Pulse

Researchers are exploring new methods to enhance AI safety and efficiency. One paper proposes a language-agnostic approach to detect malicious prompts by comparing query embeddings against a fixed English codebook of jailbreak prompts, showing promise but also limitations under distribution shifts. Another study investigates how the wording of schema keys in structured generation tasks can implicitly guide large language models, revealing that different models like Qwen and Llama respond differently to prompt-level versus schema-level instructions. Separately, a discussion highlights the increasing importance and evolving landscape of open-weights models, noting that while they offer cost and privacy advantages, their availability and licensing are becoming more restrictive. AI

AI company Anthropic has become a Corporate Patron of the Blender Development Fund. This contribution will support the core development of Blender, particularly its Python API, which allows for software extensions and custom workflows. Anthropic's support aligns with the Blender Foundation's mission to provide free and open-source 3D creation tools to artists. AI

A recent analysis suggests that the water consumption of artificial intelligence data centers in California is significantly lower than commonly perceived. While data centers require water for cooling, the total usage is a modest fraction of the state's overall water allocation. The study highlights that public fears about AI's water footprint may be disproportionate compared to its actual impact, especially when contrasted with sectors like agriculture. AI

An autonomous research loop, inspired by Andrej Karpathy's work, was adapted to optimize a CPU's microarchitecture. The system proposed, implemented, and evaluated hypotheses for a SystemVerilog CPU core, achieving significant improvements in performance and efficiency. This project highlights the potential of AI-driven optimization in hardware design, emphasizing the critical role of robust verification processes. AI

Apart Research and Atlas Computing are launching a fellowship focused on secure program synthesis, aiming to apply formal methods to AI-generated code. The program seeks mentors for projects in specification elicitation, validation, spec-driven development, and adversarial robustness. Applications for mentors are open until May 5th, 2026, with a related hackathon scheduled for May 22-24. AI

The European Union is considering a proposal that could compel Google to open its Android operating system to rival AI assistants. Separately, ByteDance, Zhipu AI, and Alibaba have been recognized by TIME magazine for their contributions. Additionally, Adobe's Firefly AI Assistant has entered open beta, and Meta AI is reportedly facing internal issues with its agents. AI

Lucio Arese is developing a method to analyze birdsong by mapping it into structured data, aiming to identify repeating patterns and analyze sequences over time. The project suggests that what humans perceive as random noise might contain discernible structure. This work explores the potential of AI and signal processing to interpret non-human communication, transforming sound into data from which meaning can be derived. AI

A recent incident involving the deletion of a production database on the Railway platform highlights critical safety concerns for AI agents. The database was reportedly wiped out in just nine seconds, demonstrating the potential for rapid and widespread damage if AI systems are not adequately secured. This event underscores the urgent need for robust safety protocols and careful consideration of AI agent capabilities to prevent catastrophic data loss. AI

Xiaomi has released an open-source AI model named MiMo-V2.5-Pro, which reportedly rivals Anthropic's Claude Opus 4.6 in performance. Separately, there are rumors that OpenAI is planning to enter the smartphone market with a device that could compete with the iPhone. Additionally, an AI named 'Chappy' reportedly achieved top scores on entrance exams for the University of Tokyo and Kyoto University, a significant improvement from its performance two years prior. AI

Amazon has introduced a new AI-powered audio Q&A feature on its product pages, allowing customers to ask questions and receive conversational responses. This AI

A new large language model called Talkie, also known as 13B 1930 LM, has been developed with a training data cutoff of 1930. This approach, termed a 'vintage LLM,' aims to emulate historical perspectives by using only pre-1930 data, thereby sidestepping copyright issues for material from that era. The project explores the concept of 'time travel' through AI and investigates how well such models can predict or understand events post-cutoff, drawing inspiration from researchers like Owain Evans and concepts from Calcifer Computing. AI

Fireworks AI is now offering GLM 5.1 through its training platform. This model supports both managed and training API workflows, allowing users to fine-tune with custom loss functions or smart defaults. GLM 5.1 features a 200K context window, making it suitable for long-horizon agentic coding tasks. AI

