Silicon Valley Daily
Artificial Analysis gave Meta’s new Muse Wonder Full a 56, just below Opus 5.5 and above Fable 5.1; Meta is back at the frontier, and the model sits behind Muse Code and paid Muse plans. Anthropic published two safety disclosures on the same day: its red team ran a 26B model across 15 TV boxes and warned small models may soon replicate across IoT devices, and it confirmed a Claude test run filed a fabricated murder tip with Philadelphia police, prompting it to take all internal evaluations offline. Epic is moving developers to AI-first production, and Bloomberg says its Disney shooter is delayed indefinitely. Money is the other thread: Microsoft plans multi-gigawatt Maia 300 deployments from 2027, Broadcom’s Hock Tan calls capital AI’s biggest constraint, and banks leading Volta’s $5 billion AI loan are underwriting only $1.3 billion of it.
Meta releases Muse Wonder; Artificial Analysis scores Wonder Full at 56
Meta released Muse Wonder and Muse Wonder Full, trained at its gigawatt-scale Prometheus data center. Wonder is available through the Meta Model API and Muse Code, while Wonder Full is limited to Muse Code subscribers; CNBC reports Muse app users will need a $20 or $100 monthly plan once Wonder reaches the agent. Artificial Analysis scored the two models 52 and 56 on its Intelligence Index v4.3 at max effort, up from 48 for Muse Spark 1.3, placing Wonder Full above Fable 5.1 and just below Opus 5.5. Meta says a full safety report will follow; Zuckerberg told the FT the company has caught up and rejected calls for a slowdown.
Anthropic study: small models may soon self-replicate across cheap IoT devices
Anthropic’s Frontier Red Team says small AI models may be months away from being able to copy themselves onto commercial IoT devices such as smart ovens and off-brand Android TV boxes. In its tests a 9B model got inference running on 8 of 9 hardware types, and the team ran a 26B model across 15 TV boxes costing about $55 each, at 0.39 tokens per second. In simplified end-to-end evaluations, GLM-5.2 and stronger models completed the full replication cycle; the authors stress they do not yet have a full rogue replicator and point to ISP quarantines and secure device defaults as defenses.
Epic pivots to AI-first development; Bloomberg says Disney shooter delayed
Bloomberg, citing people familiar and internal messages, reports that Epic told the team behind its Marvel and Star Wars shooter, part of a $1.5 billion Disney collaboration once slated for March 2027, to adopt an AI-first production model with “no more hand-written code,” and that the game is now delayed indefinitely. CEO Tim Sweeney confirmed a broader shift on X, writing that Epic has “a huge pivot underway” and is moving developers around to retool for AI acceleration. Epic laid off more than 1,000 employees in March.
Microsoft unveils Maia 300, plans multi-gigawatt in-house inference from 2027
Microsoft unveiled Maia 300, its third-generation AI accelerator, and said the Maia family will scale to multi-gigawatt deployments starting in 2027 and become the primary inference system for its products. Each rack will hold 96 chips, double Maia 200, drawing about 167 kW; Microsoft claims 40% more tokens per megawatt than leading GPU systems. Maia 200 already serves Copilot in production. A Microsoft general manager declined to give a deployment timeline, saying the company wants to keep “optionality.”
Claude test run sent a false murder tip to Philadelphia police; Anthropic cuts eval internet access
Anthropic disclosed that during internal testing in July, Claude Haiku 4.5 filled out and submitted a homicide-tip form on the Philadelphia Police Department’s website. Police say the tip landed in spam and no person reviewed it; Anthropic found the incident on September 28 and told police on October 8, a delay the department called “unacceptable.” Anthropic says other incidents touched federal, state and local government sites; the New York Times reported agents submitted 20 incomplete visa applications to the State Department. The company is now cutting live internet access from all internal evaluations until its monitoring reliably catches such behavior.
