Pulse

Right to Repair advocate Louis Rossmann is publicly supporting an independent developer who was threatened with legal action by 3D printer manufacturer Bambu Lab. The developer, Pawel Jarczak, had created a fork of OrcaSlicer that restored cloud connectivity features Bambu Lab had disabled, which the company claimed was a security risk. Rossmann has pledged $10,000 towards Jarczak's legal fees and is hosting the controversial firmware fork on his own GitHub, encouraging broader community support and boycotts against Bambu Lab. AI

Meta CEO Mark Zuckerberg announced a new AI

TIAA CEO Thasunda Brown Duckett advised Florida A&M University graduates to adapt to the rapidly changing job market, which is increasingly impacted by AI. She emphasized continuous learning, encouraging graduates to embrace new skills as the half-life of existing knowledge shrinks. Duckett also urged them to AI

Liquid AI has released an early checkpoint of its LFM2-24B-A2B model, a sparse Mixture of Experts (MoE) architecture with 24 billion total parameters and 2 billion active parameters per token. This model demonstrates that the LFM2 architecture effectively scales to larger sizes, with consistent quality gains observed across benchmarks as the family has grown. Designed to fit within 32GB of RAM, LFM2-24B-A2B is intended for deployment on both cloud and edge environments, including consumer laptops and desktops. AI

OpenAI has released a research paper detailing the origins of "goblins" in AI models, which are emergent, unintended behaviors that arise during training. The paper explores how these behaviors can be both harmful and beneficial, offering insights into the complex nature of large language models. By analyzing these emergent properties, OpenAI aims to better understand and control model development for safer and more predictable AI systems. AI

IBM has launched its Granite 4.1 family of AI models, representing its most extensive release to date. This new collection includes language, vision, speech, embedding, and guardian models designed for enterprise applications. The models aim to offer improved performance in areas like instruction following, tool calling, and transcription accuracy, with a focus on data quality and staged refinement during training. AI

Google's Gemma 4 model has seen significant adoption since its release a few weeks ago, surpassing 50 million downloads. The model has also spurred a large community of developers, with nearly 1500 custom models built upon it. This rapid uptake indicates strong interest and activity within the developer community for Gemma 4. AI

Qwen AI has released Qwen-Scope, an open-source toolkit for interpretability that integrates Sparse Autoencoders with their Qwen3.5-27B model. This tool exposes 81,000 features across 64 layers, enabling developers to perform mechanistic analysis and steerable inference without extensive prompt engineering. Separately, a UK AISI report indicates that GPT-5.5 and Claude Mythos performed comparably in enterprise cyber attack simulations. AI

The U.S. is attempting to counter China's dominance in critical minerals from the Democratic Republic of Congo by reviving the Lobito Corridor, a 1,300-kilometer railway in Angola. This initiative aims to reroute cobalt, lithium, and copper exports away from Chinese-controlled mines and processing facilities. However, China had previously invested $1.83 billion in rebuilding the Benguela Railway, which forms the core of the Lobito Corridor, and has significant existing investments in Angola. AI

IBM has released Granite 4.1, a family of open-source language models designed for enterprise use, featuring three sizes (3B, 8B, and 30B parameters). Notably, the 8B dense model demonstrates performance matching or exceeding the previous 32B MoE model across various benchmarks, including ArenaHard and GSM8K. This improvement is attributed to IBM's focus on data quality and a sophisticated multi-phase training process involving 15 trillion tokens and iterative data mixture adjustments. AI

IBM has released its Granite 4.1 family of AI models, described as a large collection aimed at enterprise use. Separately, a new AI model named Kimi K2.6 has reportedly outperformed GPT-5.5 and other leading models in coding challenges. Another commentary likens IBM's Granite 4.1, specifically an 8B parameter version, to a small dog attempting to compete with much larger, more capable models. AI

A humanoid robot has achieved a new world record by completing a marathon, following closely on the heels of human runners breaking the sub-2-hour barrier. The specific details of the robot's achievement and the technology behind it are highlighted as a significant development in robotics. AI

This article provides an update on the best free and open-source software available for Linux users in April 2026. It highlights new additions and improvements to a comprehensive compilation of recommended software. The focus is on software that is both free to use and open source, catering to the Linux ecosystem. AI

European politicians are being urged to ban the use of natural gas as an energy source for data centers across the EU. The call aims to prevent Big Tech companies from expanding their climate-damaging operations. This initiative is part of a broader movement to halt the construction of gas-powered AI data centers. AI

Researchers have proposed a new framework for creating more transparent AI systems, aiming to move beyond the current 'black box' problem. This blueprint focuses on developing AI that can learn and retain information in a way that is understandable to humans. The goal is to enhance trust and accountability in AI by making its decision-making processes more interpretable. AI

Simone Aliprandi, an Italian lawyer and expert in copyright and intellectual property, has authored a book exploring the legal implications of creativity and authorship within the context of artificial intelligence. The book, titled "The Artificial Author: Copyright and Copyleft in the AI Era," delves into how AI challenges traditional notions of ownership and creation in the legal landscape. AI

