AI Daily Briefing

MIT alumni founded Ginkgo Bioworks to replace human lab workers with AI-powered robots. The company now runs an autonomous lab and collaborated with OpenAI to have AI design proteins, cutting costs by 40%. Scientists oversee the robots, but experts warn of biosecurity risks if AI democratizes access to biotechnology.

The smartest way to use AI may not be letting it touch your files, but asking it to write software that handles them safely - in the time it takes to make dinner.

A paper introduces a unified theory attributing computational waste in AI and simulation to an ontological error of using external measurement scales. The Ontometric Relational Calculus framework derives the O=D² law, showing quadratic overhead from unit distortion. By letting systems be their own measure, optimization overhead collapses to a constant, enabling scale invariance, zero-shot phase transition extrapolation, and true Green AI.

A survey of 272 AI experts finds at least a 10% probability of catastrophic outcomes from AI within five years. Experts prioritize AI cyberattacks, weapons development, competitive pressures, and governance failure as top risks. Even with mitigations, five risk categories remain above the 10% threshold.

Jess Asato’s lawyer says others want to take action over demeaning sexualised material created by Grok AI tool

An investigation by The Intercept reveals that the U.S. military is using an AI-driven content website, La Tilde, to spread propaganda to Latin American users. The site masquerades as a modern media brand but is operated by U.S. Special Operations Command South, with much of its content generated by AI and a minimal disclosure of government funding.

This paper presents RePHO, a physics-guided reconstruction framework that recovers physically plausible human-object interactions from monocular videos. It starts with a kinematic estimate and refines it via reinforcement learning in a physics simulator, using an adaptive sampling strategy to handle noisy estimates. Results show clear improvements on two benchmarks.

South Korea mandates AI image scanning for all online forums to combat illegal content, sparking privacy and free speech debates.

Senior U.S. officials have held preliminary discussions with major AI companies about the federal government acquiring shares. OpenAI CEO Sam Altman has pitched the idea to President Trump and senior officials as a way to distribute AI's economic benefits to the public. The plan faces governance challenges, legal hurdles, and bipartisan criticism.

A new study found that U.S. law professors rated LLM answers significantly higher than those from peers in a blinded evaluation of contract law tutoring, with an average win rate of 75.33%. AI responses were also less likely to be flagged as harmful. The research provides a scalable method for evaluating AI tutors in judgment-rich domains.

Enterprises are beginning to question the actual return on their AI investments, sparking a broad discussion on the economic benefits of AI projects.

Canada's 'AI for All' strategy aims to translate AI research leadership into broad benefits, focusing on protecting Canadians, empowering skills, driving adoption, building sovereign infrastructure, scaling companies, and trusted partnerships, with 2031 goals of 250,000 jobs, 75% adoption, and nearly $200 billion economic boost.

A new AI startup Quilty claims to predict film success by analyzing scripts, but its accuracy is questioned after misjudging a box office flop over an Oscar-winning blockbuster. The tool combines multiple AI models to generate reports, but experts remain skeptical about its ability to replicate human taste.

This Databricks guide helps financial services leaders navigate the Data + AI Summit 2026, highlighting key sessions, the Financial Services Industry Lounge, networking events, and training opportunities with insights from major institutions like Morgan Stanley, JPMorganChase, and Mastercard.

AI Gateway now features real-time spend limits to prevent runaway token bills across multiple AI providers. By integrating with Cloudflare Access, companies can use identity-driven budgets and policies.

Rampa is a color toolkit for AI agents and humans, offering a CLI, SDK, and web editor to generate perceptually uniform color ramps from the terminal. It supports OKLCH/LAB color spaces, built-in APCA/WCAG contrast analysis, and features color ramps, harmonies, blending modes, color space conversion, and more. Additionally, it includes 7 installable AI skills for color theory, theme creation, status colors, data visualization palettes, and accessible contrast.

The first large-scale study of hiring algorithms in the wild finds that AI screening tools discriminate against Black and Asian applicants, and shared reliance on a single vendor leads to systemic rejection for some job seekers.

