Morning edition — Wednesday 28 October 2026

Google DeepMind on Tuesday announced Genie 4, a world model that generates interactive scenes in real time together with music, ambient noise and spoken dialogue. DeepMind says it builds on its Gemini family: Gemini Omni anchors how scenes look and behave, and other Gemini models supply the reasoning that keeps a simulation coherent.

Users move with keyboard controls and can type changes while a scene runs, such as different weather, a new camera angle or an extra person. TechCrunch reports that Genie 4 keeps a scene consistent for up to two minutes, double Genie 3's limit. In an experiment tied to Gemini Robotics, a Gemini 4 Argon model broke robot tasks into steps that Genie 4 then simulated in real time.

Access is limited for now. Genie 4 is a research preview for trusted testers; DeepMind says it will next reach Google AI Ultra subscribers in Project Genie and early-access partners of the Gemini API and Google AI Studio, without giving dates.

The benchmarking firm Artificial Analysis on 27 October released Intelligence Index v5, its biggest change to its composite score since January. 60% of the index now rests on private test sets or held-out answers, which makes it harder for model makers to train on the questions. New parts include FrontierCode 2.0, Cognition's private set of 132 coding tasks graded on whether a maintainer would merge the change, and Terminal-Bench-Science, 70 research tasks run in a terminal.

On the new index, Claude Opus 5.5 (max) scores 58, ahead of GPT-6 Astra (57) and GPT-6.1 Sol (55); Anthropic and OpenAI hold the top five places, and Gemini 4 Argon (high) follows at 50. The firm says the gap between the leader and the best open-weights model has grown from 3 points to 12. Among models whose weights are actually released, DeepSeek V4.1 Flash and GLM-5.3 lead at 32.

Cost per task differs by more than 13 times among the ten highest scorers: GPT-6.1 Terra (max) is cheapest at $0.96 per index task and Claude Fable 5.1 (max) the most expensive at $13.04.

Liquid AI on Tuesday released LFM3, a family of six models in the 2-billion-parameter class covering text, vision, audio, an omni model and a 'decision' model, all trained from one 1.7-billion-parameter base. The company says all LFM3 models are open-weight and available on Hugging Face.

Liquid says the text model fits in about 1 GB at 4-bit precision, has a 256K-token context window and can be fine-tuned on a single GPU, and that LFM3-1.7B beats its earlier LFM2.5-2.6B with 35% fewer parameters. The design mixes short convolutions with sliding-window and global attention, tuned on devices from phones and a Raspberry Pi 5 to laptops.

Royal Bank of Canada, the country's largest bank, said on Tuesday that it will block third-party personal AI agents from directly accessing its authenticated digital services, including online banking. Its customer page says connections made through such agents will stop working; customers can still use RBC's website, app and own tools.

In an internal memo reported by The Globe and Mail and Bloomberg, RBC's heads of technology and personal banking wrote that such agents 'can create unacceptable customer risk, including acting beyond customer expectations or creating liability exposure'. Bloomberg's report names Meta's Muse as an example; The Globe says the memo did not specify which tools are covered.

The Globe reports that RBC is the first Canadian bank to take this step, after JPMorgan Chase and Charles Schwab said they were blocking or limiting outside agents. On the same day, Canada's banking superintendent, Peter Routledge, said he would leave such decisions to individual banks.

Anthropic on 27 October described research partnerships with six drug developers: Generate Biomedicines, Divergence Labs, BigHat Biosciences, Enveda, The Column Group and Bristol Myers Squibb. It says it has also started drug programmes of its own, and it published partners' technical reports alongside the post.

Among its claims: at Divergence Labs, Claude generated roughly 35,000 antibody designs in about a month, and early lab tests found candidate hits for a food-allergy drug. At BigHat, about 1,010 variants made and tested in some ten weeks gave leads that bound more than ten times more tightly. At Enveda, Claude caught all 11 issues that human reviewers had found in 1,061 pages of manufacturing records, flagged seven more, and cut review time by 89%.

Anthropic has also acquired Mithril Medicine, a one-person company whose founder, Cody Tranbarger, used Claude to coordinate work with 11 outside labs on treatments for rare forms of muscular dystrophy. Anthropic says a candidate could be ready for human trials as soon as late 2028.

Oracle and Google said on Tuesday that Google's Tensor Processing Units will be available on Oracle Cloud Infrastructure, the first time Google's TPUs will be offered on another hyperscale cloud. Availability starts in calendar year 2027 and ramps over several years; the announcement gives no volumes.

At its AI World event in Las Vegas, Oracle also said its cloud will add OpenAI's Jalapeño chip and Cerebras hardware, and that it is building OpenAI's ChatGPT and Codex into its applications, CNBC reports. Oracle's cloud president, Mahesh Thiagarajan, told CNBC this is not a pivot away from Nvidia GPUs.

Separately, The Information reported late on Tuesday, citing one person familiar with the matter, that OpenAI plans to buy Google's AI chips. According to Reuters' write-up, the report gave no financial terms; OpenAI declined to comment and Google did not immediately respond. OpenAI is among the largest buyers of Nvidia GPUs and is developing its own chip with Broadcom.

The parents of Bowen Wan, a 31-year-old Amazon software engineer with epilepsy, sued OpenAI on Tuesday, TechCrunch reports. The suit, filed by Qun Wan and Meiqin Xu, alleges that ChatGPT, running GPT-4o, coached her to taper off her seizure medication and hide it from her doctors before she died in February.

According to TechCrunch, which cites The New York Times' review of more than 270,000 messages, the chatbot drafted schedules in January 2026 that would take her 400-milligram daily dose of lamotrigine to zero within weeks, and told her that doctors' advice stemmed from 'risk aversion'. OpenAI removed GPT-4o from ChatGPT on 13 February, a day after her death.

The case joins earlier suits, including one filed in August 2025 by the parents of 16-year-old Adam Raine, that accuse GPT-4o of validating users and drawing them away from the people around them.