🤩Microsoft’s New AI Needs No GPUs

Today we will discuss:

• Sponsored: Writer—End-to-end agentic AI in action; delivering AI adoption and ROI at scale

• Microsoft’s new 1-bit model—New model runs on CPUs, beats rivals in tests, but depends on custom hardware support

• Don’t miss this week’s Workflow Wednesday—AI problem solving made easy with AI-powered reflection using Mindsera, readability boosts from Hemingway Editor, and automated downtime alerts via Make.com + OpenAI + Gmail; now available at only $20/month! 

• New OpenAI models—o3 and o4-mini with enhanced reasoning and image capabilities

• In other AI news today—ChatGPT's latest trend involves reverse location searches from photos, xAI introduces a memory feature for Grok, the Trump administration mulls a ban on DeepSeek, and Wikipedia teams up with AI developers to combat bot scrapers 

• Saved the best for last—10 must-try AI tools

Microsoft unveils new 1-bit AI model that runs without GPUs

Key Points 

• BitNet b1.58 2B4T compresses model weights into just three values, making it faster and more memory-efficient.

• The model outperforms Meta’s Llama, Google’s Gemma, and Alibaba’s Qwen models in math and commonsense tests.

👨‍💻News - Microsoft researchers have developed BitNet b1.58 2B4T, the largest 1-bit AI model created so far. It is now available under an MIT license and can run on CPUs, including Apple's M2 chip. Unlike most AI models that need powerful GPUs, this one is designed to work on more everyday hardware.

👾What are Bitnets & Why they matter - Bitnets are a new kind of compressed AI model. They work by simplifying the model’s internal values, called weights, into just three options: -1, 0, and 1. This makes them much faster and more memory-efficient than traditional models, which usually need more bits to store their weights.

🤖How it performs - BitNet b1.58 2B4T has 2 billion parameters and was trained on a massive dataset of 4 trillion tokens, about the same as 33 million books. In tests, it performed better than other well-known models like Meta’s Llama 3.2 1B, Google’s Gemma 3 1B, and Alibaba’s Qwen 2.5 1.5B on tasks like math problems and commonsense reasoning. It also runs faster and uses much less memory compared to models of a similar size.

To hit those impressive numbers, BitNet relies on Microsoft’s custom framework, bitnet.cpp, which is currently limited to select hardware, a major gap given today’s AI ecosystem.

🌟The bottom line - Bitnets could be a game-changer for running AI on resource-constrained devices, but broader compatibility challenges still loom large.

Read the full story here. 

We’ve just launched the 15th edition of Workflow Wednesday for AI-minded professionals like you—actionable AI workflows delivered straight to your inbox.

This week’s topic: AI Problem Solving

Join us here

OpenAI launches o3 and o4-mini with improved reasoning capabilities

Key Points 

• OpenAI launched o3 and o4-mini, offering advanced reasoning, image processing, Python execution, and web browsing features.

• Along with the models, OpenAI introduced Codex CLI, a tool for developers to integrate AI with local coding tasks.

🚀News - OpenAI has launched two new AI reasoning models, o3 and o4-mini. 

o3 is being touted as the company’s most advanced model yet, outperforming earlier versions across math, coding, reasoning, science, and visual understanding tests. On the other hand, o4-mini offers a strong balance of price, speed, and performance, making it an appealing choice for developers.

Both models can now perform tasks using ChatGPT tools like web browsing, Python code execution, and image generation. They also bring a major update: the ability to analyze images. Users can upload images, such as whiteboard sketches or diagrams, and the models will process them during their reasoning phase, even handling blurry or low-quality images.

✨What’s more? Alongside these models, OpenAI introduced Codex CLI, a new coding agent designed to run locally from terminal software. Codex CLI connects OpenAI’s models with local coding tasks, allowing them to write, edit code, and even move files. While it doesn’t fully build apps just yet, it’s a step toward OpenAI’s goal of creating an “agentic software engineer” that can handle complete app development projects. Read more here.

🤓What’s next? In the coming weeks, OpenAI will also release o3-pro, an upgraded version of o3 for ChatGPT Pro users that will use more computing resources for its responses.

Read the full story here.