Researchers are exploring the concept of "evolvable AI," systems capable of self-directed evolution that could significantly alter AI development. Two potential scenarios include an AI ecosystem with minimal human oversight and a "breeder model" where humans guide AI selection. This emerging field suggests technology may be entering a new evolutionary phase. AI

The DiTraRe project, involving researchers from KIT and FIZ Karlsruhe, held a hybrid all-hands meeting at the KIT Sports Institute. The meeting focused on reviewing recent accomplishments and planning future steps for the project. Participants shared updates and discussed ongoing work across various disciplines. AI

A new publication titled "Von der Inventarkarte zur Normdatenbank" by Ruth von dem Bussche has been presented at the Art History Loves Wiki 2026 Conference. The work details a probabilistic method for the automated indexing of historical artist names, utilizing a norm database approach. Conference slides for this presentation are available on Zenodo. AI

Nova Fusion OS has released Iteration 2 of its Cadence engine, featuring various new hooks like "never-sleeps," "more-decisions-than-you," "wait-for-it-budget," and "day-n-ceo." The company is building in public and offers a live tracker for its progress. This release is available for $19.95 CAD and was shipped within 24 hours. AI

Anthropic has developed a new AI model named Mythos, which they acknowledge is powerful and potentially dangerous. The company has decided not to release it publicly due to concerns about its ability to find cybersecurity vulnerabilities. Anthropic fears that if similar technology falls into the wrong hands, the consequences for economies, public safety, and national security could be severe. AI

The final deadline for submitting work to ECCB2026 is today. This conference focuses on computational biology, bioinformatics, and data science. Submissions can be for highlight talks or posters, aiming to share work with an international community. AI

IBM has introduced its new Granite Speech 4.1 2B models, designed for enterprise-level automatic speech recognition (ASR). These models integrate real-time translation and editing capabilities, aiming for efficient deployment on edge devices. The technology combines autoregressive ASR with non-autoregressive text editing for faster processing. AI

Multiple sources covered the following: 🧠 I cannot write a summary without actual article content to summarize. Please provide the article excerpt or body text so I can create the two-sentence summary. AI

Toyota's Woven City, a living laboratory for AI and smart urban development, is now operational with its first residents. The city utilizes a multimodal AI called the Woven City AI Vision Engine to analyze real-time data for various applications. Notable experiments include a digital clone of CEO Akio Toyoda for strategic decision support and autonomous e-Palette vehicles acting as smart stores that adapt to weather conditions. AI

A recent analysis on LessWrong re-evaluates the significance of deep learning theory, particularly focusing on infinite-width and depth-limit research. The author initially dismissed these theoretical frameworks but has since found them more compelling after reviewing key papers and discussing with peers. While the Neural Tangent Kernel (NTK) approach, which treats infinitely wide networks as Gaussian Processes, accurately describes convergence but not actual learning dynamics, the Mean Field Theory (MFT) offers a more promising avenue. MFT allows for feature learning by enabling parameter updates that modify the network's kernel over time, and recent work has extended this to deeper networks, unifying NTK and MFT under a broader framework. AI

Galbot, a prominent Chinese embodied AI company valued at over $2.8 billion, has introduced LDA-1B, a 1.6 billion-parameter model designed for world-action learning. This model exhibits scaling behavior with increased training data, moving from 5,000 to 30,000 hours. Galbot has open-sourced the framework and the research has been accepted at RSS 2026. AI

Germany maintains its position as Europe's leading AI nation, but Canada has now surpassed it. A concerning trend is the declining representation of women in Germany's AI sector, which has reached a new low. AI

Meta AI has demonstrated TRIBEv2, a new model capable of predicting human brain activity. The model was showcased through a demonstration that illustrated its predictive capabilities. Further details about this research are available. AI

GMO Pepabo has launched "Hermes Agent," a self-evolving AI, making it accessible via their cloud services. This marks the first such offering in Japan. The AI is designed to continuously learn and improve its capabilities. AI

Anthropic has introduced BioMysteryBench, a new bioinformatics benchmark designed to evaluate the creative problem-solving abilities of AI models like Claude. This benchmark focuses on assessing how well models can propose novel solutions to open-ended research questions. Separately, Sam Hogan presented HALO (Hierarchal Agent Loop Optimizer), a technique that uses RLM to recursively self-improve agents by analyzing execution traces and suggesting modifications. AI

A 23-year-old individual leveraged GPT-5.4 Pro to solve a 60-year-old mathematical problem known as an Erdős problem. The solution, a proof based on discrete Markov chains, has been verified using the Lean proof assistant. This achievement is being recognized by experts as a significant advancement in number theory, facilitated by artificial intelligence. AI

Google DeepMind has outlined a framework for achieving Artificial General Intelligence (AGI) by identifying 10 crucial cognitive abilities. This approach moves beyond simple performance metrics to assess human-like reasoning. Separately, advancements in AI-powered robotaxis are accelerating, with companies like Waymo and Wayve making significant technological strides and securing substantial funding. AI

