bug

"Fuzzing finds hidden bugs in trusted systems, improving security, reliability, and resilience against real-world threats."

"These are solved by LLM Observability tools, which have specific observational needs: Traces: Record all inputs/outputs to the LLM. Spans: Visualize RAG search -> generation, and each step of the Agent. Evaluation: Scoring response quality, Hallucination, and relevance."

"Opus 4.7 uses more thinking tokens than previous models, and we discovered a bug where it did not properly apply rate limits for long-context requests."

"SWE-bench (real GitHub bugs): AI went from 60% to ~100% in ONE year."

"So, cyber security of tomorrow will not be like proof of work in the sense of 'more GPU wins'; instead, better models, and faster access to such models, will win."

"The biggest highlight of this session is that the AI made a 'subtraction proposal'... AI is not just a code generator, it functions as a consultation partner (sounding board) for technology selection."

"200-line stack traces become a single line: 'Promise reject is unhandled. await is missing on line 42.'"

"Based on the entered information, GitHub Actions starts, performs the development, and creates a PR."

"大規模Webサービスのリプレイスプロジェクトで、結合テストにて検出された不具合チケット約900件を、4つの根本要因に分類する必要がありました。手作業で1件ずつ読んで分類すると、1件あたり1〜2分として少なくとも丸2日。しかも人によって判断がブレます。そこでLLMを活用して自動分類に挑戦し、未分類ゼロ（100%分類） を達成しました。"

"Improvements were seen, but it did not lead to a fundamental solution. Even after switching to the Gemini 3.1 Pro model, the date discrepancies were not resolved, revealing that it was not an issue of model performance."

"3時間ググって10個の記事を比較した経験は、「どの情報が信頼できるか」を嗅ぎ分ける嗅覚を育てます。"

"I’m a PhD student in Machine Learning and Computer Vision, and I often help students and developers with Python and ML-related problems."

"I continued debugging with Claude Code + Codex, swung and missed three times, and finally reached the true cause on the fourth attempt. The true cause was an event loop self-deadlock in Spring Framework ReactorClientHttpRequest."

"A scheduled prompt fires between your turns, not while Claude is mid-response."

"Anthropic and Mozilla are collaborating, with the model discovering over 100 Firefox bugs in just two weeks, 14 of which were high severity."

"Anthropic and Mozilla collaborate, with Claude Opus 4.6 discovering over 100 bugs in Firefox in just two weeks, 14 of which were of high severity."

"If you press the 🐣 on each line, Gemini AI will explain the cause and solution in Japanese. I made this because I wanted a lightweight Mac app that you can use when you just want to look at logcat without starting Android Studio."

"It navigated the whole app autonomously through the accessibility tree and screenshots (no hardcoded coordinates), found actual bugs I missed, checked the debug logs for errors, and gave me a structured summary at the end."

"Drop GPT 5.2 Codex or Kimi 2.5 into a good enough agentic loop with full source code access, and they'll flag 20 critical bugs while you're getting coffee."

"Kanbanボードがタスクを直接AIコーディングエージェントに割り当て、そのエージェントがバグを修正してPRを作成し、結果を報告する——人間の介入なしに。"

"When multiple errors occur, you need to determine 'which is the root cause' and 'which is the secondary error.' Tracing back through the program's execution flow is highly effective for this."

"If the loss (error) repeatedly goes up and down without converging, the learning rate is often too large."

"v2.1.96 is a single bug-fix release that resolves the Bedrock authentication error (403 'Authorization header is missing') introduced in v2.1.94."

"I've found more bugs in the last few weeks with Mythos than in the rest of my entire life combined."

"In Cowork, browser operations are performed via a WebSocket bridge, utilizing a different permission flow than the side panel: Claude Desktop → WebSocket bridge (bridge.claudeusercontent.com) → Chrome Extension (Service Worker) → Browser operation."

"With LLMs, when the model hallucinates, we basically get... logs. But there's no equivalent of a stack trace that tells us WHERE in the pipeline things went wrong."

"The project is named after the 'Glasswing butterfly,' symbolizing the goal of finding vulnerabilities that are transparent and hidden within complex systems, just as the butterfly's wings are clear."

"Our evaluations show that Claude Sonnet + Rubber Duck makes up 74.7% of the performance gap between Sonnet and Opus alone, achieving better results for tackling difficult multi-file and long-running tasks."

"The cause is a combination of three elements. The tokenizer cannot compress repetitions of '拠'. While '人人' becomes one token, '拠拠' is not in the vocabulary, causing identical tokens to appear endlessly. This identical token repetition triggers a self-reinforcing loop."