Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
A group of 300+ AI builders including Chris Lattner (Creator of Swift/LLVM) and Amjad Masad (Founder of Replit) met in San Francisco at the Democratize Intelligence Summit (DEMI). DEMI is forming a counter-pressure to closed AI centralization by gathering ambitious builders creating alternative hardware chips, software frameworks, and economic models that make AI easier to build and use.
Google I/O will take place on May 20-21 at Shoreline Amphitheatre and online. Expect keynotes, product reveals, AI innovations, and hands-on sessions.
AI-generated voices can now precisely mimic human speech, raising concerns about distinguishing them from human conversations. A proposed solution is for AI voices to utilize a ring modulator, giving them a historically recognizable robotic sound. Implementing this across voice synthesis technologies would help users identify when they are interacting with AI.
Sponsored
Over half of the respondents from CallRail's 2025 business trends survey are actively exploring ways to automate and scale marketing with AI - from generating ad copy to scheduling appointments. See the
Luma AI has released an all-new model that takes image-to-video generation to the next level, delivering unprecedented natural motion, realistic physics, and coherence.
Seattle-based Ai2 has released Tulu 3 405B, an open-source AI model that outperforms both DeepSeek V3 and OpenAI's GPT-4o in certain benchmarks.
Jared Zhao founded Athenic AI in 2022 to simplify enterprise data analytics by integrating AI that adapts to each company's specific knowledge.
ebook to discover the best ways to make async collaboration a
Sponsored
In 2025, teams are rarely in the same physical space. Even with return-to-office mandates, work is distributed between physical locations (or ‘outsourced' to AI agents). Immediate, in-person interactions are the exception rather than the norm.
OpenAI highlights the rapid adoption of ChatGPT and the broader impact of AI, comparing it to past technological breakthroughs. Its Super Bowl ad positions AI as the next major tool in human progress, emphasizing its role in augmenting creativity rather than replacing human effort.
French President Macron announced a massive AI investment to boost innovation and international cooperation. He emphasized AI's role as an assistant rather than a job replacement, with particular focus on applications in healthcare.
SSI nearly quadruples its valuation from the previous round of funding.
This paper introduces a novel method for improving neural network training in regression tasks that involves framing them as classification problems using a learned target encoder-decoder pair. The approach outperforms traditional regression methods by leveraging the benefits of distributional target representation and smooth interpolation between different objectives.
3.5B model trained to do latent reasoning which also shows dramatic improvement on math problems. The model itself is a recurrent architecture.
While maximizing spectral gap has been used to address over-squashing in GNNs, this paper shows that minimizing it can also improve generalization. The authors propose three new rewiring strategies to enhance label-community alignment, homophily, and computational efficiency.
QuEST allows stable training with 1-bit weights and activations. QuEST achieves this by improving two key aspects of QAT methods: accurate and fast quantization of the (continuous) distributions of weights and activations via Hadamard normalization and MSE-optimal fitting, and a new trust gradient estimator based on the idea of explicitly minimizing the error between the noisy gradient computed over quantized states and the "true" (but unknown) full-precision gradient.
Agentic workflow for generating diagrams with language models.
This work introduces a new conversational framework for programming assistance that integrates coding history, current code, and user instructions.
Wiz researchers discovered unencrypted sensitive data in DeepSeek's open-source AI model, exposing major security vulnerabilities. Despite difficulties in contacting DeepSeek, the databases were secured within an hour after the issue was reported. Wiz highlighted that such security oversights indicate the service is not mature enough to handle sensitive data.
Jiayi Pan and his team at UC Berkeley recreated DeepSeek's R1-Zero model for just $30, showcasing a shift towards cost-efficient AI research. Their open-source model, "TinyZero," highlights potential challenges to big tech's expensive AI infrastructures. The development raises questions about the necessity of large-scale investments in AI.
