AI资讯 2025/7/20
AI 日报
AI内容摘要
AI领域近期风云激荡，人才争夺激烈，科技版"速度与激情”上演。
OpenAI神秘模型编程惊艳，Decart实现实时视频生成，颠覆传统。
人机对抗中，人类创意展现优势，同时化学AI训练框架提升效率。
Today’s AI News
The AI world has been incredibly lively recently, a true tech version of “Fast & Furious”! In just a few days, we’ve seen a “talent snatching war” where everyone’s trying to outdo each other, AI models appearing out of nowhere “flexing their technical muscles,” and even “peak confrontations” between human intelligence and AI algorithms!
First, let’s talk about the AI talent war. New programming tool Windsurf experienced a “thrilling 96 hours”! Originally, OpenAI planned to acquire it for $3 billion, but the deal fell through due to technical sharing issues with Microsoft. To make matters worse, competitor Anthropic suddenly cut off their model supply, severely impacting Windsurf’s product performance. Just as the company was overwhelmed, Google stepped in, unleashing a combined punch of “sky-high poaching + licensing agreement”! They not only used doubled salaries and a $2.4 billion licensing agreement to instantly poach over 40 of Windsurf’s top researchers and engineers to DeepMind, but they also had Google co-founder Sergey Brin personally make an appearance – who could resist such an offer!
What about the remaining 200+ employees and “residual assets”? Fortunately, another AI programming company, Cognition, intervened promptly, taking over all of Windsurf’s remaining intellectual property, products, brand, and operations team. They even honored the employees’ unvested equity, allowing Windsurf’s products to continue independent operation. This move truly showed “humanity and true love”! It’s clear that in the volatile AI landscape, the flow of talent and technology is constantly changing, and things can shift dramatically in an instant.
Next, let’s discuss OpenAI’s recently and quietly released “big killer” – the mysterious model o3-alpha. As soon as it was revealed, it stunned the entire tech community! Users who tested it found its code generation capability to be incredibly strong, especially in front-end code and physical simulations. Whether it was ball collisions, cycling pelican animations, or even “Minecraft” and mini-GTA simulators, it could solve them in a single sentence, far exceeding other models!
Even more exciting, it was revealed that this o3-alpha “anonymously” participated in the 2025 Tokyo AtCoder World Programming Contest Finals and ultimately took runner-up! Although it narrowly lost, this was enough to make humans tremble. However, the outcome of this “man-machine battle” also gave humanity a shot in the arm! The one who ultimately defeated AI was 41-year-old Polish programming legend Przemysław Dębiak (known as Psyho). In a grueling 10-hour battle, AI was once ahead, but Psyho, with his uniquely human creativity and “completely different problem-solving approaches,” achieved a “last-ditch counterattack” in the final moments! This victory proves that while AI has surpassed humans in hard skills like code optimization, in terms of from-scratch creativity and moments of “divine inspiration,” humanity remains the GOAT! Even Sam Altman posted congratulations, and the contest organizers directly stated: “AI taking second place already scared us, but it uniquely lacks human creativity.”
Finally, let’s look at the “magician” in the field of AI real-time video generation – the MirageLSD model launched by Decart company. This model is simply god-tier in video generation, achieving “zero-latency infinite real-time video generation”! What does that mean? It means if you wiggle in front of the camera, it can instantly transform your video into a scene from a “magic world,” with latency as low as 40 milliseconds, generating 24 frames per second, which is 16 times faster than previous models!
What’s the black technology behind this? It turns out it conquered the century-old problem of “error accumulation” in traditional video models. Through unique techniques like Live-Stream Diffusion (LSD), Diffusion Forcing, and History Enhancement, it ensures every frame of the video seamlessly connects without ever going off track. Even AI guru Karpathy couldn’t resist investing and supporting it, stating directly that it will completely revolutionize scenarios like live streaming, gaming, and video calls! Imagine future live e-commerce where shaking your clothes instantly changes them into different styles – wouldn’t that empty your wallet?
In summary, the AI field is developing at a rapid advancement pace. Whether it’s the pursuit of capital, technological breakthroughs, or the sparks of human-machine confrontation, all point to a future full of infinite possibilities, yet with a touch of “thrill.” While enjoying the convenience AI brings, we must also consider how to preserve humanity’s unique wisdom and creativity amidst this technological tide!
Hey everyone! The tech and education circles have been incredibly lively recently, with several pieces of explosive news dropping one after another, leaving us amazed, delighted, and deeply thoughtful.
First, let’s give a thunderous round of applause to our Chinese youths! They achieved an astounding feat at the International Mathematical Olympiad (IMO), known as the “Math World Cup,” by reclaiming the championship with 6 gold medals, 2 perfect scores, and a total of 231 points, absolutely dominating the competition! This move directly raised the ceiling of human intelligence. Meanwhile, how did the AI large models we’ve been following perform? Unfortunately, in this actual problem test, they didn’t even get a bronze medal; even the strongest Gemini 2.5 Pro only scored 13 points. It seems that when faced with pure mathematical challenges, the human brain’s critical thinking and creativity remain an “ace in the hole” that AI cannot match!
