Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Most Transformer models are normalized with a layer norm. This can be seen as equivalent to normalizing the model to a sphere. However, if you use carefully placed tanh functions, you can eliminate layer norm while maintaining stability. This is now equivalent to normalizing to a hypercube.
DeepMind has released a paper detailing scaling laws for the powerful cross data center training algorithm DiLoCo. These show how stable model training can be even with gradients synchronized across continents.
WSI repurposes the Whisper ASR encoder for multilingual speaker identification using joint loss optimization. It outperforms Pyannote, ECAPA TDNN, and Xvector in recognizing speakers across diverse languages and environments.
Sponsored
Solda builds AI Sales Agents generating millions of dollars for enterprises – better, faster, and cheaper than humans. Banks and insurers widely use Solda AI to boost revenue.
Toolkit for training VLMs to have improved grounding and reasoning capabilities.
OWL is an agentic framework that seems sensible and performant. It allows for straightforward composition and can even replicate functionality from some closed source agents.
HiRAG introduces a hierarchical knowledge-based approach to Retrieval-Augmented Generation (RAG), improving semantic understanding and indexing in domain-specific tasks.
Luma Chief Scientist Jiaming Song, who created the earliest accelerated algorithm for diffusion models, has unveiled Inductive Moment Matching (IMM)—a new multi-modal pre-training method that surpasses diffusion models with superior sample quality and 10x greater efficiency.
ThunderKittens is a powerful and simple abstraction for writing efficient and performance CUDA kernels. This post explores how to use the framework for the newest Blackwell series of GPUs from Nvidia. The main difference is thinking in terms of data flow.
Many AI models labeled as "open" come with restrictive licensing terms. Google's new Gemma 3 models and similar releases from Meta raise concerns about commercial limitations, potentially impacting smaller companies relying on these technologies.
Sponsored
Find out how German legal AI startup Xayn is developing highly capable language models while adhering to tight sovereignty, security,
Amazon's VP of Artificial General Intelligence, Vishal Sharma, has confirmed AI's integration across AWS, robotics, and Alexa, highlighting the company's extensive AI deployment.
The PANORAMA study is an international initiative evaluating AI models and radiologists in detecting pancreatic cancer in CECT scans.
Leading AI firms like OpenAI, Microsoft, and Meta are using "distillation" to create more cost-effective models by training smaller systems with a "teacher" LLM.
OpenAI is lobbying the Trump administration for minimal AI regulation, emphasizing the need for rapid innovation while raising concerns about AI competition from China. The company aims to shape the upcoming AI Action Plan drafted by the government.
Anthropic has introduced token-efficient improvements to its API, including cache-aware rate limits and enhanced prompt caching for Claude 3.7 Sonnet, reducing token usage by up to 90%.
Google's Gemini now offers personalization through its experimental Gemini 2.0 Flash Thinking model. It integrates with your Google apps, starting with Search, to tailor responses based on your past searches. Future expansions will include personalization with Photos and YouTube, enhancing AI assistance by understanding user preferences.
Cohere has trained and released an open weights model with 111B parameters. It excels in agentic, multilingual, and coding use cases. It has also been specifically tailored to enterprise tasks such as retrieval.
AnyMoLe generates motion in-between frames for arbitrary characters using video diffusion models, eliminating the need for character-specific datasets.
MMRL enhances few-shot adaptation of vision-language models by introducing a shared representation space that improves multi-modal interactions while maintaining generalization.
Sponsored
AI is only as powerful as the processes behind it. Without the right orchestration, AI initiatives can struggle to deliver real business value.
A new state-of-the-art audio understanding model built on Qwen with almost entirely synthetic data.
Agent S is a powerful and open computer use system. It achieved state-of-the-art performance on browsers, system, and even mobile tasks.
REF-VLM unifies visual decoding tasks in multimodal LLMs using a structured triplet-based representation.
Game Art Forge has introduced AI-generated templates to streamline game asset creation, enhancing speed, scalability, and creative control for developers. It allows customization while maintaining consistency and supports both indie developers and larger teams. By combining AI efficiency with human creativity, it ensures high-quality, adaptable workflows for game development.
The Open Sora effort, which has been underway since the original launch of the model, has trained a competitive model for less than $200k. It has released all of the code and weights for reproducing the results. The motions are quite compelling, although not completely state-of-the-art.
Nous Research has launched an Inference API, making its personalized, unrestricted language models accessible to developers and researchers. The API supports two models, Hermes 3 Llama 70B and DeepHermes-3 8B Preview, while a waitlist system manages demand and adds exclusivity. This launch marks Nous Research's transition from open-source releases to commercial offerings. The company is aiming to create a sustainable business model without compromising its core principles.
OpenAI's policy proposal urges the U.S. to ban Chinese AI lab DeepSeek's models, citing security risks due to alleged PRC control.
Google DeepMind's Gemini Robotics are a set of AI models that allow robots to interact with objects and environments.
Snapchat Platinum subscribers can now access AI-generated video lenses featuring animated foxes, raccoons, and flowers.
Sponsored
Ai2 is a nonprofit research institute dedicated to developing truly open AI solutions to solve the world's biggest problems. To continue innovating, they needed to scale compute performance and prepare a data center with significant room for growth.
Game Art Forge introduces AI-generated templates to streamline game asset creation, enhancing speed, scalability, and creative control for developers. It allows customization while maintaining consistency and supports both indie developers and larger teams. By combining AI efficiency with human creativity, it ensures high-quality, adaptable workflows for game development.
The AI scientist from Sakana Labs has its first reviewed and accepted publication. It generated the idea, conducted experiments, and wrote a novel paper that was accepted to an ICLR workshop (with the full consent of the conference organizers).
Google has released the weights and a technical report for the Gemma 3 model. This model, which comes in four sizes, is approximately as performant as Gemini 1.5 Pro. It also has a strong understanding of 140+ languages. It seems to be close to the state-of-the-art dense model.
Google has released the best dense models available. They score 9th in the arena, have phenomenal multilingual capabilities, and support more than 140 languages. Depending on the upcoming open launches, these may be the backbone of the community. They certainly are the best single GPU models and likely will be for some time.
TaylorSeer introduces a method to predict future timestep features in diffusion models using Taylor series expansion, significantly reducing errors in feature caching.
CATANet enhances image super-resolution by aggregating long-range content-similar tokens.
This project explores how language models perform stepwise implicit reasoning but struggle with generalization, especially when dealing with expressions containing variables as subtrahends.
Motion Anything improves conditional motion generation by introducing Attention-based Mask Modeling for better spatial and temporal control.
VideoPainter introduces a dual-stream architecture for video inpainting, significantly reducing learning complexity while improving background preservation and object generation. It also provides VPData and VPBench, the largest video inpainting dataset and benchmark to date.
PromptPex is a developer tool that treats prompts as functions and automatically generates test inputs, enabling systematic unit testing of AI model prompts.
A nice unification and simplification of diffusion models for continuous data. They use a new moment matching framework to create a few step generative model.
The Hugging Face team has announced the next update for its open replication of the DeepSeek reasoning model. It found that small models can outperform the larger ones when tuned specifically for competitive coding.
Sponsored
Oxford Languages hosted Yasmin Greenaway, founder of Dermie.AI [18], for a discussion of AI in healthcare - from data privacy to the latest
Wan is a generative image model that has recently been open sourced.
Another LoRA for the Wan video model.