News

Yesterday, Corinthian AI officially entered the "time of 3.0" with a comprehensive upgrade around the two core competencies of video and picture, focusing on the upgrading of narrative, controllability, consistency and graphics, and introducing for the first time a multitemporal integrated process of original creation. According to AI, this update covers the full chain from generation to editing and is intended to give creators a lower threshold for a film-grade classification experience. Updates include video 3.0, video 3.0 Omni, picture 3.0, picture 3.0 Omni four main modules. Clink AI indicates that this version of 3.0..

On February 4th, a password management tool, Password, published a book on February 2nd, and its security team discovered that the attackers had used the fire-fire AI smarts OpenClaw (formerly known as Clawdbot and Moltbot) to spread and implant malicious software to MacOS users. The attackers made use of OpenClaw's “Skills” files, which are usually in Markdown format and are used to guide AI on new assignments, but are disguised by hackers as a legitimate integrated curriculum. Watching..

ON 5 FEBRUARY, ACCORDING TO INFORMATION SHARED BY ALI ' S STAFF ON SOCIAL MEDIA, MAYUN APPEARED AT THE SITE OF THE ALI THOUSANDS PROJECT TEAM. ON FEBRUARY 6, THOUSANDS OF PEOPLE ASKED TO LAUNCH A 3 BILLION-BILLION-DOLLAR “KYOTCHEN GUESTS PROGRAMME”, EMPHASIZING “THE NEED TO EAT AND DRINK WITHOUT STOPPING”, AND A LARGE AMOUNT OF RED CASH, BUT NO SPECIFIC FORM OF ACTIVITY HAS BEEN OFFICIALLY DISCLOSED. PRIOR TO THIS, APP HAS BEEN CONNECTED TO ALI'S ECOLOGICAL SCENES, SUCH AS POACHING, PAYMENT, POACHING, FLYING PIGS AND GOTH, TO TEST THE AL SHOPPING FUNCTION. “THE HOPE IS TO BE INTEGRATED INTO PEOPLE'S REAL CONSUMPTION THIS SPRING”. "THE DAILY JOURNAL OF SCIENCE"

At 7:30 a.m. Beijing time, February 5, FF held the first launch of FF ' s Embodied AI robotics at a national car dealer ' s conference in Las Vegas. Before that, FF went from being listed to near bankruptcy, and last April, JACK returned to FF Co-CEO in an attempt to reshape the FF. Jiapon, with a huge debt, has been in the United States for eight years since July 2017。

It has been reported that a strike announcement on third-party inducements to share has been issued by the MS to counter irregularities such as excessive marketing, inducement to share, which is the theme of spring festival. In this context, we have received feedback and complaints from users about the proceeds of the spring festival, which has led to high-frequency sharing of images such as links to micro-messages, such as “doing the task” “red packs”, disrupting the ecological order of the platform, affecting the experience of the user and harassing the user. By way of study, the illicit link to the Yuanbao has been disposed of, limiting its direct opening within the micro-letter. The relevant disposal measures shall enter into force. We believe in public relations

According to the data, Kimi K2.5 from China was the top of the AI smart-body OpenClaw model call, making it the most popular model. According to OpenRouter data, Kimi K2.5 outnumber Gemini 3 Flash, Claude Sonet 4.5 and other top models. Until then, OpenClaw had announced that it would open the amount of Kimi K2.5 to users, and Kimi K2.5 had become an officially recommended dominant model of OpenClaw。

The OpenAI developer ' s official account, OpenAI Developments, released information on 4 February that GPT-5.2 and GPT-5.2-Codex models achieved an overall velocity increase of approximately 40% without changing the model ' s structure and parameter weights. (FIA)

According to information received on 4 February, journalists were informed that two weeks after the publication of the smart-form GLM-4.7-Flash, downloads had exceeded 1 million in Hugging Face. GLM-4.7-Flash is described as a 30B-A3B hybrid thinking model with combined performances exceeding gpt-oss-20b, Qwen3-30B-A3B-Thinking-2507 in the same and approximate size model system..

