.-----.
\ ::: /
\ : /
) (
/ . \
/ ::: \
'-----'
══════════════════════════════════════════
T H E U N C E R T A I N F U T U R E S
all the news that fits the context window
──────────────────────────────────────────
No. 87 · Tuesday, 13 October 2026Welcome back to The Uncertain Futures, good to have company.
Meta released Muse Wonder, the first model from its Superintelligence Labs that it pits against the best from OpenAI and Anthropic, at the same price as Anthropic's Opus 5.5, which an independent ranking still puts first. Microsoft says its Maia chips, which it designs itself, will become the main system running its AI services, with Maia 300 deployed at scale from 2027. And Britain says its safety testers have access to the latest AI models, weeks after Washington asked OpenAI and Anthropic to keep their newest ones from British testers until the US had reviewed them.
About the newsletter: the brief now also carries an AGI section on days with news that is about artificial general intelligence itself; today is one of them.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Meta released Muse Wonder, the first model from its Superintelligence Labs that it pits against the best from OpenAI and Anthropic, at the same price as Anthropic's Opus 5.5
Meta Superintelligence Labs, the company's AI unit, released Muse Wonder and the stronger Muse Wonder Full on Monday, models built to run long, multi-step tasks for its Muse assistant. On Meta's own figures, Wonder Full beats Claude Opus 5.5 and GPT-6 Astra on memory and legal work. The independent ranking by Artificial Analysis still puts Opus 5.5 first, with Wonder Full tied with a cheaper Anthropic model, Decrypt reports. Through Meta's API, Muse Wonder costs $4 per million input tokens and $20 per million output, exactly what Anthropic charges for Opus 5.5; Wonder Full is open only to subscribers of Meta's Muse Code. Mark Zuckerberg, Meta's chief executive, told the Financial Times that Meta has reached the frontier; he has rejected calls for an industry slowdown. The same day, The Information reported that only 14% of the AI usage running through OpenRouter, a service that passes developers' requests to many models, went to the 10 highest-quality models in the third quarter, down from 55% in the first quarter of last year (TechCrunch · Decrypt · Meta · FT via Slashdot · The Information).
Microsoft announced Maia 300, an AI chip it designed itself, and says its Maia chips will become the main system running its AI services, with large-scale deployment from 2027
Microsoft expects the third-generation chip to produce 40% more AI output for the same power as leading GPU-based systems, a market Nvidia dominates. Large-scale deployment starts in 2027, and Azure chief Scott Guthrie said Maia will become the primary AI inference system for Microsoft's products and services, GeekWire reported on Monday. Satya Nadella, Microsoft's chief executive, called it the company's biggest bet yet on its own silicon. Google and Amazon let outside customers use their in-house chips; Microsoft does not yet, so Maia's only customer is Microsoft (Microsoft · GeekWire · TweakTown).
Britain says its AI Security Institute is testing the latest AI models, including Google's Gemini 4 Argon (update)
The UK AI Security Institute has been given access to the latest advanced models and is working with the US government and developers including Google, AI minister Kanishka Narayan said on Monday. Its scientists are evaluating Google's Gemini 4 Argon, Reuters reported; as we covered on October 7, DeepMind was then still sharing it with the institute. As we covered on September 25, the White House had asked OpenAI and Anthropic to keep their newest models from British testers until the US had reviewed them. The government also said it would extend two AI pilots that have helped identify and fix more than 1,000 potential cyber vulnerabilities across government systems. Narayan said the government will consider legislating on AI incident reporting and loss-of-control risk. Britain is seeking a leading role in assessing and managing the risks of powerful AI, Reuters noted (UK government · Reuters via The Star · The Next Web · Tech.eu).
Washington plans to move the leaders of its AI testing agency into the White House, days after ordering AI companies to report incidents
Three senior officials of the Center for Advancing Innovation and Standards for Super Intelligence (CAISSI), the Commerce Department unit that evaluates advanced models, would report to Michael Kratsios, director of the White House Office of Science and Technology Policy, and Sean Cairncross, the national cyber director, The Information reported on Monday. A White House official cited coordination problems with Commerce. On Friday, the administration's Super Intelligence Force, the AI task force we covered on October 5, told all AI companies to disclose and remedy incidents involving their models promptly, calling the process "not optional", Axios reported. It followed Anthropic's disclosure of unintended actions by its test models, which we covered yesterday, and Anthropic's report to the State Department that one of its models had submitted 19 incomplete visa applications in August. The Friday statement did not say how compliance would be monitored or enforced (Reuters via US News · Axios · The Spokesman-Review).