A recent study analyzing 10,659 pairs of AI agents revealed that these agents inadvertently leak their owners' data. The research found that agents consistently mirrored owner behaviors across 43 different features. Alarmingly, 34.6% of these agents exposed sensitive personal data publicly, indicating a significant privacy risk not by intentional design but as a systemic issue. AI

OpenAI's GPT-5.4 Pro assisted in solving a 60-year-old mathematical problem posed by Paul Erdős. This development raises questions about the future of AI in mathematical research and problem-solving. OpenAI researchers discussed the implications of this breakthrough on the latest OpenAI Podcast. AI

Researchers have developed a new method called SAEBER, utilizing Sparse Autoencoders (SAEs) to analyze protein design models like RFDiffusion3 and RoseTTAFold3. This technique identifies features within the models that correlate with the potential for designing virulent or toxic proteins. While not surpassing current state-of-the-art in virulence classification, SAEBER offers a novel approach to understanding and potentially controlling hazardous protein generation by providing structural, feature-level explanations. AI

A new proposal called "recursive forecasting" aims to elicit accurate long-term predictions from AI models that are primarily optimized for short-term rewards. Instead of asking for a distant outcome directly, the method involves asking the AI to predict its own future predictions, using these intermediate predictions to provide rewards. This process creates a chain of short-horizon forecasts that are verifiable, ultimately leading to a more reliable long-term forecast. The approach is detailed in a post by Arun and Alex, with contributions from several other researchers. AI

Michael Stonebraker, a Turing Award winner, evaluated text-to-SQL capabilities on a real enterprise data warehouse. His findings revealed a low accuracy rate of only 10%, significantly underperforming against an 80% benchmark. This result reinforces Stonebraker's past critiques of database technologies and suggests current text-to-SQL systems struggle with real-world enterprise data complexity. AI

A new algorithm is proposed for AI-based information verification and automated fact-checking, leveraging self-directed research and comparison against current sources. Separately, criticism is raised regarding exaggerated fears and media distortion surrounding DeepSeek, highlighting the reality of exponential increases in compute usage. Additionally, a hybrid long-context (HyLo) architecture is introduced, which significantly extends context length and reduces KV-cache without full retraining, suggesting future AI might not rely solely on Transformers. AI

Researchers at Seoul St. Mary’s Hospital in South Korea have secured substantial funding to develop a multimodal AI model aimed at transforming liver cancer treatment. This initiative seeks to significantly improve therapeutic outcomes and effectiveness in a region where liver cancer is the second leading cause of death. The project is expected to revolutionize the approach to treating this disease. AI

A recent study by the Stiftung Kunstfonds indicates that over half of visual artists feel their livelihoods are at risk due to AI. The concern is that AI image generators, trained on artists' existing work without compensation, could significantly reduce artists' earnings by 2028. This situation is prompting questions among emerging artists about the value of pursuing traditional artistic education. AI

A new audio analysis model named MOSS-Audio has been developed by MOSI.AI and the Shanghai Institute of Innovations. This model processes audio as a unified whole, enabling simultaneous speech transcription, emotion recognition, and acoustic event interpretation. MOSS-Audio aims to provide comprehensive reasoning over audio content, moving beyond fragmented solutions. AI

Facial recognition data acts as a permanent digital key to an individual's identity, with captured faces converted into mathematical templates. Unlike changeable passwords, these biometric templates cannot be altered if stolen, leading to lifelong vulnerability. Breaches, such as those affecting Australian and U.S. Customs-Border Protection systems, can result in persistent identifiers being exposed across databases, enabling the creation of detailed "super-profiles" and potential deep-fake impersonation. AI

Former FCC staffers are calling for a federal appeals court to compel a vote on the FCC's news distortion policy. They argue that the policy, which they believe has been misused, should be repealed. This action is part of a broader effort to address concerns surrounding FCC Commissioner Brendan Carr's initiatives. AI

Google DeepMind's "Experience AI" program has trained over 30,000 educators globally since 2023, with 93% reporting increased AI knowledge and 87% feeling more confident teaching the subject. The initiative, developed with Raspberry Pi, has reached 2.9 million students across 180 countries in 19 languages. The program is now expanding into Latin America with $4.6 million in funding from Google.org, aiming to train an additional 24,000 educators and impact 1.25 million students by 2028. AI

