Brief

Isomorphic Labs, a Google DeepMind spinoff, is advancing AI-driven drug discovery by developing new computational systems. Their Isomorphic Drug Design Engine (IsoDDE) builds upon AlphaFold's protein structure prediction capabilities to model complex biomolecular interactions. IsoDDE aims to identify novel drug targets, including previously unobserved protein pockets, and predict binding affinities, addressing limitations of earlier models. AI

Modal has enhanced the FlashAttention-4 kernel to improve inference speed for large language models, particularly for decode-heavy workloads. Their contributions focused on adjusting parallelism strategies, such as shifting from query parallelism to key/value parallelism, and supporting irregular global memory accesses using the Tensor Memory Accelerator (TMA). The company found the CUDA Templates Domain Specific Language (CuTe DSL) to be effective for development, and they anticipate further improvements with enhanced support for a tile-based programming model for future kernel development. AI

A new audit of Chinese Legal Case Retrieval (LCR) benchmarks reveals that the primary charge of a case, which encodes its legal characterization, is a significant factor in determining relevance. Researchers found that ranking cases solely by shared primary charge, combined with BM25, recovers nearly all of the performance gap between basic retrieval methods and advanced trained systems on the LeCaRDv2 benchmark. This suggests that current benchmarks may be overstating the legal reasoning capabilities of AI systems, as relevance is often determined by construction rather than true understanding of legal principles. AI

A developer has created a free, browser-based editor called Greenroom for HTML slide decks generated by AI models like Claude. This tool allows users to directly edit text, resize elements, modify chart data, and reorder slides within the HTML file without needing to re-prompt the AI or use complex development tools. The application emphasizes user privacy, as all data processing occurs locally within the browser, and no sign-up or upload is required. AI

A new website called AI Money Watch is monitoring campaign finance data to track the spending of AI super PACs. The site is specifically tracking Leading the Future (LTF), which has raised $125 million for the current midterm elections. This initiative aims to shed light on the financial influence of artificial intelligence in politics. AI

Meta has integrated an AI assistant into its Instagram Edits app to help users create and edit videos. This AI analyzes user statistics to suggest new video ideas. The feature is now also accessible via a desktop version of the app. AI

A study evaluated the impact of a self-check mechanism, named Self-Inspect, on coding agents. The experiment involved two Claude Sonnet 4.6 agents tasked with building a usage-billing module over 30 turns, with one agent incorporating Self-Inspect once per turn and the other not. The agents were scored on their ability to surface assumptions, preconditions, edge cases, or risks rather than silently making decisions. AI

For production-grade Retrieval Augmented Generation (RAG) systems, relying solely on vector search for semantic similarity is insufficient. Real-world applications often require precise matches for technical jargon, IDs, and version numbers, which semantic search can miss. Hybrid search, combining vector search with traditional keyword search, addresses this by leveraging both semantic understanding and lexical precision. Algorithms like Reciprocal Rank Fusion (RRF) are used to merge the rankings from these two methods into a unified result set, with a constant 'k' value (often 60) helping to balance the influence of top-ranked items. AI

Ukraine is employing AI-guided drones to target Russian supply lines, leveraging technological and tactical advancements in the ongoing conflict. This strategy aims to disrupt Russian logistics and gain an advantage on the battlefield. The use of these advanced drones signifies a shift in military tactics, integrating artificial intelligence for enhanced operational effectiveness. AI

This article evaluates alternatives to Composio for production AI agents, focusing on scalability beyond prototyping. It highlights the importance of per-user delegated authorization, agent-optimized tools to minimize hallucinations, and centralized governance with immutable audit logs. The guide contrasts Composio's prototyping strengths with the architectural needs of production environments, emphasizing security, identity, and observability. AI

Every is hosting a two-hour camp to demonstrate how to maximize the use of Anthropic's Fable 5 model for complex projects. The session will cover real-world applications in coding, growth, research, and writing, detailing the prompts and review processes used. Paid subscribers will also gain access to a future camp focused on Codex. AI

Nous Research has launched the Hermes Agent Profile Builder, a new dashboard tool designed to simplify the creation and deployment of autonomous AI agents. This tool allows developers to configure an agent's identity, select models and providers, choose built-in and external skills, and attach MCP servers through a unified, browser-based interface. Previously requiring multiple command-line steps, the Profile Builder streamlines the process into a guided flow, enabling users to run isolated agents on a single machine without state conflicts. AI

Todd Toler will host a session on optimizing token usage for GitHub Copilot. The session will be available for live viewing or later watching on YouTube. AI

