Good Morning
DeepSeek just dropped a V3 model that runs locally, beats Claude Sonnet 3.5 in coding. Midjourney, best known for images, is now dabbling in words with a creative-writing AI. And Meta has unveiled a new tool that builds full 3D VR scenes from just one photo, taking a big step into immersive content. Here’s what’s new and why it matters.
⚔️ DeepSeek’s New Model Shakes OpenAI
DeepSeek silently released V3-0324, a 685B-parameter model. The 4-bit quantized version runs at over 20 tokens per second on a Mac Studio, delivering top-tier performance locally. It outperforms Claude Sonnet 3.5 and GPT-4o in coding tasks and is available under the permissive MIT license.
Using a Mixture of Experts architecture, the model activates only 37B parameters per task, reducing power and memory demands. The 4-bit version fits in 352GB and runs under 200 watts. It now sounds more like a tool than a companion, signaling a shift toward professional use. A reasoning-focused R2 model may be next.
DeepSeek is challenging U.S. AI giants by going fully open. Its free, high-performance models give developers a major advantage without huge budgets. This could accelerate global AI progress and pressure Western firms to embrace more openness. The AI race just tightened.
✍️ Midjourney Experiments with AI Creative Writing
Researchers from Midjourney and NYU trained models to generate creative writing by rewarding diverse answers. Their model, based on LLaMA-3.1-8B, matched GPT-4o in quality and produced story responses as varied as human writing.
They improved training methods (DPO and ORPO) to favor high-quality, unique replies. Tests showed better diversity and quality, measured by embedding distances. Human evaluations also confirmed the approach produced more diverse and sometimes higher-quality writing. The model performed well even with small datasets.
This research addresses a growing issue: LLMs becoming too generic or repetitive in creative tasks. By making diversity a training goal, the paper offers a way to keep AI-generated content fresh and original. As LLMs become more integrated into writing tools and creative apps, ensuring varied and engaging output is key to keeping them useful, inspiring, and less biased. AI startup Midjourney is famous for its image generation model, could this be its first move into creative writing?
🧊 Meta’s New Tool Builds VR Scenes from Photos
Meta Reality Labs created a way to turn a single image into a 360° 3D world. The method combines panorama generation and 3D reconstruction, producing high-quality VR scenes that beat current top models in sharpness and consistency.
The system works in two steps: it first makes a panorama using inpainting and ControlNet, then builds 3D depth with point clouds and fills gaps with a fine-tuned model. It uses Gaussian Splatting with distortion fixes for clearer results and outperforms top baselines in tests.
Although World Labs’ AI by Fei-Fei Li has already created 3D worlds from a single image, this is the first model for Meta. This model makes it easier to create VR-ready 3D worlds from one image. It cuts training time and keeps quality high. As VR grows, this tool could speed up content creation for games, art, and more.