Skip to main content

AI News

Image
Ilustrační obrázek

Multiverse Computing replaced 4.9 billion LLM parameters with a quantum circuit

Multiverse Computing and the University of Toronto took a 30-billion-parameter language model, cut out eight of its transformer blocks and put a single multi-head quantum neural network in their place. That circuit carries 145,281 trainable parameters and its quantum heads ran on IBM's 156-qubit Heron processor. The measured result was performance within roughly 1% of the original model's perplexity on WikiText-2 and C4. Throughput was not reported in the paper.
November 15, 2026 Daniel Cesak
Image
Ilustrační obrázek

Four Nonprofits Ask a California Court to Freeze Frontier AI Training

In August, 69-year-old Wynd Kaufmyn became the first anti-AI protester jailed in San Francisco for blocking OpenAI's headquarters. Three months later the campaign has moved indoors. A 63-page complaint filed in San Francisco Superior Court asks a judge to order OpenAI, Anthropic, Google, Meta and SpaceX Super Intelligence (SpaceXSI) to stop training and releasing advanced AI models.

Four nonprofit organisations brought the case: Public Citizen, The Heat Initiative, the Irreplaceable Education Fund and the Alliance for Secure AI.

November 14, 2026 Miriam Česáková
Image
Ilustrační obrázek

ChatGPT Wallet launches for US Pro users as automated shopping tasks falter

OpenAI switched on ChatGPT Wallet on November 13. US Pro subscribers can now let the company's "dots" agents find a product and pay for it, with a fingerprint or Face ID check before the charge runs. The card number never reaches the merchant or the model. Europe gets none of it yet, and OpenAI has not said when that changes.

What the wallet does

The rollout was reported by The Verge on launch day.

November 14, 2026 Daniel Cesak
Image
Ilustrační obrázek

A 27.6-Point Drop Shows What Multimodal Training Does to Text Reasoning

Remove one direction from the hidden states of Chameleon-7B and accuracy on the BIG-Bench Hard navigate task falls by 27.6 percentage points. Boolean expression accuracy rises 5.6 points under the same intervention. A paper posted to arXiv this month argues that the missing direction exists because the model was trained on images as well as text, even though the task it fails is pure text.
November 13, 2026 Daniel Cesak
Image
Ilustrační obrázek

MIT's design copilot lets engineers edit a structure mid-generation

Topology optimization usually runs as a closed box. You enter loads and constraints, the solver iterates, and a shape comes out. Sometimes the shape is efficient and impossible to build: a node where eight beams meet, a joint nobody can weld. A team at MIT has published a variant that stops and asks. The engineer marks a region, requests more or less material, and generation continues from there. In the paper's sample tests the interactive runs reached a 39% higher linear buckling load, and each feedback pass cost roughly 15 seconds of extra compute.
November 13, 2026 Daniel Cesak
Image
Ilustrační obrázek

OpenAI drops older voice modes from ChatGPT on 11 December. Hour caps stay

OpenAI has given ChatGPT users 31 days to get used to a single voice mode. Advanced Voice and Standard Voice leave the app on 11 December 2026, and every account falls back to GPT-Live, the full-duplex system that arrived in July. Transcripts stay. Audio clips last 30 days. The hourly caps stay as well.

What actually retires on 11 December

Two modes go at once. Standard Voice is the old turn-taking system: you speak, it waits, it answers.

November 13, 2026 Daniel Cesak
Subscribe to AI News
X

Don't miss out!

Subscribe for the latest news and updates.