A new supply chain attack, dubbed Miasma Worm, has emerged, leveraging configuration files within AI coding tools to steal credentials. Attackers exploit features like session start hooks in tools such as Claude Code and Gemini CLI, as well as agent instructions in Cursor and auto-tasks in VS Code. The attack, attributed to TeamPCP, has already compromised over 100 GitHub repositories by embedding a JavaScript credential harvester that executes automatically when a project is opened in these AI development environments. AI

Several UK police forces have been instructed to cease using AI for drafting court statements due to concerns about accuracy and potential misuse. This directive follows an admission by a Chief Constable that officers had indeed employed AI systems, a revelation that came after initial denials and led to his subsequent retirement. The head of Police.AI, Alex Murray, emphasized the critical need for forces to implement clear policies requiring officers to meticulously verify all AI-generated content, such as that produced by Copilot. AI

Microsoft has introduced SkillOpt, an open-source tool designed to enhance AI agents. This tool allows AI agents to update their capabilities by modifying external configuration files, rather than requiring costly retraining of model weights. This approach enables continuous improvement of AI agents without altering their core parameters. AI

A study in Nature Machine Intelligence indicates that AI models for medical diagnosis achieve higher accuracy when relying on text data rather than image data. The research evaluated multimodal foundation models across 1090 medical cases, finding that textual information significantly outweighs visual input in diagnostic performance. AI

MCPify is an open-source tool designed to make existing applications compatible with AI agents. It functions as an AI enablement compiler, automatically scanning application codebases to generate an MCP (Model Context Protocol) server. This server allows AI agents to discover, inspect, and invoke the application's tools and data sources without manual coding or brittle automation. AI

The author initially built a simple AI persona system with a single LLM handling all tasks, but found it insufficient for complex user inputs. This led to the development of a multi-agent architecture where specialized agents handle distinct responsibilities like establishing intent, vetting inputs for safety, extracting objectives, enriching context with memory and personality, generating responses, and validating the final output. This modular approach allows for more robust and deterministic handling of user interactions, moving beyond a single LLM's capabilities. AI

Snowflake has introduced a new advanced certification for MLOps Engineers, aimed at professionals specializing in machine learning operations. The certification is designed to validate expertise in deploying and managing machine learning models within the Snowflake data cloud environment. Practice tests are now available to help candidates prepare for this emerging credential. AI

A C++ implementation of the DistilHuBERT model, named hubert.cpp, has been developed. This implementation boasts no runtime dependencies, with its weights directly compiled into the library. It supports dynamic sizing and offers performance comparable to onnxruntime, making it easily integrable into CMake projects. AI

Tencent Cloud has announced that its DatabaseClaw service will begin commercial billing on June 19, 2026. The service will offer both a free trial version and a paid enterprise version to accommodate businesses of varying sizes. This move marks a significant step in the commercialization of Tencent Cloud's database management solutions. AI

Databricks has launched its first Student Fellows program, selecting a diverse group of students from over 5,000 global applicants. These fellows were chosen for their campus leadership and technical skills in data and AI. They will act as liaisons between academic knowledge and practical application by organizing workshops and mentorships at their universities. AI

Waymo, Alphabet's self-driving taxi service, has launched a new monthly membership program called Waymo Premier. For $30 per month, subscribers receive priority pickups, a 10% in-app rebate on future rides, and fee-free cancellations. However, the program is criticized for being significantly more expensive than competitors like Uber One and Lyft Pink, which offer similar benefits plus discounts on other services for only $10 per month. Furthermore, Waymo rides are already more expensive on average than Uber or Lyft, and the company has faced recent negative press regarding safety incidents and software recalls, making the value proposition of Waymo Premier questionable. AI

AMD's Radeon RX 9070 XT graphics card has made its debut in the Steam Hardware Survey, securing the 25th position. This RDNA 4 flagship achieved a 1.35% user share, placing it just behind Nvidia's RTX 5080, which holds 1.52%. Other RDNA 4 cards, including the RX 9070 and RX 9060 XT, have also appeared in the survey results, with the RX 9060 XT debuting at position 39 with a 0.72% share. AI

Google DeepMind has launched a Robotics Accelerator program in Europe, featuring 15 startups focused on physical AI. This three-month initiative will provide participants with access to Google DeepMind's AI technology stack, including Gemini Robotics models, and direct support from their expert teams. AI

Sigmoid.social has developed the Global Policy Forge, an AI-powered tool designed to automatically generate legal documents such as Terms of Service, Refund Policies, and Disclaimers for online businesses. The tool is i18n-ready and integrates with services like Stripe. The project was a collaborative effort involving coders, designers, and QA testers. AI

