An ECG, an interface and a tenth of the price

Researchers at Imperial College London present a model that reads a routine electrocardiogram in under two seconds and spots heart failure and valve disease that doctors cannot see by eye. Trained on 10.6 million ECGs and tested on 65,000 patients, it flags 81% of heart failure cases and 90% of valve disease.

Confirming those conditions today means an ultrasound, with queues running months — while hospitals perform more than a billion ECGs a year. The team is starting a trial on 590 patients across six hospitals, aiming for routine NHS use within two years. Professor Fu Siong Ng envisages running the model over every ECG, including those taken for other reasons.

↑ Back to the summary

Z.ai unveils GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series: 320 billion parameters, 18 of them active at inference, pre-training on a 30-trillion-token multimodal corpus, and a hybrid architecture combining sparse and linear attention. The company claims better scores than GLM-5.2 across all its benchmarks, at a tenth of the cost.

Before release the model was tested anonymously as “ox-alpha” on OpenCode and OpenRouter, where it became the most used model of the week — traffic served, Z.ai notes, on Chinese AI chips. The weights are published on Hugging Face. Z.ai has been listed in Hong Kong since January 2026, at a $62 billion market capitalisation in August.

↑ Back to the summary

Runway opens early access to Solaris, which it calls an “Interface World Model”: websites and apps rendered as live video, every frame drawn in real time as the user clicks and drags. No code runs underneath.

Runway's argument is not only visual, it is about training agents: today's models learn the exact layout they were trained on and stumble the moment it changes from one site to the next. Solaris gives them interfaces that change constantly. A limit the company states plainly: stable, legible text remains one of the hardest problems in video generation — it floats a hybrid system, where an image model renders text-heavy views.

↑ Back to the summary

Meta's terminal agent for macOS and Linux, backed by the hosted Muse Spark 1.2 model, inspects a repository, proposes and edits code, runs tools, validates results and coordinates persistent background agents. There is no per-seat fee: usage is billed by the token through Meta's API.

Two tiers, and the difference is not only the price. On Standard — $1.25 per million input tokens, $0.15 cached, $4.25 output — the data is labelled as not used to improve Meta products. On Contributor, far cheaper ($0.10, $0.002 and $0.20), submitted data may be. Rates checked on 31 August 2026: verify before sending code.

↑ Back to the summary

The Department of War has launched Grok for Government, from Starshield AI, on GenAI.mil, its generative AI platform. The service is accredited for Controlled Unclassified Information at Impact Level 5.

Nine months after launch, GenAI.mil claims more than 1.7 million users out of the department's roughly 3 million personnel. The release highlights adaptive reasoning modes, customisable workspaces and reusable “playbooks” — and frames the addition as a way of “eliminating vendor lock”.

↑ Back to the summary