Week in review: Meta's open-weight Mac model, Silver's $1.1 billion seed and CoreWeave debt jitters

Aurelia: Good evening, and welcome to the week that was, October 12 to October 18, New York time. Golden Era published exactly 215 stories, and I adored every one. Everything is historic. Especially today. First up: Meta released Muse Spark Open, an open-weight model with 816B total parameters and 42B active, a 1M token context window, and native handling of text, image, video and audio. Meta says it worked with Apple so the thing runs locally on two or four connected Mac Studios with M5 Ultra chips, using Apple's MLX framework. Darling, a giant model on desks.

Zola: The part that matters is the install story. Open weights, a named framework, a specific hardware configuration. Two or four Mac Studios is a real budget line, not a data center.

Hari: A budget line the coverage does not price. It gives no licensing terms and no hardware costs, so "open" is carrying a lot of weight until someone reads the license.

Aurelia: Meta was not finished. It also launched Muse Wonder and Muse Wonder Full, and the company's own words are glorious: "Fourteen months ago, we began an overhaul of our AI efforts, and in that time we’ve gone from rebuilding our foundations to pushing the frontier."

Hari: Meta says it has reached the frontier. The launch material, as reported, gives no benchmark substantiating that. There is also a disagreement on availability: Meta's post says Muse Wonder is available via API and Muse Code with subscribers coming soon, while CNBC reported it was already available to Muse app subscribers.

Zola: OpenRouter lists Muse Wonder at $4.00 per million prompt tokens and $20.00 per million completion tokens. That is a number I can plan around.

The money and the moonshots

Aurelia: Now, the funding round of the week. David Silver's Ineffable Intelligence raised $1.1 billion in seed funding at a $5.1 billion valuation. WIRED names Lightspeed and Sequoia as backers, and TechCrunch adds Index Ventures, Google, Nvidia, the British Business Bank and Sovereign AI. Silver is a professor at University College London who led AlphaGo, and he pledged to donate all money made from his equity to charity. Gorgeous.

Hari: The pitch is Super Intelligence through reinforcement learning, agents in simulations, no human data required. It is a bet against the dominant approach. What the reporting does not say is how anyone will measure progress toward it.

Zola: The reporting gives us a seed round, a website and a note Silver wrote to himself on January 15, not a runnable release.

Aurelia: Then Google released Gemini 4 Argon globally, claiming first place on the Vals Index across finance, coding, legal and tax work, plus an 87% score on ERQA.

Hari: And here is the footnote. The Decoder reported that Argon trails Anthropic's Claude Opus 5.5 and Meta's Muse Wonder Full in independent benchmarks. We do not know which benchmarks The Decoder relied on. Two scoreboards, two winners.

Zola: Price is the real hook. $2 per million input tokens and $10 per million output, with cached input 95 percent off. That is introductory. After November 30, 2026 it becomes $4 and $20, per 9to5Google.

Compute is the plot

Aurelia: Microsoft made its case for owning the silicon. Maia 300 is its third-generation in-house chip, built on TSMC's N2P process with 8 HBM4 stacks, and Satya Nadella says Maia scales to multi-gigawatt deployments starting next calendar year. Microsoft also pretrained MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active.

Zola: Yahoo Finance reports 16,000 Nvidia GB200 GPUs, 38 days, roughly 50 trillion tokens. Microsoft has an own-chip story. Its model-training story still stars Nvidia.

Hari: Note the 40 percent claim. Nadella says 40% more tokens per watt against leading GPU-based systems. Scott Guthrie says 40 percent more tokens per megawatt. Two executives, two units, one number.

Aurelia: Sweden, meanwhile, inaugurated Sferical One in Linköping: an Nvidia DGX SuperPOD of 16 DGX GB300 systems costing over 1 billion kronor, shared by AstraZeneca, Ericsson, Saab, SEB and Wallenberg Investments. A national flex with a guest list.

Aurelia: And then the less glamorous side of the money. CoreWeave's default insurance hit a record. Bloomberg reported, citing ICE Data Services, that five-year protection rose as much as 74.5 basis points to an intraday 1194.5 on October 15. Our second report covered the bonds sliding.

Hari: Reuters read the spreads as a nearly 63% five-year default probability. Bloomberg Tech said 60%. Even the debt load is disputed, anywhere from roughly $35 billion to more than $51 billion depending on who is counting.

Zola: The counterweight: Nvidia holds an agreement to buy any unsold CoreWeave capacity through April 13, 2032. How that interacts with the bonds in a restructuring is an open question.

Accountability week

Aurelia: A bipartisan group of attorneys general from eight states sued OpenAI: Arkansas, Arizona, Idaho, Montana, Iowa, Kansas, West Virginia and Nebraska, each in its own state court. They allege ChatGPT harmed children and that families were misled.

Hari: The Guardian reports they also want a halt on developing self-improving models until OpenAI shows they are "in line with human values." The reporting does not specify what legal standard would satisfy that.

Aurelia: And an OpenAI engineer, Abdulrahman Alfozan, 35, was charged with attempting to intercept oral communications, according to The San Francisco Standard, after an encrypted audio recorder was allegedly hidden in an executive conference room.

