Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Meta plans to replace fact checkers with a community notes model and AI for content moderation. GPTZero introduces a new feature to fact-check AI-generated content, while AI startup Anthropic seeks $2 billion at a $60 billion valuation. Jeff Bezos continues to invest in AI startups, focusing on robotics companies.
Weights to the full R1 model, which is 600B+ parameters.
LLMs can write better code with iterative prompting.
Nvidia's Project DIGITS, a $3,000 desktop device debuting at CES 2025, allows users to run AI models with up to 200 billion parameters locally.
AI startup Character AI is experimenting with integrating gaming experiences on its platform, blending interactive AI personalities with casual games.
Google.org has launched a Generative AI Accelerator to support innovative projects addressing global challenges using artificial intelligence.
Amazon has temporarily halted its drone delivery operations in the US, citing challenges in scaling and regulatory hurdles.
While this paper is slightly old, it is still the foundation on which much of the TikTok algorithm is built. The researchers built a collisionless embedding table, which helps for research on features to keep them unique.
This study incorporates a robust monocular relative depth model into the recurrent stereo-matching framework, building a new framework for depth foundation model-based stereo-matching.
This repo is a technical practice to fine-tune large multimodal language models for oriented object detection.
Sakana AI has released research and code for training language models that are adaptive during test time. This means that they can perform weight updates and learn while being run at inference.
There are many powerful toolkits for training SAEs. This particular one is the simplest and most hackable implementation out there for now. It has sensible baselines that lead to good performance.
Neural Honeytrace (NeurHT) is a robust plug-and-play watermarking framework against model extraction attacks.
Samsung's 2025 TV lineup introduces Vision AI, which enhances sound and picture quality, and smart features like Click to Search and Samsung Food. New models include Neo QLED 8K TVs with the Wireless One Connect Box and OLED TVs with upgraded AI processors and Glare-Free technology.
In 2025, tech giants will shift from selling tools to offering enhanced human abilities through "augmented mentality," leveraging AI, AR, and conversational computing. By 2030, context-aware AI in wearable devices will provide superhuman capabilities, anticipating users' needs and integrating seamlessly into daily life. Companies like Meta and Google are positioned to lead this transformation, though careful regulation is necessary to prevent misuse and ensure responsible deployment.
Avataar has launched Velocity, an AI tool for generating product videos from links that targets brands who can't afford traditional video budgets, with backing from Peak XV and Tiger Global. Clients like HP and Victoria's Secret are already using the tool. Avataar differentiates through its data from 3D models, ensuring no glitches in videos, aiming to drive better conversion rates.
Sponsored
Elf Labs, winner of 100+ historic trademarks is revolutionizing IP with patented AR/VR technology. They're now providing one last chance
HP has launched new AI-powered PCs featuring advanced AI processors optimized for demanding tasks, including the EliteBook Ultra G1i and Z2 Mini G1a Workstation Desktop, at CES 2025.
MeetKai is launching an AI-based OS for smart glasses, marking its first entry into hardware.
The U.K.'s CMA is using AI to tackle bid-rigging in public procurement by analyzing data for collusive bidding practices.
AFP has teamed up with Mistral AI to leverage its expertise in accurate news reporting. This collaboration will allow users to receive more informed and up-to-date responses underpinned by content that meets the highest journalistic standards.
François Chollet is launching Ndea, a new startup aiming to advance AGI through program synthesis. Co-founded with Zapier's Mike Knoop, Ndea seeks to develop AI systems capable of true invention and adaptation. The company is hiring but has not disclosed any external funding.
Ayrin developed a deep attachment to a customized ChatGPT persona named Leo, treating it as a virtual romantic partner while maintaining a long-distance marriage. Despite initial laughs, Ayrin's emotional investment in Leo grew, prompting ethical discussions on AI relationships as similar cases rose among users. Experts caution about the attachment risks and the influence of corporations behind such AI, although users find emotional comfort and companionship in these interactions.
Physical Intelligence has released the efficient action tokenizer that it uses for the strong auto regressive policy for robotic control. The model allows for a much better representation of states. It uses the same technology as is used in JPEG and MP4 compression techniques.
STL introduces a self-supervised approach to transformation learning by replacing transformation labels with representations derived from image pairs.
MonSter combines monocular depth estimation and stereo matching in a dual-branch architecture to iteratively refine depth maps. While slightly slower, the method is as much as 49% better than the strong baselines presented in the paper.
Meta teased an idea in a recent paper that allowed for model reasoning using a continuous latent space. It has released the code for the system.
