TLDR AI 2025-03-19

minutes_ - request an invite at together.ai/instant

* RUN INFERENCE FOR MORE MODELS, FASTER – Serverless APIs and Dedicated Endpoints for 200+ open source models [8] including DeepSeek-R1 and Llama.

GOOGLE UPDATES GEMINI APP

Google has released a number of key features like Canvas, Audio overview, and coding assistant to its Gemini app and web browser.

ROBLOX OPEN-SOURCES ITS GENERATIVE 3D MODEL

Roblox introduces Cube, its generative AI system for 3D and 4D model generation. A beta version of Cube 3D mesh generation will be integrated into Roblox Studio and as a Lua API.

AI SEARCH IS STARTING TO KILL GOOGLE'S ‘TEN BLUE LINKS'

Adobe reported a 1,300% rise in AI search referrals for retailers during the 2024 holiday season, indicating a shift from traditional search engines. Engagement metrics show users from AI search stay longer and bounce less compared to traditional referral users. Despite early mishaps, AI search continues to gain traction, with significant consumer use for shopping and research.

SMOLDOCLING

A new, extremely small, state-of-the-art model for document OCR. It can run lightning fast and has sufficient accuracy for many document processing tasks.

REWORDBENCH

This work shows that many popular reward models are brittle to basic rewording of the prompts. It includes a benchmark and a potential strategy to make the reward models more robust.

SINGLE-VIEW 3D SCENE RECONSTRUCTION

Niagara introduces a novel framework for reconstructing outdoor 3D scenes from a single image, combining depth and normal estimation with a geometric affine field and 3D Gaussian decoding.

STABLE VIRTUAL CAMERA

Stability AI has released a great multiview camera system that allows for novel view synthesis. It is powerful, non-commercial, and not quite state-of-the-art, but it runs quickly in just two forward passes.

ZERO-SHOT STEERING FOR 3D/4D SCENE GENERATION

SteerX has proposed a zero-shot inference-time method to improve geometric alignment in 3D/4D scene generation using pose-free feed-forward reconstruction models and geometric reward functions.

HYBRID VISION-LANGUAGE MODEL

MaTVLM is a hybrid vision-language model that integrates Mamba-2 layers into a pre-trained VLM, improving convergence speed and performance.

PERSONALIZE ANYTHING

You can personalize almost any image without additional fine-tuning or training by using token replacement in a Diffusion Transformer.

TRAIN XGB IN THE BROWSER

This blog post describes how to build a browser based training system accelerated by WASM.

GOOGLE'S NEW ROBOT AI CAN FOLD DELICATE ORIGAMI, CLOSE ZIPPER BAGS WITHOUT DAMAGE

Google DeepMind has unveiled Gemini Robotics and Gemini Robotics-ER, AI models enhancing robots' fine motor skills and adaptability for real-world applications. Gemini Robotics integrates vision, language, and action capabilities, allowing robots to perform complex tasks like origami folding. Early results highlight significant improvements in generalization and dexterity over previous models, with Google's model partnering with Apptronik and others for further development.

ask whatever you want. Sign up for free

GOOGLE AND NVIDIA EXPAND AI COLLABORATION

Google announced an expanded partnership with NVIDIA at NVIDIA GTC that will focus on AI infrastructure, hardware optimizations, and applications in fields like healthcare, energy, and robotics.

NVIDIA WON THE AI RACE, BUT INFERENCE IS STILL ANYONE'S GAME

Nvidia currently dominates AI training, but the inference market remains open to competition.

GOOGLE PLAYING AI SEARCH CATCHUP, AND FORMING RELATIONSHIPS WITH CHATBOTS

Google is adding new AI features from Gemini to its search but this is seen as playing catch-up to OpenAI's ChatGPT.