hotai.news
中文
← Home01 / AI News

AI News.

The latest AI news, ordered by when it was collected. Original publication times and summaries are included when available.

Latest news is awaiting an updateWaiting for the latest collected data
Time range
Browse news by date
Latest NewsAwaiting update
01

Vals AI research: The political stance of Chinese AI models is more complex than expected and not merely that of government mouthpieces | AI News | HotAI.news

Vals AI tested Chinese and foreign models with hundreds of questions involving China, discovering that the 'pro-government tendency' of Chinese models was generally lower than expected."Pro-government leaning"Overall performance fell short of expectations, with significant variations across models. GLM-5.3 Flash even explicitly acknowledged that evidence exists regarding China's human rights issues. Prompt language had a marked impact: asking in Simplified Chinese was more pro-government than using English, Russian, or Spanish; this trend held true for both US and Chinese models, though it was more pronounced in the latter.

↗02

Terence Tao criticizes AI methods for solving math problems: The issue lies not in proof correctness.

Rohan Paul shares Terence Tao's video perspective: Mathematicians' dissatisfaction with AI solving math problems is unrelated to the correctness of the proofs. Tao notes that while activity surrounding the Navier-Stokes equations is currently chaotic and will eventually recover to understanding the proof, this approach is highly undesirable. He calls this pattern irresponsible automated open-pit mining of problems—achieving immediate goals at the expense of the field's future development. The full video is on the New Scientist YouTube channel.

↗03

OpenAI imposes payment card and device limits on ChatGPT paid subscriptions

OpenAI has added payment card and device limits to ChatGPT paid subscriptions. Once a single card or device is bound to too many subscriptions, new ones cannot be created; when the limit is reached, the message 'Device limit reached' appears.

↗04

Google launches StreamVibe video agent to let AI decide when to speak up and help.

Google, together with teams from Stanford and Oxford, launched the StreamVibe video agent. It can distinguish between what users say explicitly, what they haven't said yet, and changes in tasks mid-way, autonomously deciding when to speak up. Its StreamVibe-VL model is based on Qwen3.5-9B, with response timing and quality improved by 10.7% compared to other models, and its memory system improving active interaction by another 12.3%. Tests show that after videos exceed five minutes, all models struggle to grasp the right timing, with key content metrics based on their own historical answers dropping as much as 61.5%.

↗05

CADI Research Institute report: China's AI inference data volume surpasses training data for the first time, with average monthly token calls exceeding 175 trillion in June.

A report released yesterday by the CADI Research Institute titled 'Top Ten Trends in Innovation and Development of New Domains and New Quality' shows that China's artificial intelligence inference data volume has exceeded its training data volume for the first time, signaling a shift from model training and R&D to scaled industrial application. As of June 2026, average monthly token calls surpassed 175 trillion, representing an increase of over 1,000 times compared to early 2024. Domestic computing power and large models have been deployed in over 1,000 application scenarios across 20 industries including steel and automotive, with the cumulative construction of more than 500 excellent-level smart factories.

↗06

The Ministry of Industry and Information Technology's Academy for the Application of New Technologies (CAICT) reports that China's AI inference data volume has surpassed its training data volume for the first time. As of June, the average daily token calls exceeded 175 trillion.

According to today's report by CCTV, the China Electronics Industry Institute (CCID) released 'Top Ten Trends in Innovation of New Domains and New Quality Productive Forces' yesterday. The data shows that China's AI inference volume has exceeded its training volume for the first time, marking a shift from the model training and R&D phase to the stage of large-scale industrial application.

↗07

ChatGPT to merge Chat and Work features, sharing a single token budget starting in December

According to OpenAI's updated support page, the Plus, Pro, and Business plans will share a unified token budget between ChatGPT and Codex starting in December. Work's agent features are being integrated into the main interface.

↗08

David Sacks reports that Anthropic disclosed in August that AI has taken on 26% of the research tasks for next-generation models.

Rohan Paul shared David Sacks' statement from the All-In Podcast. In a blog post released by Anthropic in August, it was revealed that 26% of tasks related to building next-generation models have already been completed by RSI (Recursive Self-Improvement). Sacks added that this figure reflects only August data, and progress for October remains unknown.

↗09

What is AI bringing to math education? Wolf Prize winner promises to end boring teaching methods

Could machines do mathematics? This once-philosophical question has sparked global mathematical debate and deep discussion in 2026. 'I hope that as artificial intelligence becomes increasingly adept at handling tedious and dull tasks, we can...'}, {

↗10

Ali Cloud Qwen AI Glasses OTA Update: Adds Call Translation, Media Translation, and More

The latest update introduces cross-border calling, overseas podcasts, foreign language online courses, and real-time translation of major films. Noise-canceling voice recognition has been optimized for noisy environments, while simultaneous interpretation is now more accurate and timely. Translations can be displayed in real time on the Qwen AI Wearable App.

↗