This article gives you a quick look at this week'sAI hot topics。
01. Major AI Events
70% faster than human algorithms! GoogleDeepMinduses AI to improve data sorting, lands in Nature
Google DeepMind releases three major AI tools AlphaZero, MuZero,AlphaDev, which can improve data center resource utilization, boost video compression efficiency, and discover faster algorithms.
The new algorithm discovered by AlphaDev improves the sorting efficiency of short element sequences by 70%, and its new sorting algorithm has been released into the C++ library.
BAAI releases the "Wujie" series of large models, decoding a new paradigm for interaction with the physical world
The 7th Beijing智源 Conference opened in Zhongguancun.
The Beijing Academy of Artificial Intelligence released the "Wujie" series of large models, covering the multimodal world model Emu3, the brain science model Jianwei Brainμ, the embodied intelligence brain RoboBrain 2.0, and the microscopic life model OpenComplex2, building a full-chain AI system.
Google updates the Gemini 2.5 Pro preview model to the 06-05 version, with multiple AI performance improvements
Google has updated the Gemini 2.5 Pro preview model to the 06-05 version, with improved scores on multiple AI performance benchmarks and relaxed request limits for Pro users.
The LMArena Elo score rose by 24 points, and the WebDevArena Elo score rose by 35 points to 1443 points, achieving an industry-leading position.
Google optimized the model's style and structure, making it more creative and its answer format clearer, improving the interactive experience.
Shanghai AI Laboratory creates the first "generalist" robot brain
Shanghai Artificial Intelligence Laboratory, together with multiple institutions, proposed a new general embodied intelligence brain framework VeBrain. By simultaneously integrating visual perception, spatial reasoning, and robot control capabilities, it can enable multimodal large language models (MLLM) to directly control physical entities, allowing robots to "see-think-act" like humans.
Microsoft announces a plan to cut 6,000 jobs, shifting its strategic focus to AI
Microsoft Corporation announced today the launch of a global layoff plan involving about 6,000 positions, mainly concentrated in non-technical functional departments and some subsidiary business lines.
Microsoft and India Yotta Data Services reached a strategic partnership to collaborate with participants in India's AI mission plan, government agencies, the Indian Institutes of Technology, startups, and business organizations.
2025 Global Artificial Intelligence Technology Conference, with nearly a thousand participants from government, industry, academia, and research
The 2025 Global Artificial Intelligence Technology Conference (GAITC2025), hosted by the Chinese Association for Artificial Intelligence, opened in Hangzhou Future Sci-Tech City. This year's conference is themed "Intersection, Integration, Mutual Generation, Win-Win," with more than 200 experts and scholars from around the world sharing cutting-edge reports and participants from government, industry, academia, and research.
The site showcased multiple exploratory achievements in the regional implementation of industry large models.
At the same time, Hangzhou Future Sci-Tech City launched a special support initiative for intellectual property securitization financing in the AI field, opening a new financing path for AI innovative enterprises. It is expected that within the next three years, 5 intellectual property securitization products in the AI field will be issued, reaching more than 60 enterprises and raising more than 500 million yuan.
Anthropic announces it will cut off all API access for Windsurf within 5 days, co-founder reveals the reason
Anthropic announced it will cut off all API access for Windsurf , with the specific models involved including: Claude 3.5 Sonnet, Claude 3.7 Sonnet, Claude 3.7 Sonnet Thinking.
Anthropic co-founder and Chief Science Officer Jared Kaplan It stated that the main reason the company cut off the AI programming assistant Windsurf's direct access to Claude models was market rumors that Windsurf was about to be acquired by OpenAI, Anthropic's largest competitor.
02. New Product Releases
Google launches the Veo3-Fast model, priced 80% cheaper than Veo 3
Google launched the Veo3-Fast model, which consumes 20 credits to generate a single video, one-fifth ofVeo3's.
The video resolution generated by Veo3-Fast is the same as Veo3, both 8 seconds at 720p, with automatic voiceover generation.
Jimeng Image 3.0 major update, launches intelligent reference feature
JimengImage 3.0 model updates the intelligent reference feature, performing excellently in maintaining consistency with the original image, down to the hair-thread level. Users can use simple prompts to add text, modify text, make complex designs, and perform various other operations to generate high-quality images. The feature is currently in gray-scale testing, and will be fully launched on the app and web next week.
Image editing model SeedEdit 3.0 released! Stronger preservation, higher usability rate
ByteDance's Seed team released the image editing model SeedEdit 3.0. The model is based on the text-to-image model Seedream 3.0 and adopts diverse data fusion methods and a specific reward model, significantly improving the ability to preserve image subjects, backgrounds, and details, and performing excellently in scenarios such as portrait editing and background changes.
SeedEdit 3.0 can process 4K images, accurately identify and remove irrelevant people and their shadows, and convert 2D paintings into real models while preserving details.
