Search the archive

RECONSTRUCT LARGE 3D SCENES

Momentum-GS is a method that enhances 3D Gaussian Splatting for large-scale scene reconstruction.

Abstraction

→ Check out pgai on GitHub [13]: a framework for developing RAG, semantic search, and other AI applications directly in PostgreSQL

OPEN ALIGNMENT

Open Alignment for Transformers (OAT) is a toolkit for aligning language models.

TURNING A SINGLE IMAGE INTO A 360° 3D SCENE

The PanoDreamer method transforms a single image into a complete 360° 3D scene by combining panorama and depth estimation.

STEREO MATCHING WITH MONOCULAR DEPTH INSIGHTS

Stereo Anywhere is a new framework that fuses stereo-matching techniques with priors from monocular depth models, addressing challenges like textureless regions and occlusions.

SALESFORCE AGENTFORCE: LLMS MAY HAVE A KILLER ENTERPRISE APP: ‘DIGITAL LABOR'

Salesforce's new platform, Agentforce, has rapidly gained traction with over 200 deals closed in one quarter, signaling strong enterprise interest in AI agents for sales and customer service. CEO Marc Benioff envisions AI agents leading to a vast, unlimited digital workforce, with plans to deploy a billion AI agents next year. The company foresees increased revenue, largely driven by its AI offerings, and is expanding its sales team to meet the growing demand.

DEEPMIND'S GENCAST AI IS REALLY GOOD AT FORECASTING THE WEATHER

DeepMind's GenCast AI outperforms existing weather forecasting systems, like ECMWF's, with significant improvements in accuracy and efficiency. The diffusion model, trained on 40 years of data, uses probabilistic predictions and requires less computing power than traditional methods. It has some limitations in predicting hurricane intensity. GenCast is open-source and will integrate with Google Earth.

AI HELPS RESEARCHERS DIG THROUGH OLD MAPS TO FIND LOST OIL AND GAS WELLS

Researchers at Berkeley Lab use AI to identify undocumented oil and gas wells by analyzing historical USGS maps. Their AI tool has identified over 1,300 potential undocumented wells across California and Oklahoma. This AI-driven mapping is part of a broader initiative to mitigate methane emissions and hazards from orphaned wells.

MAGEBENCH LEADERBOARD

MageBench has introduced a new benchmark to evaluate multimodal agents' reasoning and planning abilities in scenarios where visual signals are continuously updated.

PLAYSTATION BOSS: AI WILL NEVER REPLACE 'HUMAN TOUCH'

PlayStation's CEO, Hermen Hulst, says that while AI can revolutionize gaming by automating mundane tasks, it will not replace the human touch in game creation.

AI PIN MAKER HUMANE DEMOS AI SOFTWARE FOR CARS, PHONES, AND SMART SPEAKERS

Humane revealed CosmOS, an AI operating system that enhances tech devices with agent-like capabilities.

XAI'S IMAGE GENERATOR

xAI's Aurora is an advanced image generation model integrated into Grok 2.

OPENAI'S REINFORCEMENT FINE-TUNING RESEARCH PROGRAM

Reinforcement Fine-Tuning allows developers to train AI models to excel at specific, expert-level tasks by grading their responses against reference answers.

OPENAI 12 DAYS OF “SHIP-MAS”

OpenAI's 12 days of “ship-mas” has officially begun, with the company set to reveal some new features, products, and demos during all 12 days starting December 5, just a few days shy of the second anniversary of ChatGPT's explosive launch in 2022.

ALIGN3R

An improved alignment method that has consistent depth estimation for videos. It is built on Dust3r and shows strong performance in 3D estimation.

IS IMAGENET WORTH ONE LONG VIDEO?

By training an image encoded, unsupervised, on a single long walking video, this work shows how clever model changes can result in extremely powerful representations.

FLASHATTENTION ON A NAPKIN

FlashAttention is an extremely performant software implementation of attention. It is hardware aware and prevents unnecessary I/O. However, it can be challenging to understand. This paper aims to clarify and simplify the algorithm with diagrams and explanations.

Looking for the most efficient ways to run LLM inference? Lambda

compared the performance of a single instance NVIDIA GH200 Grace Hopper Superchip to a single instance NVIDIA H100 SXM GPU, running Llama 3.1 70B inference with CPU Offloading. See the full benchmark

results here

CLEARVOICE-STUDIO

Unified platform for audio separation, speech understanding, and speech enhancement.

DOCOWL

OCR-free document understanding with multimodal LLMs. It has strong chart understanding, table extraction, and more.

TRELLIS: STRUCTURED 3D LATENTS FOR SCALABLE AND VERSATILE 3D GENERATION

3D image and text generation models from Microsoft. These are the strongest models in the space today as they excel at 3D occlusions.

