AI 洞察日报 - 2025-07-28

• 日报

AI资讯 2025/7/28

AI 日报

AI内容摘要

WAIC 2025展示AI前沿突破。商汤、京东、腾讯等巨头发布具身智能、多模态大模型。
百度自动驾驶落地，AI加速融入生活。大会聚焦AI与数学融合发展。
AI助力数学研究，从验证到发现，上海成融合高地。

Today’s AI News

  1. On the stage of WAIC 2025, a thrilling drama about the future of AI unfolded, with tech giants showcasing their unique capabilities, foretelling a world that will become increasingly “smart” and even “thoughtful”!

  2. First to appear was SenseTime, presenting their sci-fi flavored “Wuneng” Embodied AI Platform. This is no ordinary AI platform; it uses a powerful Embodied World Model as its “brain,” combined with computing power from “SenseTime’s Big Device,” enabling robots and smart devices not only to “understand” the world but also to “navigate” three-dimensional space. Imagine future robots not only accurately perceiving their surroundings but even “racing” in real street scenes like in Need for Speed – of course, not actual racing, but accurate reproduction and simulation in a virtual world! Even cooler, you just need to say “go find something on the kitchen shelf,” and it can act autonomously like your loyal butler, practically lowering the threshold for human-computer interaction to the floor.

  3. Next, JD.com also made a big move, officially upgrading its large model brand to JoyAI, and showcasing an AI “Dream Team” covering multiple fields such as retail, logistics, industry, and health. Their AI large models can process models of various sizes from 3B to 750B, and understand multi-modal information like language, images, videos, and digital humans. Most striking is JD.com’s embodied intelligence brand JoyInside, which can “inject” JoyAI models into terminal devices such as robots and toys, allowing these little guys not only to move but also to become “emotionally intelligent,” for example, delivery robots and robot dogs can interact naturally with people. The AI digital humans in the retail sector are even more amazing; they can conduct live-streaming sales 24/7, at only one-tenth the cost of human anchors, yet with astonishing efficiency, truly an “ATM machine” for businesses! And the open-source JoyAgent enterprise-level intelligent agent, which gained a thousand stars on GitHub in just three days, shows that JD.com not only masters the technology themselves but also wants to share it with the world.

  4. Finally, Tencent made a grand appearance with their Hunyuan 3D World Model 1.0, announcing that this revolutionary model will be fully open-sourced! This is the industry’s first open-source world generation model that supports immersive roaming, interaction, and simulation, practically a “magic wand” tailor-made for game developers and content creators. What used to take professional teams weeks to complete 3D scene modeling now allows you to generate high-quality, freely roamable 3D worlds in minutes by simply inputting a few lines of text or uploading images! Whether it’s complex terrain and exquisite architecture in games, or immersive spaces in virtual reality, it can be easily achieved. This means that even if you’re not a professional modeler, you can create your own virtual universe, truly realizing the dream of “everyone is a creator.” Not to mention Tencent Hunyuan’s other series of open-source plans in multi-modal understanding and generation, which are practically opening up AI’s “toolbox” for anyone to use.

  5. Overall, WAIC 2025 seems to be telling us that AI is no longer just a theory in laboratories; it is stepping into our lives in various embodied and scenario-based forms, fundamentally changing the way we interact with the world. From “embodied intelligence” that can understand and simulate the real world, to AI solutions that empower various industries, and open-source models that allow ordinary people to create complex 3D worlds, we are witnessing the accelerating arrival of an era of “intelligent everything, human-machine symbiosis.” In the future, the boundaries between us and robots will become increasingly blurred, and the lines between virtual and reality will become more open. This is both exciting and thought-provoking: as AI understands us more and more, and even helps us “create” worlds, how will human creativity be redefined?

  6. WAIC 2025: AI “Flexes Its Muscles,” From Virtual Worlds to Autonomous Driving, The Future is Here!

  7. The 2025 World Artificial Intelligence Conference (WAIC 2025 World Artificial Intelligence Conference) was like a “muscle show” of future technology, with various disruptive technologies emerging one after another, dazzling the audience! At this conference, we witnessed major breakthroughs in AI across three directions, which were truly astounding.