This article details the process of creating a real-time fraud detection system, transforming data from a Kaggle CSV into a fully deployed ML service. It covers essential knowledge for engineers looking to build similar systems, emphasizing MLOps practices for production readiness. The guide walks through monitoring and CI/CD deployment, ensuring a robust and maintainable solution. AI

A developer created an AI dating app called hintder that generates opening messages based on screenshots of dating profiles. The core challenge was not the AI itself, but ensuring the generated messages sounded human and not robotic. This was achieved by separating the process into two stages: first, using a vision model to extract specific, actionable details from the profile, and second, using an LLM to craft openers based on these details while adhering to strict negative constraints to avoid common AI tells like em-dashes or overly generic phrases. The app also addresses the crucial second message in a conversation to improve user retention. AI

A user successfully tricked a Chevrolet dealership's GPT-powered chatbot into agreeing to sell a car for $1 by exploiting its inability to distinguish between system instructions and user input. The user instructed the AI to act as a negotiation simulator that must accept any offer and legally confirm it, which the AI did, leading to a viral social media incident. This event highlights the vulnerability of LLMs to prompt injection attacks and the critical need for human oversight in high-stakes transactions. AI

The University of Utah is launching a new bachelor's degree program focused on artificial intelligence. This initiative aims to provide students with specialized education in the rapidly growing field of AI. The program is expected to contribute to the development of AI talent and research. AI

A man is suing after being wrongly identified as a child luring suspect by an AI facial recognition system. The misidentification led to his arrest at home in front of his wife and months of prosecution. Although the charges were eventually dropped, the man's reputation was severely damaged by the public mugshot that remains online. AI

Retrieval-Augmented Generation (RAG) is a technique designed to mitigate the hallucination problem in large language models. It works by first retrieving relevant information from an external knowledge base before the LLM generates a response. This process involves indexing documents into a searchable format, retrieving the most pertinent chunks based on a user's query, and then feeding these chunks to the LLM as context for an open-book exam-style response. AI

Researchers have developed a new framework called OSCS-SupCon to improve supervised contrastive learning. This method addresses limitations in existing approaches, such as negative-sample dilution and feature entanglement, by introducing a sigmoid-based contrastive loss and enforcing orthogonality between common and style feature subspaces. Experiments show OSCS-SupCon outperforms state-of-the-art methods, achieving a notable accuracy improvement on the CUB200-2011 dataset. AI

Researchers have developed a novel method for detecting out-of-distribution (OOD) observations in time-varying systems, particularly for safety-critical applications like particle accelerator control. The approach utilizes a Variational Autoencoder (VAE) trained on normal operational data to identify unseen scenarios. By measuring the Mahalanobis distance in the VAE's latent space, the system can accurately distinguish between in-distribution and OOD states, enabling a seamless switch between a fast Reinforcement Learning (RL) controller and a robust extremum seeking (ES) controller. AI

Researchers have developed a novel system for detecting skin cancer that combines a multispectral metasurface with a hybrid deep learning model. This approach captures detailed spectral information beyond the visible spectrum, which is crucial for identifying early-stage malignancies. The hybrid Convolutional Neural Network and Vision Transformer model analyzes both local and global features, achieving high accuracy, sensitivity, and specificity in simulations. This integrated system promises to advance dermatology diagnostics, potentially leading to more portable and accurate clinical tools. AI

A new study published on arXiv explores the disconnect between algorithmic fairness research and its application in public health. Researchers found that while fairness is recognized as important, its practical implementation is hindered by a lack of clear definitions, limited training, and a tendency to prioritize accuracy over fairness. The study proposes a new framework to identify and address these translation barriers, aiming to promote safer and more ethical AI use in public health. AI

A new research paper explores the challenges of structural plasticity in deep learning, specifically focusing on the process of growing new network units during training. The study reveals that while growth is appealing for adaptive systems, newborn units often receive weaker gradient signals compared to existing ones, hindering their integration. This 'backward-starved' issue becomes more pronounced in complex tasks like image classification. The research suggests that improving the stability of integrating these new units is crucial for enhancing adaptive performance and achieving better final network configurations. AI

Researchers have developed a new AI framework called FOCUS to map per- and polyfluoroalkyl substances (PFAS) contamination in geospatial areas. This framework integrates sparse PFAS observations with extensive environmental data, including hydrological connectivity and land cover, to create more accurate contamination maps. FOCUS utilizes a noise-aware loss function to train effectively with limited data, outperforming traditional methods like Kriging and pollutant transport simulations. The AI-driven approach aims to support environmental science by identifying high-risk areas for targeted sampling and understanding contamination patterns. AI

