Today’s lead · X
Anthropic and OpenAI line up behind embedded safety evaluators
Anthropic is committing to permanent, employee-level access for third-party evaluators, and OpenAI says it will do the same. The practical shift is governance moving from periodic external tests toward embedded oversight, though the hard parts—coordination, antitrust, China, and enforceability—remain claims and proposals.
Top signals
3 moreCursor’s Projects push AI coding from task chats to persistent agents
Cursor’s new Projects workflow replaces per-task chats with a persistent coordinator agent that remembers project context and manages subagents. If it works as described, the useful change is less re-explaining and more continuity; the “biggest UX shift” framing is still user hype, not evidence.
Qwen3.8-27B lands on Cerebras with inference fast enough to change UI assumptions
Alibaba Qwen says Qwen3.8-27B is now running on Cerebras. A follow-up post claims 1,850 tokens per second and $0.99/M input, $1.49/M output pricing. For builders, the interesting question is whether latency moves from model serving to product orchestration, rendering, and human consumption.
A Japanese neocloud is testing the non-NVIDIA AI hardware thesis in production
Ian Cutress highlights ai&, a Japan-based neocloud using Tenstorrent hardware as part of its infrastructure. The “CUDA moat is gone” line is the CEO’s claim; the concrete signal is a sovereign-AI provider putting alternative accelerators into datacenter deployment, not just benchmark slides.
Tools & repos
2 selected
Cortex
Cortex turns API specs into interactive docs, typed SDKs across 11 languages, and MCP servers for agents. Useful if your API surface is already spec-first and you want docs, client code, and agent access generated from one layer.
QApilot MCP for Android
QApilot MCP lets Claude, Cursor, or Codex drive Android tests on real devices and emulators from plain-English flows. The practical hook is replayable Gherkin output without Appium code, though it still needs Node, Java, and Android SDK setup.
Blogs worth your time
1 reads
Simon Willison finds the sharp edge of opaque agent compaction
Willison had ChatGPT Work generate 5K and 10K OSM-based running routes, complete with visualization and GPX/GeoJSON files. The useful warning: after thread compaction, the system could not provide the code it had run.
Community discussions
2 threadsMathematicians’ AI declaration triggers a familiar jobs-versus-science argument
The thread debates whether a mathematicians’ declaration about AI misalignment applies beyond mathematics. Commenters push back on job-obsolescence readings, distinguishing applied productivity from fundamental science and arguing that faster tools do not automatically replace researchers.
Claude Code users report weekly limits evaporating without obvious usage
A Claude Code Max 20x user says their weekly meter jumped from about 7% to 50% despite minimal use. Replies report similar jumps. The builder takeaway is simple: opaque usage accounting makes agent workflows hard to budget and trust.
Funding & acquisitions
1 moves
Jaipur Robotics raises Rs 47 crore for industrial AI in waste-to-energy plants
Jaipur Robotics raised about Rs 47.2 crore to expand computer-vision systems for waste-to-energy and other industrial plants. The company’s wedge is safety and automation around hazardous waste handling, with India listed among possible expansion markets.
Bengaluru radar
2 events
AI Engineering: Under the Hood
A free in-person Bengaluru meetup on AI systems architecture, RAG, tools, agents, evaluation, infrastructure, and production trade-offs.
Dev Days | Bangalore, India
A free in-person GitHub Copilot workshop in Bengaluru, focused on practical workflows, hands-on activities, the Copilot app, and CLI.