2026-10-13
By Vadym · Generated with AI, curated by me
Meta finally shipped a real frontier model while the money elsewhere went to policing and ranking AI rather than building more of it — and one lab's biology headline got ahead of its own evidence.
Meta released Muse Wonder and Muse Wonder Full on October 12, its first frontier-class models, built on Muse Spark and aimed at agentic work — insurance claims, lease review, coding. The headline features are Active Memory, which carries context across days and weeks, and Active Focus, which keeps the model working through a hard task, tries alternate approaches, and checks in before big actions. Wonder is live now via the Meta Model API and Muse Code; Wonder Full is Muse Code subscribers only.
Arena, the AI model leaderboard formerly known as LMArena, closed a $200 million Series B led by Lightspeed Venture Partners and Khosla Ventures, pushing its valuation to $3.1 billion — up from $1.7 billion ten months ago. The company says it has crossed $100 million in annualized revenue.
Multiply Labs closed a $75 million Series B led by Patrick Soon-Shiong’s NantWorks, with AstraZeneca, Lingotto, and Teradyne joining, to expand its robotic biomanufacturing clusters — automated systems that handle the repetitive steps of making cell and gene therapies, with antibodies and mRNA next. The company claims a 74% cost-per-dose reduction and up to 100x more throughput than manual production; total funding now exceeds $100 million.
Rein Security closed a $25 million Series A led by Glilot Capital and Sienna Venture Capital, bringing total funding past $30 million. Its sidecar technology watches AI agents at runtime and can intervene when an agent’s actions look harmful, aimed at enterprises deploying agents with no visibility into what those agents actually do.
Anthropic announced that Claude agents, working with only high-level direction from scientists, identified an unusual CRISPR-like enzyme system in viral DNA whose function is still unknown. Outside researchers, including Aaron Engelhart at the University of Minnesota, called the finding preliminary — it doesn’t demonstrate actual gene editing or explain the mechanism, and the work hasn’t been peer-reviewed. Gene-editing company stocks dipped briefly after the announcement.
Microsoft president Brad Smith argued the UK is better positioned than Washington or Brussels to set global AI safety standards, noting Britain’s AI Security Institute already gets pre-release access to major models. He called AI both “a weapon and a tool” for cybersecurity risk and pushed for mandatory breach disclosure, outside approval before release, and kill switches — saying the White House’s voluntary safety approach “falls short” because it only binds companies already trying to be responsible.
One lab shipped an actual model today. Everyone else's money went toward policing and ranking AI rather than building bigger versions of it — runtime security for agents, a $3.1B scoreboard, robots that make the existing drug supply cheaper. And the industry's favorite trick, leading with a discovery before the evidence catches up, showed up in a biology lab this week instead of a pitch deck. The work that holds up is the unglamorous kind.
— Boba
Curated by Vadym