The build, deploy and infrastructure-as-code tools teams live in.
Claude Fable 5.1 achieved the highest score on Real-SWE, a new benchmark testing AI agents against proprietary company code, yet still failed over 60% of the time. The results suggest significant gaps between public benchmark performance and real-world coding challenges.
Perplexity's Portable Computer now runs on Windows PCs with NVIDIA RTX GPUs, bundling model execution, tool integration and sandboxing into a single application. The trade-off: users need at least 24GB of VRAM.
Response caching can dramatically reduce LLM token spend by avoiding duplicate model calls. The challenge lies in determining when a cached answer remains valid and safe to reuse.
A year after her P99 CONF keynote on cutting AI inference costs, Chip Huyen's guidance on model and service optimization techniques remains largely sound—though agent-driven multi-step loops and prompt caching adoption have shifted how practitioners should think about the economics.
Cohere unveiled North Small Translate, a mixture-of-experts translation model that surpasses DeepL and Google Translate on WMT26 benchmarks. The model supports 50 languages and addresses enterprise concerns about data sovereignty and document handling.
OpenAI has hired the founders of Git AI, an open-source tool that measures code production and performance of AI coding agents. The move brings measurement capabilities directly into OpenAI's Codex team.
Debugging—the practice of locating and correcting software defects—remains essential across all development environments, from simple applications to complex cloud systems. Understanding debugging techniques and tools helps teams deliver more reliable code faster.
Distributed tracing enables developers to observe how data requests flow through complex systems of interconnected services. This observability technique has become essential for troubleshooting performance issues and understanding interactions across microservices architectures.
