TLDR AI 2025-03-24

Over 20K attendees roamed downtown San Jose for 2025 GTC last week; AI developers learned from Bria.ai How to build predictable, controllable, on-brand visual content at scale while respecting artists' rights.

See a few of the key topics they covered:

• The Future of a Sustainable Economy Around Visual Generative AI is Built on Licensed Data, Not Scraped (a.k.a stolen) Data.
• Visual AI Cookbooks (GitHub projects) For Visual Generative AI Model Customization
• Explore All Bria APIs for Content Creation

Get your free API token to play with Bria today, with 1000 credits on your new account.

MIT Media Lab and OpenAI are exploring how users engage emotionally with ChatGPT, examining its impact on social and emotional well-being. The study highlights the diverse ways people use AI conversational models and sets the stage for further research on responsible AI interactions.

Hugging Face has revamped its analytics dashboard, which now features real-time updates for monitoring AI inference endpoints. The improved system ensures faster data loading and instant insights into request latency, error rates, and performance metrics.

Perplexity AI envisions a TikTok that prioritizes deep content discovery and truth-seeking powered by an advanced answer engine. It plans to enhance the platform's utility while maintaining its core function as a hub for creative expression.

Nvidia has released a new set of hybrid Mamba style models trained on 20T tokens. They achieve strong performance, especially on long context tasks where they have extreme sequence scaling benefits. This represents the first successful fp8 large scale training of a hybrid attention model.

SynCity is a novel system that generates detailed 3D worlds from text prompts without any training. It combines a 2D image generator for artistic quality and a 3D generator for accurate geometry, building scenes tile-by-tile and seamlessly blending them into a navigable environment.

NuiScene introduces a new method for generating expansive outdoor environments that utilizes uniform vector-set encoding and explicit outpainting for better compression and faster generation.

Code release for a method that takes a single image as input and creates a fully animatable 3D avatar. It is permissively licensed and seems to work quite well.

"Tokenize Image as a Set" is a new framework for image generation that uses set-based tokenization and a novel discrete diffusion method. The approach represents images as unordered token sets, enabling a unique and invertible generative modeling process.

This research investigates whether Large Language Models (LLMs) perform multi-hop reasoning – a process of connecting multiple pieces of knowledge – when presented with complex prompts. The study found evidence of this latent reasoning process, particularly in the initial step, but utilization of the connected knowledge varied and showed a scaling trend with model size, suggesting both challenges and opportunities for future LLM development.

Google's new Gemini 2.0 Flash AI model can remove watermarks from images, raising copyright concerns. The tool, available in Google's AI Studio, is experimental and not suited for production use, yet shows advanced capabilities in watermark removal. Google states that using its generative AI tools for copyright infringement violates its terms of service.

Apple is facing delays and challenges with its AI features for Siri, initially planned for iOS 18.5 but now possibly pushed to 2026. Despite showcasing features at WWDC, internal assessments reveal the technology isn't ready for public release yet. Apple remains optimistic in enhancing Siri, aiming for significant improvements by 2027 with a new infrastructure to stay competitive.

OpenAI's VP Liam Fedus is leaving to start a materials science AI company with backing from OpenAI. His new firm will compete with Google DeepMind and Microsoft in the AI materials science sector. Some experts doubt current AI can achieve groundbreaking scientific discoveries.

DynamicVis is a foundation model for dynamic visual perception in remote sensing that offers efficient parsing of ultra-large images with drastically reduced memory and computation requirements.

Gmail's updated search now uses AI to prioritize your most-clicked emails and frequent contacts for better results.

A set of models that can turn icons and basic images into SVGs.