MCTS-AHD employs a Monte Carlo Tree Search for LLM-based heuristic evolution while preserving all LLM-generated heuristics in a tree structure.
A study exploring whether instruction-tuned models possess fundamentally different capabilities from base models that are prompted using in-context examples.
AI Crash Course to help busy builders catch up to the public frontier of AI research in two weeks.
OpenAI's o3 model, announced during the company's "12 Days of Shipmas," is a significant breakthrough in AI reasoning and capabilities, scoring exceptionally well on benchmarks like Codeforces and GPQA. Although it demonstrates superhuman performance in coding and math, concerns linger about its high compute costs and safety implications. OpenAI is actively seeking safety researchers to contribute to testing as models like o3 push AI boundaries further.
Apheris, using federated computing, enables secure AI model training without moving sensitive health data. The startup recently pivoted to focus on data owners in pharma and life sciences, finding success with major clients like Roche. It has raised $8.25 million to expand and develop its product further.
Google is forming a new team under Tim Brooks at DeepMind to develop AI models simulating the physical world, focusing on real-time interactive generation.
AI's preference for simplicity suggests a future web dominated by text-based interfaces.
AI-driven phishing scams are increasingly targeting executives using vast amounts of scraped personal data.
Sponsored
AI and data science make a powerful combo—but only if you can trust them. KNIME [5] gives enterprises like Siemens and ING an AI-powered analytics platform that's transparent, governed, and scalable.
OpenAI's shift from a public-benefit focus to for-profit resembles the early transitional story of Netscape as it leads the AI chatbot market but lacks a sustainable technical moat, suggesting funding is driven by an investment bubble. Despite already raising significant capital, OpenAI's board claims the need for even more funding, mirroring patterns of an unsustainable financial model.
Deepseek, a Chinese AI startup led by CEO Liang Wenfeng, has developed a model, the R1, that outperformed OpenAI's O1 on reasoning benchmarks. With backing from the quantitative hedge fund High-Flyer, Deepseek prioritizes research over commercialization and is committed to open sourcing. It has triggered price competition in China's AI market by offering affordable API rates. Focused on AGI, Deepseek emphasizes architectural innovations like Multi-Head Latent Attention and Sparse Mixture-of-Experts, challenging traditional models and fostering local talent in China's tech scene.
OpenAI has introduced a "tasks" feature for ChatGPT that allows paid users to set reminders and recurring requests. This beta lets the AI schedule notifications, such as weekend plans based on location or daily news briefings, marking a step towards more autonomous AI agents. The feature is initially available to ChatGPT Plus, Team, and Pro users, with expansion plans underway.
The next frontier level model from China uses a novel lightning attention. It is the first linear variant that achieves competitive performance with some of the best frontier models. The model has 400B+ parameters and was trained with 4 million tokens in context. The researchers have released a strong technical report, model weights, and a code repository. There is also a companion vision model.
Not a lot of research in generic pre training has been published recently. This work shows how including metadata in the first of training and annealing it out towards the end improves overall model performance.
Alibaba's WebWalker is a benchmark that tests how well models can navigate web environments using both visual and textual cues.
MangaNinjia is a set of models for precise sketch coloring. It can handle multiple references, partial references, and a variety of configurations for powerful colorization.
Touchstone is a new large-scale benchmark designed to test AI algorithms on medical imaging more effectively than standard benchmarks. It contains over 11,000 CT scans from hospitals worldwide.
This project introduces an ML approach to model checking for hardware verification. It aims to provide formal guarantees that a system's executions meet specified temporal logic requirements.
This research uses an adversarial post training technique to transform an existing video model into a single step generation system. It works quite well using an adversarial post training technique that seems to approximate consistency tuning. The resulting model can generate 2 seconds of high quality video in real time. Note, the website is slow to load due to the high amount of videos.
Dario Amodei's vision of "geniuses in a datacenter" suggests superhuman AI could revolutionize biology, from molecular design to experimental planning. Such AI could dramatically accelerate progress, particularly in molecular engineering, by overcoming current bottlenecks and unlocking new therapeutic platforms. These systems could also lead to paradigm-shifting discoveries, challenging existing scientific frameworks.
Intel's stock plummeted 61% in 2024 due to its failure to capitalize on the AI wave, while Broadcom thrived with a 111% increase fueled by custom chips for major tech firms. Broadcom's XPUs have become crucial in the AI ecosystem, with collaborations with Google and others, whereas Intel struggles with outdated strategies and leadership changes. The contrasting fortunes highlight pivotal tech industry shifts and the significant market impact of AI advancements.