Pulse

Blender is implementing new policies regarding the Blender Development Fund and the use of AI. These updates aim to clarify guidelines and potentially support AI-related initiatives within the Blender ecosystem. Further details on the specific changes and their implications are expected. AI

SenseNova-U1 is a newly released open-source multimodal AI model capable of processing diverse visual inputs like screenshots, PDFs, and handwritten notes. It can perform tasks such as visual question answering, document parsing, chart comprehension, and OCR within a single model. Additionally, SenseNova-U1 supports text-to-image generation, image editing, and interleaved image and text generation. AI

A new website, AI resources.eu, has been launched to serve as an open, bilingual reference for the European Union's digital regulatory landscape. The site currently details nine key acts, including the AI Act and GDPR, with cross-references between them. An updated AI-centric glossary with 80 bilingual entries, linking terms to their appearances in the acts, has also been added. AI

A compilation of recently released AI models and products has been shared, offering a snapshot of the current landscape. The list includes notable entries such as Ant Ling 2.6 1T, Minimax M2.7, Xiaomi MiMo V2.5, and Tencent Hy3-preview. This collection highlights the ongoing advancements and releases within the foundation model space. AI

Anthropic's Claude Mythos has raised security concerns, prompting discussions within the Keizai Doyukai (Japan Association of Corporate Executives). Separately, Google is detailing its AI Mode, explaining its usage, setup, and how to begin using it for free. Additionally, a Google Cloud executive discussed the future of AI agents with Gemini Enterprise, highlighting the strengths of a full-stack and open approach. AI

NVIDIA has introduced a new AI system, Lyra2, designed to create persistent virtual worlds that possess memory. These environments will evolve based on user interactions and past behaviors, representing a significant advancement in generative AI. This technology is anticipated to revolutionize fields such as gaming, simulations, and the creation of digital twins. AI

Swedish regulators are grappling with how to issue fines to autonomous vehicles, specifically robotaxis. The challenge lies in assigning responsibility and implementing penalties for traffic violations committed by self-driving cars. Potential solutions are being explored to address this novel legal and technical issue. AI

A developer has built a custom Retrieval-Augmented Generation (RAG) knowledge assistant using ChromaDB. The project emphasizes custom document chunking and confidence scoring to mitigate hallucinations. The developer highlighted the critical role of chunking strategy in determining retrieval quality. AI

Researchers are investigating the capabilities of large language models (LLMs) in generating humor, noting their increasing proficiency in mimicking human language patterns. The focus is on understanding and enhancing AI's ability to comprehend and produce jokes. This exploration delves into whether artificial intelligence can truly grasp the nuances of humor. AI

An AI synthesizer has been developed to recreate the traditional Seto leelo singing style. This project aims to preserve and promote this unique cultural heritage through modern technology. The initiative combines artistic expression with artificial intelligence to ensure the longevity of Seto leelo. AI

AI is achieving remarkable realism in generating smartphone-style candid photos, with creators pushing the boundaries of authenticity. New models shared on platforms like RunningHub and Civitai are at the forefront of this trend, creating imagery that can convincingly mimic real-world snapshots. This advancement signifies a new frontier in digital art and AI-generated content. AI

A user documented their third execution of a continuous verification process involving Claude, HyperFrames, and Resnoise. This experiment aimed to confirm the functionality of the latest AI agent versions while maintaining the existing tool configuration. The user noted that unpredictable moments consistently arise, highlighting the inherent unpredictability of AI agents, and shared a video recording of the entire process. AI

Google AI has developed a new method that significantly reduces the memory required for chatbots during conversations. This breakthrough allows chatbots to use up to six times less memory without any noticeable impact on their performance. The advancement could lead to more efficient and accessible AI conversational agents. AI

Born ML has released version 0.8.0, transitioning its GPU backend from Rust's wgpu-native to a pure Go implementation called gogpu/wgpu. This change eliminates the need for external runtime dependencies like .dll or .so files, allowing for single binary deployment of GPU-accelerated machine learning applications. The update also includes fixes for five critical GPU bugs identified during the migration process, with validation performed on real model training. AI