OPENAI MAY LAUNCH SORA, ITS TEXT-TO-VIDEO MODEL, VERY SOON

OpenAI plans to introduce new AI features, including a text-to-video tool called Sora and a reasoning model, over a 12-day livestream event. Sora has faced criticism regarding data provenance, with concerns over the potential use of YouTube content without permission. Google is also developing a text-to-video tool, Veo, currently available for private review.

GOOGLE'S NEW GENERATIVE AI VIDEO MODEL IS NOW AVAILABLE

Google's Veo, a generative AI video model, is now available via Vertex AI for businesses. It creates high-quality 1080p videos from text or images. Veo includes safeguards and DeepMind's SynthID digital watermark to address copyright and misinformation concerns. Google has expanded access to Imagen 3 for text-to-image generation on Google Cloud, with new features for brand customization.

ELON MUSK'S XAI TO EXPAND COLOSSUS SUPERCOMPUTER, BOOSTING MEMPHIS AS EMERGING AI HUB

xAI is expanding its Colossus supercomputer facility in Memphis, adding one million GPUs to strengthen its AI capabilities. This move positions Memphis as a potential global hub for AI innovation, attracting major companies like Nvidia and Dell. The Greater Memphis Chamber supports this growth and has established a special team to expedite xAI's expansion.

here

OPENAI AND ANDURIL PARTNER ON DEFENSE AI APPLICATIONS

OpenAI has partnered with Anduril Industries to develop AI solutions for military applications, focusing on counter-drone defense systems.

META QUIETLY LEANS ON RIVAL GPT-4 DESPITE ZUCKERBERG'S BOLD LLAMA CLAIMS

Meta's hybrid approach combines Llama and GPT-4 to optimize generative AI capabilities.

AWS BRINGS PROMPT ROUTING AND CACHING TO ITS BEDROCK LLM SERVICE

AWS announced caching and intelligent prompt routing for its Bedrock LLM hosting service to optimize cost and performance.

Start building with Writer AI STUDIO.

OPENAI O1 SYSTEM CARD

This report outlines the safety work carried out prior to releasing OpenAI o1 and o1-mini, including external red teaming and frontier risk evaluations according to OpenAI's Preparedness Framework.

PERPLEXITY EXPANDING IT'S PUBLISHER'S PROGRAM

Perplexity has added over a dozen international news organizations to its Publishers' Program, offering tools, revenue sharing, and support to strengthen collaboration with global media.

DEEPMIND'S GENIE 2 CAN GENERATE INTERACTIVE WORLDS THAT LOOK LIKE VIDEO GAMES

DeepMind's Genie 2 is an advanced AI model that generates diverse, interactive 3D worlds from images and text descriptions. The model is particularly adept at simulating complex elements like physics and NPC behavior, setting it apart from similar models by maintaining scene consistency. Positioned as a research tool, Genie 2 is designed for prototyping interactive experiences and evaluating AI agents.

PALIGEMMA 2

Paligemma 2 is one of the best VLMs on the market today. It uses SigLIP and Gemma.

LAND COVER MAPPING WITH ASANET

The Asymmetric Semantic Aligning Network (ASANet) improves land cover classification using both SAR and RGB images.

FASTER MULTI-MODAL MODELS WITH TOKEN MERGING

Researchers have developed a training-free method to make multi-modal language models (LLMs) more efficient without sacrificing much performance. Their approach reduces computational demands by up to 7 times by smartly merging and pruning visual data tokens.

continuous change. Read the guide

GRAPHCAST

DeepMind has open sourced its GraphCast algorithm, which is substantially better and faster at localized weather prediction at 36 hours. It also runs in a fraction of the time.

CREATING OPTICAL ILLUSIONS

This project advances visual anagram generation - images that change appearance when flipped or rotated - by refining how diffusion models handle multiple viewpoints.

3D SCENE COMPLETION FOR AUTONOMOUS VEHICLES

ScoreLiDAR is a new method that speeds up 3D LiDAR scene completion for autonomous vehicles.

NEW FISH AUDIO MODEL

The newest Fish Audio 1.5 is currently #2 on Text To Speech Leaderboards behind ElevenLabs. It supports voice cloning and is fast to run, although the output quality seems hit or miss.

OPENAI PARTNERS WITH ANDURIL

OpenAI and defense company Anduril have partnered to create a national security-based AI.

KEY LEADERS BEHIND GOOGLE'S VIRAL NOTEBOOKLM ARE LEAVING TO CREATE THEIR OWN STARTUP

Three key Google NotebookLM members have left to start a new stealth AI venture. The startup aims to leverage the latest AI models to create consumer-facing, user-first AI products. It remains in its early stages without a clear focus or announced funding.