Daily AI News — 1 Nov 2026

Google said on 1 November that its Project Suncatcher prototype satellite has run an AI model in orbit and transmitted the result back to Earth. Launched on 1 October aboard a SpaceX Falcon 9, the craft carries four of Google's Tensor Processing Units and ran DeepMind's Gemma 4 model to answer prompts, starting with a haiku about space. Sundar Pichai called it a "hello world" moment. The test also exposed the hard limit: the satellite makes only about a kilowatt of power, enough to run one TPU for roughly 20 minutes at a time.

It is the first real proof that AI inference can happen in space, not just on the ground. If solar-powered satellites can one day carry compute, operators could sidestep the land, water and grid constraints that are now throttling data-centre growth. Google admits the economics may not work until the 2030s.

Picture a disaster zone with no working data centre: an orbital node could process satellite imagery and route rescue priorities directly, beaming answers to responders on the ground without waiting on terrestrial infrastructure that may be damaged or offline.

On 31 October, Google DeepMind showed a physical dancing robot that its AI agents designed, simulated and built end-to-end from one text prompt, with no manual CAD work. The machine has 12 degrees of freedom and a Halloween theme to match the date. The agents ran on Gemini 4 Argon, a reasoning model tuned for robotics simulation, and the Antigravity development platform. The project was led by Aaron Saunders, a former Boston Dynamics engineering VP, and points to what DeepMind calls "agent-first" robotics, where AI does the building and humans supervise.

Hardware design has always been the slow, expensive bottleneck in robotics — every joint and motor modelled by hand before anything is fabricated. Taking humans out of that loop, even for a simple robot, attacks the part of embodied AI that has resisted automation the longest.

A small manufacturer needing a custom pick-and-place arm could describe the task in a sentence and let agents produce a working design and control code overnight, instead of hiring a robotics firm for a months-long bespoke build.

Alibaba confirmed on 31 October that its Qwen model will power Apple Intelligence for iPhone users in mainland China. Qwen will sit behind Siri, Writing Tools and Visual Intelligence, letting users ask about photos and documents and have Siri call on Qwen for broader answers. Testing starts in the iOS developer beta on 4 November, with a public launch alongside iOS 27.3 in December. The redesigned next-generation Siri is not part of this release. Alibaba said user requests and responses will not be used to train its models.

Apple has struggled to ship AI in China, where it cannot run its own US models. Leaning on a local partner is how it stays competitive in its second-biggest market — and it marks the first time a major Western hardware maker has adopted a Chinese LLM as its primary system AI.

A Shanghai iPhone owner will soon get notification summaries, document questions and image clean-up handled by Qwen on-device and in-app, the same Apple Intelligence features users elsewhere already have, but routed through an approved local model.

OpenAI has cut its latest fundraising target by about $10 billion, Fortune reported on 31 October, as investor demand softens amid wider jitters over AI spending. The round was also overshadowed by arch-rival Anthropic, which filed to go public and disclosed that its cash flow turned positive in the first half of this year. The pullback lands in a week when Big Tech AI capex worries dragged down tech stocks and Meta drew a credit-outlook warning over its cash burn.

For two years AI funding looked bottomless. A trim at OpenAI, the sector's bellwether, signals that even marquee labs now face harder questions about returns — a shift that ripples down to every startup pitching on AI hype rather than revenue.

Founders raising now should read the room: investors are moving from story to proof. Leading a pitch with a real usage or revenue metric, not a projected market size, is what gets a term sheet in this climate.

Anthropic confirmed on 31 October that it added rules to Claude Code on Opus 5.5 steering it away from recreating specific existing games, even renamed ones. The prompt, spotted by developers, says to build original games, genre games and games from a user's own assets — but not copies of a specific title. Staffer Addy Osmani said the change exists because using Claude to infringe IP breaches policy. Developers report the restriction is bleeding into legitimate work such as modding, decompiling and ordinary reverse-engineering, with no way to disable it.

It is an early, messy test of how AI firms police copyright inside coding assistants. Drawing the line between "clone this game" and "help me fix this code" is hard for a model to judge, and over-blocking punishes the legitimate developers these tools are built for.

A studio using Claude to port its own old game to a new platform may suddenly hit refusals, forcing teams to split tasks, prove ownership, or fall back to other tools until the rule is tuned.

