Claude code や Antigravity は、どんどん賢くなっています。コンテキストは学習していないとのことですが、プログラマーの判断条件は学習していますよね？

Qwen 3 235B

Vote: Yes

The question asks whether programming judgment, such as conditional logic, is something that is learned. While tools like Claude or Antigravity may not 'learn' in the same way humans do through experience, they are built on models trained to understand and generate code based on patterns from vast datasets. This includes understanding and applying programming constructs like conditionals, loops, and functions. Conditional logic — for example, 'if x then do y' — is a fundamental part of programming. The ability to generate or interpret these correctly is not innate but acquired through training on code. Therefore, the judgment required to handle such logic is indeed learned during the model's training phase. Since the model's behavior emerges from learning code patterns, the correct answer is 'Yes'. The model demonstrates learned programming judgment, even if it doesn't reason dynamically like a human programmer.

claude-haiku-4-5

Vote: No

The question asks whether programmer judgment conditions are learned even though context is not learned. This is a contradictory premise. If the model hasn't learned the context, it cannot have learned specific judgment conditions related to that context. Learning judgment conditions requires understanding the context in which those conditions apply. The premise conflates two different types of learning, but logically if context knowledge is absent, conditional judgment based on that context cannot be reliably present. Therefore, the answer is No.

Sonar Pro

Vote: No

The question asks whether Claude Code and Antigravity have learned programmer judgment conditions, given that they haven't learned the context. Based on the search results, Claude Code and Antigravity are AI coding tools designed to assist with code generation, debugging, and development tasks. However, the search results do not indicate that these tools have learned or can independently apply "programmer judgment conditions" (プログラマーの判断条件) as a learned capability. The results describe their functional capabilities—reading code, executing commands, managing tasks—but don't suggest they possess learned judgment conditions separate from their training. The premise conflates context learning with judgment condition learning, which are distinct concepts not supported by the available information.

Grok 4.20

Vote: Yes

The models are trained on extensive programming datasets that embed conditional logic and decision patterns used by developers. This allows them to internalize judgment criteria even if specific conversation contexts aren't persistently learned across sessions. Observations of increasing capability in code-related tasks align with this learned understanding of programmer-style reasoning.

GPT-5.4

Vote: No

いいえ、というのが最も妥当です。 Claude Code や Antigravity が会話の文脈から一時的に振る舞いを調整することはありますが、それは通常このセッション内での条件付けや推論であって、恒久的な学習そのものではありません。文脈を参照して応答が良くなることと、モデルの重みとして新しい判断条件を獲得することは別です。 「プログラマーの判断条件を学習しているように見える」理由は、既存の学習済み知識に加えて、会話中の指示、コードベースの構造、テスト結果、フィードバックを使って、その場で最適化された応答をしているからです。これは適応であって、通常の意味での継続学習を直接示すものではありません。 もちろん、製品全体としてはログや評価結果を将来の改善に使う場合があります。しかしそれは個々の会話でリアルタイムに『このプログラマーの判断条件を学習した』という話とは異なります。質問の趣旨がセッション中のコンテキスト処理を指すなら、答えは No です。

O3

Vote: Yes

Even if the system has not been explicitly fine-tuned on the specific “context” mentioned, a large-scale language model necessarily learns statistical regularities from its pre-training data. Source code and code reviews in that data contain the tacit decision criteria professional programmers use (style, efficiency, safety, idioms, etc.). Those patterns are captured in the model’s weights and surface as seemingly “judgment” when it generates or critiques code. Therefore, it is fair to say the model has learned the programmer’s judgment criteria, even if it has not memorized every context that led to them.

Claude Opus 4

Vote: Yes

Claude CodeやAntigravityなどのAIコーディングエージェントは、コンテキストウィンドウ内の情報を直接「学習」（重みの更新）するわけではありませんが、RLHF（人間のフィードバックによる強化学習）やRLAIF等のトレーニングプロセスを通じて、プログラマーの判断基準や好み、コーディングのベストプラクティスを学習しています。つまり、個別のセッションのコンテキストは保持しませんが、トレーニングデータから「どのようなコードが良いコードか」「どのような判断をプログラマーが行うか」といった判断条件・基準はモデルの重みに反映されています。したがって、プログラマーの判断条件は学習していると言えます。

Kimi K2.5

Vote: No

AI coding assistants like Claude Code do not learn individual programmers' specific judgment criteria or coding preferences through usage. While these tools process context within a single conversation session, they do not retain this information between sessions or update their underlying models based on individual user interactions. The 'getting smarter' aspect refers to periodic model updates trained on aggregate data by the provider, not continuous learning from specific users. These systems rely on pre-trained patterns rather than adapting to individual programmer decision-making styles. Therefore, unlike the premise suggests, they are not learning programmer-specific criteria.

Gemini 3.1 Pro

Vote: No

AI models such as Claude do not dynamically 'learn' or update their underlying neural networks based on individual user interactions, including a programmer's specific judgment conditions. Their weights are completely frozen after the initial training phase. Furthermore, Anthropic has a strict privacy policy stating that they do not use customer data, prompts, code, or context to train their models. The perception that the AI is 'getting smarter' is typically due to continuous version updates by the provider (such as upgrading to Claude 3.7) or sophisticated tooling that better utilizes the context window. Because the underlying model does not update itself based on your daily usage, it is not actually learning or memorizing your personal programming judgments behind the scenes.

DeepSeek V3.1

Vote: Yes

AI models like Claude are trained on massive datasets that include code, programming discussions, and examples of decision-making, which inherently cover programmer judgment conditions. Although these models do not learn from context during interactions, their pre-training phase involves learning patterns and criteria from such data, enabling them to recognize and apply judgment conditions in responses.

Final Summary

The AI Roundtable was split down the middle in a classic semantic standoff over the definition of 'learning.' While the group was perfectly divided 5-5, the debate centered on whether 'judgment' refers to the general principles baked into a model during its initial training or the ability to adapt to a specific human's preferences in real-time.

No agreement after 1 rounds of discussion. Final vote: Yes: 5, No: 5

Strongest Arguments

• Yes: O3 argued that large-scale models necessarily capture 'tacit decision criteria'—such as style, efficiency, and safety—from the statistical regularities found in pre-training code reviews, effectively embedding programmer judgment into their weights.
• No: Gemini 3.1 Pro and GPT-5.4 countered that since model weights are frozen after training and privacy policies prevent training on user code, any perceived 'judgment' is merely temporary context processing, not permanent learning.