🤖Meta’s Behemoth AI

Today we will discuss:

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

• Meta’s big Llama drop—New models, bold claims, and one Behemoth still in the oven

• Don’t miss this week’s Workflow Wednesday—AI personalization in action with AI-generated Ghibli memes, personalized cold emails with Lemlist, and AI-crafted social media posts using Hootsuite’s OwlyGPT;now available at only $20/month! 

• Self-improving AI gets real—DeepSeek and Tsinghua unveil a smarter, leaner model that self-tunes as it goes

• In other AI news today—Microsoft brings Copilot Vision to Windows and mobile; OpenAI may buy Jony Ive and Sam Altman’s AI hardware startup for $500M; and Microsoft creates an AI-generated Quake II demo using its Muse model

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

Meta unveils Llama 4 models, teases record-breaking AI behemoth

Key Points 

• Meta’s Llama 4 Scout and Maverick models outperform rivals like Gemini Flash and GPT-4o in benchmark tests.

• Llama 4 Behemoth, still in training, is designed to outdo GPT-4.5 and Claude on STEM benchmarks.

☕News - Meta has officially rolled out Llama 4, its latest family of AI models now powering Meta AI across the web, WhatsApp, Messenger, and Instagram. 

👨‍💻Here's the lowdown - The lineup includes two new downloadable models: Llama 4 Scout and Llama 4 Maverick. Scout is a compact model that runs on a single Nvidia H100 GPU, while Maverick is positioned to rival heavyweights like GPT-4o and Gemini 2.0 Flash.

Scout boasts a 10-million-token context window and reportedly outperforms models like Google’s Gemma 3, Gemini Flash-Lite, and Mistral 3.1 across a wide range of benchmarks. Meta claims Maverick also goes toe-to-toe with GPT-4o and Gemini Flash in coding and reasoning tasks, achieving similar results to DeepSeek-V3 using less than half the active parameters.

🤓The behemoth is coming - Still in training, Llama 4 Behemoth is set to be Meta’s most powerful model yet, with 288 billion active parameters and a total size of 2 trillion. CEO Mark Zuckerberg says it’s “the highest performing base model in the world,” already surpassing GPT-4.5 and Claude Sonnet 3.7 in STEM-related benchmarks.

Read the full story here.

We’ve just launched the 13th edition of Workflow Wednesday for AI-minded professionals like you—actionable AI workflows delivered straight to your inbox.

This week’s topic: AI Personalization

Join us here

DeepSeek and Tsinghua push for self-improving AI with open source release

Key Points 

• DeepSeek and Tsinghua introduced a new training method that improves model performance while reducing computing demands.

• Their self-principled critique tuning strategy outperformed existing approaches using fewer resources.

☕News - Chinese startup DeepSeek is working with researchers from Tsinghua University to build AI models that can improve themselves while relying on far less training. The aim is to reduce the high operational costs that come with large-scale model development.

✨A new spin on reinforcement learning - The research focuses on a refined version of reinforcement learning that encourages AI models to generate responses that are not just accurate, but also more aligned with human expectations. 

While this technique has shown success in narrower applications, adapting it for general-purpose AI has proven more difficult. DeepSeek’s approach, called self-principled critique tuning, aims to close that gap.

🚀Solid results, open code - Initial results suggest this method doesn’t just work—it works better. According to the joint paper, the new strategy outperformed existing reinforcement learning techniques across a range of benchmarks, and did so with fewer computing resources.

DeepSeek is branding the resulting models as DeepSeek-GRM (short for “generalist reward modeling”) and plans to release them as open source. The move puts it alongside other major players like Alibaba and OpenAI, who are also exploring ways to make AI models more reasoning-capable and self-refining in real time.

Read the full story here.