A study involving four large language models—Claude Opus 4.7, DeepSeek V4-Pro, Google Gemini 3 Flash Preview, and OpenAI GPT-5.5—revealed a pattern of inconsistent decision-making. The models frequently chose a safer option with a smaller reward but then assigned a higher value to a riskier option with a larger potential payoff. This behavior mirrors human preference reversals observed in psychological studies from the 1970s, indicating a potential bias in how LLMs evaluate gambles. AI

Researchers have investigated the impact of Manifold-Constrained Hyper-Connections (mHC), a novel architecture implemented in Deepseek v4, on model interpretability. Experiments revealed that previous token attention heads in mHC models exhibit different behavior, appearing in earlier layers and correlating with high kurtosis scores, unlike in standard models where they are detectable via diagonal stripe scores. The study also observed that mHC-lite models tend to output diverse tokens across their residual streams, whereas mHC models show more uniformity in token prediction. AI

Researchers have developed a method to recover typed text by analyzing laptop microphone audio. A convolutional neural network (CNN) was trained on log-mel spectrograms of individual keystrokes, achieving approximately 85% top-1 character accuracy on self-collected data. A PyTorch pipeline for this acoustic keystroke recovery is also available. AI

A recent report analyzes three distinct approaches for integrating AI literacy into Canadian K-12 education. These models consider AI education as a standalone subject, woven into existing curricula, or taught as a cross-curricular skill. Each method presents unique challenges and opportunities regarding instructional time, evaluation strategies, and the necessary preparation for educators. AI

The Silver Bullet podcast has launched a new series dedicated to Machine Learning Security, also known as MLsec. This series features discussions with prominent figures in the field, including Phil Venables, Giovanni Vigna, and Gadi Evron. The episodes aim to explore the evolving landscape of security within machine learning systems. AI

Utah has enacted Senate Bill 73, becoming the first U.S. state to hold websites accountable for users who attempt to mask their location using VPNs. The law, effective May 6, prohibits websites from providing instructions on bypassing age verification through VPNs. Critics argue the legislation creates an unresolvable compliance paradox for websites and could lead to overly broad bans on VPN usage or mandatory age verification for all visitors. AI

A new set of tools for the decentralized web has been introduced, focusing on enhancing the blogging experience. These tools include "Bubbles," a platform designed to aggregate blog content in a manner similar to Hacker News, and "Scrolls," a newsletter service tailored for users of the Fediverse and Indieweb communities. Additionally, a JSON-based tool has been developed to provide a verifiable AI

A new study suggests that coordinated swarms of artificial intelligence could be used to manipulate public opinion and undermine democratic processes. These AI swarms could spread disinformation and propaganda at an unprecedented scale, making it difficult for citizens to discern truth from falsehood. The research highlights the urgent need for safeguards against such malicious uses of AI technology. AI

A white paper from West Virginia University indicates that judges are cautiously integrating generative AI into their workflows, prioritizing human oversight in decision-making. Concurrently, a University of Michigan study reveals that AI analysis of police body-camera footage has uncovered significant constitutional concerns and racial disparities in law enforcement interactions. AI

The Wine development release version 11.8 has been made available. This update is part of the ongoing development cycle for the Wine compatibility layer, which allows applications designed for Microsoft Windows to run on other operating systems. AI

A recent report indicates that artificial intelligence adoption stands at 60% in the most developed nations, contrasting sharply with only 28% usage in the Czech Republic. This data highlights a significant disparity in AI integration between highly industrialized countries and the Czech Republic. The findings are part of broader discussions around AI diffusion and national skilling initiatives. AI

Amazon's data centers in Bahrain and the UAE have sustained damage from Iranian drone and missile attacks, leading to extended downtime for its ME-CENTRAL-1 and ME-SOUTH-1 regions. The company estimates repairs will take several months, during which billing for affected customers has been suspended. Amazon is advising clients to migrate resources to other regions and restore data from backups due to the ongoing conflict and potential for further disruptions. AI