  8. First up was tech giant Tencent, who didn’t just hype concepts but directly unveiled their “1+3+N” panoramic blueprint for the AI era! The core of this system is their self-developed “Hunyuan Large Model” – doesn’t that sound like a formidable internal martial art from a wuxia novel? It can not only support the super cool “Hunyuan 3D World Model” (imagine, the future metaverse might be built like this!), but also drive the Embodied AI Platform Tairos (which sounds incredibly hardcore!). In simple terms, Tencent aims to use AI to help you build a virtual world that can see, hear, and move, and even make robots smarter, learning to understand and act like humans. Isn’t this bringing the dream of “Westworld” into reality ahead of schedule?

  9. Next, there’s a “video magician” – ShengShu Technology. Their Vidu Q1 “Reference Video Generation” feature is simply a godsend for video creators! In the future, you won’t need to struggle with storyboarding; just throw in a few reference images and text prompts, like “Have Zhuge Liang, Churchill, and Napoleon chat in a conference room,” and Vidu can “imagine” a complete video for you, with character appearances that won’t “change face,” completely solving the “subject consistency issue” that plagues video models! This technology is like giving AI a “memory chip,” letting it know what you want and ensuring consistent generation every time. Even more impressively, ShengShu Technology also collaborated with Tsinghua University on an Embodied AI Model Vidar. This means they can not only bring virtual worlds to life but also teach robots to watch videos and then imitate the actions within them, using AI’s “brain” to control the robot’s “body.” Isn’t this a real-life “Iron Man” training camp?

  10. Finally, let’s talk about a “future” we can directly experience – autonomous driving. At the WAIC venue, Shanghai grandly issued autonomous driving demonstration operation licenses to companies like Apollo Go (you heard that right, the cute name, Baidu’s autonomous driving service!). This means that in some areas of Shanghai’s Pudong, you’ll soon be able to hail “AI taxis” without drivers! Apollo Go also specially provided L4 autonomous driving shuttle services at the conference, allowing attendees to personally experience the thrill of “driverless driving.” Imagine, in the future, sitting in a car, no longer needing to stare at the road conditions, but instead being able to safely scroll on your phone or enjoy the scenery – this is truly a blessing for the lazy!

  11. This WAIC not only showcased AI’s grand narratives (such as 3D world models, embodied intelligence) but also brought tangible convenience at the micro-application level (video production, autonomous driving). AI is no longer a laboratory concept; it is striding into our lives, changing every aspect of our work. Of course, it also brings an interesting thought: when AI becomes more “human-like” and even more efficient than humans, what then is the unique value of humanity? Perhaps, pondering how to coexist with AI and how to use AI to unleash our creativity is the truly interesting new subject!

  12. Hey guys! Imagine what kind of sparks will fly when Artificial Intelligence, this “all-rounder,” meets Mathematics, this “ancient wisdom”? Two recent events have provided the perfect footnote for this “marriage” of AI and mathematics!

  13. First, if you’re a developer or curious about AI applications, you absolutely must check out two freshly released “martial arts manuals”:

    • One is “awesome-llm-apps” (github.com/Shubhamsaboo/awesome-llm-apps), which is simply a “treasure trove” of Large Language Model (LLM) applications! It contains various ways to play with models from “big players” like OpenAI, Anthropic, and Gemini, as well as many clever uses of open-source models, especially those with AI Agents and Retrieval-Augmented Generation (RAG) features, practically helping you unleash the full potential of AI assistants!
    • The other is “ai-cookbook” (github.com/daveebbelaar/ai-cookbook), as the name suggests, this is a recipe book for “cooking” AI systems, with various examples and tutorials that guide you step-by-step to build AI from scratch, truly a boon for beginners!

  14. However, the real show was the high-energy forum titled “AI for Math” that recently took place at the 2025 World Artificial Intelligence Conference (WAIC) in Shanghai. This was no ordinary academic seminar; it directly put AI to the test, allowing it to “compete” on the same stage with top human mathematicians!