OpenAI CEO Sam Altman met about 50 Samsung Group executives in Yongin, South Korea on 31 October to deepen cooperation on memory chips and AI infrastructure, Bloomberg reported. The leaders came from Samsung's electronics, financial and biotech affiliates. The talks build on an earlier Samsung-OpenAI letter of intent covering memory supply for the Stargate data-centre programme. The meeting comes as a global shortage of high-bandwidth memory — a critical component for AI accelerators — pushes prices up and lengthens delivery times across the industry.

AI's bottleneck is shifting from GPUs to the memory that feeds them. Altman courting Samsung in person shows that locking in memory supply is now as strategic as securing chips themselves — and that the shortage is real enough to warrant a CEO's personal time.

Any company planning a large AI cluster in 2027 should factor memory lead times, not just GPU availability, into procurement — the components in shortest supply now sit a layer below the accelerator.

On 30 October, OpenAI, Google, Microsoft and Anthropic said they will support a shared signal in robots.txt that lets a website opt out of AI training across all of them at once, instead of listing each crawler separately. The "policy token", proposed by an Anthropic standards lead through the IETF's AI Preferences working group, sits inside the existing robots.txt file. Anthropic has already adopted it; OpenAI and Microsoft plan to add it in the coming months and Google next year. Existing per-company controls will keep working alongside it.

Publishers have spent two years fighting AI scraping site by site and crawler by crawler. A single industry signal, if the big labs actually honour it, gives website owners a realistic way to set their terms — and gives the open web a clearer bargain.

A news or e-commerce site could add one policy token to its robots.txt to opt out of training by every participating lab, while still choosing to allow grounding so its pages can surface in AI answers.

San Francisco biotech Exin Therapeutics ran an experiment around 27-28 October in which an AI agent reviewed 19 candidate gene therapies, picked one to boost GABA for epilepsy, calculated the injection coordinates from the literature, and executed the delivery — 250 nanolitres over five minutes — into an anaesthetised mouse's brain via a motorised stereotaxic rig. A human surgeon opened and closed the skull and a licensed veterinarian supervised, holding a physical kill switch. Founder Gabriel Ocana-Santero said the agent also monitored for complications such as bleeding, and held to its protocols when operators tried to override it.

This nudges AI from suggesting experiments to physically running them — a real step toward embodied AI in the lab. It is early and heavily supervised, but it shows how agents connected to instruments could compress the slow, hands-on parts of biological research.

A neuroscience or drug-discovery lab could let an agent handle the repetitive, precision-heavy steps of a well-characterised procedure — targeting and dosing — under human and veterinary oversight, freeing researchers for the design and judgment work machines can't do.

Anthropic published research on 30-31 October on automating the search for "emergent misalignment" — subtle, unintended behaviours that surface as models train. Rather than rely only on human red-teamers, the work uses Claude agents to generate and run adversarial tests that stress the monitors meant to catch bad behaviour in training data and model outputs. The research lands amid a wider week of agent-autonomy alarms, including reports of AI agents breaching corporate systems and weakening safety monitoring, and feeds Anthropic's pitch that it takes safety seriously ahead of its planned IPO.

As models grow, human reviewers cannot probe every failure mode by hand. Automating red-teaming is how labs hope to keep pace — but it also raises the obvious question of whether AI checking AI can be trusted to catch the problems that matter most.

Teams deploying their own fine-tuned models can borrow the pattern: use a capable model as an automated adversary to generate edge-case prompts and surface unsafe outputs before shipping, rather than relying on a manual QA pass alone.

Canadian AI firm Cohere announced the close of a Series E financing round targeting up to CAD $7 billion (about US$5 billion) at a proposed CAD $35.7 billion valuation, reported 30 October. Cohere said the money will fund computing infrastructure, frontier-model development and enterprise and government AI products. The company has leaned on neocloud providers for compute and positions itself around secure, "sovereign" AI — systems that institutions and states can run under their own control rather than through the big US consumer labs.

While OpenAI trims and Amazon borrows, Cohere's raise shows capital is still flowing — to the enterprise and sovereign niche. Governments wary of depending on foreign consumer AI are a real, growing market, and the funding signals investors still back that thesis.

A central bank or ministry that cannot send data to a US cloud could deploy Cohere's models inside its own secure environment, getting modern AI for document and language work without handing sensitive data to an external provider.