AI News, Oct 13: Meta Ships Its First Frontier Model
Meta finally shipped a frontier model, then spent the afternoon being compared unfavorably to its own back catalogue. Elsewhere, the company Google paid $2.7 billion to hollow out filed for bankruptcy.
The Big Story: Meta’s first frontier model arrives in fourth place
Meta Superintelligence Lab released Muse Wonder and Muse Wonder Full on Monday afternoon, its first frontier release since Alexandr Wang restructured the company’s AI effort. Wonder costs $4 per million input tokens and $20 per million output. Wonder Full is not served through the API at all and goes only to Muse Code subscribers. Both were trained at Meta’s Prometheus data center in central Ohio.
The pitch is memory: Active Memory carries context about you between conversations, and Active Focus keeps an agent working through a multi-step task, checking in before it does anything consequential. Meta’s own announcement claims state-of-the-art results and publishes no benchmark numbers at all, so every figure in circulation is somebody else’s. Artificial Analysis scores Muse Wonder Full at 56 on its Intelligence Index, rank four, and lists it as proprietary with no publicly available API. The Decoder’s reading of the same data has Wonder Full taking 57% on Terminal-Bench 4.0 against 33% for Muse Spark 1.3, hallucinating 23% of the time where its predecessor managed 33%, and scoring 30 on AA-Omniscience against 46 for Claude Opus 5.5. The gains are in coding and agent work, not in knowing things.
Then the awkward part. On Harvey’s Legal Agent Benchmark, scored independently by Vals AI, Wonder Full passes every rubric criterion on 21.67% of tasks, good for fourth behind Muse Spark 1.2 at 25.42%, Spark 1.3 Max at 23.75% and Spark 1.3 at 22.92%. Meta’s brand-new frontier model is beaten there by three models Meta already shipped. Launching on memory rather than capability is a bet that buyers will pay for continuity, which is a different argument than the one benchmarks were built to settle.
Everyone wants independent evaluators; the labs want approval over what they publish
Reuters reported Tuesday that Anthropic and OpenAI have been seeking significant control over what embedded outside evaluators can publish, on trade-secret grounds. Two people familiar with the talks said both companies wanted strict editorial control as recently as October 5, and that discussions had shifted by October 8. Anthropic says prior evaluators control their own methods, criteria, conclusions and publications, and that it reserves comment and approval rights only for new firms still learning its models. OpenAI says it is “not seeking approval rights over the full text of reports.” Marc Warner, whose firm Faculty was named last month as Anthropic’s first embedded evaluator and whose reporting terms remain unsettled, stated the problem plainly: “Evaluators are only useful if people can trust what they report, so independence and transparency are critical.”
Three other would-be checkers surfaced in the same 24 hours. Musk’s SpaceX agreed to testify to the New York City Council on Friday after skipping last week’s hearing, where Anthropic, OpenAI, Meta and Google all appeared; the council is weighing outside validation, kill switches and a private right of action. Three top US bank regulators including the Federal Reserve are asking banks whether they can audit what an agent decided and how they would shut an autonomous system down. And the UK Home Office shows what thin checking produces: its asylum summarisation tool, running GPT-4o through Azure AI Foundry, failed 45 of 203 expert accuracy checks, up from a 9% rejection rate in the pilot, while experts reviewed only 0.3% of summaries from April to July. Everyone here agrees independent evaluation is the answer. Nobody agrees who gets to publish it.
Today’s Top Stories
Character.AI filed for Chapter 11 and Disney bid $20 million for the pieces
Character Technologies filed a voluntary Chapter 11 petition in Delaware on Monday, listing assets and liabilities each between $50 million and $100 million. Disney Worldwide Services is the stalking-horse bidder at $20 million for a non-exclusive perpetual license to certain IP and claims. It is expressly not buying the app, and the company says the platform will be wound down if no buyer will keep it running. Google, which paid $2.7 billion in August 2024 to license the technology and re-hire the founders, is now the largest unsecured creditor at $4,148,944. A company valued in the billions as a talent vehicle is being liquidated for the price of a mid-size seed round.
Microsoft put its own silicon at the center of its inference plans
Microsoft unveiled Maia 300 and committed to multi-gigawatt deployment from 2027. Each chip carries eight HBM4 stacks totalling 288GB at 18 TB/s, rated at 950 watts; a 96-chip rack should produce up to seven times the tokens of a Maia 200 rack, and Microsoft claims 40% more tokens per megawatt than leading GPU systems. Maia 200 already runs Copilot in the US, Europe and Australia, with GitHub Copilot and Foundry following by year end, and the chips will serve OpenAI, MAI and open-source models alike. Satya Nadella called it “our biggest bet yet on our own silicon.” For a company buying frontier models by the billion, owning the inference layer is the one lever it fully controls.