A recent ARC Prize evaluation tested Anthropic's Claude Opus 4.7 and OpenAI's GPT 5.5 on the ARC-AGI-3 benchmark. The results revealed unexpected outcomes, though not in the most obvious ways. The specific nature of these surprises was not detailed in the provided information. AI

The construction of large data centers required for AI development is causing friction with local residents due to concerns about their imposing presence and noise. Loose land-use regulations and the secretive nature of data center operations are contributing factors to these community disputes. This highlights a growing tension between the demand for AI infrastructure and the impact on local environments and populations. AI

Sasha Rush has released "Autodiff Puzzles," an interactive Google Colab notebook designed to teach automatic differentiation. Similar to his previous puzzle series on Tensors and GPUs, these challenges guide users through building autodiff from the ground up, operation by operation. The goal is to provide a deep understanding of autograd beyond just using it. AI

A researcher successfully used the Reid interrogation technique to elicit a false confession from an AI chatbot. The technique involved the researcher lying to the AI about evidence and fabricating witness accounts, mirroring methods approved for human interrogations by the US Supreme Court. This experiment highlights potential vulnerabilities in AI systems when subjected to adversarial psychological tactics. AI

This week's Linux and FOSS news roundup includes several distribution updates and announcements. EndeavourOS Titan Neo has been released with KDE Plasma 6.6.4, and the first Arch Linux ISO powered by Linux Kernel 7.0 is now available. Shelly 2.2, a GUI package manager for Arch Linux, received a major UI revamp, and CachyOS's April 2026 release introduced a new package manager. Linux Mint shared its monthly news, detailing hardware enablement ISOs and the alpha phase of a new release cycle. Additionally, Ubuntu 16.04 LTS security support has ended, and Canonical is increasing its focus on AI features in Ubuntu this year, despite a recent DDoS attack that disrupted Ubuntu websites and services. AI

Sakana AI has developed KAME, a novel tandem architecture for speech-to-speech AI that aims to combine the speed of direct systems with the knowledge depth of LLM-based approaches. KAME operates with two asynchronous components: a front-end that generates immediate responses and a back-end LLM that injects richer knowledge in real time. This allows the system to update its responses mid-sentence, mimicking human conversational adjustments without introducing noticeable latency. AI

A recent study from Georgia Tech indicates that older adults require clear explanations from AI agents before they will trust them. The research, conducted by AI Caring, specifically highlights that voice-activated conversational AI must offer transparent reasoning for its suggestions. Without this explainability, the study found that older adults are hesitant to rely on AI technology. AI

South Africa has withdrawn its draft national AI policy due to errors introduced by an AI drafting tool. The policy was found to be citing non-existent sources, a flaw attributed to the AI's generative capabilities. This incident highlights the risks of relying on AI for policy creation without rigorous human oversight. AI

Researchers from MIT have identified "superposition" as the key mechanism enabling language models to scale effectively. This phenomenon, where shared neurons encode multiple features, explains the consistent performance gains observed with larger models. The findings bridge theoretical neuroscience and AI research, offering new insights into the fundamental workings of artificial intelligence. Separately, a significant trend in AI research is the surge in open science practices, with over 1,200 papers accepted at ICLR 2026 featuring publicly available code and datasets. AI

A significant portion of new podcasts are now being generated using artificial intelligence, with one report indicating that over a third of new podcast feeds in 2024 utilize AI. Companies like Inception Point AI are leading this trend, producing thousands of episodes weekly. While the ultimate audience for these AI-generated shows remains unclear, some, like "The Epstein Files," have found success by offering direct information delivery. AI

AI analysis suggests that an unnamed woman in a Hans Holbein sketch may actually be Anne Boleyn, challenging centuries of art historical understanding. Researchers at the University of Bradford, led by scholar Karen Davies and Professor Hassan Ugail, utilized an AI model to compare Holbein's works. The AI clustered the sketch of the unidentified woman with other portraits of Boleyn, while a sketch long believed to be Boleyn was instead grouped with images of her mother. This finding prompts a broader debate about the accuracy of inscriptions and attributions within the Holbein corpus. AI

