AI code generation

Microsoft has released MXC, a sandboxed code execution system, on GitHub. The tool lets developers run untrusted code in isolated environments, a growing need as AI agents and automated pipelines increasingly execute generated code. Developers are weighing in on its design and how it compares with existing sandboxing options.

Wired has published a hands-on piece in which its writer used Google's AI Playground to build a series of games, all of which turned out badly. The article reflects wider curiosity about whether current generative AI tools can handle creative coding tasks, and the verdict here is that results remain rough and unreliable despite rapid progress.

JetBrains has published a developer diary describing how it built a retrieval-augmented generation pipeline for semantic code search, sharing field notes and practical lessons from the process. Developers are discussing the technical trade-offs of applying RAG to codebases, a topic of growing interest as AI-assisted coding tools become mainstream in software engineering workflows.

New research reported by Ars Technica finds that AI coding agents do generate more code, but software output overall does not increase. The study concludes that productivity gains are absorbed by a human review bottleneck, as developers must still verify, test and integrate machine-written code. Readers are debating whether review capacity, not generation, is now the limiting factor in software development.

Linux creator Linus Torvalds says he "really likes using AI", a notable stance as the Linux kernel community debates how to handle LLM-generated code and content in contributions. The comments come amid discussions over disclosure rules for AI-assisted patches, with maintainers weighing quality and accountability concerns against growing use of AI coding tools. Torvalds' pragmatic view contrasts with stricter proposals from parts of the developer community.

Commenters are arguing that research on AI productivity in software development has focused on the wrong question. The claim is that for AI-assisted coding to deliver real gains, another language model should review most of the code an AI generates, with humans acting mainly as judges and moderators rather than line-by-line reviewers.

A developer writing on Dev.to argues that software patches should not merge just because continuous integration jobs pass. Instead, merging should require richer evidence: every changed symbol carrying property credit, fixture hashes matching the lock, and flaky-test freezes contributing nothing. The piece contends CI compresses multiple distinct quality signals into a single pass-fail bit, letting flawed code through.

A new study reported by Ars Technica finds that AI coding agents dramatically increase the amount of code generated, yet overall software output does not improve. The efficiency gains are reportedly absorbed by a human review bottleneck, as developers must still spend time evaluating, testing and integrating the machine-written code. The findings challenge assumptions that AI coding tools automatically translate into faster software delivery and are prompting debate about where the real limits of AI productivity lie.

A new Stack Overflow developer survey shows widespread AI use among programmers, but most remain skeptical of the technology's accuracy and judgment. 79% of respondents say source attribution matters when AI tools answer coding questions — a notable point for Stack Overflow, which licenses its Q&A archive to AI model builders. The findings highlight a gap between adoption and trust.

AI coding assistants are complicating software bill of materials (SBOM) practices, as generated code and dependencies often fall outside documented component inventories. The concern lands just as the EU Cyber Resilience Act approaches, a law that will require manufacturers to maintain accurate SBOMs and manage security risks across products. Security teams are being urged to rethink how they track AI-generated dependencies before compliance deadlines arrive.