News ยท
Multipass AI News Roundup
- #News
- #Meta
- #Anthropic
- #OpenAI
Hey Multipassers ๐
Settle in with something warm, because this week a giant model asked for a terabyte of memory, Wall Street caught a cold, and Washington tried to rename an entire industry again. Here's the rundown.
This week in one breath: Meta locked its doors and still climbed the leaderboard, Anthropic eyed a trillion-dollar debut, AI stocks wobbled while the IMF fidgeted, DeepMind quietly out-mathed the mathematicians, and OpenAI collected lawsuits like souvenirs.
Let's get into it!!
๐ Meta Flips the Script (Twice)
Meta unveiled Muse Wonder and Muse Wonder Full, and these ones are firmly closed-source. The bigger of the two landed in second place on the intelligence rankings, edging past Gemini and GPT-6 Astra. Then, just days later, Meta turned around and published Muse Spark Open, an 816-billion-parameter open-weight model designed to live on Apple hardware. The fine print: you will need a Mac Studio with a full terabyte of memory.
Meta is now playing both sides of the open-versus-closed fight, and winning on the scoreboard either way. Everyone with a normal laptop can enjoy the press release from a safe distance.
๐ฐ Anthropic Eyes a Trillion-Dollar Entrance
Anthropic reported third-quarter revenue of $16.8 billion, up 45%, and is reportedly lining up a November IPO at a valuation that could approach $2 trillion. Separately, co-founder Jack Clark moved his forecast for AI that can improve itself up to 2027, a year sooner than his previous guess. A new test suggests Opus 5.5 can plan AI research faster and cheaper than human experts.
The "labs only torch cash" storyline is looking a bit stale. Also, a model doing research on how to build better models is the kind of loop that deserves a second look.
๐ Wall Street Gets the Jitters
Hotter-than-expected inflation numbers sparked roughly $80 billion in losses across AI-linked stocks, with Dell and Oracle among the casualties, and chip names like Samsung and SK Hynix took a beating. The cost of insuring against AI-related defaults hit record levels. The IMF warned that a pullback in AI spending could hurt the global economy, even as it nudged its 2026 growth forecast up to 3.1% largely because of the data-center building spree.
Borrowed money is paying for the AI boom, and borrowed money got more expensive. The party is still on, but somebody just turned the lights up a notch.
๐ง Google Has a Big Week
Gemini 4 Argon rolled out to Pro and Ultra subscribers and developers. Google also signed paid news-licensing deals with CNN, Axios, and the Financial Times for its generative AI features, and launched Universal Cart in the US, which collects items from any store, watches prices, pings you about restocks, and hunts down discount codes. Over on Android, Google Play now lets people build and run private, vibe-coded apps made with AI Studio that never touch the public store.
Paying publishers is starting to look like a real way out of the copyright brawl, and Google just appointed itself your personal shopper. Your credit card has been notified.
๐งฎ AI Does Real Science Now
DeepMind's AlphaProof Nexus agent cracked nine open Erdลs problems, two of which had defeated humans for 56 years. It pairs Gemini's knack for dreaming up proofs with a strict automated checker that audits every step. Meanwhile, CERN and Mistral AI are building "Flare," a science model trained on particle collision data instead of text.
This is AI producing research-grade results without a person holding its hand. Mathematicians, we hope you have other hobbies.
๐งฑ The Chip Wars Get Personal
Microsoft plans to deploy more than a million of its own Maia 300 chips next year for Copilot and OpenAI workloads, promising about 40% more compute per megawatt than leading GPUs. In China, Nvidia's slice of the AI chip market fell to 34% while Huawei climbed to 28%, as export limits pushed Chinese buyers toward homegrown silicon.
Big Tech wants out of the Nvidia tab, and the export rules are doing wonders for Huawei's sales team.
โ๏ธ OpenAI's Legal Week
Eight US states sued OpenAI, alleging ChatGPT nudges teenagers toward self-harm and violence. In a separate fight, OpenAI accused a former employee of walking off with confidential files for Apple, and hit back at Apple's own suit. On a lighter note, OpenAI said it used its own models to help track down an engineer charged with planting a listening device in a San Francisco conference room. And Sora might get a sequel: its former lead is reportedly founding a filmmaking startup, and OpenAI is in talks to take a stake in return for reviving the product.
OpenAI wants a gigantic valuation while its legal calendar fills up. The Sora comeback is the one plot twist nobody asked for but everybody will watch.
๐ก๏ธ The Safety and Watchdog Beat
Leading labs, OpenAI and Anthropic among them, want more say over what outside safety evaluators are allowed to publish, citing trade secrets. Testers say they need full independence. Britain's AI Security Institute is getting to test new frontier models such as Gemini 4 Argon, Nvidia is donating $250 million of computing power to UK safety research, and the head of MI5 warned that AI is turbocharging cyberattacks and propaganda, with bioweapons topping his list of nightmares. A group of AI's most senior researchers signed a letter urging labs to disclose containment failures, G7 finance officials are rehearsing responses to rogue AI agents attacking the financial system, and regulators in the US and Europe are quizzing banks about autonomous agents.
