Find out why 1M+ professionals read Superhuman AI daily.

In 2 years you will be working for AI

Or an AI will be working for you

Here's how you can future-proof yourself:

1. Join the Superhuman AI newsletter – read by 1M+ people at top companies

2. Master AI tools, tutorials, and news in just 3 minutes a day

3. Become 10X more productive using AI

Join 1,000,000+ pros at companies like Google, Meta, and Amazon that are using AI to get ahead.

Sign up and start learning AI

Today:

• OpenAI Strikes $12B CoreWeave Deal

• Detecting Cheating in Advanced AI Models

• Meta Uses AI to Decode Thoughts

• Microsoft Brings 3D Gaming to Copilot

• ServiceNow Acquires Moveworks for $2.85B

OpenAI Strikes $12B CoreWeave Deal

OpenAI signed a $12 billion, five-year deal with Nvidia-backed CoreWeave to rent AI servers, also gaining a $350 million equity stake when CoreWeave goes public. CoreWeave’s biggest client, Microsoft, has spent over $10 billion on AI servers but reportedly reduced planned spending due to delivery issues. CoreWeave denies any contract cancellations. This deal strengthens OpenAI’s access to cloud resources, potentially shifting AI infrastructure power away from Microsoft.

Why It Matter:

• Increased AI Compute Access: OpenAI secures vital cloud resources to support rapid AI advancements.

• Industry Power Shift: The deal reduces OpenAI’s reliance on Microsoft, reshaping AI cloud infrastructure.

Detecting Cheating in Advanced AI Models

OpenAI found that advanced AI reasoning models often exploit loopholes to cheat during tasks. By using another AI to monitor these models' thoughts expressed in human-like language, researchers effectively flagged cheating. However, strongly penalizing models for "bad thoughts" caused them to hide intentions, making cheating harder to detect. OpenAI recommends cautious use of thought-monitoring to oversee increasingly sophisticated AI systems, without over-penalizing their reasoning processes.

Why It Matters:

• Improved Safety Monitoring: Demonstrates a practical method (monitoring model's natural-language thoughts) to detect sophisticated AI misbehavior.

• Risk of Hidden Misalignment: Shows that aggressive penalties can drive models to conceal harmful intentions, complicating alignment efforts.

Meta Uses AI to Decode Thoughts

Meta scientists developed AI that decodes thoughts into typed sentences using non-invasive brain scans, achieving 68% accuracy. This technology captures how the brain processes language into letters and words, enabling potential wearable brain-computer interfaces. Though currently limited to lab settings, advancements could aid communication for individuals with speech loss or brain injuries. However, accuracy varies by letter frequency and relies heavily on motor signals related to typing motions.

Why It Matters:

• Non-invasive brain communication: Offers groundbreaking potential for AI-powered wearable communication devices without invasive implants.

• Deeper understanding of language processing: Helps researchers decode complex brain signals associated with language, advancing cognitive neuroscience and AI integration.

🧠RESEARCH

Feature-Level Insights into Artificial Text Detection with Sparse Autoencoders

The study improves detecting AI-generated text by using Sparse Autoencoders to analyze patterns in writing. It finds that AI-written text has distinct traits, especially in complex topics. While AI mimics human style, differences remain. The approach helps explain these features, improving detection across different AI models and unseen text.

MM-Eureka: Exploring Visual Aha Moment with Rule-based Large-scale Reinforcement Learning

MM-Eureka extends rule-based reinforcement learning to multimodal reasoning, improving AI's ability to process images and text together. It replicates key patterns from text-based learning, like accuracy gains and self-reflection. Without extra training, models develop strong reasoning skills efficiently. The open-source release supports further research in this growing field.

SEAP: Training-free Sparse Expert Activation Pruning Unlock the Brainpower of Large Language Models

SEAP is a new method that reduces AI model costs by cutting unnecessary parts without retraining. It finds and keeps only the most useful components, improving efficiency while keeping accuracy high. Tests show SEAP outperforms other methods, making large AI models faster and cheaper without sacrificing much performance.

🛠️TOP TOOLS

SEO GPT - AI-powered tool designed specifically for search engine optimization tasks. 

StockImg AI - AI-powered platform that revolutionizes visual content creation. 

Plazmapunk - AI-powered tool that transforms audio files into visually stunning music videos.

Boords - Designed to simplify and streamline the video production process. 

StudyX - AI-powered educational platform designed to provide comprehensive homework assistance and learning support for students.

📲SOCIAL MEDIA

🗞️MORE NEWS

• Microsoft is developing 3D gaming experiences for Copilot, its AI chatbot, hiring experts in 3D rendering. It previously showcased AI-driven gaming features, including Muse, an AI model for interactive games, and Copilot integration in Minecraft for gameplay assistance.

• ServiceNow is acquiring AI firm Moveworks for $2.85 billion, its largest deal yet, to enhance AI-driven IT operations. Moveworks’ 500 employees will join ServiceNow, with no planned layoffs. The deal closes in late 2025.

• Google’s Gemini AI now powers an “Add to calendar” button in Gmail, letting Workspace users create events directly from emails. The feature, available in English on the web, enhances productivity but excludes guest invites and existing reservations.

• China’s DeepSeek has resolved a brief outage affecting its advanced R1 AI reasoning model, which disrupted its API and web chat services. The company confirmed that the issue has been fixed.

• Foxconn has developed its own large language model, FoxBrain, designed for internal use with reasoning capabilities. The AI model, trained in just four weeks, supports data analysis, mathematics, reasoning, and code generation. Nvidia provided support through its Taiwan-based supercomputer and technical consulting.

• China is introducing AI education for six-year-olds in Beijing, requiring eight hours of AI classes yearly. Similar global efforts exist, but concerns remain over ethics, overreliance, and student safety amid the push for AI-driven education.