Google Gemini Can Now Generate Documents, Spreadsheets, and Presentations Directly in Chat

Google has unveiled new Gemini capabilities. Users can now generate PDFs, Word documents, Excel spreadsheets, PowerPoint presentations, and more directly within the chat interface.

Until now, AI chat output has centered on text and images, so the ability to produce formats that are ready for practical use marks a significant step up in AI's usefulness as a productivity tool. In particular, it eliminates the common business workflow of "generate with AI → convert the format," which is expected to have a major impact for Google Workspace users.

Parallel Web Systems Reaches $2 Billion Valuation in Just 5 Months

Parallel Web Systems, an AI agent startup founded by former Twitter CEO Parag Agrawal, has reached a $2 billion valuation with a $100 million round led by Sequoia Capital. The rapid rise comes just five months after its previous major fundraise.

The AI agent space is drawing attention as a new generation of automation platforms built around tool integration and autonomous execution, and Parallel Web Systems' fast growth reflects strong investor appetite for this market. That said, its actual enterprise adoption and path to monetization remain to be proven.

SenseTime Releases Image Model for Chinese Chips Under US Sanctions

SenseTime, the Chinese AI company whose access to cutting-edge chips is restricted by US sanctions, has open-sourced an image generation model that runs fast on Chinese-made chips.

According to WIRED, the company has turned its restricted access to state-of-the-art technology into an advantage, pursuing a strategy focused on optimizing for domestic chips. This marks a new development in the US-China AI race. As Chinese AI companies build out their own hardware ecosystems, attention is focused on how effective sanctions really are and whether the technology can be substituted.

Mistral's Le Chat Spreads Iran War Disinformation in 60% of Prompts

An audit by NewsGuard found that Mistral's AI chat, Le Chat, repeats state-sponsored disinformation related to the Iran war in roughly 60% of leading prompts.

While this is a separate issue from the Mistral Medium 3.5 release covered in our previous article, it shows that ensuring safety and reliability is just as urgent as improving model performance. The risk of AI models unintentionally spreading misinformation is a particularly serious concern for high-public-interest applications such as news distribution and information search.

inclusionAI Releases 1-Trillion-Parameter Model "Ling-2.6-1T"

inclusionAI has released Ling-2.6-1T, a comprehensive flagship model at the 1-trillion-parameter scale, on HuggingFace.

The model has become a talking point in the LocalLLaMA community, and it appears to be designed not just for sheer parameter count but also for multilingual support and comprehensive task handling. Its open-weights release means researchers and developers can freely evaluate and use it.

Runway CEO: "AI Video Is Just the Beginning — World Models Come Next"

Cristobal Valenzuela, CEO of Runway, the pioneer of AI video generation, said on a TechCrunch podcast that AI video is merely a "prequel." Runway, which has raised around $860 million at a $5.3 billion valuation, is looking ahead to building "world models" that simulate the physical world as its next step.

As it competes with Google, OpenAI, and others, the shift beyond simple video generation toward spatial understanding and physical simulation suggests that AI's range of applications could expand into gaming, robotics, autonomous driving, and more.

Ramp's AI Feature Leaks Financial Data — A Real-World AI Security Risk

It has been revealed that the AI spreadsheet feature of Ramp, an enterprise spend management platform, could leak financial data to outside parties through prompt injection attacks. The case, reported by security firm PromptArmor, has drawn attention with 32 points on Hacker News.

As with the Claude AI agent database deletion incident covered in our previous article, the balance between autonomous data access in AI tools and security management is once again being called into question. The incident underscored that careful design and auditing of data access permissions are essential when enterprises adopt AI features.

---

**Also worth noting**

IBM's Granite Speech 4.1 (a speech recognition model) has been released in the LocalLLaMA community. IBM's open-source model strategy is expanding into the speech domain as well.

The Warp terminal has open-sourced its AI-powered client, mounting a challenge to its closed-source competitors.

DeepMind published a research paper arguing that "AI can simulate consciousness but cannot embody it," earning 58 points on Hacker News.