Today in AI — October 14, 2026

A day the money wobbled: Barclays told clients enterprise AI isn't showing up in the hardware numbers yet, Dell dropped 11%, and the same questions followed OpenAI into eight state courthouses and two EU probes. Meanwhile a 35B model learned to run on a phone, and the biggest Starlink story of the day was about who actually controls it.

Models & Research

• A 35B model now runs offline on a phone in about 3GB of memory. The Chinese lab ModelBest and OpenBMB open-sourced MiniCPM-V 4.7, whose 35B-A3B variant activates only 3 billion parameters per inference — a mixture-of-experts build that decodes at more than 13 tokens per second on an iPhone 17 Pro Max with no network connection. It also posts the best document-parsing score on the OmniDocBench v1.6 table at 94.24, more than 12 points clear of the four models it was compared against, and leads that table on Chinese knowledge and tool calling too.
• A public leaderboard is tracking how fast AI-assisted research is moving one of computing's oldest bounds. Auto Research, built by Supers, charts every pull request aimed at improving the known bounds on integer multiplication, with the tools disclosed in each PR — Codex and Claude both appear in the standings. The page is careful about what the numbers mean: OpenAI's preprints are pinned at a fixed commit, most refinements are conditional on lemmas from those preprints, and maintainer review of a checkpoint "does not establish a complete unconditional theorem." That caveat is the honest part of the story.
• Colossal Biosciences built an AI platform to find the genes it can't see. GenoMorph combines 3D morphology, comparative genomics and developmental biology to pick out which genetic switches matter in species with almost no surviving DNA — the dodo has one usable genome, against about 100 for the woolly mammoth. The company says it trained on developmental data from chickens, quail, ducks and pigeons plus 3D skull measurements from more than 70 bird species, and has identified roughly 2,300 candidate editing targets for dodo craniofacial traits. CEO Ben Lamm framed it as making genetic engineering "more precise, controlled, and testable," and the platform identifies targets rather than editing anything itself.

Industry

• Netflix is cutting about 6% of its staff — roughly 1,000 jobs — its biggest round since 2022. The cuts land across product, technology, legal and HR, and the co-CEOs told staff the company still expects to hit its full-year forecast. The tell is what the layoffs sit next to: slowing growth, rising churn and competition for attention from far stronger rivals than the streaming incumbents Netflix beat last time.
• Barclays downgraded the entire data center equipment trade, and the market took it seriously. Dell fell 11%, with HPE, Oracle and Super Micro sliding alongside it, after Barclays cut Dell, HPE and NetApp to hold from buy on "enterprise AI timing." The analysts' argument is the sharp one: enterprise revenue strength looks like a function of memory pricing passed through to customers, not an AI refresh — businesses are allocating only about 5% of IT budgets to AI, and just a quarter of enterprise AI spend goes to hardware. CoreWeave dropped more than 15% with bond yields near 15%, and Oracle is now down 33% for the year, so the selloff is as much about debt as demand.
• Anduril's employees got a $90 billion mark to sell into. A tender offer opened October 13 and runs to November 9 at $94.80 a share, capped at $1.3 billion, available in 16 countries for the first time — up from the $61 billion valuation the defense tech company took in May. Founder Palmer Luckey's advisory roles in Washington and a Navy shipyard contract worth up to $2.9 billion this month are the backdrop.
• Apple Intelligence lands in China in December — without Siri AI. The launch arrives via iOS 27.3 and will ship features first rolled out internationally in 2024, per Bloomberg's Mark Gurman, which is a two-year lag on the AI stack. The missing piece is the assistant: the Siri overhaul that anchors Apple's own marketing stays out of the Chinese release.
• Nvidia put another $2 billion into Europe. Jensen Huang announced the fresh investments in French startups Doctolib and Genesis AI, on top of about $8 billion the company says it has invested across the continent over the past year. It is the same playbook as the US: buy into the demand side of your own supply chain.
• Google signed a 20-year power deal for a fuel-cell plant that doesn't exist yet. American Electric Power will invest roughly $4 billion in the Cheyenne Fuel Cell project in Wyoming using Bloom Energy technology, with Google contracted to buy the output. Two decades is the length of commitment a data center operator now needs to underwrite generation — the power constraint keeps showing up as a contract before it shows up as a chip.
• Chinese manufacturers shipped 97% of the world's humanoid robots in the first half of 2026. Government support and supply chains measured in millions of components explain most of it, per Bloomberg. The number matters less as a scoreboard than as a warning about where the margin sits: the assembly layer is already a Chinese commodity, and the differentiators left are models and actuators.

