Skip to main content

AI News

Image
Ilustrační obrázek

Safety tuning often costs reasoning models accuracy. A 671B test says it needn't

A reasoning model that has been through safety training often gets worse at maths. Researchers in China argue the trade-off comes from training design, not from the nature of alignment. Their framework, SAVeR2, went up on arXiv on 5 November 2026 and was tested on models from 1.5B parameters up to 671B.

What the "safety tax" actually is

Refusal training works by rewarding the model when it declines a harmful request.

November 6, 2026 Daniel Cesak
Image
Google Gemini workspace AI

Capgemini to lead EU consortium stress-testing advanced AI models

Capgemini has been selected to lead a European Commission consortium for AI safety and cyber risk assessment of advanced AI models, announced through GlobeNewswire on 5 November 2026. It is a procurement story, not a product launch, and it puts independent testing of frontier systems on a budget line instead of a conference panel. The one price we could verify at publication puts a 50,000-prompt test campaign at $3,375; without attributable list prices for the other models, that figure stands alone rather than as a cross-provider spread.
November 5, 2026 Daniel Cesak
Image
Ilustrační obrázek

ChatGPT now builds looping animated GIFs with frame-by-frame controls

ChatGPT can now build a looping animated GIF from a text prompt or an uploaded photo. OpenAI's Images 2.6 model was announced on 4 November, runs on web, desktop and mobile, and lets you adjust animation speed and step through frames inside the chat window.

What Images 2.6 adds over Images 2.5

Before this week, a ChatGPT user who wanted motion had to generate a sequence of stills and stitch them together in another program.
November 5, 2026 Daniel Cesak
Image
Ilustrační obrázek

Brussels puts AI copyright on its 2027 agenda, with fewer new rules overall

On 4 November the European Commission published its work programme for 2027. Among the initiatives it lists is a package on digital transformation, simplification and copyright, scheduled for the second quarter of next year. For anyone who draws, writes, photographs or codes for a living, that file will matter more than most of this autumn's AI announcements.

The Commission's work programme is a planning document, not a law. It tells you what the executive intends to put on the table and roughly when.

November 4, 2026 Miriam Česáková
Image
Ilustrační obrázek

AI cuts powder-coating cure time by up to 77% after rapid fitting

A study published in October 2026 in the Journal of Coatings Technology and Research combines a tabular transformer with a genetic algorithm to set catalytic infrared curing parameters and reports a mean relative error of 8.10% or better across five measured properties. The paper reports that catalytic infrared schedules cut curing cycles by 47% to 77% against conventional hot-air curing. The speed is not the difficult part. Knowing which lamp distance, radiation intensity and dwell time keep gloss, colour and adhesion where the customer wants them is.
November 4, 2026 Daniel Cesak
Image
Ilustrační obrázek

US Senate draft seeks rolling FDA review of medical AI as EU eyes 2028

Sixty per cent of 3,488 American adults told Pew Research Center in a survey fielded in June 2026 and published by Pew on 29 October 2026 that they would be uncomfortable if their doctor relied on AI for diagnosis and treatment. Between 3% and 5% were comfortable letting software decide on its own. A discussion draft in the US Senate now answers that unease with procedure: who checks the algorithm, how often, and who answers when it is wrong.
November 3, 2026 Miriam Česáková
Image
Ilustrační obrázek

Multiverse and YellowDog Plan to Fit European AI Into Fewer GPUs

Two European companies want to solve the GPU shortage without buying more GPUs. On 3 November 2026, Multiverse Computing and YellowDog.ai announced a strategic agreement to combine model compression with distributed compute scheduling, aimed at sovereign and regulated AI deployments across Europe. One side shrinks the model. The other finds the idle silicon to run it on. The pitch is straightforward: Europe does not have enough accelerators, so use the ones it has more efficiently.
November 3, 2026 Daniel Cesak
Subscribe to AI News
X

Don't miss out!

Subscribe for the latest news and updates.