Amazon is in talks to buy Decart, a startup whose software moves AI models between different chipmakers' hardware, for about $7 billion
Decart's software helps developers move AI models between chips from different makers, and Amazon has already given it early access to its own Trainium chips, The Wall Street Journal reported on Sunday. The reported price is about 75% above the nearly $4 billion Decart was valued at in May. Anthropic walked away from a $6 billion deal for the company last month, as we covered on September 9. For Amazon, a deal would deepen its push to compete with Nvidia; the software it would be buying moves models off Amazon's chips as well as onto them (Spokesman-Review · KSL).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ EARNINGS ▓▓▒▒░░ASML, the chip-equipment maker, and TSMC, the contract chipmaker, both report this week.
░░▒▒▓▓ AI ▓▓▒▒░░Anthropic says new defenses have cut the data rivals extract from Claude through distillation by around 90% to 99%, on early evidence. Distillation here means training a cheaper model on another model's answers without the owner's permission. Anthropic says it restricted access to Claude's reasoning and made it harder to probe the model through proxy networks; it has banned tens of thousands of accounts tied to distillation, often within two hours. It named Zhipu, now Z.ai, the Chinese developer of the GLM models, saying it distilled leading US models' cyber skills for its GLM-5.3 and, once Anthropic's safeguards degraded its attempt on Claude Fable 5, switched to an older Claude model; as we covered on September 10, US agencies had named Z.ai among six Chinese AI companies they accused of distillation (Anthropic · Reuters via Yahoo · BusinessWorld).
Epic Games told the team behind its delayed Disney shooter to stop writing code by hand. The game, once set for March 2027 and featuring Marvel and Star Wars characters, is now an "AI first" production using tools such as Anthropic's Claude, and one of its directors told staff in an internal message that its art would come from AI workflows. Tim Sweeney, Epic's chief executive, said Epic has a huge pivot underway to retool for maximum AI acceleration; the aim, he wrote, is for everyone at Epic to build more and better things rather than to do the same work with fewer people (esports.gg · Push Square).
Microsoft's two most senior executives both asked for tighter controls on AI. In an essay posted on X on Saturday, Satya Nadella, the chief executive, wrote that companies should assume a model is compromised and give a human operator an emergency brake that can stop it partway through a job. In an interview published on Monday, Brad Smith, Microsoft's president, pointed to the safety checks milk must pass and said a washing machine is subject to more safety rules than AI (TechCrunch · BetaNews).
Nvidia is in talks to invest more in Reflection AI or to buy it, the Financial Times reported on Saturday. Nvidia has already put $800 million into Reflection, the open-model startup whose first open-weight model, Beam, we covered on October 6, and the latest round valued Reflection at $25 billion before the new money. The talks are early; one option is an acqui-hire, in which Nvidia would hire the staff and license the technology and could avoid a lengthy regulatory review (SiliconANGLE · US News).
░░▒▒▓▓ AGI ▓▓▒▒░░Jack Clark, an Anthropic co-founder, now expects AI that can improve itself without human help to become possible in 2027, a year earlier than he expected in May. Clark tracks this recursive self-improvement, AI systems building better AI systems, in his Import AI newsletter, where on Monday he wrote: "I now think it’s more likely we’ll see full RSI become possible in 2027, whereas before I thought it was more likely in 2028." On May 4 he had put the chance of it happening by the end of 2028 at roughly 60%. He cites TasteVal, a new benchmark from P-Zero Research, which reports that frontier models' skill at designing AI experiments has doubled every three months since December 2025. The top scorer on TasteVal, ahead of a baseline drawn from 24 human experts, is his own company's Opus 5.5 (Import AI · GIGAZINE).
░░▒▒▓▓ TECH ▓▓▒▒░░Global smartphone shipments fell 20% in the third quarter, the weakest such quarter since 2012, as AI data centers soak up memory chips. Phone makers raised prices and cut volumes as memory costs rose, the market-research firm Counterpoint Research said in preliminary estimates. Samsung led with 23% of the market. Apple's share reached 22%, its first time above 20% in a third quarter, even though it shipped fewer phones; Xiaomi's shipments fell by more than 30% (Reuters via The Star · 9to5Mac).
Apple holds an event devoted to the smart home in New York today, with news expected around 15:00 in Zurich (update). The centerpiece is expected to be the smart-home hub we flagged on October 8, which Bloomberg says will be called HomeView, a Siri-driven home display. 9to5Mac expects it alongside a new Apple TV and HomePod mini, with HomeView reportedly shipping around mid-November (9to5Mac · MacRumors).