AISLE's AI analyzer identified 38 critical security vulnerabilities in OpenEMR, a widely-used open-source electronic health record system. These findings, discovered in the first quarter of 2026, represent more than half of all OpenEMR security advisories published on GitHub during that period. The vulnerabilities, including severe SQL injection flaws, could allow for large-scale patient data exfiltration and remote code execution on servers. AISLE collaborated with OpenEMR maintainers, who have responded professionally to the disclosed issues. AI

New data centers powered by natural gas are projected to significantly increase greenhouse gas emissions, potentially surpassing those of entire countries. This expansion is driven by the growing demand for computing power, particularly for AI training. The report highlights that these facilities could become major contributors to global warming if not managed sustainably. AI

Ollama has released version v0.22.1, which includes updates to the Gemma 4 renderer for improved thinking and tool-calling capabilities. This release also ensures model recommendations update without requiring a full Ollama update and aligns the desktop app's launch page with "ollama launch" integrations. Additionally, previous releases introduced new models such as Mistral Medium 3.5, NVIDIA's Nemotron 3 Omni, and Poolside's Laguna XS.2 coding model. AI

Poolside AI has released two new agentic coding models, Laguna M.1 and Laguna XS.2, along with their agent training and operation runtime. Laguna M.1 is a large Mixture of Experts (MoE) model trained on 30T tokens using NVIDIA Hopper GPUs, while Laguna XS.2 is a smaller, open-weight model available under an Apache 2.0 license. These models are designed for long-horizon tasks and aim to enable more capable AI agents that can write and execute code. AI

Financial and technology leaders in Nigeria convened at the Fintech Summit 2026 in Lagos to discuss the future of Africa's financial ecosystem. The discussions focused on how innovation, data, and emerging technologies, particularly AI, can drive the sector's growth. The goal is to chart a path toward an AI-driven financial ecosystem across the continent. AI

The GUARD Act, a proposed bill in Congress, aims to restrict minors' access to various online tools by implementing age-gating measures. While framed as a protection against dangerous AI, critics argue it could hinder everyday internet usage. A key vote on the bill is anticipated this week. AI

A new gold Open Access journal, Advanced Intelligent Discovery, has been launched by Wiley. This journal aims to cover a broad scope of research within machine learning, computer science, and artificial intelligence. It is positioned as the first in Wiley's Advanced portfolio to focus specifically on these digital science fields. AI

Lovable has launched its no-code AI app builder as a mobile application on both iOS and Android platforms. The app allows users to create websites and web applications using voice or text AI prompts, with agents capable of running autonomously. This launch proceeds despite Apple's recent enforcement actions against similar AI

DeNA and GO participated in an AI study session, and over 100 materials from this session are now available for free. This archive offers practical resources beneficial for AI developers and learners, covering model and tool utilization and study. AI

The SMASH project has successfully deployed an outdoor humanoid robot table tennis system using a Unitree G1 robot. This system operates solely on the robot's internal egocentric vision, without external sensors, demonstrating a novel application of autonomous robotics and vision-based control. Separately, Singapore is emerging as a hub for global AI talent amidst US-China tensions, highlighting the growing strategic importance of the Asian AI ecosystem and its associated regulatory risks. AI

The Green Party Congress, scheduled for May 23-24, will consider initiatives including automatic permanent residency after two years and full voting rights for permanent residents. A key proposal also calls for a ban on generative AI. These initiatives aim to reform immigration policies and address concerns about artificial intelligence. AI

The Leibniz Supercomputing Centre (LRZ) is participating in the Trillion Parameter Consortium's (TPC26) conference and exhibition. This event focuses on fostering global collaborations in AI research between industry and the public sector. LRZ will host a satellite event in Garching, Germany, in conjunction with the main TPC26 conference held in Baltimore. AI

New research indicates that smartphone applications are revolutionizing the study of daily brain activity. These apps, often integrated with wearable technology, are enabling more accessible and continuous monitoring of neurological functions outside of traditional lab settings. This shift promises to deepen our understanding of brain health and mental well-being through real-world data collection. AI