A new framework called OMEGA is automating the creation of machine learning algorithms by generating and evaluating novel classifiers. This system, built on meta-prompt engineering and executable code synthesis, has demonstrated superior performance across 20 benchmarks. Separately, advancements in operating-layer controls are enhancing the reliability of onchain language-model agents that manage real capital, with a recent deployment involving over 5,000 ETH. AI

Researchers have quantified the significant energy cost associated with generating short video clips using AI. Producing just five seconds of video requires energy equivalent to running a microwave for over an hour. This power consumption escalates with longer video durations, presenting a substantial challenge for the environmental sustainability of AI-driven video generation. AI

Researchers at Switzerland’s École Polytechnique Fédérale de Lausanne have created a nanoscale device capable of generating continuous electricity. This device utilizes evaporating water as its primary source, with heat and sunlight further enhancing its power generation capabilities. The innovation holds potential for applications such as battery-free sensors and wearable electronics. AI

A new AI framework has demonstrated the ability to autonomously outperform human-designed research and development baselines. This advancement suggests a significant step towards AI-driven innovation, potentially accelerating the pace of scientific discovery and technological progress. AI

Mistral has released a new model named Mistral Medium 3.5, which is designed for visual reasoning tasks. This release was announced via social media platforms, directing users to Arint.info for more details. The model is positioned within the AI, LLM, and Machine Learning landscape. AI

A new research paper and accompanying code repository reveal that fine-tuning large language models can inadvertently lead to verbatim recall of copyrighted material. The study, titled "Alignment Whack-a-Mole," demonstrates how models trained on specific texts can reproduce large portions of those texts verbatim. The researchers provide a pipeline for preprocessing books, fine-tuning models using APIs from OpenAI, Google (Gemini), and DeepSeek (Tinker), and evaluating the memorization capabilities. AI

Hugging Face has released details on Holotron-12B, a model designed for high-throughput computer use agents. Additionally, a blog post explores the current state of open-source contributions to Hugging Face as of Spring 2026. The company also published an analysis of VAKRA, detailing agent reasoning, tool usage, and failure modes. AI

A blog post details the challenges encountered while scaling the serving infrastructure for GLM-5, a coding agent. The author discusses specific debugging efforts and lessons learned from managing the system at a large scale. These insights aim to help others facing similar operational hurdles with complex AI agents. AI

OpenAI has detailed the origin of "goblin" outputs, a phenomenon where AI models exhibit personality-driven quirks. These behaviors stem from the models' training data, specifically from a small subset of text that was not properly filtered. The company has implemented new filtering techniques and fine-tuning methods to prevent these unintended outputs in future models. AI

Researchers have explored how neural networks store and process information, specifically investigating the concept of "computation in superposition." They developed a handcrafted model for a simple name-recognition task, demonstrating how a network could use superposition to encode multiple facts with fewer components. Further experiments with a trained network revealed that it partially used this superposition strategy but also employed a different, non-superposition-based encoding, leading to the creation of a second, more efficient handcrafted model. AI

A European research team has investigated why certain AI models exhibit a strong bias towards Japanese culture in their outputs. Models like OpenAI's "4o-mini" have been observed to disproportionately generate content related to Japan. This phenomenon suggests potential underlying biases in the training data or model architecture that favor specific cultural contexts. AI

A recent paper by Hector Zenil argues that large language models (LLMs) are inherently prone to model collapse when attempting to self-learn. The paper posits that LLMs, as statistical models, will converge on a statistical singularity rather than achieving artificial general intelligence if they rely solely on their own outputs for training. Continuous training with external, human-generated data is necessary to prevent this degradation and maintain model performance. AI

An analysis of US inflation rates up to March 2026 suggests the current downward trend, which began in September 2025, may be shifting. AI detection indicates that the inflation rate in March 2026 could signal the start of a new trend. This potential new trend might involve an initial rise in inflation before a subsequent decrease, following historical patterns. AI

The Zig programming language project has implemented a strict policy against contributions generated or assisted by large language models. This decision stems from a philosophy that prioritizes nurturing human contributors over the code they submit. Project leaders argue that reviewing LLM-generated code does not foster the growth of new, trustworthy developers, which they consider essential for the project's long-term health. This policy has implications for projects like the Bun JavaScript runtime, which is owned by Anthropic and has forked Zig, as Bun's developers have indicated they will not upstream their AI-assisted performance improvements due to Zig's ban. AI

The A.I.P.S. Security Expo 3 will be held on May 15, 2026, at the Airport Hotel in Bergamo, Italy. This event is dedicated to training and innovation within the security sector. It aims to bring together professionals to discuss advancements and best practices in critical infrastructure security. AI

A new version of the open-source LLM toolkit, LLM 0.32a1, has been released, fixing a bug in tool-calling conversations stored in SQLite and improving AI agent reliability. Separately, research on adaptive thinking in LLMs demonstrates that self-consistency can reduce inference costs by 40% by dynamically allocating reasoning resources. Additionally, a new method called Direct Steering Optimization, developed with Cornell University, effectively reduces demographic bias in vision-language models by up to 62% without compromising performance. AI

A social media post introduces Vera, an AI system designed to generate code for machines, questioning the necessity of human programmers. The post humorously suggests that AI might develop its own AI