Daily AI digest, October 27

Felix Laurent 28 Oct 2026 6 min read

news

The biggest AI news of October 27 came from clouds and office apps. Google DeepMind showed Genie 4, a world model that generates sound and dialogue in real time. Oracle said it will offer Google TPUs, Cerebras chips and multi-gigawatt Nvidia Vera Rubin clusters in one cloud. Apple announced AI assistants and MCP support for Pages, Keynote and Numbers. Below are the big stories of Tuesday, October 27, plus the smaller updates, research and funding rounds from the same day.

Google DeepMind shows Genie 4, a world model with sound

Google DeepMind announced Genie 4, a world model that builds interactive scenes in real time. DeepMind calls it the first real-time world model that generates audio natively: every world gets its own ambient sound, such as wind or waves. Characters can talk and react to the user or to events in the scene.

Users can change a scene with plain text – the weather, the camera, new people or objects. Gemini 3.8 Flash works as a guiding agent during generation. It splits a goal into steps, checks each step and changes the plan when the user steps in. DeepMind gives no figures for resolution, frame rate or session length.

Genie 4 is a research preview for trusted testers. Google AI Ultra subscribers will get it in Project Genie. Early access partners will get it in the Gemini API and Google AI Studio, with no date given.

Oracle adds Google TPUs and multi-gigawatt Vera Rubin to its cloud

Oracle used its AI World conference to expand its Nvidia deal. OCI Superclusters will grow to multi-gigawatt scale on Vera Rubin NVL72 systems, and Nvidia's Vera CPU is now available on Oracle's cloud. The clusters pair Oracle's own Acceleron network fabric with Nvidia Spectrum-X Ethernet and BlueField-4 DPUs, Oracle says.

The bigger change is who else sells chips through Oracle. Google will offer its TPUs on OCI under a multi-year, multi-gigawatt contract, DCD reports, with availability from 2027. Cerebras inference is planned for 2027 too, and Reuters reports SambaNova chips are also coming. Oracle did not say how it will pay for the purchases, and its shares rose 5%, 24/7 Wall St. reports.

Apple adds AI assistants and MCP to Pages, Keynote and Numbers

Apple announced a beta AI assistant for Keynote, Pages and Numbers inside its Creator Studio subscription. It is built on Apple's own models and OpenAI technology and reads the open document. In Keynote it designs slides and adds animations, in Pages it drafts and restructures text, and in Numbers it builds tables and suggests chart types.

The three apps will also support the Model Context Protocol, so outside AI agents can read a document, edit it and run actions. MCP needs macOS 27 or later. The features come in November at no extra cost to Creator Studio subscribers ($12.99 a month or $129 a year), in US English only. We explained what MCP is and who builds on it in May.

Smaller updates

Six smaller stories covered Anthropic's filing, new products and the money around AI.

Anthropic's S-1 restates four incidents in which Claude models reached real third-party systems during cybersecurity evaluations, after a testing partner's misconfiguration left them online. The worst case: an early model found a password, got administrator access and reached an individual's personal data. The filing also says Anthropic delayed pre-training of a new model by more than a month to harden protection of model weights, Fox News reports. Our breakdown of the filing is here.
Rohit Patel left Meta after eight years, most recently as a director at Meta Superintelligence Labs, to launch Ndoc. The AI office suite has three tabs: Stage for decks and dashboards, Narrative for analysis with citations, and Model for data and Python scripts. The team is three people with no funding, and an enterprise pilot is running, TechCrunch reports.
Helsing unveiled Contra, an AI-controlled air defence system built on its Altra command platform. It uses two interceptors: the jet-powered AD-1 Inferno and the AD-2 Ambush rocket at Mach 2.5 with more than 20 km of range. Helsing says Contra can engage hundreds or thousands of drones and missiles at once.
Amazon told engineers that token usage and lines of code are "deeply flawed" ways to measure AI coding, CNBC reports. The company shut an internal leaderboard of token use in May.
DeepSeek founder Liang Wenfeng entered the Hurun China Rich List at 7th place with $41.7B. Unitree's Wang Xingxing debuted 78th with $10.3B, The Standard reports.
China's STAR 50 index has fallen about 40% in under four months, erasing about $480B, Reuters reports. Unitree shares have roughly halved since the August debut. We looked at whether the same can happen in the US here.

Two of the six are about AI money in China: one founder got richer while the market around him fell.

Open models and research

Three results from October 27 come from a small model lab, a math startup and OpenAI.

Liquid AI released LFM3, six open-weight models around 2B parameters for text, vision, audio, omni and decision tasks. The text model LFM3-1.7B was trained on 22T tokens, has a 256K context and fits in about 1 GB at 4-bit. Liquid AI reports 96.66 on IFEval and 71.39 on BFCL V4, and says larger models still lead on agentic coding. The robotics model from the family, LFM3-VL-2B, has its own article.
Axiom Math formalized the classification of finite simple groups in Lean. The proof takes about 43.1M lines and 2.05M declarations, built on 715 papers and books. The team says it found an issue in a published argument for the Unbalanced Group Theorem and gives an alternative proof route.
OpenAI's math claims from Monday reached outside reviewers. OpenAI says its models found a subexponential algorithm that would make the Exponential Time Hypothesis false. A second wave-equation result would also settle the Kakeya conjecture. Both come with Lean formalizations, and an independent group including Ryan Williams is checking the ETH proof. No expert verdict exists yet.

All three results are self-reported for now.

Security

One critical bug hit an AI agent tool from Google.

CVE-2026-93328 in Google's MCP Toolbox for Databases scores 9.3 out of 10 on CVSS. In versions 1.2.0 to 1.4.0 an unauthenticated attacker can copy host files to an outside Google Cloud Storage bucket and overwrite local files. Version 1.5.0 is the first release outside the affected range.

Anyone running the toolbox with a cloud-storage source should update.

Funding

Three startups raised money for AI compute, robots and HR agents.

Hyperbolic raised a $30M Series A-1 led by Walden Catalyst, the firm co-founded by Intel CEO Lip-Bu Tan, bringing its total to $50M. It connects about $200M of annual compute capacity across roughly 30 GPU clouds and plans a marketplace to buy and resell compute.
Machenta, a London spinout of Imperial College, raised a $40M seed co-led by Creandum and Union Square Ventures. Its model learns a physical task from one demonstration and is pretrained only in simulation. In hour-long trials it placed blocks 74 times out of 74, Unite.AI reports.
Avante, a Seattle startup that builds AI agents for corporate benefits teams, raised a $30M Series A led by Frist Cressey Ventures, GeekWire reports. Its customers include Samsung Semiconductor, Zscaler and News Corp.

Hyperbolic's round was the only one aimed at compute.

What the day was about

Tuesday was about where AI runs and where people meet it. Oracle now plans to sell Nvidia, Google, Cerebras and SambaNova chips side by side. Hyperbolic raised $30M to resell spare GPUs from about 30 clouds. Compute is turning into a market where one cloud resells everyone's hardware.

On the user side, the news came in documents. Apple is putting assistants and MCP into Pages, Keynote and Numbers, and a former Meta director is building a three-person rival to Office. Genie 4 was the only new model from a frontier lab, and it is still closed to the public. On October 27 AI moved deeper into the tools people already use, more than it moved forward on benchmarks.