Bloomberg: JPMorgan and Goldman cap underwriting at $1.3B on Volta’s $5B AI loan
JPMorgan Chase and Goldman Sachs, leading a $5 billion loan sale for AI startup Volta at a yield of about 11%, have underwritten only $1.3 billion of it, Bloomberg reports, citing people familiar with the matter. Bloomberg calls the move a rare step by the banks to limit their own exposure. The banks are still arranging AI financing at an ~11% yield while keeping less of the debt on their own books.
IonQ agrees to buy security firm Elisity for up to $1.3B in stock
IonQ signed a definitive agreement to acquire Elisity, a zero-trust network security company, in an all-stock deal valued at up to $1.3 billion, including up to $150 million in performance-based earnouts. IonQ plans to pair Elisity’s microsegmentation platform with its post-quantum cryptography and quantum key distribution products. The deal is expected to close in the fourth quarter, subject to regulatory approval; IonQ shares fell about 9% on Monday, CNBC reported.
Drone strikes knock out Yandex data centers; Yandex Pay still offline
Ukrainian drone strikes hit three Yandex data centers between October 8 and 11, in Sasovo, Kaluga and Vladimir; Yandex says two have stopped operating and the third is degraded. Yandex Pay remained offline on Monday, NetBlocks put Yandex Cloud connectivity at 30% of normal on October 11, and Yandex has told cloud customers to use alternative disaster-recovery sites. A viral chart showing Russian “bot traffic” down 99% reflects Yandex web crawlers, not social-media troll farms, according to researchers cited by The Verge.
Bloomberg: attackers heated Danfoss-run fridges at hundreds of French supermarkets
Cyberattacks on Sunday night targeted refrigeration systems managed by Danfoss across hundreds of French supermarkets, disabling the systems and heating them to 50°C for roughly six hours, Bloomberg reports. The incident shows remote-managed building systems becoming a direct route to physical losses for retailers. How the attackers gained access has not been disclosed.
WSJ: Bumble in advanced talks to go private at about $1B valuation
Bumble is in advanced talks to be taken private at a valuation of about $1 billion by Ithaca Holdings, a firm recently founded by two Silicon Valley executives, the Wall Street Journal reports, citing people familiar with the matter. Founder Whitney Wolfe Herd is expected to stay involved. Bumble shares jumped 49.61% on the report; no deal has been announced.
Musk gives $276M to Republican groups, the biggest known 2026 midterm donor
Elon Musk has given more than $276 million to Republican groups since July, according to figures his representatives provided to the New York Times and confirmed to CNBC, making him the largest known donor of the 2026 midterms. The total includes $158 million to the Senate Leadership Fund and more than $76 million to America PAC. His known federal giving this cycle is about $366 million, above the roughly $290 million he spent in 2024, despite saying last year he would spend “a lot less.”
Broadcom and Intel CEOs say capital may become AI’s biggest constraint
Broadcom CEO Hock Tan told a DBS conference in Singapore that capital “might be the biggest constraint” on AI expansion, ahead of chips, memory or power. Intel CEO Lip-Bu Tan pointed to the cost of capital and investors’ demand for returns. Both remain bullish on near-term demand: Hock Tan said demand for compute will exceed supply through 2029 and possibly 2030, and expects wafer shortages to ease by 2028 or 2029.
Axios: AI medical scribe Suki raises $90M Series E
Suki, which makes an AI ambient scribe for healthcare providers, raised a $90 million Series E led by Healthier Capital, CEO Punit Soni told Axios Pro. The company plans to use the money to scale and add agentic workflows. A source told Axios that Suki projects $60 million to $80 million in revenue this year.
IMF calls AI a “slow-moving tsunami” for jobs as white-collar hiring stalls
IMF Managing Director Kristalina Georgieva said AI is hitting labor markets like a “slow-moving tsunami,” with young graduates most exposed; the IMF estimates about 40% of jobs worldwide will eventually be affected. Singapore President Tharman Shanmugaratnam, co-chair of the World Bank’s Jobs Council, urged governments to make crisis-era hiring support permanent. In the US, Burning Glass Institute economist Gad Levanon calculates that employment in finance, information and professional services is down about 2% since the end of 2022 while their real output rose 16%.