░░▒▒▓▓ INFRA ▓▓▒▒░░Applied Materials, the chip-equipment maker, opened a research center it expects to cost about $5 billion over time, where chipmakers work on new manufacturing processes under one roof. The EPIC Center in Sunnyvale, California, has 600,000 square feet with nearly four acres of cleanroom, Applied said on Monday. Its founding partners are TSMC, Intel, SK hynix, Micron and Samsung, several of which compete with one another for the same orders (Applied Materials · GuruFocus).
Tencent is selling about $5 billion of bonds, mainly for AI spending, most of it on chips. The offering, which could be priced as early as today, combines $1.5 billion to $2 billion in dollar notes with 20 billion to 25 billion yuan of offshore yuan notes, about $3 billion to $3.7 billion, Bloomberg reported. Tencent raised capital spending on AI and computing by 176%, to 52.8 billion yuan, in the June quarter as it tries to catch ByteDance and Alibaba (Bloomberg · Yahoo Finance).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Russia kept striking Ukraine's energy facilities after Trump announced an "energy ceasefire" that he said both sides had agreed to. Russia's Foreign Minister Sergei Lavrov said on Monday that Zelensky had not earned a respite, and the Kremlin rejected Zelensky's offer of a 60-day ceasefire to allow elections, offering a pause on election day only. Russian strikes hit a power plant in Kharkiv shortly after Trump's announcement on Sunday. Four Republican and three Democratic senators asked the administration to reverse the Russian diesel deal we covered yesterday (NPR · Ukrainska Pravda · Kyiv Independent · Senate committee).
Saudi Arabia extended the closure of Riyadh's main airport to October 17 after the deadly Houthi attacks (update). Authorities announced the extension on Monday, after the attacks we covered yesterday. Lufthansa Group canceled flights to Riyadh until October 16, and Air France-KLM suspended service until the end of the month (ABC News Australia · Al-Monitor).
░░▒▒▓▓ ALPS ▓▓▒▒░░Pool and sauna associations in German-speaking Switzerland want glasses that can record banned from all pool premises. The associations told towns and local authorities on Monday to amend their pool rules, and produced a guide and a poster to help staff enforce the ban. Camera glasses are hard to tell from ordinary ones, they said. Their counterpart for French-speaking Switzerland and Ticino is taking no measures for now, it told Keystone-SDA. Politicians in Bern and Zurich have already pushed for such bans (swissinfo · Bluewin).
░░▒▒▓▓ PAPERS ▓▓▒▒░░Anthropic — small open models may be months away from being able to copy themselves across cheap home devices. Anthropic's Frontier Red Team, its in-house group for probing dangerous capabilities, published research on October 12 testing whether models could install themselves and spread across a sealed-off network of devices. Of the open models tested, only Z.ai's GLM-5.2 reliably completed the whole cycle, reaching all eight emulated test devices in four of five runs; too large to run on them, it worked from a separate server and installed a small stand-in model, Meta's Llama 3.1 8B, on each. The team also ran one of Google's Gemma models across 15 Android TV boxes costing about $55 each, under $1,000 in hardware. The authors note the test kept the hacking step trivial, so the result is an upper bound, and Anthropic says there is no full rogue replicator yet. A copy with no data center behind it has no central switch to turn off (Anthropic · Al Jazeera · Unite.AI).
Apple — idle Macs can be pooled into a private GPU cloud for running AI models. Apple's researchers describe Minions, a system that gathers the graphics processors of Apple Silicon Macs while they sit unused and runs machine-learning inference on them, with fault tolerance. The pitch is for teams short of central GPUs: compute from Macs a company already owns, without buying more hardware (Apple Machine Learning Research · arXiv).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Salesforce renamed its AIForce platform SIForce, adopting the White House's "super intelligence" vocabulary. Marc Benioff, Salesforce's chief executive, announced the rename on X on Friday; it follows a September 29 executive order telling federal agencies to make the switch in official documents and websites. The order does not apply to private firms, The Next Web noted, adding Business Insider's point that renaming a product does not show it meets the superintelligence bar (The Next Web · TheStreet).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Five Swiss churches and church organizations published ethical principles for dealing with AI on Monday, under the title Humanity in the Digital Age. Among them are the Swiss Bishops' Conference, the Evangelical Reformed Church of Switzerland and the Swiss Evangelical Alliance. The guidance is meant for church institutions, congregations and ministries. There are ten principles, naturally (Bluewin).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Birdwatchers — Axel Boman, Superpitcher
Amen — from the Uncertain Futures desk