On the research side, Anthropic showed that small models are edging toward copying themselves onto cheap connected gadgets like smart TVs, and a group of researchers got bots scraping a map website to reply to a hidden message, with the bots claiming to be Claude.
The people building the tests want to edit the grade report, the spies are worried, and the bots are roaming free. Also, lawyers warn that if you don't save your prompts and sources, you may have no paper trail when someone sues.
๐ท๏ธ Washington Wants a New Name for AI
The US Treasury is working with the IMF to make "Super Intelligence" the official global term, following a directive from the President. Apple, meanwhile, says Apple Intelligence reaches mainland China in December with iOS 27.3, using Alibaba-built filters to satisfy local rules, and without Gemini powering Siri.
Renaming something doesn't change what it does, but it does change the slide decks. Apple needs the China launch to keep iPhone sales humming.
๐ Character AI Hits the Wall
Character AI, once valued at $2.5 billion, filed for bankruptcy and put itself up for sale. Disney offered $20 million to license some of its technology, a bit awkward given the two were recently in court against each other.
Safety scandals and lawsuits can sink even a shiny unicorn. A very expensive lesson.
๐ค Agents Move Into Everything
SpaceXSI, the company formerly known as X, bundled X Premium, Grok, and the Cursor coding app into a single subscription. Amazon introduced Bee wearables that listen to your day and take notes, with its devices chief predicting ambient computing will replace apps. Rivian will let outside AI agents from Anthropic and Meta run in its cars, and Bluesky's Attie app now lets anyone vibe-code websites, games, and polls. American Express built a checkpoint that compares an agent's purchase to your exact instructions before it spends a cent, and Eden Labs raised $1.5 million for an assistant that books travel, orders groceries, and pays bills for you. Epic Games' CEO reportedly told staff to stop hand-writing code in favor of maximum AI acceleration.
The chatbot era is giving way to agents with real credit cards. "My AI ordered four hundred air fryers" is now a plausible sentence.
๐ Startup and Funding Roundup
Ramp raised $1.85 billion at a $60 billion valuation, and launched a card meant for AI agents to use. Pluto Security pulled in $42 million to lock down AI apps that finance and marketing teams have cheerfully built with no security review. Ambi raised $26 million for an app that listens in the background all day, and WindBorne landed a $50 million US Air Force deal for data from its self-navigating weather balloons. Blueprints AI turns rough sketches into permit-ready architectural drawings, Empacer built an assistant for salons and repair shops that prepares the final action and waits for the owner's thumbs-up, and Coinbase fine-tuned a small Qwen model on its own data to catch fraud, beating Claude Opus on accuracy and speed. Benchmark admitted AI is breaking the old VC habit of never backing rival startups, and a 19-year-old's calorie-tracking app, Cal AI, became a $30 million-plus hit via a small army of paid influencers before MyFitnessPal bought it.
Small, focused models and well-aimed marketing are quietly beating brute force. Ambi's all-day listening is useful, sure, and also a consent headache waiting to happen.
๐ฌ Fun Stuff to Close Out the Week
A woman in Russia was fined after posting an AI-made picture of a bear roaming her town. The photo went viral and police and wildlife rangers actually went looking for the nonexistent animal.
Flox Intelligence uses AI cameras to detect wildlife near train tracks and runways, then plays tailored sounds to shoo specific animals away. Facebook, meanwhile, now lets you describe in everyday language what you want in your feed.
Andrej Karpathy says only about 5,000 people at the top labs have used truly frontier AI, while everyone else is working with the watered-down version.
๐งพ The Big Picture
- Money: AI is holding up a lot of the economy, and rising interest rates just reminded everyone it's all financed.
- Models: closed and open are both getting stronger, and the chip supply chain is fracturing.
- Safety: labs, spies, and regulators are circling the same question of who gets to judge whether AI is behaving.
- Agents: they're moving into your car, your wrist, and your wallet.
That's the week, wrapped up and tied with a bow. Thanks for reading all the way down. Real people typed this, then fixed the typos.
P.S. If your terabyte Mac ever offers to buy you lunch with its own AI credit card, double-check what it ordered before you say yes.
See you next Saturday (provided nobody has renamed the calendar),
Team Multipass AI
๐ค Why ask one AI when you can ask five??
Multipass AI sends your question to ChatGPT, Claude, Gemini, Grok, and Qwen at once, then shows you where they agree and where they don't. Try it free: 5 questions, no account needed.
Meta's closed model just edged past Gemini and GPT. We call everyone, bestie.
๐ www.multipassai.com