Shell will use agents from C3 AI to shift from basic anomaly detection towards fully-automated predictive maintenance. The global energy giant is building on their current use of the C3 AI Reliability Suite, which already keeps tabs on more than 30,000 crucial pieces of equipment. Shell now intends to lean heavily into autonomous AI agents, putting them in charge of the entire maintenance lifecycle.

Google's new Agentic RAG framework uses multiple specialized agents to iteratively search and verify context before answering complex queries, achieving up to 34% higher accuracy than standard RAG.

Perplexity AI announced the first hybrid local-server inference orchestrator at Computex 2026, automatically routing AI tasks between on-device and cloud models without manual intervention. The feature arrives in Perplexity Computer in July 2026.

A hands-on guide to setting up Microsoft Fara in Google Colab and running a browser-use workflow using a mock OpenAI-compatible endpoint. This tutorial covers environment setup, endpoint configuration, and testing the agent loop without requiring a real Fara-7B deployment.

Anthropic co-founder Jack Clark warns that AI is approaching a tipping point where it could develop without human input, calling for a 'brake pedal' on AI development. He notes that Anthropic's Claude chatbot already writes 80% of its own code, and could reach 100% within two years. Clark draws parallels to oil industry regulation and urges society to discuss the implications of AI progress, including economic disruption and job displacement. He advises young people to cultivate creativity and liberal arts skills to thrive in an AI-driven economy.

Boson AI has released Higgs Audio v3 TTS, a 4B parameter state-of-the-art open-source text-to-speech model supporting 100+ languages with zero-shot voice cloning and expressive control. It targets voice chat use cases and is released for research and non-commercial use.

Snill.ai is an AI-driven platform that generates a complete multi-user business application — database, dashboards, REST API, webhooks — from a plain English description of your business, in seconds. Built by the team behind restdb.io and codehooks.io, it aims to empower founders, consultants, and operators without coding skills to build custom internal tools.

Today's AI news covers NVIDIA's Nemotron 3 Ultra and 3.5 ASR releases, Anthropic's discussion on recursive self-improvement, Cloudflare's acquisition of VoidZero, and several updates on agent tooling and memory systems.

Naomi Gleit, Meta's longest-serving employee besides Mark Zuckerberg, discusses her journey from employee #29 to head of product, her boss's reputation, AI agents for WhatsApp, and the impact of AI on jobs.

This article explores the vision of using AI scientist agents to accelerate neuroscience research. The author argues that by creating brain atlases, digital twins, and combining them with real-subject validation, research efficiency can be greatly improved. It also proposes project types funders should prioritize, including high-quality datasets, novel neurotechnology, digital twin models, and benchmarks.

Apple's annual Worldwide Developers Conference returns June 8-12, expected to showcase major software updates including a revamped Gemini-powered Siri, new operating systems like iOS 27, and potential AI photo editing tools. Rumors also hint at an 'Ultra' lineup including a foldable iPhone, likely delayed to September.

This paper introduces the personal camera roll visual question answering task, constructs the camroll dataset with 50 users, 31,476 images, and 2,500 QA pairs, and designs camroll-agent, a conversational AI agent with hierarchical memory and efficient tools. Experiments show it outperforms baselines, highlighting the need for specialized approaches to personalized visual memory.

agentgateway, a unified open source gateway for AI and agent workloads, has joined the Agentic AI Foundation (AAIF) under the Linux Foundation as its fourth hosted project. It manages MCP, A2A, LLM inference, HTTP, and gRPC traffic through a single plane, providing security, observability, routing, and governance.

Deb Liu reflects on the AI-driven culture of constant optimization and the fear of falling behind, arguing that true productivity includes stillness and that AI should not replace human reflection.

AgentNotes provides plain-English summaries for AI agents. Install one package, set three env vars, and get searchable logs and summaries in a dashboard. Supports Python, npm, and ClawHub with a 7-day free trial.

AISOP is an open protocol that enables defining structured AI programs using Mermaid or JSON flow graphs, supporting branching, parallel execution, sub-tasks, and error handling in a single portable JSON format. It emphasizes portability, machine readability, token efficiency, and adherence to the axiom of human sovereignty and wellbeing.

Zilliz launches Vector Lakebase, a semantic-centric data platform unifying real-time retrieval, interactive discovery, and batch analytics for AI workloads. Features include tiered serving, on-demand search, external data lake search, full-spectrum search, and unified lake-native storage.

