join.dev
AI news · 10 stories
Google previews Gemini 4 Carbon for cybersecurity and long tasks · Microsoft previews its Copilot harness in Foundry Agent Service · NVIDIA releases Nemotron 3.5 Transcribe and Granary v2
Google announced Gemini 4 Carbon on October 28, reporting improvements in coding, cybersecurity and sustained agent workflows. Access begins with selected cyber defenders and trusted partners while Google gathers safety feedback before broader availability.
Read at Google
Microsoft introduced its Copilot harness in Foundry Agent Service in public preview on October 28. The managed runtime combines model choice, tools and execution controls so developers can build agents for complex, long-running work without assembling every layer themselves.
Read at Microsoft
NVIDIA released an open-weight speech model combining transcription in 40 languages with translation, speaker attribution and audio question answering. The accompanying Granary v2 data pipeline and adaptation recipes give developers tools to customize speech workflows.
Read at Hugging Face
Anthropic launched a gallery of legal artifacts on October 28 for tasks including contract review, citation checks and regulatory tracking. Teams can adapt these interactive tools to their documents and connectors, with source links supporting lawyers’ review of the results.
Read at Anthropic
Anthropic described collaborations with six medicine-development organizations and its own drug-discovery programs on October 27. Most of the work remains preclinical, offering an early look at how Claude could support research rather than evidence of approved treatments.
Read at Anthropic
Fragment launched on October 28 with $56 million in total funding, including a Series A led by 8VC. Its AI platform investigates purchasing and payment exceptions across existing systems, targeting the manual reconciliation work that slows physical-goods businesses.
Read at GlobeNewswire
Alphabet’s October 28 earnings remarks reported 25% year-over-year revenue growth and 103% growth at Google Cloud, with AI infrastructure and solutions cited as major cloud drivers. The company said it is adding capacity to meet demand, highlighting the commercial scale of AI spending.
Read at Google
Ai2 released a benchmark, data and code on October 27 to test whether language models can judge research novelty. Across six models, small evaluation changes could sharply alter verdicts, underscoring the need to test judging methods before relying on them to rank AI-generated ideas.
Read at Hugging Face
Google Research’s October 27 update describes a sleep framework that alternates active learning with offline memory consolidation and dreaming. The research aims to turn temporary context into lasting model knowledge while reducing the loss of previously learned capabilities.
Read at Google Research
GitHub’s October 28 Octoverse report records 17.9 billion code pushes over the past year and more than 225 million developers on the platform. It also measures AI involvement across programming languages, providing evidence of how agents and automation are changing software development.
Read at GitHub