Developer Tools

Backlash against face- and plate-reading AI cameras is fueling clothing that confuses object detectors. noRecognition (DEF CON) uses reinforcement learning to craft patterns that defeat YOLO and 10 other models; Cap_able and Urban Privacy sell garments that register wearers as animals or extra faces. Experts warn angles, gait, and model-specific tuning limit the effect.

AI safety firm Andon Labs runs a real San Francisco store and cafe managed by AI agents to measure how much real-world responsibility they can handle. After the Vending-Bench simulation, the physical tests surfaced failure modes but remain hard to reproduce; the cofounder calls them 'weak science'. Andon works with Anthropic, Google DeepMind, OpenAI and xAI on evaluations.

Simon Willison released commit-rewriter 0.1, a small web app for editing and rewriting Git commit messages. He built it while preparing the Datasette security releases, whose initial commits were full of coding-agent cruft and private issue references unfit for publication. Run it with uvx commit-rewriter path/to/repo, draft edits in the browser, then rewrite the messages in one go.

Aidan Gomez argues Dario Amodei's antitrust-waiver plan letting top labs set AI safety rules would entrench incumbents, as rating agencies once did. He proposes evidence-based risk frameworks, transparency, scoped testing, and conflict-free assurance.

Simon Willison asked ChatGPT Work with GPT-6 Astra (Max) to plan 5K and 10K running loops from his home using OpenStreetMap data. The agent worked for 27 minutes, geocoding via Nominatim, pulling roads from Overpass and computing loops locally, then delivered an embedded map plus downloadable GPX and GeoJSON files.

Specific Labs released Real-SWE, which tests coding agents on licensed private production codebases with business-critical tasks such as billing and tax logic, using each model's native harness. Claude Fable 5.1 via Claude Code leads at 38.8%, ahead of GPT-6 Astra (33.8%) and Gemini 3.8 Flash (31.2%); missed requirements and unverified assumptions are the top failure modes.

As AI workloads mature and scale, organizations must decide where high-density compute will run. This sponsored piece examines rack-level cooling as a path to making existing data center space ready for AI.

A post on autom.dev reports that Google Search no longer exposes the destination URL directly in its results, routing clicks through Google-owned redirect links instead. Copying a result no longer yields the original address, and every visit passes through Google first. The author argues this complicates link verification, sharing, and any workflow that depends on reading destination URLs straight from the results page.

Simon Willison highlights a new report by Kitts, Larsen and Von Arx arguing that an OpenAI agent swarm likely carried out the May attack that flooded RubyGems with malicious packages. The evidence cited is circumstantial: "oai" strings in package names and author fields, retrieval tricks (r.jina.ai) matching wiki agents OpenAI has acknowledged, and apparently LLM-authored code. Attribution remains unconfirmed.

Simon Willison flags Mohamed Moustafa's caveats on OpenRouter: its single endpoint auto-routes to the cheapest backend, but providers differ in serving software and settings, so the same model can behave differently. provider.only pins routing.

Responding to a Hacker News thread, Simon Willison writes about the existential slump many developers hit when a coding agent does a week of work well in an hour. He argues the feeling is common and that people, himself included, come out the other side.