ByteDance Unveils Goku

Hey there 👋

We hope you're excited to discover what's new and trending in AI, ML, and data science.

Here is your 5-minute pulse...

print("News & Trends")

ByteDance unveils Goku AI image and video creation (5 min. read)

ByteDance and the University of Hong Kong introduce Goku, a state-of-the-art image and video generation model using rectified flow transformers. Trained on 160M images and 36M videos, it achieves industry-leading quality, excelling across text-to-image and text-to-video benchmarks. Goku sets new standards for AI-driven content creation.

U.S. Asserts AI Dominance in Europe Amid Regulatory Tensions (9 min. read)

At an AI summit in Paris, Vice President JD Vance declared the U.S. will lead AI development, controlling chips, software, and regulations. He urged Europe to ease regulations or risk falling behind. The speech underscored a widening divide between U.S. innovation-first policies and Europe's regulatory approach​.

Perplexity Unveils Sonar: A Faster, More Accurate AI Model (8 min. read)

Perplexity introduces Sonar, a blazing-fast in-house AI model optimized for answer quality and user experience. Built on Llama 3.3 70B, it surpasses GPT-4o mini and Claude 3.5 Haiku, approaching GPT-4o performance at 10x the speed. Powered by Cerebras, Sonar delivers instant, factual, and highly readable answers​.

Open source model DeepScaleR surpasses OpenAI’s O1-Preview (11 min. read)

UC Berkeley researchers have introduced DeepScaleR-1.5B-Preview, a language model fine-tuned from DeepSeek-R1-Distilled-Qwen-1.5B using reinforcement learning. It achieves a 43.1% Pass@1 accuracy on AIME 2024, surpassing OpenAI's O1-Preview with just 1.5 billion parameters. The team has open-sourced their dataset, code, and training logs to encourage further advancements in scaling intelligence with reinforcement learning.

Snap Unveils AI Text-to-Image Model for Mobile Devices (4 min. read)

Snap introduced an on-device AI text-to-image model, generating high-resolution images in 1.4 seconds on an iPhone 16 Pro Max. The efficient diffusion model lowers computational costs while powering AI Snaps, Bitmoji Backgrounds, and more. This marks Snap’s deeper push into AI, reducing reliance on external models​.

print("Applications & Insights")

Virtualization and Containers for Data Science Newbies (12 min. read)
Virtualization and containers are game-changers for data science, ensuring consistency and scalability. Virtual machines create isolated environments, while containers streamline deployment with lightweight efficiency. This tutorial shows you how to set up your own virtual machine on your laptop in just a few minutes — even if you have never heard of Cloud Computing or containers before.

How to Measure the Reliability of a Large Language Model’s Response (11 min. read)
How do you measure if an LLM’s response is actually reliable? This guide explores key evaluation techniques like uncertainty estimation, confidence scores, and calibration methods to help you assess model trustworthiness. Learn practical approaches to make AI responses more dependable!

The AI Science and Society Conference 2025 (Video)
The AI, Science, and Society conference kicks off as a prelude to the AI Action Summit in Paris, uniting top researchers, industry leaders, and policymakers. Discussions focus on AI’s role in science, economics, and ethics, tackling challenges like bias, privacy, and regulation, shaping the future of global AI policy and innovation. A major topic of discussion was that of Meta’s Chief AI Scientist Yann LeCun who stated that AGI will only be achieved through world models like JEPA—not LLMs.

print("Research & Advancements")

TOP RESEARCH PAPERS

Building Bridges between Regression, Clustering, and Classification (23 min. read)
This paper introduces a novel method for improving neural network training in regression tasks that involves framing them as classification problems using a learned target encoder-decoder pair. This approach improves regression performance across real-world datasets, offering a fresh take on training models for continuous targets.

Private Federated Learning In Real World Application – A Case Study (15 min. read)
Boosting AI privacy, this study introduces a novel Private Federated Learning (PFL) framework for training models on edge devices without exposing user data. By leveraging neural networks with attention mechanisms, the approach adapts to user behavior while maintaining privacy. Real-world tests confirm its potential for secure, personalized AI applications.

Agency is Frame Dependent (35 min. read)
Agency—the ability of a system to steer outcomes toward a goal—is tricky to define. This paper argues that agency is frame-dependent, meaning its measurement depends on the reference frame. Using reinforcement learning, the authors explore how core properties of agency shift with perspective, reshaping how we study AI and cognition.

TOP REPOS

Data Visualization
microsoft/Data Formulator
☆ 5,5K stars
Microsoft's Data Formulator leverages AI to transform data into compelling visualizations, streamlining the data visualization process. The latest update enhances support for various models, including OpenAI, Azure, Ollama, and Anthropic, broadening its AI capabilities.

Machine Learning Engineering
block/Goose
☆ 7,4K stars
Goose is an open-source feature store for real-time machine learning, built to simplify feature engineering and data management at scale. It enables fast, low-latency feature retrieval and supports batch, streaming, and on-demand data sources. Developed by Block, it aims to streamline ML model deployment and feature versioning.

Knowledge Retrieval
labring/FastGPT
☆ 20,8K stars
FastGPT is an open-source chatbot framework optimized for custom knowledge retrieval and large-scale data handling. It supports RAG (Retrieval-Augmented Generation), enabling enterprises to build AI-powered assistants with multi-turn memory, document indexing, and API integration. Designed for high-speed responses, it's ideal for business intelligence and knowledge management.

That’s it for today!

Before you go we’d love to know what you thought of today's newsletter to help us improve the pulse experience for you.

See you soon,

Andres