AI & ML
- Harvard Medical School developed COMPASS, an AI model that improves prediction of cancer immunotherapy response rates by 8.5% over existing approaches. The model analyzes tumor gene activity and explains its predictions, potentially enabling better patient selection for trials and new drug discovery. Results published in _Nature Medicine_ July 3. Read more
- Anthropic launched Claude Sonnet 5 as an accessible, broadly available model for everyday work and agentic tasks, signaling a shift toward bifurcated model portfolios where highly capable or sensitive models remain restricted while powerful agentic models become the enterprise layer. Read more
- Meta superintelligence chief Alexandr Wang announced the company's upcoming Watermelon model matches OpenAI's GPT-5.5 on key benchmarks and uses significantly more compute than the current Muse Spark. An update to Muse Spark is coming soon with major coding and agentic capability improvements. Read more
- Chinese firms released two new frontier models—Zhipu AI's GLM 5.2 (outperforms Anthropic Opus and OpenAI GPT-5.5 on vulnerability-finding benchmarks at $0.17 per vulnerability) and 360 Security's Tulongfeng—raising concerns that AI-powered exploit creation is accelerating the attack-defense gap. Read more
Browser & Web Platform
- Chrome 151 beta is now available with new features and improvements. The release includes better Async Clipboard API defaults through selective format read, allowing developers to control which data formats users can paste. Read more
Security
- Bad Epoll (CVE-2026-46242) is a newly disclosed Linux kernel flaw allowing unprivileged users to gain root access on Linux desktops, servers, and Android devices. Researcher Jaeyoung Chung found and exploited the race condition; it sits in kernel code where Anthropic's Mythos model missed it despite catching other bugs. A public proof-of-concept exists. Read more
Open Source
- headroomlabs-ai/headroom is a token compression tool that reduces LLM input by 60–95% for tool outputs, logs, files, and RAG chunks. Available as a library, proxy, or MCP server to lower inference costs. Read more
- getpaseo/paseo orchestrates multiple coding agents in parallel on your own machines, supporting self-hosted execution with full dev environments from desktop or mobile. Read more
Dev Tools & IDEs
- PhpStorm 2026.2 Beta was released with performance improvements targeting high CPU activity in control flow analysis and type inference on large files with complex conditions. Read more