There’s a reason 400,000 professionals read this daily.

Join The AI Report, trusted by 400,000+ professionals at Google, Microsoft, and OpenAI. Get daily insights, tools, and strategies to master practical AI skills that drive results.

Sign up now for free and work smarter, not harder.

Today:

• DeepSeek Mimics OpenAI Technology

• Microsoft Integrates DeepSeek’s R1 Model

• SoftBank Invests $500M in Skild AI

• ChatGPT Pro Drives OpenAI Revenue

• BrightHeart Uses Meta's DINOv2 for Screening

DeepSeek Mimics OpenAI Technology

OpenAI has expressed concerns that Chinese companies, including DeepSeek, are using its AI models to develop competing products at lower costs. DeepSeek, a Chinese app mimicking ChatGPT's performance, is under investigation for potentially using unauthorized data from OpenAI. The U.S. government is also examining national security risks linked to the app, with the U.S. Navy advising against its use. Meanwhile, DeepSeek has reported cyberattacks and is temporarily limiting new registrations.

Microsoft Integrates DeepSeek’s R1 Model

Microsoft has quickly integrated DeepSeek's R1 AI model into its Azure AI Foundry and GitHub platforms, allowing developers to experiment and incorporate it into their applications. The R1 model, which is cheaper and more efficient to train than competing models, has caused significant market disruption. Microsoft is also investigating whether DeepSeek used OpenAI's data for training. Additionally, a distilled version of R1 will soon be available for local use on Copilot Plus PCs.

SoftBank Invests $500M in Skild AI

SoftBank is negotiating a $500 million investment in Skild AI, a robotics startup focused on developing foundational AI models for robotics. The company, valued at $4 billion, raised $300 million last year. Skild’s AI model is versatile, adaptable across different robotic domains. This investment follows increasing interest in AI-powered robotics, with major figures like Jeff Bezos also backing similar ventures, such as Physical Intelligence and Figure AI, reflecting growing confidence in the sector’s potential.

ChatGPT Pro Drives OpenAI Revenue

OpenAI's ChatGPT Pro subscriptions, which launched seven weeks ago at $200 per month, have quickly become a major revenue driver, surpassing the income generated from its business team subscriptions. The Pro tier is estimated to generate at least $25 million per month, or $300 million annually. This early success could help OpenAI meet its projected $12 billion revenue goal for 2025, following a $3 billion revenue estimate for 2024.

BrightHeart Uses Meta's DINOv2 for Screening

BrightHeart, a medical tech company, uses Meta's DINOv2 AI model to improve the detection of congenital heart defects (CHDs) in fetuses. By leveraging DINOv2’s self-supervised learning, BrightHeart can analyze ultrasound videos more accurately and quickly, helping clinicians identify potential CHDs before birth. This approach led to FDA 510(k) clearance in just two years. The AI-powered software aims to boost prenatal CHD detection, ultimately improving outcomes for newborns with heart defects.

🧠RESEARCH

SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training

This paper compares supervised fine-tuning (SFT) and reinforcement learning (RL) in enhancing model generalization. It finds that RL excels at generalizing across text and visual tasks, while SFT tends to memorize data. RL improves visual recognition and generalization, but SFT is crucial for stabilizing models before RL training.

Optimizing Large Language Model Training Using FP4 Quantization

This paper introduces a new FP4 quantization framework for training large language models (LLMs), addressing challenges like quantization errors. It combines a differentiable quantization estimator and outlier clamping to maintain stability. The framework achieves accuracy similar to BF16 and FP8, offering an efficient solution for ultra-low precision training on next-gen hardware.

Over-Tokenized Transformer: Vocabulary is Generally Worth Scaling

This paper introduces Over-Tokenized Transformers, a framework that improves language modeling by scaling input vocabularies with multi-gram tokens. It shows that larger input vocabularies reduce training loss and enhance model performance without increasing costs. The findings highlight tokenization’s importance in scaling laws and offer insights for better tokenizer design.

🛠️TOP TOOLS

SNAPVID - AI-driven video editing platform designed to simplify the creation of professional-quality short videos for social media.

Automatic 1111- Open-source, user-friendly interface for the Stable Diffusion AI image generation model.

Casper AI - Chrome extension that leverages OpenAI’s GPT models to simplify and enhance digital workflows.

Dubbing AI Voice Changer - Transform users’ voices instantly across various platforms.

Shap-e - Text-to-3D generative AI model developed by OpenAI that creates high-fidelity 3D meshes directly from short written descriptions.

📲SOCIAL MEDIA

🗞️MORE NEWS

• Oumi, an open-source AI platform, was launched by ex-Google and Apple engineers to provide transparent, collaborative tools for AI development. It aims to simplify model building and reduce costs using distributed computing across universities.

• UVeye, an AI-driven vehicle inspection startup, raised $191 million to expand in North America and Europe. Its technology detects 96% of vehicle issues, surpassing manual inspections, and is already used by Amazon and car dealerships.

• SecurityPal, a startup based in Nepal, automates the tedious process of security questionnaires for tech companies like OpenAI and Figma using AI and human expertise. With $21M in funding, it aims to scale globally, focusing on quality and efficiency.

• DeepSeek optimized AI training by using Nvidia's PTX programming instead of CUDA, achieving 10x higher efficiency. This innovation allows faster model training and reduced hardware costs, shaking up the AI industry and challenging Nvidia’s dominance.

• OpenAI and Microsoft blocked suspicious accounts linked to Chinese AI startup DeepSeek last fall, suspecting the unauthorized use of OpenAI's models for distillation, a technique to enhance smaller models.

• DeepSeek’s AI chatbot has made waves in the industry but is found to refuse answering 85% of prompts on sensitive topics related to China, such as the Tiananmen Square protests and Taiwan. The AI’s responses often carry a nationalistic tone, and the model is reportedly easy to manipulate.