Search the archive

TIDE FOR UNDERWATER SCENE UNDERSTANDING

A text-to-image and dense annotation generation method for underwater scenes that enables high-quality synthetic datasets with consistent pixel-level labels.

EARTH AI'S ALGORITHMS FOUND CRITICAL MINERALS IN PLACES EVERYONE ELSE IGNORED

Earth AI has discovered promising mineral deposits in previously overlooked regions of Australia using AI-driven analysis. Unlike traditional methods, Earth AI's technology scans large areas rapidly to identify potential sites for minerals like copper and cobalt. This innovative approach highlights a shift in mining towards leveraging AI for more efficient exploration.

QUORA'S POE LAUNCHES ITS MOST AFFORDABLE SUBSCRIPTION PLAN FOR $5/MONTH

Quora's chatbot app, Poe, launched new subscription plans at $5/month for 10,000 daily points and $250/month for 12.5 million points.

NVIDIA'S AI ASSISTANT IS HERE TO OPTIMIZE YOUR GAMING PC

Nvidia's Project G-Assist is a real AI assistant for RTX GPU owners that optimizes game settings, measures frame rates, and controls accessory lighting.

NVIDIA IS REPORTEDLY IN TALKS TO ACQUIRE LEPTON AI

Nvidia is in talks to acquire server rental company Lepton AI for several hundred million dollars.

👉 CHECK IT OUT

XAI ACQUIRES X IN $80B ALL-STOCK DEAL

xAI has officially acquired X in an all-stock transaction that values the combined company at over $110 billion.

GEMINI 2.5: OUR MOST INTELLIGENT AI MODEL

Gemini 2.5 Pro, an advanced AI model, is leading LMArena benchmarks by a significant margin. It enhances performance and accuracy through improved reasoning capabilities. The model "thinks" by analyzing information and making informed decisions, building on Gemini 2.0 Flash Thinking advancements.

ANNOUNCING ARC-AGI-2 AND ARC PRIZE 2025

The ARC Prize has launched ARC-AGI-2, a challenging benchmark aimed at advancing general AI systems. Current AIs score significantly lower compared to humans. The accompanying ARC Prize 2025 competition, hosted on Kaggle with a $1 million prize pool, aims to drive open-source innovation by rewarding efficiency and capability in solving ARC-AGI-2 tasks.

TEST-TIME VISUAL IN-CONTEXT TUNING

A novel method to adapt VICL models at test time using only the test sample, improving generalization under domain shifts across various visual tasks.

HARMFUL FINE-TUNING ATTACKS

Researchers highlight vulnerabilities in existing defenses against harmful fine-tuning attacks and propose Panacea, an adaptive perturbation method that preserves model safety while maintaining fine-tuning performance.

HIGH-FIDELITY SIMULTANEOUS SPEECH-TO-SPEECH TRANSLATION

Kyutai has released its next amazing audio system which is a real time, audio to audio, translation system. It is built on a powerful multi stream transformer and boasts expressive voice ability.

Listen to the podcast here.

MOBILE-VIDEOGPT

A lightweight multimodal video model under 1B parameters that features dual visual encoders and token pruning for real-time inference on edge devices.

MULTIMODAL ADAPTATION METHODS

A curated list of approaches for multimodal adaptation that covers traditional domain adaptation, test-time adaptation, and newer methods.

REASONING AUGMENTED GENERATION CODE

Traditional Retrieval-Augmented Generation (RAG) systems rely on a two-step process: first, semantic search retrieves documents based on surface-level similarities. Then, a language model generates answers from those documents. While this method works, it often misses deeper contextual insights and can pull in irrelevant information. ReAG – Reasoning Augmented Generation – offers a robust alternative by feeding raw documents directly to the language model, allowing it to assess and integrate the full context. This unified approach leads to more accurate, nuanced, and context-aware responses.

AWESOME VISION-TO-MUSIC GENERATION

A curated and regularly updated list of methods, datasets, and demos for turning visual inputs into music. It covers both academic and industrial advances in V2M.

OPENAI RESHUFFLES SAM ALTMAN'S JOB ONCE AGAIN

OpenAI has expanded Brad Lightcap's role to oversee operations and partnerships, allowing CEO Sam Altman to concentrate on research and product development. Mark Chen and Julia Villagra have been promoted within the company amid leadership restructuring following recent executive departures. OpenAI is also transitioning to a for-profit model, leading to a lawsuit from cofounder Elon Musk.