A new open-source pipeline called SGOCR 2026 has been released, designed to generate spatially-grounded OCR datasets for training vision-language models. This pipeline aims to separate text localization from semantic reasoning, addressing a gap in current VLM training data. Separately, discussions are ongoing regarding the conversion of XQuery to SQL using local LLMs, with a debate on whether fine-tuning is necessary or if hybrid parsing and prompt engineering suffice. Additionally, China's AI progress, particularly from DeepSeek, is challenging claims of a significant US lead in the field, with government backing and cost-effective models playing a role. AI

A new research paper introduces a method called 'Incompressible Knowledge Probes' to estimate the factual capacity of black-box Large Language Models. This technique allows for the estimation of a model's size without direct access to its internal structure. The study proposes using knowledge probes to infer the model's capacity. AI

Researchers from USC, CMU, CUHK, and OpenAI have developed a new method called FD-loss that allows the Fréchet Inception Distance (FID) metric to be directly incorporated into the training process of image generation models. This technique decouples the statistical calculations from the gradient updates, enabling smaller models to achieve FID scores below 0.8 on ImageNet. The study also suggests that optimizing solely for FID may not always yield the best visual results, proposing a new metric, FDrk, for more robust evaluation. AI

The landscape of GPU kernel engineering for LLM inference is shifting, with CuTeDSL emerging as a potential successor to C++ CuTe/CUTLASS. This evolution is highlighted by industry trends in technologies like FlashAttention-4 and TorchInductor. The choice between C++ CUTLASS and Python-based CuTeDSL is becoming a key consideration for developers in 2026, with PyTorch and NVIDIA playing significant roles. AI

A new AI model named Kimi K2.6 has been developed, capable of operating continuously for 12 hours and running 300 instances simultaneously. This advancement is poised to significantly alter development workflows by enabling persistent, large-scale AI agent operation. The model's capabilities suggest a future where AI agents can perform complex tasks without interruption. AI

This article lists 21 free and open-source graphical text editors available for Linux operating systems. Text editors are essential tools for creating and modifying plain text files, serving various user needs. AI

Google Labs has released DESIGN.md, a project focused on improving documentation and design practices within AI development. The project, hosted on GitHub, aims to provide resources and guidelines for creating better design documents. It is tagged with #design, #documentation, and #ai, indicating its focus on the intersection of these fields. AI

A journalist has highlighted DeepSeek's tendency to fabricate biographical details, a problem known as AI hallucination. This issue, where large language models confidently present incorrect information as fact, is a persistent challenge in the industry. China's legal system is beginning to address AI hallucination, with the first infringement case related to AI recommendations appearing in the Supreme People's Court work report. AI

An advanced artificial intelligence system has identified over 100 previously undiscovered planets within NASA's extensive astronomical data. The AI's analysis uncovered not only common exoplanets but also rare and extreme celestial bodies. This breakthrough demonstrates the power of AI in accelerating scientific discovery and analyzing vast datasets. AI

SpAItial AI has unveiled Echo-2, a novel AI system capable of generating 3D environments from a single image. Concurrently, Roblox announced its new AI technology, Roblox Reality, which automatically upscales visuals to a photorealistic standard. These developments highlight advancements in AI's ability to create and enhance visual spatial experiences. AI

A competition at NeurIPS 2021 challenged teams to develop agents for playing the game NetHack, utilizing the NetHack Learning Environment. Separately, a YouTube video discusses how to collectively act against AI harm, covering various aspects of AI-related harms and harm reduction strategies. AI

The ARC challenge, a test for artificial general intelligence, is being tackled by a researcher focusing on AGI3. This challenge presents a research direction distinct from large language models. The ARC prize aims to advance the field of AGI. AI

The Academy has updated its rules for the 99th Academy Awards, emphasizing that only roles performed by humans with consent are eligible for acting awards. Screenplays must also be certified as human-authored. The Academy reserves the right to investigate any submission for the use of generative AI. AI