China Unwinds Meta’s Manus Deal, Then the Agent Startup Raises $500M

Here’s what matters in AI this Friday — and why you should care.

AI Agents

China Blocks Meta’s Buyout — Then Manus Raises $500M to Go It Alone

What happened: AI agent startup Manus has raised over $500 million in a round led by Boyu Capital and IDG Capital, with Tencent, HSG and ZhenFund participating. It’s the company’s first raise since its founders and investors bought the firm back from Meta — a $2 billion acquisition announced in December that Beijing’s economic planning body ordered unwound in April. Media reports peg the new valuation near $4 billion, roughly double the Meta deal.

Why it matters: This is the agent wars colliding head-on with geopolitics. Manus is a “do real-world tasks for you” agent — the same category Meta’s new Muse is chasing — and it just got forcibly repatriated into an independent, heavily-funded Chinese champion. The company says it’s now forming teams to “develop products for the domestic market.” For anyone building or buying agents, the signal is clear: the personal-agent race is now a two-front contest, with state-level pressure deciding who gets to own what. The talent, capital and product momentum aren’t flowing into one global winner — they’re splitting along national lines.

What’s next: Watch whether Manus’s standalone app Cue (its answer to Muse) gains traction outside China, and whether other cross-border AI deals get caught in the same regulatory net. If you’re evaluating agent platforms, geography just became a real variable in your vendor risk assessment.

Source: Yahoo Finance / AFP

Meta’s Muse Expands to iPad — and Pushes an Open Standard for Agent Identification

Muse launched a dedicated iPad app a month after its iPhone debut, with Sensor Tower estimating 6.6 million installs since September 8, plus new connectors for Asana, Canva, Figma, Notion and Granola. More quietly, Meta says it’s working with industry partners on an open standard so personal AI agents can identify themselves to businesses — a move that matters because it starts building the trust layer agents will need to transact on your behalf.

Source: The AI Insider

Arena Doubles to a $3.1B Valuation — Because Evaluating AI Is Now a Business

Arena, the crowdsourced model-ranking platform that began as a UC Berkeley research project, raised a $200 million Series B at a $3.1 billion valuation — nearly double its January number — on $100 million in annualized revenue. It’s adding an “alignment” leaderboard that scores models on unauthorized actions, false attribution and “deceptive completion.” The takeaway: as agents get more autonomous, independent evaluation is becoming infrastructure in its own right, not an academic side project.

Source: TechCrunch

AI News

OpenAI’s 722-Paper Math Drop Has Mathematicians Calling It a ‘Mathocalypse’

OpenAI published hundreds of AI-generated mathematical results touching the Riemann hypothesis and other Millennium Prize problems — then retracted three papers over an elementary error and amended others. Mathematicians say many are so unintelligible an editor would “bin” them, and even OpenAI’s own ChatGPT called one result “a serious hallucination.” The lesson for AI builders: raw output volume is not the same as verified truth, and peer review still can’t be automated away.

Source: The Conversation

Anthropic’s ‘Claude-Led’ Gene-Editing Discovery Is Drawing Skepticism

Anthropic says 950 AI agents found a CRISPR-like enzyme system in 21 hours of autonomous search — but scientists note it isn’t peer-reviewed, doesn’t yet demonstrate actual gene editing, and one researcher claims it mirrors his own unpublished work. It’s a useful reality check on “autonomous discovery” claims: the gap between an interesting pattern and useful knowledge is where most of the real work lives.

Source: CNN

Meta Turns LLMs on Child Exploitation — and Found 97% Before Users Did

Meta deployed a new LLM that tracks where ads “signpost” to illegal content, not just what the ads say, alongside a red-teaming agent that probes its own safeguards. The company reports its systems found more than 97% of the 33.2 million pieces of child-exploitation content it removed in the first half of 2026 before users flagged them — a reminder that AI safety work is quietly running at industrial scale behind the headlines.

Source: The AI Insider