Google unveils a real-time world simulator, Delaware's AI-run company plan draws fire
Google shows off a world simulation model, a US lawmaker challenges Delaware's plan to let AI agents run companies, and new Spanish survey data shows AI use is still rare among smaller firms.
Every line links to a primary source in the full briefing.
01Google DeepMind previews Genie 4, a real-time interactive world model
Google DeepMind introduced Genie 4, a model that generates interactive simulated worlds in real time, complete with sound and responsive characters, built on top of its Gemini models. It is rolling out as a research preview to trusted testers first, with wider access planned later for subscribers and developers.
This is a research tool for now, not something a business can use today, but it signals where big labs are investing: synthetic environments for training robots and software agents rather than consumer chat products.
Source: Google DeepMind
02Lawmaker investigates Delaware plan to let AI agents legally run companies
A US House committee chairman sent Delaware's governor a letter challenging a proposed new corporate structure that would let AI agents operate a business with no human in charge. He argues the structure could let company owners and investors escape liability for reckless or harmful decisions made by the AI, and is requesting documents on how the plan was developed.
Any owner considering heavier use of autonomous agents should watch this closely, since it previews the liability questions regulators will eventually force onto every business that lets software make real decisions.
Source: Ars Technica
03Spanish survey finds AI use still rare among small and mid-sized firms
New Spanish coverage reports that only 4.1% of the small and mid-sized companies analyzed show identifiable AI use, even as adoption headlines accumulate elsewhere. The figure points to a gap between public enthusiasm and actual deployment on the ground.
A business owner reading constant AI news can take some comfort, and some warning, from the fact that most peers have not actually implemented anything yet, so being an early, careful adopter is still a real advantage.
Source: Revista Cloud Computing
04Liquid AI releases LFM3, small models built to run on local hardware
Liquid AI launched LFM3, a family of compact models (around 2 billion parameters) covering text, vision, audio and decision tasks on a shared architecture. The company says the models can be fine-tuned on a single GPU and deployed on devices ranging from phones to laptops, with faster performance on longer inputs than its previous generation.
Smaller, cheaper models that run on ordinary hardware lower the cost of building a narrow, specific AI tool in-house instead of paying for a large general-purpose service.
Source: Hugging Face
05Study questions whether AI can reliably judge if a research idea is new
Researchers published a benchmark testing how consistently large language models judge whether a scientific idea is genuinely novel versus a repackaging of existing work. The study found that choices like which model, prompt or evaluation method is used can materially change the verdict, meaning there is no standard, trustworthy way to do this yet.
This matters more for research and product teams than most SMEs, but it is a caution against trusting an AI tool's confident judgment on originality, whether for research, branding or patent-adjacent claims, without a human check.
Source: Hugging Face
06Security researchers report a sandbox escape flaw in DeepSeek's coding harness
A newly disclosed vulnerability, tracked as CVE-2026-82533, reportedly allows a one-command escape from the sandboxed environment used by DeepSeek's coding harness. Details come from a single security outlet so far and have not been independently corroborated.
Any business running AI coding tools in production should treat this as a reminder to patch and sandbox properly rather than trust default isolation, especially with open-source or lower-cost model providers moving fast.
Source: OrcaRouter
Check who is actually accountable before handing an AI agent a real decision
With Delaware's proposal to let AI agents legally run companies drawing federal scrutiny over liability, this is a good moment to write down, in plain terms, who signs off on what an AI agent does in your business and who is responsible if it goes wrong. It costs nothing and takes under an hour, but it closes a gap regulators are actively arguing about right now.
Sources
6 links- 01La inteligencia artificial avanza en España, pero solo el 4,1 % de las pymes analizadas muestra un uso identificableRevista Cloud Computing
- 02Build specialized multimodal models with LFM3Hugging Face
- 03Can we trust LLMs to judge whether a research idea is novel?Hugging Face
- 04Genie 4: A new generation of world simulationGoogle DeepMind
- 05"No human at the helm": Delaware plan would let AI agents run entire businessesArs Technica
- 06DeepSeek Harness CVE-2026-82533: one-command sandbox escapeOrcaRouter