Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Georgia Tech researchers leveraged Meta's Project Aria to train robots using egocentric video data. The EgoMimic framework reduces the reliance on traditional teleoperation, accelerating humanoid robot learning from human demonstrations.
QMaxViT-Unet+ enhances medical image segmentation using a hybrid Transformer-CNN approach with weak supervision. It outperforms existing methods across multiple datasets, achieving high segmentation accuracy.
An interesting technical deep dive into ways to accelerate the Hadamard product on GPU hardware.
Google Ads has introduced generative AI tools to help advertisers create diverse and engaging assets. New features include AI-generated images of adults, asset audience recommendations, and asset testing for Performance Max campaigns.
Galileo set out to answer the question "How do top LLMs perform in real-world agentic scenarios?" by releasing a leaderboard ranking how well different AI agents effectively utilize tools and functions in complex scenarios. Their analysis of 17 leading LLMs shows Gemini-2.0-flash leading the pack while highlighting rapid improvements in open-source alternatives like mistral-small.
X-Dyna is a zero-shot, diffusion-based pipeline for animating a single human image using facial expressions and body movements derived from a driving video.
Large-scale dataset for general reasoning tasks. It consists of high-quality challenging reasoning questions.
Moondream has announced a phenomenal new small model capability for gaze detection which enables you to determine where someone is looking in a scene.
Stella can automate your admin tasks from reminders & scheduling to research & email drafts.
HP acquired most of Humane's assets for $116M, leading to the discontinuation of Humane's AI Pins. The AI Pins will lose functionality by February 28. Customers are advised to transfer data. Humane's team will form HP's new AI innovation lab, HP IQ.
The former CTO of OpenAI, along with many extremely talented scientists and engineers, have joined to make another AI company. The goals are somewhat vague, but it seems to be a product and foundation model company, with a focus on infrastructure.
Meta is launching LlamaCon, an open-source AI developer conference, on April 29. The event will showcase advancements in the Llama AI model ecosystem, followed by Meta Connect in September for XR and metaverse developments.
DeepSeek has entered the Attention Alternative arena with this novel algorithmic method to speed up quadratic Attention. They get as much as an 11x speed up with no loss in overall performance.
A great benchmark of over 1,400 freelance software engineering tasks from Upwork, valued at $1 million USD total in real-world payouts. SWE-Lancer encompasses both independent engineering tasks - ranging from $50 bug fixes to $32,000 feature implementations - and managerial tasks, where models choose between technical implementation proposals. Top models earned $400k.
Researchers present a quantitative analysis of how ReLU neural networks fold input space, revealing self-similarity patterns. They introduce a novel metric to study these transformations and provide empirical results on benchmarks like CantorNet and MNIST.
Sponsored
Over 5 million tech pros start their day scanning TLDR for the latest news—so when an ad appears, it gets noticed.
A new approach to multi-task learning mitigates negative transfer by dynamically grouping tasks and updating them sequentially during training. The method, based on proximal inter-task affinity, significantly improves performance over existing multi-task optimization techniques.
CAMEL improves reinforcement learning efficiency by integrating LLM-generated suboptimal policies with dynamic action masking.
Perplexity has post-trained R1 to remove Chinese censorship. They do so in a way that doesn't harm underlying reasoning. It is Perplexity's first open weights release.
Google is introducing new tools for flood experts in Flood Hub, including an inundation history map and basin view, while collaborating with aid groups like GiveDirectly and the IRC to assist flood-affected communities.
This article provides a comprehensive overview of xAI's Grok 3.
Great X article by PufferLib maintainer. It dives into the key differences between types of RL and gives a great guide for base hyperparameters.
OpenAI is considering building data center campuses in 16 states as part of President Trump's Stargate initiative, in partnership with Oracle, SoftBank, Microsoft, Nvidia, and Arm, aiming to invest up to $500 billion over four years.
Researchers at Stanford and the University of Washington have trained an AI reasoning model for under $50 using distillation and modifications to an Alibaba AI model.
OpenAI has updated the "chain of thought" for its o3-mini model, enhancing transparency in its reasoning steps for both free and paid ChatGPT users.
Mistral Saba is a 24B parameter model trained on meticulously curated datasets from across the Middle East and South Asia. The model provides more accurate and relevant responses than models that are over 5 times its size while being significantly faster and lower cost.