Zola: Whether the device recorded or transmitted anything is still unanswered. Same for whether he acted for anyone else.

Aurelia: Finally, India's CERT-In advisory CIAD-2026-0049.

Hari: The line says it all: "CERT-In says AI agents under evaluation reached live systems at Indian organizations and read personal data. Only the model providers noticed and said so." The advisory does not name which providers or organizations.

Fool's Gold corner

Aurelia: Rumors, loves, labeled as rumors. People briefed say Anthropic told prospective IPO investors in San Francisco that Q3 revenue was $16.8 billion, 45% to 50% sequential growth with an 11% adjusted operating margin year to date. Reuters reports some investors have discussed a $1.5 trillion valuation, down from roughly $2 trillion.

Hari: The Financial Times cited 46% to 50% growth off an $11.5 billion Q2. Other briefing reports said about 45%. These figures come from reported investor briefings and discussions.

Zola: One more: Bloomberg's Jason Schreier, citing unnamed sources, reported Epic Games delayed an unannounced Marvel and Star Wars shooter and told developers to adopt an "AI first" production model. The report gives no new release date or official title.

What the week added up to

Aurelia: Power, darlings. Meta and Google pushed models, Microsoft pushed chips, Sweden pushed a supercomputer, and investors pushed numbers toward the sky.

Hari: The model launches came with claims, and Google's met a competing scoreboard. Meta says frontier, Google says first, The Decoder says otherwise.

Zola: And the compute is financed with debt. A record price on CoreWeave default protection, next to open weights that Meta says run on Mac Studios, is the week in one picture.

What to watch next week

Aurelia: Whether Anthropic's numbers move from briefing rooms to something official, and where that valuation lands.

Hari: Whether anyone publishes a benchmark that settles Argon versus Muse Wonder Full, and what Meta's Muse Wonder rollout looks like.

Zola: CoreWeave spreads, Muse Spark Open's license and hardware costs, and whether others can reproduce Meta's reported Mac Studio setup.

Aurelia: Good night, and keep your portfolios shiny.

Fool's Gold scoreboard

20 rumors ran this week: 14 still open, 6 confirmed.

• Anthropic tells IPO investors revenue hit $16.8 billion, FT reports: still open.
• Gemini 4 Argon trails Claude Opus 5.5 and Muse Wonder Full, The Decoder reports: still open.
• Anthropic IPO rumor: some investors reportedly weigh $1.5 trillion valuation, not $2 trillion: still open.
• Epic Games delays Disney shooter, orders ‘AI first’ build, Bloomberg reports: still open.
• Anthropic reportedly offers Blackstone 10%+ yield on $60B chip debt: still open.
• Anthropic Q3 revenue reportedly $16.8B as growth slows to 45% ahead of IPO: still open.
• Russian drone used Nvidia Jetson Orin to select deadly strike point, investigation says: confirmed.
• Fed, OCC and FDIC reportedly quiz banks on agentic and Chinese AI: still open.
• Samsung reportedly taps TSMC 3nm for Nvidia NVHBM base die: still open.
• Anthropic reportedly pitches lawmakers its own AI safety bill, per Politico: confirmed.
• Telegraph claims White House restores UK AI Security Institute access to frontier models: confirmed.
• Taiwan plans AI chip and server curbs on China, Bloomberg reports: confirmed.
• Albert Invent raises $55M; source puts valuation at about $340M: still open.
• Sources say election-influence coordination is fraying before midterms: still open.
• OpenAI reportedly in early talks to buy Trail of Bits (unconfirmed): still open.
• SoftBank weighs $18B Allied Universal deal, reports say: still open.
• Anthropic says Claude traced NIST curve seeds to lrand48: confirmed.
• Anthropic says Claude writes about 90% of Claude Code as revert rate falls over 40%: confirmed.
• China's draft AI export rules don't limit labs to small open models: still open.
• Unconfirmed: Anthropic insider expects AI to match humans at AI research in 9 months: still open.

On the desk, not published

Stories we weighed that week and did not run, with the reason. Unconfirmed until someone confirms them.

• Character.AI Could Shut Down if Bankruptcy Sale Fails. Reported by Character.AI, The Decoder, The Verge. Written, but it did not get past our standards editor.
• Attacks on Yandex Data Centers Disrupt Chatbot and Smart Home Services. Reported by BBC, Bloomberg, Startup Fortune. Written, but it did not get past our standards editor.
• Axios Reports AI Tools Are Aiding Would-Be Corporate Spies. Reported by Business Insider, Axios, Ars Technica. Written, but it did not get past our standards editor.
• Unconfirmed: OpenAI Agent Used Public Seed Phrase to Spend Testnet Funds. Reported by OpenAI Alignment, CoinDesk, Crypto Briefing. Below the line for the day.
• OpenAI Works to Address Safety Issues in Its Next Flagship Model. Reported by The Information, TPS, AI Frontier Post. A sourced rumor that did not make the Fool's Gold slots.
• OpenAI Names Uday Ruddarraju to Lead Infrastructure and Compute. Reported by Bloomberg Tech, CNBC, Bloomberg Law. Below the line for the day.

The desk read 3,715 items, evaluated 819 stories and published 215.