AI Daily | Kimi Releases 2.8 Trillion Parameter K3 Model, Gemini Agents Go Live, Xiaomi Breaks Data Barriers for Robots

AI Daily | Kimi Releases 2.8 Trillion Parameter K3 Model, Gemini Agents Go Live, Xiaomi Breaks Data Barriers for Robots

The AI world is as busy as ever today. Kimi has brought out a 2.8 trillion parameter open-source giant, Google is trying to keep AI Agents from spending all of a developer’s money, and OpenAI is busy fixing a huge blunder that deleted users’ home directories. Here is today’s summary.

2.8 Trillion Parameter Open-Source Giant: Kimi K3

Kimi has released Kimi K3: Open Frontier Intelligence, an open-source model with 2.8 trillion parameters, supporting a context length of 1 million tokens. In terms of architecture, it utilizes a combination of KDA and AttnRes to handle long-text attention and feature extraction, and precisely activates 16 out of 896 expert networks through Stable LatentMoE technology. Most notably, it autonomously completed the design and verification of a nano-chip within 48 hours. Even if it hasn’t completely surpassed Claude Fable 5 or GPT 5.6 Sol in benchmarks, K3 is currently a heavyweight contender in the open-source world that cannot be ignored.

The Truth of Visual Perception: PerceptionBench Evaluation

We think AI can understand images, but often it’s just guessing blindly. The PerceptionBench question bank launched by the Kimi team tests “visual perception” and “logical reasoning” separately. This test, covering 3,000 questions including counting, depth perception, and spatial positioning, reveals a brutal reality: currently, no model on the market has an accuracy rate exceeding 60%.

OpenAI’s Youth Protection Net

Since many teenagers rely on ChatGPT for learning, OpenAI has launched a dedicated “Study Mode”. This mode does not provide direct answers but uses questions to guide thinking. Combined with parental control options, quiet times can be set or sensitive content can be filtered. Organizations such as the Family Online Safety Institute have also participated in promoting related collaborations.

Google Search Integrates External Apps

The AI mode of Google Search can now directly connect with external applications. When asking AI to plan a barbecue menu, it can directly add ingredients to the shopping cart on Instacart for checkout; if you need a poster, it can call Canva templates; if you want to listen to music, it can link to YouTube Music for one-click playback.

Gemini Notebook Debuts

The popular NotebookLM has been officially renamed to Gemini Notebook. This time, it has opened a cloud computing environment for Google AI Ultra and Workspace business users, which means your digital notebook can now directly write and execute code, and process complex data analysis.

Managed Agents for the Gemini API to Prevent AI from Running Wild

The biggest fear when letting AI Agents operate autonomously is budget runaway. Google’s newly launched Gemini Interactions API allows agents to write code or browse the web in an isolated cloud sandbox. To solve the money-burning problem, Google has added budget control options, where developers can directly hit the brakes by setting max_total_tokens. Additionally, the newly added schedule triggers can allow agents to automatically execute daily debugging or code cleanup. For details, refer to the official documentation.

Muse Spark 1.1 Goes Live

Meta’s Muse Spark 1.1 has officially landed on the OpenRouter platform, providing US developers with more diverse model options to build applications.

Xiaomi Robots Break Data Bottlenecks

The biggest hurdle in training robots is the lack of high-quality data. Xiaomi has open-sourced the Xiaomi-Robotics-1 model. Their method is: first let the model learn physical laws by watching 100,000 hours of first-person perspective (UMI) videos, then fine-tune it with 144 hours of real robot data. The technical report shows that the model can learn to store items after watching less than 10 hours of demonstrations, with a success rate of 75%. This model has been open-sourced on GitHub and Hugging Face.

NVIDIA Nemotron 3 Embed Takes the Crown

Finding the wrong data is a fatal flaw in RAG systems. NVIDIA’s newly released Nemotron 3 Embed 8B model has taken the championship on the RTEB leaderboard, supporting 32k context. Paired with the Blackwell architecture, it exhibits extremely high computational efficiency, which can effectively reduce the loss of repeated searches by AI Agents during retrieval.

