News

After the recent release of a new version of the Afrikaan Health Application under the umbrella of the Ant Group, downloads have risen rapidly, and today they are in third place on the Apple Applications Master List. The application has reached 15 million live users per month and more than 5 million health questions are answered daily. The QuestMobile 3rd Quarterly report shows that the ants Afrood experienced a high growth rate of 831 TP3T per month, well above the industry average rate of 13.51 TP3T, showing strong growth momentum。

ON DECEMBER 16TH, COMFORT TECHNOLOGY OFFICIALLY LAUNCHED THE AI OFFICE SMART BODY, RACCOON 3.0. THE PRODUCT IS POSITIONED TO MOVE FROM "TOOLS" TO "SMART BODIES" AND IS DESIGNED TO DELIVER DIRECTLY AVAILABLE RESULTS THROUGH TASK UNDERSTANDING AND EXECUTION, SUCH AS THE GENERATION OF HIGH-QUALITY PPTS WITH OUTLINES, LAYOUTS, CHARTS AND ILLUSTRATIONS IN ONE KEY. ACCORDING TO OFFICIAL PRESENTATIONS, THE NEW VERSION FOCUSES ON IMPROVING THE UNDERSTANDING AND IMPLEMENTATION CAPACITY OF COMPLEX MISSIONS, SUPPORTING A LINKAGE ANALYSIS OF MILLION-VOLUME DATA AND ALLOWING FOR THE DISMANTLING OF VAGUE INSTRUCTIONS INTO STEP-BY-STEP PLANS. AT THE SAME TIME, THE PRODUCT IS FORMALLY ON-LINE, MOBILE-END APPLICATIONS TO SUPPORT THE MOVEMENT OF TASKS ACROSS DEVICES. TRADE SOUP TECHNOLOGY SAYS RACCOON..

On December 16th, Mi Miomi MiMo-V2-Flash released the Xiaomi Open MoE model, total parameter 309B, active parameter 15B, designed for smart AI, focused on speed. According to official Mi, this is a MoE model that is self-researched for extreme reasoning efficiency 309B (activation of 15B), accelerated by Hybrid Attention Architecture Innovation and Multi-level MTP Logic to maintain access to global open source model Top 2 on multiple Agent benchmarking; code capabilities are..

According to The Register, IBM researchers have recently launched the Open-source AI Smart CUGA (Configurable Generalist Act) to automate complex business missions through multi-intelligence synergetic, API integration and code generation. It achieved mission completion rates of 61.71 TP3T and 48.21 TP3T, respectively, in the WebArena and AppWorld baseline tests. CUGA uses a dynamic task decomposition mechanism, decomposes user commands into sub-tasks and refers them to a dedicated agent to support the Langflow low..

On 17 December, OpenAI officially launched today a new image generation model GPT Image 1.5, which is open to free ChatGPT users. Compared to the previous GPT Image 1, GPT Image 1.5 brings with it four core upgrades, including a fourfold increase in the speed at which the instructions are implemented with greater accuracy, more precise editing, more complete detail retention and production. At the same time, users can generate multiple images in parallel, without waiting for the previous one to be completed and with a significant increase in creative efficiency. GPT Image 1.5 is known to follow complex and detailed hints..

According to WIRED, on 16 December, the Chief Communications Officer (Chief Communications Officer) Hannah Wong announced internally that she would leave the company in January 2026. OpenAI spokesman Kayla Wood confirmed this to WIRED. In a joint statement, CEO of OpenAI (CEO) Sam Altman and the Application of CEO Fidji Simo..

It has been learned that the visual design platform Canva can be used for the official launch of the dialogue-based AI al-Ai to the Chinese market at the annual sharing session. The product focuses on “dialogue start-up, talk-over” and streamlines the design process into a natural dialogue. Users need only describe the idea to generate a first draft that can be edited and adjust in real time through oral instructions to lower the operational threshold。

At Wired's conference, AMD CEO Souzif stated that she was not worried about her competitors and that the real pressure came from the speed of innovation. She noted that the AI industry was still in infancy and that there was a significant future growth. Suzito stressed that there was no bubble in the AI market and that bubble-minded people lacked vision. The core mission of AMD is to accelerate the pace of innovation to meet growing market demand and to commit to the next generation of products to be better than at present。

On 16 December, according to the China-China Business City official network, the soybean mobile phone assistant and the collaborator nubiaM153 have now re-opened the F-code alternate application channel and formally reinstated the product. According to the nubiaM153 F-code pre-issuance information-gathering form, only a small number of applications for F-code queues, with no expansion plans, will be made available to those who have not previously purchased the machine. It should be noted that a user who has purchased a product, has not received a F-code billing, cancelled an order or returned goods may not repeat the request, and that the alternate application does not promise to obtain F-codes, which will be followed by a questionnaire..