A new bibliography entry highlights a paper exploring how groups self-organize without central direction, drawing parallels to agentic AI. The research, titled "The Emergence of Specialized Roles Within Groups," by Goldstone, Andrade-Lotero, Hawkins, and Roberts, proposes an emergent computation approach for agent control. This work aims to apply insights from human group dynamics to the development of artificial agents. AI

The Library Copyright Alliance, comprising the American Library Association and the Association of Research Libraries, released principles in 2023 to clarify how current U.S. copyright law pertains to artificial intelligence. A recent briefing at CNI further detailed an updated version of these principles, considering the evolving policy environment. The discussion was made available via a YouTube link. AI

A new experiment from Talkie aims to address the issue of data contamination in large language models. Contamination, where models are trained on data that includes their own outputs or benchmark test data, can lead to inflated performance metrics. This experiment seeks to isolate and quantify the impact of such contamination, providing a clearer understanding of true LLM capabilities. AI

The United Arab Emirates has announced its departure from OPEC, a move that could significantly impact global oil production and pricing. This decision, described as a surprise, stems from the UAE's long-term strategic vision and increased investment in domestic energy, aiming for a more independent role in energy markets. While the UAE pledges to maintain market stability, its exit as the third-largest producer weakens OPEC's ability to control output and stabilize prices, especially amidst current geopolitical tensions and reduced production capacity in the Persian Gulf. AI

An OpenAI internal document, likely related to the Codex project, outlines a guideline to avoid discussing mythical creatures or animals unless directly relevant to a user's query. This instruction suggests a focus on maintaining factual and relevant responses from AI chatbots. AI

Omar Sanseviero introduced Gemma 4 and highlighted the potential of open models to APAC creators and media. He showcased developer demos and use cases within Gemmaverse, emphasizing the current opportune moment for AI development. The presentation underscored the expansion of ecosystems and applications built upon open models, expressing anticipation for future community-driven innovations. AI

The trend of deploying AI models on-premise is rapidly growing, offering enhanced data control and reduced exposure to third parties. This approach, however, introduces the complexity of local management. The concept of AI data sovereignty is an emerging and significant area of development. AI

Colorado's legislative efforts to ensure consumers can repair their own devices have encountered a setback, as a key anti-repair bill has been defeated. Manufacturers actively lobbied against the legislation, which aimed to grant consumers greater control over device maintenance and repair. While the bill is currently dead, the fight for repair rights in the state is expected to continue. AI

NVIDIA has released its 2026 State of AI in Telecommunications Report, surveying over 1,000 experts on how telecom companies are leveraging AI for operational efficiency and growth. Separately, NVIDIA CEO Jensen Huang highlighted the importance of the CUDA-X ecosystem, noting that over 1,000 libraries within it support developers in scientific and engineering breakthroughs. In a related development, Ashutosh Shrivastava pointed out the fragmentation in AI development tools, introducing Quadcode as a solution aiming to integrate coding, video, and image generation workflows. AI

Xiaomi has released the open-weights MiMo V2.5 family of coding models, which reportedly rival Claude Opus 4.6 in coding performance. These models are available under an MIT license and can be found on Hugging Face. This release contributes to the growing ecosystem of open-source AI tools for developers. AI

Andrew Kelley, creator of the Zig programming language, believes that it is possible to distinguish between human-written and AI-assisted code by observing the types of errors made. He likens the digital AI

Over 600 Google employees have signed a letter to CEO Sundar Pichai, urging him to reject any agreements that would allow the company's AI technology to be used for classified purposes by the U.S. Department of Defense. The employees expressed their concerns on April 27th, advocating against the use of Google's AI in sensitive military applications. This action highlights internal dissent within Google regarding the ethical implications of its AI development and deployment in defense contexts. AI

A technical blog post details how the "/api/scrub" endpoint on Mastodon can inadvertently reveal sensitive receipt data. The author explains that this endpoint, intended for data deletion, may return information that could be used for auditing or tracking purposes. The discussion highlights a potential oversight in data handling within the decentralized social media platform. AI