  15. Why must AI and Mathematics form a “strong alliance”? Simply put, AI development has hit a bottleneck, urgently needing mathematics as its “foundation.” For instance, large models now have trillions of parameters, making traditional optimization methods difficult; this requires theoretical guidance from convex optimization and non-convex optimization. The attention mechanism in the Transformer architecture is essentially the result of probability and statistics and information theory. To make AI safer and multi-modal integration smoother, it’s inseparable from advanced mathematics like game theory, cryptography, and topology. Conversely, AI can also “supercharge” mathematics. For example, DeepMind’s AlphaGeometry can solve geometry problems using AI, even outperforming international Olympiad participants! This “AI feeding back into mathematics” model is quietly changing the way mathematical research is conducted.

  16. How did AI “conquer” the math world at WAIC? This forum was practically the “Olympics” of the AI mathematics world:

    • Fields Medalist personally “set problems”: The most explosive moment was mathematical master Professor Shing-Tung Yau personally designing challenging problems for AI, with several top AI large models solving them on the spot, their reasoning processes displayed in real-time on a large screen – wasn’t this a real-life “man-machine battle”?!
    • Large models “answering live” were dazzling:
      • Shanghai AI Lab’s Intern-IMO system successfully tackled the first problem of the 2025 International Mathematical Olympiad (IMO), reportedly showing excellent mathematical intuition and clear thinking!
      • SenseTime’s “SenseNova” large model demonstrated “multi-path reasoning,” solving geometry problems using two methods, not only reaching solutions but also revealing the mathematical beauty behind the problems, truly an “academic ace” performance!
      • StepAhead’s model showcased the ability to “seek help,” correcting its own mistakes by actively calling external tools to perform formal verification – impressive AI EQ!
      • MiniMax’s M1 was even more formidable, not only solving problems correctly but also extrapolating to answer open-ended questions like “If conditions are weakened, does the conclusion still hold?,” proving more rigorous than a certain internationally renowned large model (GPT-4o)!
    • Experts’ “Critiques” and “Praises”: Mathematicians were amazed by AI’s complete logical chain in probabilistic recursion but also pointed out that for problems requiring “counter-intuitive” constructions, human mathematicians still hold the main stage. This is like AI being able to recite the multiplication table perfectly, but creating the “Pythagorean theorem” from scratch still requires human input.
    • “Genius” Pairings, Promising Future: The event also saw mathematics and AI students from top global universities like Oxford, Cambridge, and Toronto, paired with leading Chinese universities, collaborating on research. This is truly building up young talent for future “mathematical intelligence”!
    • Keynote Speeches: Academician Xu Zongben of the Chinese Academy of Sciences proposed that AI architecture design should start from “infinite-dimensional” intelligent problems, and while current technology is still “finite-dimensional,” the focus should be on the essence. Academician Torsten Hoefler of the European Academy of Sciences revealed AI’s leap from “next word prediction” to “tree of thought reasoning,” which can double computing power utilization.
    • International Mathematics Labs “Land” in Shanghai: Two top international labs – the Hitchin–Ngo Lab focusing on algebraic geometry and mathematical physics and the Fefferman Lab specializing in fluid mechanics – were officially unveiled in Shanghai. This means that the world’s most powerful minds and AI computing power will deeply integrate here to jointly tackle century-old problems like the Navier-Stokes equations!
    • Clash of Ideas: Can AI lead to “AGI (Artificial General Intelligence)”: Some scholars believe that AI’s “hallucinations” might be “seeds of imagination,” and humanity’s value lies in extracting truth from “counter-intuitive” insights. Others proposed that AI’s path to AGI might depend on its ability to develop “social intelligence” in multi-agent systems, just like humans.