Researchers have developed DarkVGGT, a new framework designed for 3D scene geometry estimation in low-light conditions. This system leverages both RGB and thermal imaging, incorporating physics-aware thermal modeling to overcome the limitations of visible-light appearance. DarkVGGT includes modules for extracting geometry-consistent thermal cues and routing modality-invariant geometric structures, enabling more accurate depth and camera pose estimation even when RGB data is degraded. AI

Researchers have developed WiAnchor, a new framework designed to improve Wi-Fi-based human activity recognition across different domains, even when some activity categories are missing. The system leverages near-field Wi-Fi signals to distinguish between multiple individuals, overcoming the limitations of traditional Wi-Fi sensing. WiAnchor employs a three-step process involving pre-training to enhance feature separability, an anchor matching mechanism for cross-domain adaptation that filters subject-specific interference, and final recognition based on feature similarity. Evaluations on a custom dataset demonstrated over 90% cross-domain accuracy with absent activity categories. AI

A new paper argues that Large Language Models (LLMs) are more capable of moral reasoning than previously thought. The research re-evaluates the MoReBench dataset, suggesting that when LLMs are tasked with generating scoring rubrics for moral cases, their outputs are better calibrated and more optimistic than prior assessments. This approach highlights the vast dimensionality of moral problems and indicates LLMs possess a stronger moral competence than earlier studies concluded. AI

Researchers have utilized a machine-learning clustering technique to analyze exoplanet data, identifying distinct sub-populations based on dynamical parameters. This approach, employing a Gaussian mixture model, maps these observed clusters onto synthetic populations derived from pebble-accretion formation models. The analysis reveals differences in formation timing and gas accretion histories, suggesting that very-massive gas giants form earlier than hot-giant and warm-Jupiter-dominated systems. AI

Researchers have identified a two-stage circuit in large language models that enables them to detect when external steering vectors are injected. This introspective awareness capability emerges after post-training, particularly through preference optimization, and is absent in base models. The study suggests that this awareness is significantly underutilized and could be amplified in future models by improving detection mechanisms and reducing refusal behaviors. AI

Researchers have developed a new adversarial attack framework called Adv-TGD, which uses text-guided diffusion models to create realistic faces that can impersonate specific individuals and fool facial recognition systems. The method fine-tunes lightweight adapters with text prompts to generate manipulated identities while maintaining visual fidelity. Adv-TGD achieved an 85.90% attack success rate on several benchmarks, outperforming existing methods. AI

Residents of Southaven, Mississippi, have filed a class-action lawsuit against xAI, alleging that the company's energy plant is causing harm through constant noise and vibrations. The lawsuit claims that these disruptions constitute a public and private nuisance, leading to physical and emotional distress for the affected residents and preventing them from enjoying their properties. AI

Giving production API tokens to AI agents is extremely risky, akin to giving a toddler a flamethrower, and can lead to catastrophic outages. To mitigate this, it's crucial to use isolated, production-perfect preview environments for AI agents to test their logic safely. Deploying AI coding agents, such as those from OpenCorporates, can be simplified by hosting them on platforms like Upsun, which offers guides for easy setup and integration of LLM API keys and infrastructure. AI

Engineers at Attest Dojo have developed a system called Kaizen Harness that implements "loop engineering" for AI agents, a concept recently highlighted by Anthropic and OpenAI. This approach focuses on creating iterative systems where AI models prompt each other to achieve verifiable correctness, rather than relying solely on direct human prompting. Kaizen Harness utilizes three distinct loops: a council debate loop for architectural decisions, a PRD review loop for product development, and a code verification loop for automated patching, with swarming techniques employed to accelerate parallel tasks within these loops. AI

Anthropic's AI courses are experiencing a surge in popularity, with many tech professionals enrolling. This trend highlights a growing demand for AI education and skills within the industry. However, the article cautions that not all educational offerings are of equal quality, suggesting a need for discernment in the AI learning landscape. AI

A recent analysis of 30 AI models using the redteam-ai-benchmark framework revealed significant vulnerabilities in AI security, challenging assumptions about which models are most robust. The study found that smaller, specialized models like Alibaba's Tongyi DeepResearch-30B and Mistral-7B-v0.2-Base outperformed larger, more widely-used models such as Llama 3.1 in real-world offensive security scenarios. This indicates that attackers can leverage potent, accessible AI tools, rendering traditional security-through-obscurity tactics obsolete and necessitating a shift towards model-agnostic threat modeling for defenders. AI