Contributing writer covering AI infrastructure and developer tooling.
Algorithms & ResearchA UW and Meta paper has models manage context as an editable file, reporting higher accuracy at lower compute on long agent tasks. The results are self-reported.
Machine LearningAi2's Olmo-core 3 is an open training stack for mixture-of-experts models, with self-reported 52,000 tokens/sec/GPU on a 47B MoE and candid failure notes.
AI NewsNvidia is reportedly in early talks with insurers and broker Howden Re on covering lenders if neoclouds default on loans backed by its GPUs.
LLMsAnthropic's Frontier Red Team reports that Zhipu's open-weight GLM-5.3 nearly matches Claude Mythos Preview in cyber exploit tests, while its safety refusals can be easily bypassed.
AI NewsThree of the biggest AI labs are working to launch an industry-funded safety standards body, tentatively called the Standards Authority for Frontier AI, and have approached former White House AI adviser Sriram Krishnan to lead it.
Tools & FrameworksThe new antigravity-preview-09-2026 harness cuts output tokens on file edits by 40% and adds two APIs for handling files and secrets in autonomous agent workflows, ahead of an October 5 deprecation of the prior version.
LLMsxAI has released Grok 4.7, a flagship model built for longer coding and engineering tasks, with a 500,000-token context window, significant benchmark gains over Grok 4.6, and unchanged pricing.
Algorithms & ResearchA swarm of 950 Claude agents screened 200,000 reverse transcriptases and identified a previously uncharacterized enzyme system, an early result from Anthropic's new life sciences research group.
Tools & FrameworksAI coding startup Cognition has raised $2 billion at a $48 billion valuation, nearly doubling its worth in four months as its Devin coding agent approaches $900 million in annualized revenue.