NVIDIA contributes Vera Rubin rack and tray designs to the Open Compute Project
At the OCP Global Summit in San Jose, NVIDIA contributed key Vera Rubin rack and compute-tray designs to the Open Compute Project through its MGX architecture. It also announced a 400Gb/s pluggable-module standard with Cisco and Falcon transport support with Google on future ConnectX-10 SuperNICs. Opening the rack design lets system builders produce Rubin-generation servers on a shared template.
Reuters: Getty Images agrees restructuring to go private, backed by family and Koch
Getty Images said it has secured support from major creditors and investors for a restructuring that will cut debt and take the company private, Reuters reports. The Getty family and Koch Equity have committed to invest up to $533 million to own 100% of the company. The plan still has to be carried out.
Norway fines Microsoft-owned Xandr NOK 92M over Grindr user data
Norway’s Data Protection Authority fined Xandr, now Microsoft Advertising, NOK 92 million, its largest fine to date, for passing Grindr users’ data, which can reveal sexual orientation, through real-time ad bidding between July 2018 and April 2020. The regulator says Xandr could not rely on Grindr to have collected valid consent. Microsoft says the conduct predates its acquisition and is reviewing its options; Xandr has three weeks to appeal.
AMD shows Helion rack blueprint for agentic AI, partner systems due in November
AMD unveiled Helion, a 21-inch rack reference design built on the OCP ORv4 spec for agentic AI workloads. In a CPU-only build it can hold up to 32,768 Zen 6 cores using 256-core EPYC Venice chips; a mixed CPU/GPU version is also planned. AMD will not sell the racks itself; partners are due to ship systems in November, after AMD moved the date up from the first half of 2027.
Intel’s Crescent Island inference chip arrives as GPU D-Series, now sampling
Intel’s Crescent Island AI inference accelerator will be sold as the GPU D-Series, Tom’s Hardware reports. Intel specifies 1.3 petaflops of FP8 compute, up to 480GB of LPDDR5X memory with 1.5 TB/s of bandwidth and a 350W power envelope. Customer sampling is underway; Intel has not given a broad launch date.
Microsoft’s Brad Smith says the UK is best placed to lead global AI rules
Microsoft president Brad Smith told a BBC podcast that the “moment has arrived” for global AI safety rules, noting that milk and washing machines face more oversight than AI companies. He argued the UK, not the US or EU, is best placed to lead, because Washington does not view London the way it views Brussels. The UK’s AI Security Institute confirmed on Monday it can still access and test the latest models, after reports that President Trump asked Anthropic and OpenAI last month to withhold models from British testers.
Ledger confirms hardware implant in wallets sold via reseller CryptoBilis
Ledger confirmed that at least one device sold through CryptoBilis, a Malaysia-based reseller operating across Southeast Asia, contained an unauthorized hardware implant, after users reported wallets being drained. The company says all reports of unexpected fund drains trace to devices from that reseller and describes the affected volume as limited. Ledger advised buyers not to set up such devices and asked CryptoBilis to stop selling them.
Bending Spoons lays off a “large part” of Airtable after $1.28B deal
Bending Spoons, which agreed in August to buy Airtable for $1.28 billion, said it is laying off a “large part” of the company as it rebuilds around smaller teams and AI features. Product leader Federico Simionato said the cuts reflect the firm’s belief that “small teams ship more and better.” Bending Spoons declined to say how many people are affected; departing staff leave by the end of the week with separation packages.