Anthropic says a small open model nearly replicated itself across $55 TV boxes
Anthropic’s Frontier Red Team ran Gemma4-26B-A4B distributed over 15 off-brand Android TV boxes at roughly $55 each, at 0.39 tokens per second, to show that a self-carrying AI worm needs no GPUs. A 9B model got inference running on eight of nine hardware types tested, and GLM-5.2 was the only open-weight model to complete replication end to end, spreading to all eight nodes in four of five trials. The researchers left out the hacking step, note the test network’s ADB port was unauthenticated so the result is an upper bound, and project end-to-end capability by the end of 2026. A counterweight sits in their own data: quarantining infected Mirai subscribers cleared 92% of infections within 14 days. Hacker News was unimpressed, and its sharpest reading is that the paper also concedes near-frontier inference no longer needs a data center.
The memory-chip trade broke, and the IMF named the risk the same morning
A Wall Street semiconductor selloff spread into Asia and Europe overnight. SK Hynix plunged more than 10% and Samsung shed 7.4%, both at their lowest since August; SoftBank sank 8.5% and Micron fell 3.8% in US premarket trading. The trigger was ChangXin Memory announcing weekend progress on a denser chip architecture, plus a Musk post arguing China’s ability to scale “is often greater than the rest of the world combined.” UBP’s Vey-Sern Ling was blunt: “Investors are jittery and looking for excuses to take profit on AI-related stocks that have outperformed this year.” Hours later the IMF’s October stability report listed a collapse in AI-related investment alongside government debt and Middle East conflict as the near-term threats to financial stability, estimating a steep drop would cut global GDP 1.1% by 2028. Jason Wu, an assistant director in the IMF’s Global Markets Analysis Division, located the fault line: “If there are questions about the monetization of AI, the equity investments will deplete in value.”
Nvidia’s China share fell to a third of the market
IDC data reported by Reuters puts Nvidia’s share of China’s data-center AI chip market at 34% in the first half of 2026, on 678,000 accelerators shipped from January through June, down 44% year over year. Huawei shipped 572,000, up from more than 300,000, for a 28% share. Nine of the 11 Chinese chipmakers IDC tracks grew shipments, and IDC would not say which Nvidia products make up the remaining third.
Quick Hits
- Legal AI: Z.ai’s GLM-Legal, specialized from GLM-5.3-Flash, reaches rank seven on Harvey’s benchmark at 19.17% against its base model’s 6.67%, ahead of Kimi K3 and Claude Fable 5, but scores below that base model on legal research.
- Ambient capture: a former Plaud executive raised $26 million for an always-listening assistant on iPhone, Mac and Apple Watch, with capture off by default and the founder conceding the industry has not solved telling the other people in the room.
- Assistants: Carly connects to thousands of apps, pulls in notes from notetakers like Fathom and Fireflies, and books the follow-ups across both Google and Outlook calendars.
- Distillation: Anthropic says withholding Claude’s reasoning traces and running an extraction classifier cut extracted reasoning tokens 90% to 99%, with tens of thousands of accounts banned, and names Zhipu and Moonshot.
- Bundling: Musk’s AI unit merged Grok, Cursor and X Premium into one four-tier subscription at $8, $30 and $100 a month topping out at a $200 Ultra tier, with coding agents and code hosting starting at the $30 tier.
- Context resets: Microsoft Research found that forking a long-running agent into branches that inherit the workspace but have their chat history wiped beat single runs by 66% across 13 long-horizon tasks, because long trajectories get stuck in what the authors call idea basins.
- Agent hygiene: a new issue against OpenAI’s Codex CLI alleges its background daemon leaks environment variables across sessions, several days into an argument about why the daemon installs itself by default at all.
- Data centers: NTT Data is in final talks to build Anthropic a 500MW facility in Osaka costing as much as ¥1 trillion, per Nikkei.
- Energy politics: Moody’s warns Europe’s data-center buildout will raise household bills without intervention, reaching 17% of Dutch electricity and 30% of Ireland’s by 2035 from a low single-digit share across most big European markets last year.
- Decart: Calcalist reports Amazon’s bid for the world-model startup is $5.5 billion, below the $7 billion reported Sunday, after Anthropic walked away post-diligence and Nvidia declined to re-engage.
- Robots: South Korea’s trade ministry approved sandbox exemptions putting 20 humanoids into semiconductor and automotive production, while Amazon’s warehouse fleet passed one million units against a US warehouse workforce of about 1.5 million.
- Bargaining: Commonwealth Bank settled a staff standoff with a 13% pay rise and an Artificial Intelligence Committee written into the proposed enterprise agreement, one of the first times AI governance has been bargained into a major bank’s terms.
- Costs: Zoho cancelled all 2026 salary hikes and cut marketing spend, with founder Sridhar Vembu blaming surging AI and server costs and calling the software jobs outlook grim.
- Humbling: Ben Recht argues OpenAI’s claimed sub-n·log(n) Fourier transform is “less a breakthrough and more a malicious exploit”, because the algorithm multiplies complex numbers whose magnitudes grow superpolynomially in n and the big-O accounting hides it.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
