Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
OpenAI has introduced "Deep Research," an autonomous research agent within ChatGPT capable of performing multi-step research by synthesizing vast online sources. It is powered by an optimized version of the upcoming OpenAI o3 model.
A new paper from the Anthropic Safeguards Research Team describes a method that defends AI models against universal jailbreaks. A prototype version of the method was robust to thousands of hours of human red teaming for universal jailbreaks, albeit with high over-refusal rates and compute overhead. An updated version achieved similar robustness on synthetic evaluations and did so with a 0.38% increase in refusal rates and moderate additional compute costs.
DeepSeek's AI models, which are significantly cheaper to train compared to other leading models, have disrupted the AI market, potentially challenging Nvidia and other tech giants by showcasing efficient use of resources. This has shaken investor confidence in the AI sector, which traditionally believed that more spending equated to better performance. DeepSeek's success suggests that innovation, rather than just financial investment, could redefine the competitive landscape.
A great and thorough paper that explores how to encourage models to use more thinking tokens. One of the main findings is that with an extremely high-quality curated dataset of 1k examples and by appending “wait” to the end of a thinking sequence, you can encourage models to think longer, which leads to substantially improved performance on math and reasoning tasks.
The modern workhorse of multimodal understanding and generation is learned Tokenizers. These models are typically autoencoder style with a learned discrete codebook. They tend to work well but are very hard to train and require careful tuning of many auxiliary losses. This work shows that with a single diffusion loss, image tokenization is stable, scalable, and higher quality than many traditional methods.
DeepMind researchers analyzed how language models can perform regression tasks by decoding numeric predictions as text and found them as effective as traditional regression models while also enabling flexible density estimation.
Kron is a new optimizer that is making the rounds as a strong alternative to second-order methods. It dramatically outperforms Adam on a number of baselines. This code is a drop-in optimizer for PyTorch.
Oumi is a fully open-source platform that streamlines the entire lifecycle of foundation models, from data preparation and training to evaluation and deployment. Whether you're developing on a laptop, launching large-scale experiments on a cluster, or deploying models in production, Oumi provides the tools and workflows you need.
RaySplats enhances 3D Gaussian Splatting by integrating ray tracing, improving the handling of light and shadows in 3D object rendering while maintaining fast training and rendering speeds.
Great chapter on many policy gradient methods like PPO and GRPO, which can be used for tuning generative auto-regressive models.
OpenAI plans a $500 billion investment in "Stargate," a project aimed at creating AI infrastructure, while economists debate the job-creation claim as automation may perform most computer-based tasks. DeepSeek has made significant advances with self-improving reinforcement learning, potentially leading to rapid AI capability improvements. This underscores China's quick progress in AI, highlighting geopolitical stakes in the technology race.
Nepenthes is a tarpit malware designed to trap and poison AI web crawlers that ignore robots.txt rules. The release of Nepenthes has inspired other tools like Iocaine that aim to hinder AI data collection and impact the industry financially.
DeepSeek, a Chinese AI lab, leveraged tens of thousands of NVIDIA H100 GPUs to develop its R1 model, which competes with top AI models like OpenAI's o1 and Meta's Llama.
David Sacks accused Chinese AI firm DeepSeek of using OpenAI's models to train its own, likening the act to theft.
Jack Dorsey's Block has developed an open-source AI agent named "codename goose" for automating engineering tasks using popular LLMs.
Sponsored
It's never been easier to experience the power of AI-assisted coding. Simply open Copilot in your IDE, choose your AI model, and you're good to go.
OpenAI has released two variants of the o3 model, o3-mini and o3-mini-high. The newest, most cost-efficient models in the reasoning series, they are available in both ChatGPT and the API.
Stanford's "Virtual Lab" uses AI agents as collaborators in scientific research, aiming to tackle complex problems through interdisciplinary efforts. Researchers demonstrate its potential in projects like developing COVID-19 treatments by simulating expert interactions between AI agents. The framework allows scientists to create AI-based expertise, offering a new approach to collaborative research and discovery.
Alibaba's Qwen team has unveiled Qwen2.5-VL, a new AI model rivaling OpenAI's and Google's offerings that excels in text, image analysis, and software interaction. The model is available on Alibaba's Qwen Chat app and Hugging Face, but has restrictions that align it with Chinese regulatory standards. The flagship Qwen2.5-VL-72B requires Alibaba's permission for commercial use by large entities.
PPTAgent offers presentation generation through a two-stage, edit-based approach inspired by human workflows.
1.58-bit FLUX successfully quantizes the FLUX.1-dev text-to-image model using minimal weights, maintaining performance. The method operates without image data, relying on self-supervision. It significantly reduces model storage and memory usage while improving inference latency.
Microsoft has released the benchmark topping synthetic data models on Hugging Face for commercial use due to the MIT license.
