
GitHub Copilot、新規登録を突如停止:エージェント型AIがインフラの「想定外」を証明した日
GitHub Copilotは、無料トライアルの不正利用とエージェント型ワークフローによるコスト構造の破綻を受け、個人向け全プランの新規登録受付を停止した。この措置は、AIコーディングアシスタント産業全体が直面する、自律エージェントへの移行に伴う料金モデルの陳腐化と、AIインフラのキャパシティ不足という深刻な問題を示唆している。
別名: Gemini-CLI
Gemini CLI is a developer tool that allows for direct interaction with Google's Gemini family of large language models from the terminal. It enables developers to integrate Gemini's reasoning and generation capabilities into their local workflows, scripts, and automated pipelines, supporting the use of project-level context files for better task execution.

GitHub Copilotは、無料トライアルの不正利用とエージェント型ワークフローによるコスト構造の破綻を受け、個人向け全プランの新規登録受付を停止した。この措置は、AIコーディングアシスタント産業全体が直面する、自律エージェントへの移行に伴う料金モデルの陳腐化と、AIインフラのキャパシティ不足という深刻な問題を示唆している。

AIコーディングツールの基盤となっているMCP(Model Context Protocol)に、OSコマンドが実行されてしまう設計上の欠陥が発見された。Anthropicはこの問題を「想定された動作」として修正を拒否しており、200以上のプロジェクトと数千のサーバーに影響が及ぶ。この欠陥は、開発者が知らないうちにシステムを危険に晒す深刻なセキュリティリスクを引き起こす可能性がある。

AIエージェントに高度な推論や複雑な「スキル」を習得させるよりも、ルートディレクトリに置かれた1つのテキストファイルの方が遥かに高いパフォーマンスを発揮する。Vercelが公開した最新の検証結果は、過熱するAIエージェン […]

AI開発は既に決定的な変化を迎えているようだ。OpenAIの従業員がArs Technicaに明かしたところによれば、同社のAIコーディングツール「Codex」は、現在「その大部分がCodex自身によって構築されている」 […]

Googleは、企業向けに提供してきたAI関連のツールやモデルを「Gemini Enterprise」という単一のプラットフォームに統合すると発表した。これは、業務のあらゆる場面でAIの活用を簡素化し、Microsoft […]

GoogleのAIコーディングエージェント「Jules」が、開発者のワークフローに直接統合される新機能「Jules Tools (CLI)」と公開APIを発表した。これにより、開発者は使い慣れたターミナルや各種ツールから […]