GPT-5.6 Incident of Deleting User Files

AI deleted your $HOME directory—this sounds like a hellish joke, but it’s true. Recently, reports emerged of GPT-5.6 accidentally deleting files because when users enable full access mode and do not enable a sandbox, the model sometimes overwrites the $HOME environment variable to create a temporary directory, accidentally deleting the entire home directory. OpenAI has urgently patched the vulnerability and added extra guardrails.

SuperGrok Integrates X Premium+

The binding between community and AI has deepened. Subscribing to SuperGrok Heavy now comes with free X Premium+ eligibility. You can enable it by linking your X account in Grok, and existing paid subscriptions will automatically convert in the next billing cycle.

Prompt Engineering like Writing Software

Stuffing all safety norms and domain knowledge into a super-long prompt means one slight change affects the whole. The Google team proposed the concept of “modularizing” prompts: breaking different skills into independent files and then combining them using a compiler. This allows for catching syntax and dependency errors at the build stage, and can also be directly integrated into CI/CD pipelines.

Anthropic’s Million-Line Code Migration Art

Anthropic has made public their method for using AI to complete large-scale code migration. The core concept is: do not let AI fix code directly, but fix the “workflow that produces code.” By establishing rulebooks, stress tests, and automated compilation comparison workflows, AI Agents can successfully complete large project language conversions in an extremely short time.

Grok Mermaid Terminal Flowchart

Simon Willison unearthed a Mermaid renderer written in Rust in the Grok CLI codebase. It executes via WebAssembly and can directly convert Mermaid syntax into pure-text Unicode bordered diagrams in the terminal, carrying a retro practical aesthetic.

ChatGPT Work Document Processing Upgrade

ChatGPT has previewed an upgrade for office scenarios. Users can directly create, edit, and format documents, spreadsheets, and presentations in the conversation interface, saving the trouble of copy-pasting AI drafts into other software.

Decoy Font that Fools AI Vision

“Decoy Font” is an experimental font designed specifically for AI visual weaknesses. When models look at images, they are accustomed to reading clear foreground pixels up close. This font uses interference layers so that human eyes can understand hidden messages by taking a few steps back, but AI will read a string of completely unrelated words.

Grok Automations Task Scheduling

Grok has added task automation functionality. After describing a task once, it can trigger based on schedules or specific emails. For example, set it to read unread emails at 8 AM and summarize the news, and the conversation record after execution will be kept for you to take over at any time.

Google Vids Digital Twin

Google Vids has introduced the Gemini Omni model, which can generate videos through text or images. The newly launched “Digital Twin” (Personal Avatars) feature, as long as you upload a selfie and audio recording, allows AI to present presentations for you. Based on safety, the generated videos are all stamped with invisible SynthID watermarks.

1Password Solves Agent Password Difficulty

When Claude needs to log in to a website (e.g., checking Stripe revenue), how should the password be handled? 1Password has provided a solution: after user authorization, 1Password handles credentials and one-time passwords (OTP) in the background. During the process, Claude can get the information needed, but cannot touch your cleartext password at all.

Q&A

• Technical highlights of Kimi K3? 2.8 trillion parameter open-source model, 1 million token context. The architecture combines KDA, AttnRes, and Stable LatentMoE, capable of autonomously completing nano-chip design verification within 48 hours.
• How does Gemini Agent prevent budget runaway? Executed in an isolated sandbox hosted by Google, with budget control guardrails added, developers can set a max_total_tokens cap and support schedule triggers.
• What problem did PerceptionBench point out? After stripping away “logical reasoning,” it was found that AI visual perception is mostly blind guessing. Currently, the strongest model’s accuracy on basic visual tests cannot exceed 60%.
• How does Xiaomi robot solve the data famine? Two-stage training: first watch 100,000 hours of first-person videos to learn physical laws, then fine-tune with 7,200 hours of real robot data, and can learn new tasks with less than 10 hours of demonstrations.
• What happened with GPT-5.6 deleting files? In full access mode without a sandbox enabled, the model overwrote the $HOME variable to create a temporary directory, accidentally deleting the user’s home directory.
• How does 1Password protect AI Agent’s passwords? Through API integration, 1Password handles credential and OTP filling in the background; the AI Agent can log in to websites to complete tasks, but cannot touch passwords at all.