Sponsored
💬 Are your users typing? Help them type 30-50% less with features like smart autocomplete, real-time autocorrect, advanced grammar, and magic replies. Typewise [11] runs natively on all platforms, works offline, ensures data privacy by design, supports 40+ languages, and delivers 3-4x better performance than Apple and Google. Try Typewise
GuardReasoner introduces a reasoning-based safeguard for LLMs, improving explainability and generalizability in safety-critical applications. It outperforms GPT-4o+CoT and LLaMA Guard 3 across multiple benchmarks. Its training data, models, and code have been made publicly available.
This project presents a new approach to generating accurate 3D bounding box annotations for traffic lights and road signs, crucial for self-driving cars.
This study explores how language models handle long-document contexts by comparing various extension methods using a controlled evaluation. It highlights that perplexity remains a key performance metric, while approximate attention methods struggle with longer contexts.
Marc Andreessen suggests AI should "crash" wages to achieve an economic utopia, emphasizing productivity gains and reduced consumer prices. His views reflect a broader tech industry logic prioritizing economic transformation over mitigating job market disruptions. Critics highlight the disconnect in tech leaders' visions, which often lack immediate solutions for affected workers.
The announcement of DeepSeek, a Chinese AI app that claims superior performance at lower costs, caused significant drops in tech stocks, including Nvidia. This development raises concerns about the U.S. falling behind in AI, impacting investors and VCs heavily invested in AI startups. As DeepSeek's model challenges existing AI giants, questions about future funding and competitiveness in the global AI race arise.
A great synthetic reasoning dataset from R1 of 114k examples of reasoning tasks that can be used to train strong reasoners via distillation or as a seed to RL cold start.
Quartz has been using AI to generate articles, aggregating content from outlets like CNN and TechCrunch under "Quartz Intelligence Newsroom."
DeepSeek's open-source R1 LLM exhibits strong benchmark performance but struggles with self-identification and erratic responses.
Meta plans to invest up to $65 billion in AI projects, build a massive data center, and expand AI teams by 2025.
Sponsored
Yes, we said it. Customer interviews & financial analysis found that a composite organization experienced benefits of $23.5 million over three years and an ROI of 413% with Dataiku [4]. Plus, 80% time savings on manual processes, reduced costs, and improved decision making on key business activities.
AI2's Tulu-3, a 405B parameter open-weight language model, surpasses DeepSeek V3 and even OpenAI's GPT-4o on key benchmarks.
Mistral has released a very powerful 24B model that achieves strong performance, especially in multilingual data. It is the perfect size for deployment and strength.
Figure AI is establishing the Center for the Advancement of Humanoid Safety to address gaps in safety for robots in workplaces. Led by former Amazon Robotics safety engineer Rob Gruendel, the initiative will focus on testing and certifying robots to industrial safety standards. The company aims to provide transparency with quarterly updates on testing processes and improvements.
Acoupi is an open-source Python framework that simplifies the deployment of AI-based bioacoustic monitoring on low-cost devices. It integrates recording, processing, and real-time messaging.
SliceOcc introduces a novel vertical slice representation for 3D semantic occupancy prediction in dense indoor environments. It achieves state-of-the-art performance using an RGB camera-based model.
Reqo is a new query optimization model that leverages Bi-GNN and probabilistic ML to improve cost estimation accuracy. It introduces an explainability technique that highlights the contribution of query subgraphs.
Sponsored
Tech inboxes are crowded, but TLDR stands out—and so do the ads inside. With over 5 million software devs, founders and other tech decision-makers tuning in daily, this is where ideas, tools, and
VIRUS is a method designed for generating adversarial data that can bypass moderation systems and disrupt the safety alignment of large language models.
Researchers demonstrate that crowdsourced voting on Chatbot Arena can be manipulated to boost or lower model rankings using strategic rigging techniques, impacting the leaderboard's reliability.
Qwen2.5-VL, an amazing new vision language model, has a companion set of cookbooks that show how to use the model for various different tasks.
Generative AI adoption is surging globally, with 66% of leaders prioritizing AI skills over traditional experience. However, access disparities hinder adoption in developing regions, where only a small percentage can leverage GenAI due to limited digital infrastructure. Addressing infrastructure and education gaps is crucial to prevent AI from widening global inequalities.
Researchers from Google DeepMind and LSE conducted a study using a text-based game to explore AI "sentience," testing LLMs by having them choose between options with varying pain and pleasure associations. Findings revealed some models prioritized avoiding pain over scoring points, suggesting a potential framework for assessing AI consciousness.
The second wave of AI coding is advancing, allowing models to prototype, test, and debug code, potentially moving developers into more supervisory roles. OpenAI has entered longevity science with a model that designs proteins to transform cells into stem cells, claiming results surpassing human efforts. Cleaner jet fuels from alternative sources are gaining momentum, promising significant emission reductions and prompting industry shifts.
Omi has raised $2M to develop an AI wearable that enhances mind and productivity.