OpenAI tested ChatGPT's persuasiveness against Reddit's ChangeMyView forum and found that its current AI models rank as more persuasive than humans in 82% of comparisons but still fall short of "superhuman" levels. The company warns of potential risks if AI achieves near-zero-cost human-level persuasive writing, as this could aid biased journalism and scams. OpenAI is implementing measures to monitor and mitigate AI-based persuasion, particularly regarding political influence and large-scale manipulation.
Sponsored
Join Elastic and Unstructured to learn the best practices for scaling the data layer that fuels your vector stores, knowledge graphs, and training pipelines. Get your GenAI apps thriving on clean canonical JSON with less stress and more efficiency for your team. Register to
Anthropic requires applicants not to use AI assistants for writing job applications to evaluate personal interest and unaided communication skills.
Meta's new Frontier AI Framework outlines conditions under which it would not release certain AI systems deemed "high-risk" or "critical-risk."
Chinese AI startup DeepSeek demonstrated its ability to replicate OpenAI's chatbots at a significantly lower cost, reigniting the open-source debate in the AI industry.
Blazingly fast new 1,000 words per second chat assistant from Mistral. It uses Mistral's powerful state-of-the-art coding models and a lovely canvas to assist in many tasks.
Pika Labs has unveiled Pikadditions, a video-to-video AI tool that seamlessly integrates objects and characters into videos while preserving realism.
Google has outlined a policy framework with actionable steps for policymakers to accelerate scientific discovery using AI that emphasizes responsible deployment and collaboration in the research community.
This paper introduces a novel approach to tracking how features discovered by sparse autoencoders evolve across consecutive layers of large language models using a data-free cosine similarity technique that maps feature persistence, transformation, and emergence. The paper demonstrates how the resulting cross-layer feature maps enable direct behavioral control of the models through feature manipulation while providing mechanistic insights into model computations through granular flow graphs.
QLASS introduces a Q-guided stepwise search method for language agents that improves decision-making by providing intermediate rewards. This approach enhances inference efficiency and reduces reliance on annotated data.
FedELC is a two-stage framework designed to enhance federated learning by addressing the issue of label noise in client datasets.
Sponsored
Speed through your next coding project with prediction and precision. GitHub Copilot's [10] new next edit suggestions feature keeps track as you code and predicts the next edit... anywhere in your open file. Try it now, with 2,000 code completions and 50 chat messages a month for FREE!
This repository contains models that score or rate audio files based on a number of metrics. It can be used for retrieval or even as a signal for RL reward.
Facebook developed a toolkit for training human and robot collaboration systems.
CAPA is a metric that assesses model similarity based on shared errors.
Andrej Karpathy has released another amazingly educational video that dives deep into many aspects of developing language models that covers pre-training, hallucination mitigation, and post training.
A new object-centric dataset for 3D deep learning and 3D generative AI.
Dynamic content insertion into videos. DynVFX achieves strong performance for a variety of dynamic content inclusion like water and smoke. It still has a ways to go for more character based content insertion.
Sponsored
Elf Labs holds some of the world's most valuable IP (e.g. Snow White, Cinderella) and is monetizing it with patented AR/VR & AI technology.
Kiva AI produces high quality, scalable, and cost-effective data labeling, annotation, and RLHF for enterprises.
OpenAI announces plans to establish a new office in Munich in the coming months.
The Fine Personas dataset is a huge 21m person dataset extracted from fine-web-edu.
John Schulman, a key figure behind ChatGPT, has left Anthropic after just five months. While his reasons remain unclear, the departure marks a significant shift in the AI research landscape.
Lyft is collaborating with Anthropic to integrate Claude-powered AI solutions, reducing customer service resolution times by 87% and enhancing the overall rideshare experience through AI-driven automation and engineering improvements.
DeepSeek's claim of training a top-performing AI model for $5.6m shook the US market, impacting Nvidia's valuation and raising questions about hefty AI infrastructure investments from the likes of Meta and Microsoft. Despite skepticism, leading tech firms including OpenAI remain committed to major spending on AI advancements. Meanwhile, OpenAI plans a new, free ChatGPT version, maintaining high spending levels alongside ambitious AI projects like Stargate.
Kyutai has released an 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.