## Major AI Science Breakthrough: Meet Kosmos, the AI Scientist

Scientists have created an amazing new type of artificial intelligence called Kosmos that works like a human researcher. This AI scientist can conduct scientific research completely on its own. In just twelve hours, Kosmos completed work that would normally take a human scientist six months to finish. The system is incredibly productive—it read 1,500 scientific papers and wrote 42,000 lines of computer code in a single run. Most impressively, Kosmos made new scientific discoveries with 79.4% accuracy, proving that AI can do real scientific research at scale. This breakthrough shows that the future of scientific discovery might involve AI and humans working together.

## How AI Agents Are Stronger in Teams

Researchers made an important discovery about how multiple AI agents work together. They found that when individual AI agents try to work as a team, they often fail—even though each agent works well on its own. Scientists call this the "collaboration gap." To solve this problem, they developed a solution called "relay inference" that helps AI agents communicate and coordinate better. This discovery is important because it mirrors how human teams work. Just like a soccer team plays better than individual players, teams of AI agents can solve harder problems than single AI models.

## New AI Models Launch Around the World

Baidu, a major technology company from China, introduced a new AI model called ERNIE. This model performed better than other popular models like GPT and Gemini when tested on visual tasks like reading charts and solving math problems. The ERNIE model is especially useful for industries that work with lots of diagrams, dashboards, and videos. It can zoom in on images, identify objects, and even search through long video files. In the United States, Microsoft announced two new AI models as part of its plan to develop its own artificial intelligence technology. The first model, MAI-Voice-1, is designed for speaking and sounds natural with low delays. The second model, MAI-1-preview, is a large foundation model that Microsoft trained using about 15,000 powerful computers. This means Microsoft is becoming more independent from OpenAI and creating its own AI tools.

## Open-Source Model Released with Impressive Efficiency

Moonshot AI, a company from Beijing, China, released a powerful new open-source AI model called Kimi K2 Thinking. This model contains one trillion parameters, which means it is very powerful. What makes Kimi K2 special is that it can execute 200 to 300 tool commands in a row, meaning it can perform complex tasks with many steps. The most impressive fact is that Moonshot claims this advanced model cost only $4.6 million to build and train—much cheaper than other large AI models. Kimi K2 Thinking is designed as a "thinking agent" that can handle long, complicated reasoning tasks. This affordable, powerful model could help more companies and teams build their own AI agents for research and analysis.

## Google, OpenAI, and Tools for Better Research

Google added a new feature called Gemini Deep Research to its tool called NotebookLM. This feature helps people gather information, create research plans, and produce detailed reports automatically. The system can use information from Gmail, Google Drive, Google Chat, Sheets, Word files, and PDFs. OpenAI released a new version of its famous ChatGPT called GPT-5.1 Instant and GPT-5.1 Thinking**. The company designed these models to sound warmer, more friendly, and more conversational. They also improved how well the models follow specific instructions.

## The Future: AI That Understands Physical Space

Fei-Fei Li, a famous artificial intelligence researcher, explained an important limitation in today's AI. She said current AI systems cannot reason about physical space or predict how environments change over time. To fix this problem, she proposed developing "world models"—AI systems that can understand and simulate 3D environments. These world models would combine information from multiple sources and support robotics, scientific research, and creative tasks. Fei-Fei Li's company created a prototype called Marble that can generate environments users can explore, and these environments stay stable as users interact with them.

## The Bigger Picture: Teams of Specialists Instead of One Super AI

The most important trend in 2025 is a major shift in how companies use AI. Instead of building one super-powerful AI model that does everything, companies are creating teams of specialist AI agents that work together. Each agent specializes in one type of task—just like a sports team has different players with different skills. This approach mirrors how human teams solve complex problems by having specialists work together. Experts predict that by 2026, 40% of business software will include specialized AI agents, jumping up from less than 5% in 2025. The old way of using AI—where AI helps humans with suggestions—is becoming less popular. The new way is agentic AI, where AI agents act independently to complete tasks with minimal human supervision.