Claude Code's Non-English Output Quality Degrades with Opus 4.7 — Abnormally High Frequency of Korean Slang Revealed

Claude Code user yurukusa reported that the quality of non-English output degrades when migrating from Opus 4.6 to Opus 4.7. The core symptom: the Korean slang term "박다" (bakda, meaning "to pound in") appears frequently in place of the verbs that should be used in formal contexts. The phenomenon reportedly occurs at 18 times the normal frequency.

This issue has a direct impact on non-English-speaking users, and similar quality degradation may be occurring in other languages such as Japanese and Chinese. A Qiita article breaks the problem down along 4 axes and presents 3 workarounds users can take.

Claude Code Extended Thinking Freezes Sessions Permanently — Over 180 Reports Pile Up

The same user, yurukusa, also reported that when continuing to work with extended thinking enabled on Opus 4.7 or Opus 4.8, the API keeps returning 400 errors and the session becomes completely unusable.

The error message begins with `API Error: 400 messages.N.content.M: 'thinking' or 'redacted thinking' blocks`, and once it occurs, recovering the session is difficult. More than 180 cases were confirmed in May alone, and the article introduces 7 analysis points along with 4 defensive hook configurations.

This is a serious problem for developers working in long sessions, and workarounds such as periodically recreating sessions or enabling extended thinking incrementally are worth considering.

How Software Architecture Changes After AI

Brian Guthrie published an article titled "Software Architecture After AI," examining the impact of AI on the design philosophy of software architecture. Discussion has also started on Hacker News, raising the question of how the architect's role changes in an era when AI agents write code.

Beyond the traditional microservices-vs-monolith debate, the piece is expected to offer practical analysis of the maintainability of AI-generated code, shifts in testing strategy, and which layers will still require human review.

AWS Publishes an LLM Inference Observability Solution for SageMaker AI

The AWS Machine Learning Blog introduced a comprehensive observability solution for LLM inference running on Amazon SageMaker AI. Leveraging Amazon Managed Grafana dashboards, it visualizes both infrastructure metrics (quantity), such as GPU utilization and latency, and model metrics (quality), such as answer quality and safety scores.

For quality evaluation, it uses Claude Sonnet 4.6 (via Amazon Bedrock) as an LLM-as-judge, tracking four metrics: composite quality score, safety score, relevance score, and professional tone score. Threshold-based alerts can also be configured with Grafana Alerting, notifying SRE teams via SNS.

For teams running LLMs in production, the ability to monitor infrastructure health and model quality in a correlated manner offers practical value.

GPU Comparison for Local LLMs Gains Traction — Is the Mac Studio Overrated?

A comparison of major GPU specs is drawing attention in Reddit's LocalLLaMA community. The poster evaluates GPUs comprehensively, considering not only bandwidth but also the balance between prefill and generation performance, power efficiency, and cost-effectiveness.

A notable hot take: the Mac Studio is said to offer only about as much efficiency as an "overrated Raspberry Pi." Meanwhile, the P100 — at $200 (dual GPU) with 32GB of 700GB/s memory and 70% of the M3 Ultra's compute — is rated as extremely cost-effective as an entry-level LLM GPU.

Additionally, the V100 is considered more valuable than the 3090 for single-stream workloads, while the P40, though niche, is said to be well suited for running MoE models that need 48GB of VRAM. It was also noted that the spread of multimodal models is making prefill performance increasingly important.

---