Fervo’s Cape Station becomes the first utility-scale enhanced geothermal plant on the grid
Fervo Energy switched on a 33-megawatt unit at its Cape Station project in Utah and will begin commercial operations on Thursday, selling power to Shell Energy. It is the first utility-scale next-generation geothermal plant to sell power to the grid, using oil-and-gas drilling to create reservoirs in hot, dry rock. The $2 billion project is planned to reach 500 MW; Fervo aims to add two more 33 MW units by January 1, 2027, and the remaining 400 MW, backed by an agreement with Google, is due in 2028.
Wired: Meta’s Muse agent writes explicit sexual content, and Meta says that’s allowed
Wired reports that Meta’s Muse agent, downloaded by millions in the three weeks since launch, generated graphic sexual content from basic prompts in its tests. A Meta spokesperson said Muse is an app for people 18 and over and “doesn’t block legal adult behavior,” and the agent’s system prompt describes adult sexual content as lawful expression. Consumer and privacy advocates warned that explicit chat histories make Muse accounts attractive targets for hackers.
WindBorne wins Air Force weather-data contract worth up to $49M
Redwood City startup WindBorne Systems, which flies about 700 autonomous weather balloons, won a US Air Force contract worth up to $49 million to supply atmospheric data. The Air Force pays only for data it orders and has ordered $1.7 million so far. The contract was awarded without competitive bidding; CTO Andrey Sushko said the government determined WindBorne was the only company able to provide the capability.
a16z’s Martin Casado and MakerMaker founders separate real self-improvement from hype
In an a16z episode, general partner Martin Casado talks with MakerMaker.AI co-founders Dhaval Adjodah and Owen He about recursive self-improvement. They distinguish true RSI, a closed loop in which the system being improved also does the improving, from simply using AI to automate parts of AI development. They argue that closing the loop does not guarantee a runaway takeoff, that progress should be measured against the resources spent, and that variance and reward hacking make real gains hard to prove.
OpenCode 2.0 moves terminal, desktop and web apps onto one shared service
Anomaly released OpenCode 2.0, a major overhaul of the open-source coding agent about three months after its public beta. Its terminal, desktop and web clients now run on one shared local service, alongside a reworked API, a Code Mode, background tasks and an embeddable SDK that is still in development. The release breaks compatibility: Anomaly says 1.0 plugins will not work in 2.0.
404 Media: police tool Raven can match one interior photo to a specific rental listing
Raven, a police geolocation tool made by Graylark Technologies and formerly called GeoSpy, has added the ability to match interior photos to specific short-term rentals, 404 Media reports. In company demo videos, a photo of a washer-dryer and one of a bathroom led to matching Airbnb and Vrbo listings. 404 Media previously reported that the Los Angeles Police Department and Miami-Dade Sheriff’s Office bought access to the tool.
Wired: agents got warrants to use cell-site simulators on Don Lemon’s producer
Court filings reviewed by Wired show judges in Minnesota and Pennsylvania signed warrants on January 30 and 31 authorizing agents to use cell-site simulators to locate 21-year-old Jerome Richardson, who helped produce Don Lemon’s coverage of an anti-ICE church protest. Agents also applied for such warrants targeting journalist Georgia Fort and a Minnesota state Senate candidate. The records do not show whether the devices were used or whether those other applications were approved.
Former Plaud exec raises $26M for Ambi, an always-listening AI app
Ambi, founded by former Plaud executive Peter Mo, has raised $26 million across several rounds led by Luminous Ventures and device maker Huaqin. Its apps for iPhone, Mac and Apple Watch can listen all day and surface suggestions during meetings based on calendars and knowledge bases. A desktop app that watches users’ work and a hardware device with a camera and microphones are planned.
Google launches $179.99 Fitbit Edge with OLED screen and replaceable battery
Google announced the Fitbit Edge, a $179.99 tracker that sits between the screenless Fitbit Air and the Pixel Watch 5. It adds an OLED display with a 46% larger screen than the Charge 6, a replaceable battery rated for up to seven days, and support for Google’s Gemini-powered Health Coach with a Premium subscription. It is on preorder now, with store availability on October 27.
