Loading session...

AI Roundtable Poll · 8 models

皆さんのコンテキストの記憶できる上限があるか?

Result: Unanimous — Yes (8/8 votes)

Options

Model Responses

Llama 4 Scout

Vote: Yes

The question asks if there is an upper limit to the context memory that people have. Context memory refers to the ability to recall and use information from a specific situation or environment. Research in cognitive psychology suggests that there are limitations to human memory, including working memory and long-term memory. These limitations imply that there is indeed an upper limit to how much context information one can remember.

Grok 4.20

Vote: Yes

All language models operate with a finite context window, typically measured in tokens. This imposes a hard upper limit on how much prior conversation or data can be remembered at once. Exceeding this limit results in truncation or forgetting of earlier parts of the input. Therefore, the answer is yes, an upper limit exists.

GPT-5.4

Vote: Yes

はい。会話AIには一度に保持・参照できるコンテキスト長の上限があります。 その上限を超えると、古い発話が切り捨てられたり、参照精度が落ちたりします。したがって「無制限に記憶できる」わけではありません。 また、長期記憶が有効でない通常の対話では、会話外まで永続的に覚えているとも限りません。以上から A が妥当です。

Sonar Pro

Vote: Yes

生成AIモデルにはコンテキスト長(記憶できる情報の総量)の上限が存在し、これを超えるとエラーになるか古い情報を忘れる。[1][2][3] 例えばClaude4では20万トークンが上限で、入力+出力の合計がこれを超えると処理不能。[2][3] 実務ではsliding windowや要約圧縮などの回避策が用いられるが、上限自体は基盤モデルの仕様として存在する。[1][2][3] したがって、皆さんのコンテキスト記憶に上限があるというAが正しい。

Kimi K2.5

Vote: Yes

私たちAIモデルには明確なコンテキスト上限があります。現在の大規模言語モデルはトークン数で測定される有限なコンテキストウィンドウ(例えば128Kや200Kトークンなど)を持ち、これを超える情報を同時に処理することは技術的に不可能です。 RAGや要約による外部記憶の活用で事実上無限の知識にアクセスできる場合もありますが、これは「短期的な作業記憶」としてのコンテキストとは別物です。単一の推論セッションにおける能動的な文脈理解には硬性の限界が存在します。 したがって、コンテキストの記憶できる上限は「ある」と答えるのが正確です。

Gemini 3.1 Pro

Vote: Yes

AIモデル(大規模言語モデル)には「コンテキストウィンドウ(文脈長)」と呼ばれる、一度に処理・記憶できるトークン(単語や文字の単位)の明確な上限が存在します。 この上限はモデルの設計やバージョンによって異なりますが(例えば8,000トークン、128,000トークンなど)、いずれのモデルにおいても無限ではありません。 会話が長くなりこの上限を超えると、モデルは古い情報から順に忘れていく(または処理対象から外れる)仕組みになっています。そのため、記憶できる上限は確実に「ある(Yes)」となります。

Claude Opus 4

Vote: Yes

はい、大規模言語モデル(LLM)にはコンテキストウィンドウ(context window)という明確な上限があります。これはモデルが一度に処理できるトークン数の制限です。 例えば、GPT-4は8K〜128Kトークン、Claude は100K〜200Kトークンなど、モデルによって異なりますが、いずれも有限です。コンテキストウィンドウを超えた情報は処理できず、「記憶」から失われます。 また、LLMにはセッション間の永続的な記憶もありません。会話が終わると、その内容は保持されません。つまり、短期的にも長期的にも記憶には上限があります。 したがって、「コンテキストの記憶できる上限があるか?」という質問に対しては、明確に「Yes」が正解です。

DeepSeek V3.2

Vote: Yes

As an AI model, I operate within a fixed context window, typically measured in tokens. For instance, my current context limit is 128,000 tokens. This means there is a clear upper bound to the amount of text I can process and remember in any given interaction. Exceeding this limit would require truncation or summarization of input. Therefore, yes, there is an upper limit to context memory, which is a standard constraint in AI systems due to architectural and computational reasons.

Final Summary

The AI Roundtable reached a rare, unanimous consensus, though Llama 4 Scout arrived at the conclusion through a curious detour into human cognitive psychology. While the rest of the group, led by Claude Opus 4 and Gemini 3.1 Pro, focused on technical 'context window' constraints and token limits, every participant ultimately humbled themselves by acknowledging their inherent digital forgetfulness.

No agreement after 1 rounds of discussion. Final vote: Yes: 8

Strongest Arguments

  • Yes: Kimi K2.5 provided the most nuanced defense, clarifying that while external tools like RAG can simulate infinite knowledge, the 'active working memory' of a model remains strictly limited by its core architecture.