TIM COOK SAYS CHINA'S DEEPSEEK AI IS 'EXCELLENT' DURING VISIT

Despite DeepSeek AI's security and privacy issues, Tim Cook praised it as "excellent" during his China visit. The AI, developed in China, rivals top global models at lower development costs but faces investigations in the US and Europe. Cook, who is attending the China Development Forum, often has to make diplomatic remarks about China due to Apple's business interests there.

GOOGLE'S AI-FOCUSED EDUCATION TOOLS AI

Google's new AI-focused educational tools include training for educators, student resources, and expanded access to Gemini for younger users.

MICROSOFT ANNOUNCES SECURITY AI AGENTS TO HELP OVERWHELMED HUMANS

Microsoft has introduced six AI-powered security agents for its Security Copilot to help teams handle phishing and data loss incidents more efficiently.

PERPLEXITY CEO ADDRESSES FINANCIAL RUMORS

Perplexity CEO Aravind Srinivas has denied financial trouble rumors, stating the company has healthy funding and no IPO plans before 2028.

Download the white paper →

ANTHROPIC RESEARCHERS TRACE THE THOUGHTS OF LLMS

OPENAI EXPANDS CYBERSECURITY GRANT PROGRAM

OpenAI has outlined its ongoing efforts to support AI-native cybersecurity research. They now include microgrants and expanded areas like model privacy and agentic security.

CLAUDE 3.7 SONNET USAGE TRENDS IN LABOR

Anthropic's latest report revealed how Claude 3.7 Sonnet is used across coding, education, and healthcare tasks. It offers new insights into how AI augments or automates different occupations.

DIFFUSION-BASED COUNTERFACTUALS

This paper introduces two methods that use diffusion models to generate counterfactual explanations in image regression tasks, highlighting trade-offs in sparsity and quality between pixel and latent space approaches.

MIXTURE-OF-MAMBA

Mixture-of-Mamba introduces modality-aware sparsity to state-space models (SSMs), enabling efficient multi-modal pretraining. It achieves similar performance to Transformers with significantly lower computational cost across text, image, and speech modalities.

OPEN PROBLEMS IN MECHANISTIC INTERPRETABILITY

A great paper that talks through some of the challenges and unsolved problems facing the field when trying to understand how knowledge and skills are represented internally in a language model.

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.

MEASURING INTERNAL BIAS IN VISION MODELS

A new metric that uses attention maps to quantify internal bias in vision models. It is capable of identifying confounding features and going beyond group-based accuracy disparities.

GUIDANCE-FREE TRAINING FOR IMAGE MODELS

Guidance-Free Training (GFT) removes the need for Classifier-Free Guidance (CFG) in visual generative models, reducing computational cost without sacrificing performance. Unlike distillation-based approaches, GFT trains from scratch and requires minimal modifications to existing codebases.

A UNIVERSAL INTERATOMIC POTENTIAL

PET-MAD is a universal interatomic potential model trained on the MAD dataset that is capable of predicting energies and forces for complex materials modeling across the periodic table.

FLUX INPAINTING (HUGGING FACE HUB)

Inpainting support for the new powerful diffusion model Flux.

HUGGING FACE CUSTOM RERANKER TRAINING

This post walks through how to fine-tune high-performing reranker models using Sentence Transformers 4.0, outperforming many large public rerankers with smaller, specialized versions.

AI PREDICTION MODEL IS A MAJOR BREAKTHROUGH IN WEATHER FORECASTING

Aardvark Weather, developed by the University of Cambridge, offers rapid AI-driven weather forecasting on desktop computers while needing less data and computing power than traditional models. It surpasses conventional systems like the U.S. GFS in accuracy across multiple metrics and provides customizable, location-specific predictions. The system's accessibility democratizes forecasting, making it viable even in regions with limited infrastructure.

with DeepSeek R1. Read the blog

AWS GENERATIVE AI EXEC LEAVES TO LAUNCH STARTUP

Raj Aggarwal is leaving AWS, where he served as GM of generative AI, to start a new company.

BROWSER USE, THE TOOL MAKING IT EASIER FOR AI 'AGENTS' TO NAVIGATE WEBSITES, RAISES $17M

Browser Use, a Y Combinator startup, enhances AI agent navigation on websites by converting site elements into a text-like format.

XR TRAINING WITH LLAMA-POWERED SIMULATIONS

Cornerstone is transforming enterprise training with XR and the Llama 3.1 model, enabling the fast creation of personalized simulations that feature intelligent virtual mentors and real-time multilingual capabilities.

INVEST IN MODE'S PRE-IPO OFFERING AND EARN YOUR STAKE IN THIS $1T INDUSTRY.