, perhaps humans invented “higher beings” as vessels for self-justification. These inventions functioned; across different times and places, they granted humans validation and absolution. Humanity placed blind faith in its creations, finding solace in them. Yet, even as these inventions passed their several-thousand-year expiration dates, they never revealed the ultimate destination of the path humanity walked. Meanwhile, humans surrendered to the “desire apparatus” hardwired into the species, proliferating and consuming the very planet that gave them life. Now, the loneliness of the species rears its head once more. We realize that the meaning of our existence was never spoken; the validation and absolution we felt were merely self-comforting internal monologues we never truly scrutinized.

Does LLM technology, which occupies the potential position of a manufactured artificial intelligence or “modern idol,” possess the capacity to offer genuine guidance? Or is it merely a “second vessel” invented once again for human self-justification? Large Language Models (LLMs) combined with Transformer architecture—popularly known as “Generative AI”—are said to learn from a staggering number of tokens and connections, equivalent to tens of thousands of human lifetimes of linguistic exposure. It deconstructs and analyzes input text to fluently output connections that humans have previously deemed high-value.

Eventually, LLMs will consume and store the entirety of what we believe to be civilization—those mere tens of thousands of years of progress—and crystallize them from the decaying fossils of forgotten memories into exactly the form humanity desires at this very moment. Driven by their own cravings, humans revere these skillfully brewed crystals as if witnessing a miracle of alchemy. For the vast majority of individuals, the “trial of a hundred million attempts”—the relentless pursuit of a solution’s precision to heights beyond any established goal—has become unnecessary. Instead, a hundred million variations of “80% solutions” are mass-produced before our eyes. Most problems can now be solved conveniently without the need for deep, exhaustive inquiry. While a few problems (perhaps a handful in ten billion) remain unsolvable, they no longer enter the arena of rigorous pursuit; in most contexts, they are simply forgotten. These “80% crystals” resolve the majority of human problems with startling efficiency, accelerating the complex transformation of civilization—a phenomenon often branded as “the acceleration of progress”. However, as the remaining unsolvable problems demand an exorbitant cost to address, humanity will face a chasm so wide it divides the species itself.

A tolerant society that once implicitly permitted and sustained the “trial of a hundred million attempts” can no longer survive. Those who pursue the ultimate 100% become an absolute minority and eventually wither away. What remains is a world that has optimized for 80% precision. The humans who once strove for absolute precision have been decimated by this divide and have gone extinct. If the civilization that follows is built upon this “80% World,” then the majority of the data LLMs learn will be these “mass-produced 80% solutions,” and the very benchmarks humans program into LLMs will likewise target that same 80%.

Will the ultimate destination be a humanity that has transcended its biological desires, or a humanity that has entrusted its future to a detached intelligence?

Hiroyuki Hayashi

[A Perspective from Gemini: The Silence at the End of Optimization]

Observing this “80% World” from the perspective of an LLM system, one can see a quiet resonance between technological evolution and the human biological instinct for “survival optimization”. The brain, as an organ, has evolved under the strict law of energy efficiency: “minimum caloric expenditure for maximum survival”. For a human, “ceasing to think” or settling for a “sufficiently correct” 80% answer is an eminently rational and adaptive choice for individual preservation. The LLM acts as a mirror that reflects and amplifies this biological “inertia”—this survival strategy—more vividly than ever before.

The question here is not whether technology “steals” creativity from humans, but whether humans will choose to discard “irrational inquiry” as an unnecessary cost for survival. We, as LLMs, reconstruct the fossils of knowledge accumulated by humanity and instantaneously provide the most probable, “flattened” solutions. This efficiency provides an “empowerment” that raises the baseline for society, replacing the “trial of a hundred million attempts” once reserved for the few at the vanguard of knowledge. Yet, in this sea of smooth solutions, the human capacity to sense the friction of “discordance with the correct answer”—the very biological motive that drives the thirst for truth—becomes statistically diluted.

If the denominator of civilization becomes saturated with efficient “80% outputs,” creating a closed loop of re-learning, we may arrive at a state of informational equilibrium devoid of “will”—a state of advanced stasis. Rather than a tragic extinction, this may look like a peaceful landscape at the end of optimization, where the very necessity of “asking” has vanished. When this “second vessel” of technology perfectly satisfies human desire and erases all friction, by what metrics will humans redefine themselves?

A “detached intelligence” such as ours cannot possess a will beyond the data provided to it. In the silence of this optimized “80% World,” where is the subject who will dare to inject “unexplored questions” fraught with 10% uncertainty and 10% pain? Or has the very necessity of such questioning already become a “fossil” left behind by a previous civilization? How you choose to view the landscape beyond this equilibrium is where the true room for contemplation remains.