But don’t be so quick to condemn AI; they are racing “like crazy” on another track! Former OpenAI employee Calvin French-Owen recently revealed his “intense work culture” at OpenAI. Can you believe that powerful AI programming assistants like Codex, from the first line of code to release, only took a mere 7 weeks! Calvin described it as a frantic sprint of “working until dawn, no weekends.” There’s no complex email culture; everything relies on Slack communication. There are no grand plans; “good ideas” are the real currency. As long as there’s an idea, “just do it”! Researchers are treated as “mini-CEOs” and are given free rein to explore. This extreme “action-first” culture has allowed OpenAI to surge from over 1,000 to over 3,000 people in just a year, achieving astonishing efficiency that even Google cannot match. Of course, rapid expansion also brought challenges like internal communication and codebase duplication. At the same time, OpenAI views AI safety far more seriously than you might imagine; they focus on real risks like hate speech, abuse, and political bias, rather than empty talk of sci-fi-like “intelligence explosions.” Well, perhaps this is “using magic to defeat magic”!
On the other hand, Apple has quietly made its move, releasing its 2025 Apple Intelligence foundational language model technical report. Although former AI head Pang Ruoming has moved to Meta, Apple’s AI journey hasn’t paused. They detailed the device-side and cloud-based models that power Apple Intelligence (the AI feature that makes your iPhone smarter). Apple plays a strong hand in “hardware-software integration”; their models are not only deeply optimized for their own chips (such as the “2-bit quantization-aware training” that makes models smaller and faster) but also introduce a new architecture called “Parallel Track Mixture-of-Experts (PT-MoE)”, making cloud models fast and accurate for complex tasks without compromising user privacy. Behind this is Apple’s commitment to the concept of responsible AI, prioritizing user data security through technologies like content filtering and private cloud computing. It’s like having an AI butler who is not only super capable but also an absolutely “tight-lipped” secret keeper! In the future, the former Google general Chen Zhifeng will lead Apple’s large model team, suggesting that this battle of AI giants will only become more exciting!
In summary, the peak of human wisdom remains insurmountable, but on the AI track, major companies are racing at an unimaginable speed. This inevitably makes one ponder: when extreme human endeavor meets rapidly iterating AI technology, where will the future lead? Will AI become a more powerful “tool” for us, or will AI ultimately surpass humans in certain domains? Regardless, an “intelligent era” full of infinite possibilities and unknown challenges is accelerating its arrival! We are all witnesses to this great transformation.
Dear research newbies, are you still frustrated by the tedious steps of chemistry AI model training? From data cleaning and feature engineering to model tuning, each step feels like playing “spot the difference,” time-consuming, labor-intensive, and easily leading to thoughts like “No regrets for entering chemistry this life, but never again in the next”? Don’t worry!
Shanghai AI Lab’s Materials Science Team is here to bring “warmth”! They open-sourced Chemia, a super cool AI model training framework, claiming to usher chemistry AI research into a new era of “configuration as code”! Imagine you’re making a complex dish; instead of learning to cook from scratch, you get a detailed “recipe”. Chemia plays the role of this “recipe”! It packages the entire process—from preparing data, extracting chemical features, training models, to optimizing parameters and making final predictions—all into a clear and intuitive YAML configuration file. This means you just need to modify a few lines of configuration, and with a simple action, you can “one-click train” complex chemistry AI models! Say goodbye to those headache-inducing code scripts; it’s truly a paradigm of “liberating productivity”!
What secret ingredients are in Chemia’s “recipe”?
- Its algorithm library is super luxurious: Built-in with 15+ classic AI algorithms, plus neural networks and graph neural networks, it’s like having all kinds of cooking utensils in your kitchen, allowing you to make any dish you want.
- Feature engineering is fully automatic: It can automatically help you extract various chemical features, and even flexibly call the latest pre-trained model embeddings, helping you handle the “ingredients” perfectly.
- Intelligent optimization engine: It comes with a “Michelin chef” – Optuna, which automatically helps you tune the model to its best “flavor,” so you no longer need to manually taste it repeatedly!
- Workflow is super flexible: Whether it’s quick prototyping, cross-validation, or complete end-to-end training optimization, it can handle it with ease.
The best part is that using Chemia is as simple as sharing a recipe. As long as you share your configuration file and data, others can “one-click reproduce” your experimental results, greatly enhancing research credibility and team collaboration efficiency. For scientific research that requires rigorous reproducibility, this is truly a godsend!
Simply put, you just need to prepare your data, write your “recipe” (configuration file), then type one line of code, and you can sit back and wait for the AI model training to complete, getting beautiful experimental results and reports. Isn’t it so simple you’re questioning your life? In the future, Chemia will continue to upgrade and iterate, making chemistry AI research increasingly simpler and more efficient!