OpenAI Introduces GPT-4.5

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")

OpenAI Introduces GPT-4.5: Smarter, More Intuitive AI (7 min. read)

OpenAI has introduced GPT-4.5, its largest and most advanced AI language model to date. Designed to enhance pattern recognition and generate creative insights without explicit reasoning, GPT-4.5 offers a more natural conversational experience with improved emotional intelligence. While it excels in tasks like writing and problem-solving, it does not introduce significant new capabilities over previous models. Rolling out now to all ChatGPT Pro users, followed by Plus and Team users next week, then Enterprise and Edu users the following week.

Google Releases Free Gemini Code Assist (4 min. read)

Google has launched a free version of Gemini Code Assist, an AI-powered coding assistant designed to help developers write, debug, and optimize code more efficiently. It provides real-time code completions, suggestions, and AI-driven reviews within popular IDEs like VS Code, JetBrains, and GitHub. By leveraging Gemini 2.0, it enhances productivity by automating repetitive coding tasks and improving code quality.

DeepSeek Open-Sources DeepGEMM, the Library Powering V3/R1 Training and Inference (7 min. read)

DeepSeek has open-sourced DeepGEMM, the library that optimized V3/R1 training and inference on NVIDIA's Hopper GPUs. This CUDA-based library delivers highly efficient FP8 General Matrix Multiplication (GEMM) kernels with fine-grained scaling, supporting both standard and Mix-of-Experts (MoE) grouped GEMMs. DeepGEMM compiles kernels at runtime via a lightweight Just-In-Time (JIT) module, eliminating the need for pre-compilation. Designed for simplicity, its core kernel function is just ~300 lines of code, making it easy to use and integrate.

Alibaba Unveils Wan2.1: Open-Source Video Generation Model Surpassing OpenAI's Sora (9 min. read)

Alibaba has released Wan2.1, an open-source video generation model that achieves 2.5 times faster generation speeds compared to OpenAI's Sora. Wan2.1 can generate videos based on text, images, and other control signals, offering a versatile tool for creators. This release underscores Alibaba's commitment to advancing AI technology and providing accessible solutions for developers.

ElevenLabs Launches, Scribe, the Most Accurate Speech-to-Text Model (3 min. read)

ElevenLabs introduces Scribe, a cutting-edge automatic speech recognition (ASR) model supporting 99 languages with word-level timestamps, speaker diarization, and event tagging. It outperforms Whisper Large V3, Gemini 2.0 Flash, and Deepgram Nova-3 in benchmarks like FLEURS and Common Voice, achieving 96%+ accuracy in English, Italian, and underserved languages. Available via API and ElevenLabs' dashboard, with a low-latency version in development.

print("Applications & Insights")

How I use LLMs (Video)
Andrej Karpathy explores the many ways he integrates LLMs into his daily workflow, from basic interactions and deep research to advanced applications like image processing, coding assistance, and video generation. He covers ChatGPT, Claude, NotebookLM, and tools like DALL·E and Sora, offering a comprehensive look at real-world LLM use cases.

Evaluating AI Agents (Course)
DeepLearning.AI offers a short course, "Evaluating AI Agents," in partnership with Arize AI. The course teaches systematic assessment and improvement of AI agents through observability, component-wise evaluations, and structured experimentation. Participants will learn to build agents, visualize their processes, and implement effective debugging strategies.

Managing LLM inplementation projects (8 min. read)
Piotr Jurowiec discusses strategies for implementing Large Language Models (LLMs) from initial scoping to deployment. The article addresses challenges such as data handling, model selection, and workflow optimization, providing insights into efficient application development with LLMs.

Automate Jira project management with Gemini 2.0 and Crew AI for efficient data extraction and updates (7 min. read)
Explore ways to automate Jira project management using Gemini 2.0 and Crew AI. It details how AI agents can autonomously extract data from Jira, perform analyses, and send status updates via Slack, enhancing efficiency in managing complex projects.

print("Research & Advancements")

TOP RESEARCH PAPERS

Minions: the rise of small, on-device LMs (18 min. read)
Minions is a protocol that enables small on-device LMs to collaborate with powerful cloud models, reducing AI computing costs. By offloading tasks locally, the approach achieves 97.9% of cloud-only accuracy at 17.5% of the cost, making AI cheaper, faster, and more accessible for real-world applications.

System 2 Thinking in LLMs (28 min. read)
A survey on reasoning LLMs like OpenAI's o1/o3 and DeepSeek's R1 that analyzes their step-by-step logical reasoning capabilities and benchmarks their performance against human cognitive abilities.

SmolVLM2: Bringing Video Understanding to Every Device (4 min. read)
Hugging Face introduces SmolVLM2, a family of lightweight video language models with 2.2B, 500M, and 256M parameters, optimized for broad accessibility. The 2.2B model excels in math, text recognition, and diagram comprehension, outperforming peers. Fully compatible with Transformers and MLX, it’s now available on Hugging Face.

TOP REPOSITORIES

AI Research Tools
langchain-ai/open_deep_research
☆ 1.7K stars
Open Deep Research is a customizable web research assistant that generates comprehensive reports on any topic, allowing users to tailor models, prompts, report structures, search APIs, and research depth.

Conversational Data Analysis
sinaptik-ai/pandas-ai
☆ 16.8K stars
PandasAI enables conversational data analysis by integrating large language models (LLMs) with your database or data lake, supporting SQL, CSV, and Parquet formats.

Multi-Agent Systems
geekan/MetaGPT
☆ 48.3K stars
MetaGPT is a multi-agent framework that assigns different roles to GPTs, forming a collaborative entity capable of tackling complex tasks, moving towards natural language programming.

Model Training Frameworks
jingyaogong/minimind
☆ 13.5K stars
MiniMind provides a framework to train a 26-million-parameter GPT model from scratch in just 2 hours, offering an accessible approach to building small-scale language models.

GUI Automation Tools
microsoft/OmniParser
☆ 18.2K stars
OmniParser transforms UI screenshots into structured, machine-readable elements, enabling LLMs to interact with graphical interfaces more effectively. The latest V2.0 update introduces a cleaner, expanded dataset, 60% lower latency, and a 39.6% accuracy boost on the ScreenSpot Pro benchmark, making it a powerful tool for automating GUI-based AI workflows.

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