  17. How has AI progressively “reshaped” mathematical research? This didn’t happen overnight; AI’s role in mathematics has undergone a splendid “triple jump”:

    1. “Super Proofreader” (1970s-2000s): During this stage, computers acted as rigorous “formal verification” assistants, transforming mathematical proofs from “paper discussions” into “machine code verification.” The most famous example is the proof of the “Four Color Theorem,” where computers helped verify thousands of configurations, allowing humanity to “confidently” believe this complex proof for the first time. Later, the Coq proof assistant and the “Flyspeck Project” verifying Kepler’s conjecture pushed this rigor to the extreme; computers could not only confirm correctness but also identify overlooked details that humans took for granted.
    2. “Logic Shredder” (2010-2020): With algorithms and computing power soaring, AI began to process ultra-large-scale logical reasoning, like SAT (Boolean Satisfiability Problem) and SMT (Satisfiability Modulo Theories) solvers, handling problems that humans simply couldn’t compute manually. For instance, solving the “Boolean Pythagorean Triples Problem” took a computer four years, generating 200TB of data to reach a conclusion – a proof that humans could never “read through,” yet it’s correct! AI proved that some mathematical truths can only be reached by machines.
    3. “Pattern Explorer” (2020-Present): Now, with the rise of deep learning and large language models, AI is no longer just a “tool,” it can actively participate in the discovery of mathematical patterns and the generation of conjectures!
      • It can deeply integrate with formal proof tools like Lean, where Terry Tao’s team used AI to find redundancies in proofs and extract more general proof frameworks, significantly improving research efficiency.
      • AI can also discover mathematical correlations in vast datasets that are imperceptible to the human eye. For example, in knot theory, AI analyzed nearly 2 million data points to find an analytical relationship between seemingly unrelated parameters and “signature values,” with human mathematicians ultimately completing the proof – this is the new model of “machine discovers patterns, human proves patterns”!
      • While large language models occasionally “hallucinate” (meaning they can convincingly state incorrect information), frameworks like DeepMind’s FunSearch with its “generate-and-verify” loop, and AlphaGeometry combining symbolic reasoning and neural networks, have enabled AI to demonstrate superhuman capabilities in areas like combinatorial problems and geometric proofs. It can even construct unimaginably complex mathematical objects through reinforcement learning, and even found faster matrix multiplication algorithms! Of course, there are times when AI “can’t handle it.” As mathematicians from Cambridge University pointed out, AI can write beautiful proof steps, but it currently cannot propose grand theories like “Langlands Program” that unify multiple mathematical fields – original breakthroughs from “0 to 1” remain the unique talent of human mathematicians!

  18. Shanghai: A “Feng Shui Treasure Land” for the “Marriage” of AI and Mathematics! Why was this grand event held in Shanghai? Because this city is simply a “chosen land” for the integrated development of AI and mathematics!

    • Excellent City Genes: Shanghai not only boasts world-leading computing infrastructure like the Huawei 384 Super Node real machine but also a complete AI industry ecosystem such as Zhangjiang AI Island. Simultaneously, mathematics disciplines at universities like Fudan and Jiaotong are among the nation’s best. This combination of “AI industrial clusters + top mathematics disciplines” is simply perfect!
    • Talent “Magnet” Effect: Professor Shing-Tung Yau established a research center in Shanghai, and WAIC annually attracts global AI luminaries. This talent aggregation effect provides continuous intellectual support for breakthroughs in AI mathematics integration.
    • Guiding Frontier Directions: Shanghai is comprehensively deploying around three major directions: breakthroughs in basic theory, AI-assisted mathematical research, and industrial scenario transformation. It strives to provide “axiomatic support” for AI through mathematics and enable AI to find the best solutions via “mathematical models + AI optimization” for complex problems across various industries.
    • Full-Chain Guarantee: From establishing “AI + Mathematics” interdisciplinary programs to cultivate composite talents, to building a “Mathematical Intelligence Innovation Platform” for sharing global resources, and setting up a “Mathematical Intelligence Innovation Fund” to incentivize scientific research and innovation, Shanghai is constructing a complete ecological closed-loop from talent to platform, from policy to application.

  19. In conclusion, the “AI for Math” grand show at WAIC not only demonstrated AI’s astonishing capabilities in the field of mathematics but also profoundly revealed the future landscape of human-AI co-evolution. When rigorous mathematical thinking meets powerful AI technology, they will jointly drive breakthroughs in fundamental theories, sketching out a future that is more intelligent and full of infinite possibilities! Shanghai is undoubtedly becoming the “originating place and demonstration window” for this great transformation, and its “Shanghai Solution” is worth looking forward to!