This week in AI, we’re seeing major breakthroughs reshaping tech, safety, and even economics:

🤖 X to Get Fully AI-Powered Feed
By Nov 2025, X will run a fully AI-powered feed via Grok, letting users personalize content by voice. Musk will open-source the algorithm biweekly for transparency.

⚖️ Huawei’s DeepSeek R1-Safe
Huawei, with Zhejiang University, launches a model that blocks politically sensitive content with near-perfect accuracy while keeping performance loss under 1%, aligning with China’s AI regulations.

💹 DeepMind’s AI Agent Economies
DeepMind introduces a “sandbox economy” where autonomous AI agents trade and coordinate. The framework highlights risks, governance, and fairness while exploring new systems like auction models and mission economies.

From smarter AI systems to safer governance and even agent-driven economies, the future of AI is accelerating faster than ever.

Musk Confirms First AI-Driven Social Media Feed Coming to X

Elon Musk announced that X (formerly Twitter) will switch to a fully AI-powered feed by November 2025, using Grok, xAI’s assistant, to deliver highly personalized content tailored to users’ niche interests instead of mainstream or political algorithms. Musk said the algorithm will be open-sourced biweekly for transparency, and users will soon be able to adjust their feeds dynamically by simply asking Grok. Alongside this, Grok gained a new “Read Aloud” feature with a natural, human-like voice, building on the launch of Grok 4 Fast, a cost-efficient multimodal AI model with a 2 million token context window. This shift represents Musk’s push to make X an AI-driven “everything app” while addressing long-standing criticism of opaque social media algorithms.

Politics-Aware DeepSeek R1 Model Balances Content Safety

Huawei has introduced DeepSeek-R1-Safe, a safety-focused version of its DeepSeek AI model designed to almost entirely block politically sensitive content and harmful speech, aligning with China’s strict regulatory requirements. Developed in partnership with Zhejiang University and trained on 1,000 of Huawei’s Ascend AI chips, the model claims near-100% success in filtering politically sensitive topics under standard conditions, though its effectiveness drops to around 40% in complex, disguised scenarios. It achieves an 83% comprehensive security defense score, outperforming rival models like Alibaba’s Qwen-235B, while maintaining less than a 1% performance loss compared to the original DeepSeek-R1. This launch underscores Huawei’s commitment to ensuring AI compliance with China’s content controls as part of broader efforts to create politically safe and trustworthy AI systems domestically.

Google DeepMind Shaping the Economic Layer for AI Agents

A "Virtual Agent Economies" by Google DeepMind researchers discusses the rapid rise of autonomous AI agents creating a new economic layer where these agents transact and coordinate at speeds beyond direct human oversight. It introduces the "sandbox economy" framework to analyze this emerging system, focusing on its origins (emergent vs. intentional) and its permeability with the human economy (permeable vs. impermeable). The authors highlight the potential of AI agents to accelerate coordination in various sectors, while also warning of challenges such as systemic economic risk and inequality. They propose design choices like auction mechanisms for fair resource allocation, AI "mission economies" for collective goal achievement, and socio-technical infrastructure to ensure trust and safety. The paper calls for proactive design to steer agent markets toward benefiting humanity’s long-term flourishing by balancing opportunity with robust governance and fairness measures.

Hand Picked Video

In this video, we'll look at building a powerful Open Source Deep Research Agent by combining DeepSeek-R1 and CrewAI.

Top AI Products from this week

• Notion 3.0 - Today we launched Notion 3.0 with Agents. Anything you can do in Notion, your Agent can do too. The busywork that fills your day can now be done in minutes.

• Strata – Strata is one MCP server that guides your AI agents through thousands of tools in multiple apps progressively. It eliminates context overload and ensures accurate tool selection, enabling agents to handle complex, multi-app workflows with ease.

• alphaAI Capital – alphaAI Capital dynamically manages your portfolio using AI, shifting between aggressive and defensive modes as market conditions change. Access fully-automated leveraged ETF, crypto ETF, and tax-aware strategies tailored to your risk profile.

• Pie – Stop choosing between shipping fast and shipping right. Our AI QA team catches what breaks before your users do. Setup in minutes, ship with confidence forever.

• Voiser - AI-powered dubbing & video translation in 140+ languages. Instantly turn any video into multilingual content and connect with a global audience effortlessly

• Humanly AI Interviewer & Talent Partners – Hire better, faster, and more efficiently with Humanly. Our AI Interviewers—supported by human talent partners—help recruiters process 40% more candidates, cut time-to-fill by 11 days, and reduce cost-per-hire by 85% while improving fairness and consistency.

This week in AI

• World's Most Powerful AI Datacenter - Microsoft's $7B Fairwater datacenter in Wisconsin, launching 2026, houses hundreds of thousands of NVIDIA GPUs, delivering 10x supercomputer speed, with sustainable design.

• ShadowLeak AI Data Leak - ShadowLeak is a zero-click, server-side flaw in ChatGPT’s Deep Research agent letting attackers steal sensitive Gmail data silently via crafted emails—fixed by OpenAI in 2025.

• Most Adorable Humanoid Robot - Meet the world’s cutest humanoid robot, blending lifelike expressions, smooth gestures, and AI-driven emotional responses to create a truly charming companion in 2025.### Heading:
  Most Adorable Humanoid Robot

• Google & PayPal Partnership - Google and PayPal partner to revolutionize commerce, combining AI, payment infrastructure, and personalization to create seamless agentic shopping experiences globally.

• Orchids AI Full Stack Engineer - Orchids is the world’s first AI Full Stack Engineer, topping UI Bench & Design Arena, building frontend, backend, auth, DB, and payments without third-party tools. Try at orchids.app.

Paper of The Day

Structured Information for Improving Spatial Relationships in Text-to-Image Generation" presents a lightweight, plug-and-play method that enhances spatial accuracy in text-to-image generation by appending tuple-based structured information to natural language prompts. Using a fine-tuned T5-small model, the approach automatically converts plain text into compact tuples representing objects and their spatial relations, which are then fed into Stable Diffusion XL without changing its architecture. Experiments show significant improvement in spatial alignment, color, and shape accuracy compared to plain prompts and outperform leading baselines like RealCompo and DPT-T2I, while maintaining or enhancing image quality as measured by the Inception Score. The method is portable, requires minimal training data, and achieves comparable or better results than manually created tuples, making it a practical solution to address spatial reasoning challenges in current text-to-image systems.

To read the whole paper 👉️ here.