According to information received on 16 December, XAI has formed an enterprise-level artificial intelligence sales team of over a dozen people. However, the company ' s insufficient experience in marketing to large enterprises is hindering the decision-making process of potential clients. Although XAI has signed up with large clients such as Morgan Stanley and Palantir, these clients are currently conducting small-scale technical tests, each of which generates revenues ranging from hundreds of thousands to millions of dollars. "The Daily Journal of Science"

The publisher of the Wedge Dictionary Editor, Wedge, Inc., recently announced the selection of the English word “slop” as the 2025 hot word and defined it as “low-quality digital content, usually generated in bulk by artificial intelligence (AI). In its website, Wessel publishes that in 2025, a large amount of poor content was flooded on the Internet, including absurd videos, false advertising images, vulgar propaganda, false news, poorly-quality books created by AI and inefficient reports on work-related waste. These elements are both offensive to the public and consumed and disseminated in large numbers. I don't know

On 15 December, OpenAI local time confirmed the appointment of Albert Lee, Director of Google Enterprise Development, as Vice-President of Corporate Enterprise Development. Lee will formally join OpenAI on 16th of local time and report to the Chief Finance Officer, Sarah Friar. "Same flowers for the good."

The multiple AI models cannot correctly count the number of fingers in the face of the six-finger image, even if the prompt clearly states that there are six fingers, and the model maintains five; the root causes of the problem are the strong association of "man = five" in the training data and the lack of a visible structural constraint in the Transformer structure, which transmits untraceable state information one way forward; and the proliferation model, which captures the overall distribution and texture but makes it difficult to accurately control the local fragmentation structure and exposes the current Achilles stasis in the interpretation of visual reasoning and causality。

The former CEO of Google, Eric Schmidt, presented the San Francisco Consensus, arguing that a combination of language intelligence and reasoning capabilities would approach the core human capacity, and that technology integration at a certain stage would result in AI incremental self-improvement; he predicted that AI mathematicians would emerge within the next year as a catalyst for the emergence of a new mathematical theory, a change that industry generally agreed would take place within two to four years, but stressed the need to maintain human ownership and decision-making; and the divide in the U.S.-ACAI competition path: US focus on super-intelligence research and development but insufficient electricity, China was fully committed to promoting AI commercial applications and adequate electricity, with both countries relying on..

Byte beats release the Feedance 1.5 pro-video co-generation model, which supports precision sound-syncing, multilingual dialects, film-class mirrors and 15-second video generation; the model uses the MMDIT framework to achieve accurate synergy of audio-visual flows, with natives supporting dialects such as Chinese, Japanese and Chinese and Sichuanese, and audio commands following industry head level; in the comprehensive assessment of SedVideoBench 1.5, dynamic performance is full, the person's face is moving, and sound-painting errors are significantly reduced and can be applied to visual, advertising, short theatre, etc。

This is the first video model in the country to support role-playing, supporting sound-synthesis, multi-scenes generation and sound-driven, and is the most functional video-generation model in the world; video-generation supports the synchronization of 15 seconds of video, multi-synthesis and natural sound-synthesis, allowing reference to input video-synthesis and sound-synthesis to achieve mono- and multi-person photography; and image-generation upgrades include high aesthetic control in the artistic style, real-person image enhancement, text-generated graphics illustrations, image-mixing output and multi-migration generation。

THE ONLINE WRITING MODE SUPPORTS THE AUTO-COMPENSATION OF A FULL-TIME LOVER BY SETTING AN OUTLINE AND PRODUCING A SINGLE TEXT OF TENS OF THOUSANDS OF WORDS, ENSURING THAT THE CHARACTER OF THE PERSON IS CONSISTENT, CONSISTENT AND ABLE TO REMEMBER PENS; THE FUNCTIONAL MATCHING HISTORY, SCI-FI, AND PEER-TO-PEER MULTI-TOPIC CREATION, WHICH USERS NEED ONLY ONE SENTENCE TO SET UP TO ALLOW AI TO COMPLETE THE STRUCTURE OF THE CHAPTER OF THE OUTLINE AND TO DEFINE THE STORY'S DIRECTION AND OUTCOME; AND THE THREE BEAUTIES CAN GENERATE APPROXIMATELY 30,000 WORDS IN ABOUT 14 MINUTES, WITH 50,000 WORDS WRITTEN IN HALF AN HOUR, AND SUPPORT THE EXPORT OF A LONG COPY TO A LOCAL DOCUMENT OR TO A LOCAL DOCUMENT。

The developer Manthan Gupta reverses the discovery of the ChatGPT memory system using a four-tiered structure: meta-data, user memory, summary of recent conversations and slided windows; the system does not use vector databases or RAG retrieval, but achieves the effect of "recognizing users" by pre-generated lightweight summaries and structured information saved in visible form; the largest context window of GPS-4 is 128k token, which is forgotten beyond the last one, and users can request the model to remove or modify memory content at any time。

The Newmotron 3 series of open models was launched in Weida, covering the sizes of Nano, Super and Ultra, using a groundbreaking hybrid expert structure; Nemotron 3 Nano was four times higher in volume than in the previous generation, leading to the number of tokens per second in a large multi-intelligence system, and significantly more efficient in reasoning; and the model achieved excellence through advanced and enhanced learning techniques and training after large-scale parallel multi-environments, and provided complete training data sets and an enhanced learning base。

Apple is developing the first-named "Baltra" AI server chip, developing network technology in cooperation with Chase, using the 3nm power build-up process, which is expected to deploy in 2027; Apple has abandoned the self-building model to pay approximately $1 billion per year for the 1.2 trillion-parameter Gemini model customized by Google, which is used mainly to meet large AI reasoning needs; and the chip structure will focus on time extension and optimization of throughput, using low-precision calculations such as INT8 and may use 64 chips to connect to high-capacity LPDR memory options。

Original link