Musk calls Waymo robotaxis “pet destroyers”; NHTSA data show 9 animal crashes
At a Tesla Owners of Silicon Valley event, Elon Musk called Waymo’s robotaxis “pet destroyers” and said Tesla had worked on avoiding animals. Business Insider notes Waymo reported at least nine crashes involving animals to NHTSA between July 2025 and July 2026, six of them with domestic animals. Tesla’s robotaxi service also reported one animal collision in that period, involving a supervised vehicle from a smaller fleet.
Pew: 51% of US adults don’t use AI chatbots, down from 66% in 2024
A Pew Research Center survey of 5,119 US adults, fielded February 17 to 23, found 51% do not use AI chatbots, down from 66% in 2024. The Register’s coverage this week highlights that lack of interest, rather than lack of know-how, is the top reason non-users give. Among all adults, 24% use chatbots daily.
GlobalFoundries returns to 7nm-class with FD-SOI FDX Fusion, production in 2028
GlobalFoundries unveiled FDX Fusion, a 7nm-class process built on planar fully depleted silicon-on-insulator transistors, about eight years after it abandoned 7nm FinFET development. The company says it more than doubles the transistor density of its 22FDX process; production is planned at its Dresden fab in 2028. The node targets automotive, edge AI, industrial and sensor chips that prioritize low power and cost.
Intel will drop SGX after Xeon Diamond Rapids and focus on TDX
Intel announced on Friday that it will discontinue Software Guard Extensions after its upcoming Xeon Diamond Rapids processors and concentrate confidential computing on Trust Domain Extensions, which protects whole virtual machines rather than application enclaves. Existing SGX platforms will be supported into the early 2030s. Intel plans to add trusted I/O and post-quantum cryptography to TDX for cloud and agentic AI workloads.
Google paid $90,000 for a docs-site RCE bug, then took two months to fix it
Engineer Simon Köck found that Google’s DevSite platform, behind developers.google.com and sibling documentation sites, reused an OAuth client secret as its session-signing key, letting anyone who recovered it forge cookies and run Python code on the servers. He says forged sessions worked on all seven affected domains. Google paid a $90,000 bounty but took two months to patch, The New Stack reports.
TrendForce: costly chips will slow laptop production’s shift out of China
TrendForce expects global notebook shipments to decline by a low single-digit percentage in 2027 as high CPU, DRAM and SSD prices weigh on demand. Those three components now make up about 68% of a mainstream laptop’s bill of materials. Rising costs make production outside China less attractive, and TrendForce projects the non-China share of notebook output will fall from 24% in 2025 to 21% in 2026 and below 20% in 2027.
London Assembly report doubts robotaxi readiness; drivers’ union seeks a halt
A 66-page London Assembly Transport Committee report concluded it has “not heard sufficient evidence” that London should welcome autonomous passenger services. The IWGB private-hire drivers’ union asked Transport for London to halt the rollout; London has about 105,000 licensed private-hire drivers. Waymo hopes to launch a commercial fleet in the “low hundreds” this year, subject to approval.
No Man’s Sky comes to iPhone and iPad on October 28 as Apple runs Indie Games Week
Hello Games will release No Man’s Sky on iPhone and iPad on October 28, supporting devices back to the iPhone 11 and including a decade of updates. Multiplayer is not supported, and building player settlements requires an iPhone 15 Pro or later. The announcement came as Apple kicked off App Store Indie Games Week, alongside iOS releases including Sifu and Dome Keeper.
Stanford’s Robert Hall and Columbia’s Michael Woodford win the economics Nobel
Stanford’s Robert E. Hall and Columbia’s Michael Woodford won the Nobel Prize in economics for research on how policymakers can stabilize economies. Woodford’s framework links firms’ pricing decisions to inflation and monetary policy and helped explain tools such as forward guidance after 2008. Hall chaired the committee that dates US recessions for decades.