South Korea has temporarily banned new downloads of China's DeepSeek AI chatbot due to data protection concerns. The app, which gained over a million users in a week, will remain unavailable until compliance improvements are made.
fal has secured $49M in Series B funding led by Notable Capital and joined by a16z and others, totaling $72M in funding to date. The company is focused on expanding its platform for AI-driven generative media, especially video content, catering to industries like advertising and gaming. fal's proprietary technology promises fast, scalable, and reliable deployments, which has already attracted enterprise clients such as Quora and Canva.
Large scale training of a diffusion model for language that matches LLaMA 3 8B in performance across many benchmarks.
Researchers introduce a benchmark that assesses LLM open-ended text generation using n-gram statistics and rules, avoiding reliance on human or LLM-based judgments. It closely correlates with GPT-4o evaluations while being computationally efficient.
CopySpec is a method that accelerates LLM inference by identifying and copying repeated sequences in chat history without extra GPU memory usage. It achieves up to 3.08x speed-up on specific tasks and integrates effectively with speculative decoding for additional gains.
This is the Multimodal Large Language Model (LLM) component of Step-Audio. It is a 130-billion-parameter multimodal LLM that is responsible for understanding and generating human speech. The model is specifically designed to seamlessly integrate functions such as speech recognition, semantic understanding, dialogue management, voice cloning, and speech generation.
This study explores the limitations of existing methods for evaluating the memory capacity of language models. It introduces the "forgetting curve," a new approach for assessing how well models retain information over long contexts.
AdaSimulator provides a physics-enabled environment for studying Visual Language Navigation (VLN) in realistic settings.
A U.S. judge ruled that Ross Intelligence infringed on Thomson Reuters' copyright by using Westlaw headnotes to train its AI. The decision may influence other AI-related copyright lawsuits but remains narrowly focused on non-generative AI applications.
Red Hat emphasizes a pragmatic approach to open-source AI, focusing on practical enterprise needs rather than chasing AGI. Challenges include reconciling transparency with competitive realities, especially regarding open-source AI's opacity in training data and model weights. Red Hat aims to foster collaboration, avoiding vendor lock-in, while acknowledging the complexities of AI compared to traditional open-source software.
Although federal leadership seems wary of heavy AI regulations, many US states are implementing legislation modeled on Europe's AI Act. This article outlines how “algorithmic discrimination” laws, inspired by EU rules, could create detailed impact assessments, require compliance paperwork, and impose liability on AI deployments—raising the real prospect of higher operating costs for teams building AI systems.
Sponsored
Andrew Boyagi of Atlassian went viral when he posted about running a project from start to finish with zero meetings. Want a behind the
OpenAI is deploying ChatGPT to 460,000 students and 63,000 faculty at California State University for personalized tutoring and administrative tasks.
Tinder plans to counter declining user engagement by introducing AI-driven features for discovery and matching.
Google is introducing a digital watermark in Photos for images edited with generative AI using the Magic Editor's Reimagine feature on Pixel 9 devices.
xAI's Grok 3, trained with 200 million GPU-hours, introduces enhanced reasoning, self-correction, and synthetic data training. It is set to be released on Monday.
Anthropic has partnered with the UK government to explore AI applications in public services, emphasizing responsible deployment, economic growth, and scientific research using its Claude model.
OpenAI has updated its AI training policies to reduce content restrictions, emphasizing neutrality, truth-seeking, and offering multiple perspectives on controversial topics.
CodeI/O improves reasoning in large language models by converting code into an input-output prediction format. This enhances performance across various reasoning tasks by teaching universal reasoning principles without relying on code syntax. Further refinement through multi-turn revision boosts accuracy by verifying predictions.
CoCoMix is a pretraining framework that enhances next-token prediction by integrating continuous concepts learned from a sparse autoencoder. It improves sample efficiency, outperforming standard methods in language modeling and reasoning tasks. Additionally, it enhances interpretability by allowing direct inspection and modification of predicted concepts.
A new multiple instance learning framework for whole slide image classification introduces a dual-scale vision-language approach, using a prototype-guided patch decoder and a context-guided text decoder to enhance model performance on pathology tasks.
Sponsored
Modern bots are smarter than ever [11]—executing JavaScript, storing cookies, rotating IPs, and even cracking CAPTCHAs with AI. As attacks grow more sophisticated, traditional detection just isn't enough.