On 4 February, it was reported that Chief Executive Officer Sam Ottman of OpenAI had indicated that the company might consider investing in companies or providing subsidies for the development of new drugs or treatments using its artificial intelligence technology in large quantities and that it might receive returns through the collection of patents or royalties. According to Otman, artificial intelligence developers can choose to work with pharmaceutical companies to share the cost of using their artificial intelligence models and subsequently obtain a “specific share” from the company's research. He noted that they had not yet had such cooperation. (FIA)

According to informed sources, the artificial intelligence firm Anthropic is moving forward with an equity deal that allows some employees to sell their shares at a minimum of $350 billion; at the same time, the company is conducting a round of fund-raising, which is expected to exceed $20 billion. This offer will allow Anthropic employees to realize part of their equity, and the company is one of the most valued artificially intelligent start-ups in the world. According to the above-mentioned sources, the estimate of $350 billion is consistent with the valuation of the ongoing fundraising negotiations of the company and, for the pre-investment valuation, is not included in the proposed funding. I don't know

Google released PaperSmart, which created a multi-intelligence collaboration framework to generate-respond-optimize, aimed at freeing scientists from time-consuming essay illustrations; a system that involves five roles of searcher, planners, stylist, visualist and commentator working together to improve on key indicators such as simplicity, readability, integrated aesthetics; and limitations of text distortion or wiring errors in dealing with the complex architecture of dense nodes, with future introduction of code proliferation model mapping and human interfaces。

Open-source music generation model ACE-Step 1.5 received first-day support at ComfyUI, using a hybrid LM+DiT structure to produce a full four-minute song on RTX 5090 for only about one second; model support for 50+ language commands, which can be run without less than 4GB visibility, with a music consistency rating of 4.72 above most business models; and support for LoRA to fine-tune to personalize style, which is about to support music re-engineering Cover and clip restoration of Repaint functionality, and to work locally to ensure data security。

Apple releases an Xcode 26.3 version, formally introducing "smart body programming" support, and developers can directly call on AI bodies such as Claude Agent in Anthropic and Codex in OpenAI; integrated AI intelligents can browse through the entire project structure, read-write-editing files, build projects according to instructions, and automatically capture apple official documents to solve problems; user feedback is polarized, with some claiming comfort, while others reflect the existence of card death, poor Diff mechanisms and unstable cross-file re-engineering。

The Skywork desktop version was released by Kunlun Skywork, which does not require the uploading of clouds directly to local missions, allows reading of big local files for collation and generation of new products, and supports multitasking parallels; supports Claude Opus 4.5, Sonnet 4.5 and Gemini 3 Pro models to switch to the inside 100+Skills, which covers the Office three sets, web pages, pictures, video generation, etc.; and gives priority to Windows systems, with better image video generation and all operations to secure data in the local virtual machine isolation environment。

MiniCPM-o 4.5, an all-work capability of only 9B parameters to "look, listen, speak" becomes the first industry "free and instant" large model; models use end-to-end "free dialogue" full-format structures that automatically determine whether or not to speak at a 1Hz frequency, through time-sharing and active interaction mechanisms; sound generation is fully upgraded to support voice cloning and role-playing, long-speech synthesis is stable, with open source Allama.cpp-omni high end..

The site line, known as Rentahuman.ai, is located on "the body level of AI", which can employ humans through the MCP agreement or REST API to perform sub-line tasks such as the delivery of goods, movement sign-ups, field surveys, etc.; the web line, which reaches 20,000 people with less than 48 hours of manpower, allows humans to spend their time and is free from gossip; and the published tasks, which include taking photographs, testing restaurants, collecting packages, etc.; the website triggers a discussion of responsibility, the authenticity of the mission, and the ethics of AI hiring human beings, which is also seen as another reflection of the value of the MCP agreement。

Ali Open-source Qwen3-Coder-Next model, total parameter 80B activated parameter only 3B, achieving over 70% problem resolution on SWE-Bench Verified, a model that is 10-20 times larger; model that supports 256k context with long-term reasoning, use of complex tools and recovery from implementation failures, and seamless integration with multiple IDE platforms such as Cline, Claude Code; simultaneous publication of the SWE-Universe framework, which extends the real world multilingual SWE environment to near..

HPC-Ops, CUDA and Cue from zero, QPM up 30%, DeepSeek up 17%; Attention up by 2.22 times the single account performance, GroupGEM up by 1.88 times the DeepGEMM and FusedMoE up by 1.49 times the TensorRT-LLM

Ying Wei Da is close to completing a $20 billion investment agreement for OpenAI's latest financing, and will be the largest single investment Ying Wei Da's ever made. This is a very good investment”; OpenAI's total target for this round of financing is $100 billion, the Amazon project is investing up to $50 billion, and the soft bank is thinking of investing $30 billion, with an estimate of about $83 billion; this investment means that AI infrastructure is being tied to head model developers, and capital is accelerating to a very small number of superplayers。

Ali Babala has announced the launch of Qwen3-Coder-Next, an open-weight language model for coding agents and local development. The model is based on the Qwen3-Next-80B-A3B-Base, with a new structure of mixed attention and MoE; smart body training through mass mission-capable synthesis, environmental interaction and enhanced learning, with strong programming and intelligent body capabilities, while significantly reducing reasoning costs. Qwen3-Coder-Next does not rely on simple parameter extensions, but focuses on expanding intelligence..

Original link