Cursor 1.0 released: BugBot code review, Memories, one-click MCP installation, and more
Cursor 1.0 is officially released, adding the BugBot automatic code review tool, which can automatically review PRs and accurately identify potential issues. After discovering problems, it can comment on GitHub PRs to provide fix suggestions.
It launched the beta Memories feature, which can store key information from conversations by project; supports one-click installation of MCP servers and simplifies the verification process; the chat interface adds visualization support, and the settings and dashboard interfaces have been optimized; the pricing model has been simplified.
OpenAI released two ChatGPT new features: "Connectors" and "Record Mode"
OpenAI launched two new features for ChatGPT: "Connectors" and "Record Mode."
Connectors can integrate with platforms such as GitHub,Google Drive, to enable data querying and organization, helping enterprise knowledge management.
Record Mode can automatically record audio, take notes, and generate meeting minutes, improving meeting efficiency.
Manus launches text-to-video feature, challenging competitors such as OpenAI and Sora
Manuslaunched a native "text-to-video" feature, currently only available to paid members.
The feature can quickly convert text commands into ordered video stories. Manus offers three membership services: Basic, Plus, and Pro. The Pro membership costs $199 per month.
Kimi launches a brand-new AI academic research assistant: Academic Search
Kimi launched a brand-new AI academic research assistant, "Academic Search," which helps users improve efficiency and quality in academic research and paper writing through deep reasoning search, adherence to user instructions, and real-time interaction.
Users can use it after registering and logging in on the official website, and it is suitable for various scenarios such as academic research support and technical problem solving.
AI video tool PixVerse's domestic version "Paiwo AI" officially launched
PixVerse's domestic version "Paiwo AI" is officially launched, simultaneously released on web and mobile, and supports version V4.5.
Users can download and experience it by visiting the Paiwo AI official website and major app stores. The product offers innovative features such as AI effects and first and last frames, helping users create unique videos.
China's Water Conservancy Standards AI Large Model Officially Released
The 'Water Conservancy Standards AI Large Model' independently developed by the China Institute of Water Resources and Hydropower Research under the organization of the Department of International Cooperation, Science and Technology of the Ministry of Water Resources has been officially released and put into application.
The model is based on the 'SkyLIM' system and adopts a 'massive knowledge base +DeepSeek/Qwen dual-model + standard service' architecture. It integrates multi-source corpora such as more than 1,800 water conservancy and hydropower standards and more than 500 laws and regulations, and has multi-dimensional functions including water conservancy standard duplicate checking and comparison, query, drafting, review, and evaluation.
Microsoft Bing launches AI video generation tool Bing Video Creator, free to use the Sora model
The Microsoft Bing team launched Bing Video Creatora free AI video generation tool, powered by OpenAI's Sora model.
Users only need to enter a detailed text description in the Bing mobile app to generate a 5-second vertical-screen video (9:16 format).
It offers two generation speeds, fast and standard. Initially there are 10 free fast generation opportunities, after which users can redeem with points or switch to standard speed.
ElevenLabsRelease of the latest TTS model — Eleven v3
ElevenLabs, the unicorn specializing in AI speech synthesis, released its latest TTS model —Eleven v3, firmly mastering emotional control.
Eleven v3 not only supports more than 70 languages (including Chinese), but can also conduct multi-person conversational chats, during which each person's emotions, tone, and other expressions are very vivid.
The company confidently stated that this is 'the most expressive text-to-speech model to date'.
Other Developments
MetasoThe 'Learn Something Today' platform launched a PPT export feature. Users can click the 'Export PPT' button on the video explanation page to download a complete PPT containing images and text, voice explanations, and a verbatim transcript.
StepFun's To C product 'Maopao Ya' will cease operation, and the team will be merged into the conversational product 'Step AI'. Since 2025, StepFun has shifted its strategic focus from the C-end to model research and development and ToB/G, with ToB Agent products as a key layout direction.
OpenAI announced that it willChatGPT's memory feature gradually open to free users. Free users can experience a lightweight version of the memory feature, which supports short-term conversation continuity, while paid users can reference older conversation records.
OpenAI upgraded the voice feature forChatGPTpaid users. Its voice is more natural, with more delicate intonation and more realistic rhythm (including pauses and emphasis), and it expresses certain emotions more accurately, including empathy and sarcasm. In translation, it is also more efficient and can continuously perform multiple rounds of translation.
CRIC launched China's first AI Agent in the vertical real estate field —CRIC Deep IntelligenceAlliance. Based on CRIC's 20 years of industry experience and massive data, combined with multimodal large model technology, it provides intelligent solutions for the real estate industry. CRIC Deep Intelligence Alliance helps real estate practitioners improve work efficiency and decision quality, and is known as the 'second brain' of real estate professionals.