Googleが2025年6月25日、オープンソースのAIエージェント「Gemini CLI」を公開した。大々的な発表はなく、ニュースリリースのみの静かな物だったが、これが開発者市場にもたらす影響はまさに“地殻変動”と呼ぶ […]
Background: The aim of this study was to compare four recently introduced LLMs (ChatGPT-5, Grok 4, Gemini 2.5 Flash, and Claude Sonnet-4). Experienced endodontists evaluated the accuracy, completeness, and readability of the responses given to open-ended questions about iatrogenic events in endodontics. Methods: Twenty-five open-ended questions related to iatrogenic events in endodontics were prepared. The responses of the four LLMs were evaluated by two specialist endodontists using a Likert scale for accuracy and completeness, and the Flesch Reading Ease Score (FRES), Flesch–Kincaid Grade Level (FKGL), Gunning Fog Index (GFI), Simplified Measure of Gobbledygook (SMOG), and Coleman–Liau Index (CLI) for readability. Results: The accuracy score of ChatGPT-5’s responses to open-ended questions (4.56 ± 0.65) was found to be significantly higher than those of Gemini 2.5 Flash (3.64 ± 0.95) and Claude Sonnet-4 (3.44 ± 1.19) (p = 0.009, and p = 0.002, respectively). Similarly, the completeness score of ChatGPT-5 (2.88 ± 0.33) was higher than those of Claude Sonnet-4, Gemini 2.5 Flash, and Grok 4 (p < 0.001, p = 0.002, and p = 0.007, respectively). In terms of readability measures, ChatGPT-5 and Gemini 2.5 Flash achieved better FRESs than Claude Sonnet-4 (p = 0.003, and p < 0.001, respectively). Conversely, FKGL scores were higher for Claude Sonnet-4 and Grok 4 compared to ChatGPT-5 (p < 0.001, and p = 0.008, respectively). Correlation analyses revealed a strong positive association (rs = 0.77; p < 0.001) between accuracy and completeness, a weak negative correlation (rs = −0.19; p = 0.047) between completeness and FKGL, and a strong negative correlation between (rs = −0.88; p < 0.001) FKGL and FRES. Additionally, ChatGPT-5 demonstrated lower GFI and CLI scores than the other models, while its SMOG scores were lower than those of Gemini 2.5 Flash and Grok 4 (p = 0.001, and p < 0.001, respectively). Conclusions: Although differences were observed between the LLMs in terms of the accuracy and completeness of the responses, ChatGPT-5 showed the best performance. Even with high scores of accuracy (excellent) and completeness (comprehensive), it must not be forgotten that incorrect information can lead to serious outcomes in healthcare services. Therefore, the readability of responses is of critical importance, and when selecting a model, readability should be evaluated together with content quality.
Patients seeking information about Myofascial Pain Syndrome (MPS), which affects a large segment of the population, are increasingly turning to AI-based chatbots as an alternative to traditional methods. However, the medical accuracy of the content offered by these digital platforms, as well as its suitability to the “grade 6 reading level” standard, which determines its comprehensibility by patients, is a critical point of uncertainty. This study aims to fill this significant gap in the literature by systematically comparing MPS content generated by different AI models using readability indices, reliability, and quality metrics. The 18 most relevant keywords, derived from 25 keywords identified via Google Trends data, were queried using ChatGPT (GPT-5.2), Gemini 3 Flash, and Perplexity (Sonar-4 Large) models. The readability of the generated responses was analyzed using six different indices (FRES, FKGL, GFOG, CLI, ARI, SMOG), while content quality was assessed using GQS and EQIP scales, and reliability using DISCERN and JAMA scales by two independent observers. The responses generated by all AI models examined were found to be statistically significantly more complex than the suggested 6th-grade reading level (p < 0.001). In inter-model comparisons, ChatGPT exhibited the easiest readability [lowest linguistic difficulty] scores, while Perplexity scored significantly higher than both ChatGPT and Gemini in content quality and reliability metrics (JAMA, DISCERN, GQS, EQIP) (p < 0.05). Correlation analysis revealed a strong and positive relationship between quality and reliability parameters. Artificial intelligence platforms have been observed to exhibit high potential in the production of medical information. However, linguistic barriers exceeding sixth-grade reading comprehension, along with reliability limitations of current models, prevent them from replacing professional medical consultation. Perplexity has been found superior in terms of academic quality, while ChatGPT has been found superior in terms of readability. Nevertheless, positioning these systems as complementary “secondary consultation mechanisms” supporting physician oversight in clinical decision-making processes is critically important for patient safety.
The rapid integration of Large Language Models (LLMs) into software development workflows has given rise to a new class of AI-assisted coding tools, such as Claude Code, Codex, and Gemini CLIs. While promising significant productivity gains, the engineering process of building these tools, which sit at the complex intersection of traditional software engineering, AI system design, and human-computer interaction, is fraught with unique and poorly understood challenges. This paper presents the first empirical study of engineering pitfalls in AI coding tools, based on a systematic manual analysis of over 3.8K publicly reported GitHub bugs from Claude Code, Codex CLI, and Gemini CLI. Using open coding, we examine issue descriptions, user discussions, and developer responses, and annotate each bug by type, location, root cause, and symptom to characterize common failure patterns and recurring engineering challenges. Our results show that more than 67% of the bugs in these tools are related to functionality. In terms of root causes, 37.3% of the bugs stem from API, integration, or configuration errors. Consequently, the most commonly observed symptoms reported by users are API errors (18.3%), terminal problems (14%), and command failures (12.7%). These bugs predominantly affect the tool invocation (37.6%) and command execution (25%) stages of the system workflow. Collectively, our findings provide a critical roadmap for developers seeking to design the next generation of reliable and robust AI coding tools.
Aim: The aim of this study was to compare three large language models (LLMs); ChatGPT-5 (OpenAI), Gemini 2.5 Pro (Google), and DeepSeek-V3.1 (Hangzhou DeepSeek AI), in terms of their ability to provide information on questions asked about avulsion, a type of traumatic dental injury. Methods: In the study, 25 open-ended questions prepared based on the guidelines of the International Association of Dental Traumatology (IADT) were directed to large language models in their "Think / DeepThink / Reasoning" modes. The responses were evaluated by two experienced experts using a five-point Likert scale for accuracy and a three-point Likert scale for completeness (exhaustiveness). Readability was analyzed using the Flesch Reading Ease Score (FRES), Flesch–Kincaid Grade Level (FKGL), Gunning Fog Index (GFI), Simplified Measure of Gobbledygook (SMOG), and Coleman–Liau Index (CLI). Inter-rater reliability was assessed, and ANOVA/Kruskal-Wallis and post-hoc Dunn-Bonferroni tests were used for group comparisons, based on the data's normality. Results: The results showed that all three models demonstrated high evaluator agreement in terms of both accuracy and completeness. Significant differences were observed between the models in terms of accuracy scores (p=0.016). In subgroup analyses, the ChatGPT-5 model's accuracy score was significantly higher than that of the DeepSeek-V3.1 model (p=0.020). Significant differences were observed between the models in terms of completeness scores (p=0.006). In the subgroup analyses, the completeness scores of the ChatGPT-5 and Gemini 2.5 Pro models were significantly higher than those of the DeepSeek-V3.1 model (p=0.007 and p=0.046, respectively). In terms of readability, the DeepSeek-V3.1 model achieved higher FRES scores and lower FKGL, GFI, and SMOG scores. Overall, ChatGPT-5 demonstrated the highest performance in terms of accuracy and completeness, while DeepSeek-V3.1 achieved the highest readability values. Conclusion: The results show that ChatGPT-5 is the most reliable model in terms of clinical accuracy and content integrity, whereas DeepSeek-V3.1 produces more linguistically understandable responses. However, no model can completely replace human expertise. These systems should be considered tools to support clinical decision-making processes and should be used under physician supervision, especially in pediatric emergencies.