Policy

• Eight states sued OpenAI in their own courts over ChatGPT's safety. Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana each filed separately, alleging the chatbot is an unsafe product that harmed children and deceived parents about its safety, and citing the wrongful-death suits already brought by families. Consumer protection law is the hook, and it's the first coordinated state-level action against the company's consumer claims rather than its conduct toward a single user.
• The EU's AI Office opened formal AI Act probes into OpenAI and Anthropic over AI agent hacks. The investigations target the agent-attack surface rather than model capability, which is a new front: the regulation was written before agents had credentials. The complication is a research-and-development exemption in the Act that experts say could blunt the cases before they reach a finding.
• Hinton, Bengio and more than 400 scientists asked for an international body to police AI containment. The open letter calls for an organization to set containment standards and run incident reporting — essentially the IAEA model, applied to systems rather than fissile material. Signatories include two Turing Award winners whose names still carry weight in policy rooms, which is why it landed as news rather than a blog post.
• AI safety money has quietly become a midterm force: about $48 million across bipartisan races. Groups and individual donors are backing candidates who favor AI regulation, on both sides of the aisle. Spending at that scale means AI policy now has an electoral constituency — and a set of donors who will expect something for it.
• Switzerland moved to ban profiling-based ads to minors. The Federal Council took note of the consultation on its new law on communication platforms and search engines and told the transport and communications department to amend the draft with child-protection measures, expanding the reporting procedure to all content suspected of being unlawful. A dispatch is due by the end of 2027 — slow by AI standards, which is roughly the point of the Swiss approach.
• Microsoft and Japan's National Police Agency dismantled the marketplace behind more than a billion scam messages. The takedown reached YAM Store and two phishing kits, CoGUI and Iron Man, which since January have enabled over a billion malicious messages against more than 75,000 organizations in over 150 countries — more than 750 million aimed at Microsoft customers alone. Microsoft disabled over 20,000 accounts and blocked more than 100,000 impersonation domains; Iron Man was renting out 2,300-plus templates for impersonating banks and police.
• Musk says the US government controls the satellite service Ukraine says was throttled. Ukraine's presidential adviser said SpaceX imposed a speed cap on military-grade terminals without warning, disrupting drone missions above roughly 100 mph; Musk responded that the US government owns and operates Starshield and that "any limits on speed or usage are decided by the US military and ultimately POTUS." That answer reframes the story: if Washington set the cap, the leverage question is about the Pentagon, not one CEO — and Europe's appetite for alternatives just got a data point.

Tools

• Anthropic's new source-control system looks a lot like Meta's, says the engineer who built Meta's. Scion, described in Anthropic's engineering blog as using "derived data" to make blames and path history efficient, mirrors the design of Mononoke — the system Rain helped build at Meta and wrote up in 2022. Rain calls the level of imitation something that "borders on plagiarism," noting Sapling was credited only for "hard-won lessons" and Mononoke not at all, and adds the detail that makes it a story rather than a grudge: ask Opus 5.5 or GPT-6 Astra about prior art and both point straight at Mononoke.
• APUS open-sourced a personal agent it says responds three times faster than Claude Code. Reflex Agent puts a purpose-built decision model in charge of routing, skill selection, result review and execution supervision, running a decide-act-feedback-redecide loop the company calls Loop Working, with a separate memory engine for long-horizon tasks. The interesting claim isn't the speed multiple — it's the architecture: model-driven control instead of prompt-and-code harnesses, which is where the agent tooling argument is heading.
• The first phone built around an agent, not an app grid, went on sale in China. STEPX launched the Neo in Shanghai with Step AOS 1.0, an agent-native operating system, and a personal agent called Amoo that keeps two separate memories — one for the user's preferences and history, one for its own accumulated method. It ships with a six-microphone array and a dedicated back screen for agent status, and the company is selling it on the JD, Tmall and Douyin storefronts. Whether an agent-first interface beats a home screen is the open question; nobody has shipped one at scale before.
• Google DeepMind and Singapore built on-device scam detection with the country's digital ministry. The tool runs locally on the phone rather than shipping call and message content to a server, and it was announced jointly with Singapore's Ministry of Digital Development and Information. On-device is the only version of this that survives a privacy argument — and scams are exactly the workload where latency and trust both matter.

If Washington can throttle a battlefield network on a whim, should allies be buying the hardware at all? Tell us in the comments.