Galbot releases the world's first product-grade end-to-end embodied FSD large modelTrackVLAThe model possesses pure vision-based environmental perception, language instruction-driven operation, autonomous reasoning, and zero-shot generalization capabilities, without the need for prior mapping or teleoperation control, enabling a closed-loop motion of "listen → see → understand → walk." It has eight core capabilities, including natural language understanding and target recognition, accurate following in complex scenarios, target recovery after loss, and autonomous navigation in unfamiliar environments.
03. Open Source Projects
Alibaba Open-Sources New Qwen3 Model Embedding and Reranker
Qwen3 adds a new Embedding series, designed specifically for text representation, retrieval, and ranking tasks, available in three sizes: 0.6B/4B/8B. The 8B version ranks first on the MTEB multilingual Leaderboard, surpassing numerous commercial API services in performance. A Reranker series of ranking models was also released.
Kuaishou Open-Sources "Auto Think" Large Model, Automatically Adjusts Thinking Based on the QuestionDeep
The Kuaishou Kwaipilot team has open-sourced the KwaiCoder-AutoThink-preview automatic thinking large model. Aiming at the "overthinking" problem of deep thinking large models, the model proposes an automatic thinking model training paradigm that integrates "thinking" and "non-thinking" capabilities, and can automatically switch thinking modes based on question difficulty.
Open-Source AI Design Agent — Jaaz, a Free Alternative Project to Lovart
Open-source AI design agent — Jaaz, with powerful AI design capabilities, can intelligently generate design prompts and batch-generate images, posters, storyboards, etc.
Jaaz supports local image and language models such as Ollama,Stable Diffusion, Flux Dev, enabling free image generation.
Xiaomi's Multimodal Large Model MiMo-VL Open-Sourced, Leading Qwen2.5-VL-7B in Multiple Aspects
Xiaomi has officially open-sourced the multimodal large model MiMo-VL. It significantly leads the benchmark multimodal model of the same size, Qwen2.5-VL-7B, in multiple tasks such as general question answering and understanding reasoning for images, videos, and language, and rivals specialized models on the GUI Grounding task.
Google Open-Sources Deep Research Framework
Google has open-sourced a Deep Research system template, detailing how to build a research-enhanced conversational AI based on the Google Gemini 2.5 model and the LangGraph framework, aiming to dynamically generate search terms, use Google Search to query the web, reflect on results to identify knowledge gaps, and iteratively refine searches until it can provide well-supported answers with citations.
Google DeepMind Launches Sign Language Translation Model: SignGemma, Breaking Down Sign Language Communication Barriers
Google DeepMind launches SignGemma, a powerful sign language translation model that can convert sign language into spoken text, primarily deeply optimized for American Sign Language (ASL) and English. It is an open-source model and will join the Gemma model family later this year. DeepMind hopes to use this to break down communication barriers for sign language users and help them participate more smoothly in work, study, and social interactions.
ModelBest MiniCPM 4.0 Open-Sourced, On-Device Inference Typically Accelerated 5x
ModelBest重磅推出MiniCPM 4.0 — an extremely efficient on-device large model that, through its self-developed CPM.cu inference framework, can achieve an extreme speedup of 220x, with a typical 5x acceleration.
This time, two parameter-scale versions, 8B and 0.5B, are core-released in the open-source community, both achieving the best performance in comparisons with models of the same level.
04. Investment and Financing
Turing Award Winner Bengio Starts Another Business: Raised $30 Million in Initial Funding
Turing Award winner and one of the three giants of deep learning, Yoshua Bengio, has started another business, founding the nonprofit organization LawZero, and has raised $30 million in initial funding.
LawZero builds "secure by design" AI systems. It does not build Agents, but supervises Agents, with the core goal of understanding and learning about the world, and provides verifiable real answers.
AMD acquires AI software optimization startup Brium
AMD announced the acquisition of Brium. Brium focuses on developing machine learning inference technology that can adapt to multiple hardware architectures, enabling AI models originally designed for specific hardware such as NVIDIA to run effectively on platforms from other brands. AMD stated that this move is advancing the building of a "high-performance, open AI software ecosystem" and supports developer innovation.05. Opinion Insights
Sam Altman advises business leaders: just start directly, don't wait for the next version
At the Snowflake Summit 2025 held in San Francisco this week, OpenAI CEO Sam Altman said bluntly that for business leaders still waiting and watching in 2025, his advice is just one sentence: just start directly, don't wait for the next version.
He said that if companies always wait for better models and more mature solutions, they will instead miss out on the real first-mover advantage.What truly achieves breakthroughs are the teams willing to bet early and move quickly in small steps. Technology is evolving rapidly, and victory or defeat depends on who has the ability to quickly trial and error and quickly learn.
Snowflake CEO Sridhar Ramaswamy strongly agrees with this. He believes that curiosity is the most overlooked scarce capability in companies today. Many ways of working that were once taken for granted now need to be questioned.
Original link:AI Weekly丨AI industry developments you must not miss this week (6.2-6.8)