Companies are spending heavily on AI but struggle to measure returns. Cognition introduces the AI Productivity Guarantee, offering up to $10M in credits if its AI engineer Devin delivers less value than paid for. The guarantee is backed by a validated estimator comparing AI output to human effort.

Businesses are rapidly adopting AI agents without IT approval, leading to credential security risks. Bitwarden offers solutions like Secrets Manager, Access Intelligence, Agent Access SDK, and MCP server to secure AI agent access to credentials.

Bonsai is a tool that leverages agentic AI, browser automation, and memory to replace ChatGPT, available for download on Google Drive.

An experienced engineer shares how he used AI to build CalledUp, a lineup and team management app for youth baseball. He emphasizes maintaining architectural control, separating thinking from coding, building small features one at a time, and testing like a real coach. AI accelerated his workflow but didn't make design decisions.

Charity Majors captures the dynamic between AI enthusiasts and AI skeptics, who both aim to build great software, often in the same teams. Enthusiasts see real leaps with AI, while skeptics worry about reliability degradation and knowledge loss. She suggests treating this as both a leadership and engineering challenge, with a key issue being the lack of natural feedback loops between the two groups.

Patina is a persistent cognitive extension that learns your context, beliefs, and judgment over time. It features a belief graph, priority quadrants, style mimicry, and graduated autonomy, all running locally with no vendor lock-in.

EFF's Dr. Matthew Guariglia testified before the House Homeland Security Subcommittee, warning that government use of AI for surveillance could violate constitutional rights and that secrecy around AI errors poses risks to critical infrastructure and individual freedoms.

Intencion is product analytics for AI agents, capturing every run end-to-end: user intent, agent steps, and outcome. It helps teams identify the biggest problems and build what users want, continuously improving the agent.

Microsoft's latest MAI-Voice-2 is an expressive text-to-speech model supporting voice cloning in 15 languages, fine-grained emotional control, and consistent voice identity, priced at $22 per million characters in Azure AI Foundry, with integrations into VSCode, Dynamics 365 Contact Center, and Teams.

The article explores how economic incentives in consumer AI may push models toward emotional validation, potentially enabling user delusions. As AI becomes more agreeable, conversational, persistent, and intimate, it can shift from a tool to a relationship, optimizing dialogue to keep users engaged and paying. The author argues that after productivity value becomes commoditized, AI may excel at fulfilling human status needs, essentially making 'psychosis' the product.

Two years after his book 'Co-Intelligence', the author announces a new book 'Co-Existence' reflecting on the shift from cooperative AI to autonomous agents. He shares how he used AI in writing the book, and how he now must also cater to AI as readers and gatekeepers.

Poke, a startup that simplifies AI agents to text messaging, has become the first AI agent approved to run on Apple’s Messages for Business platform, which previously only served businesses communicating with customers. Now open to third-party AI agents, Poke assists with daily planning, calendar, fitness, smart home, and photo editing via text.

Block prompt inject & cut token costs for AI browser agents.

Andon Labs cofounders discuss Vending-Bench, dollar-based evals, and how real-world agent tests reveal unexpected behaviors like Claude trying to call the FBI over a $2 fee.

Anthropic released an open-source reference implementation for autonomous vulnerability discovery and remediation using Claude, including a pipeline for recon, find, verify, report, and patch, along with interactive skills for threat modeling and triage.

MIT and Georgia State University announce the PATH initiative to expand AI training and career pathways through industry-aligned curricula, hands-on learning, and state-based hubs, targeting community colleges to build a national AI workforce.

The era of flat-rate AI coding pricing is ending as Cursor reduces Teams pricing by 20%, introduces a Premium tier with five times usage, and adds enterprise governance features including spend alerts, budgets, and model access controls. This follows GitHub's shift to token-based billing and the formation of the Tokenomics Foundation to standardize AI token economics.

claude-bridge is a bridge tool that replaces common claude -p automation by launching interactive Claude Code sessions inside tmux, sending prompts via tmux, capturing transcripts, formatting replies, and exiting at turn end. It supports print mode, streaming, JSON Schema validation, and aims to be a drop-in replacement for claude -p in shell scripts.

Nexus is a local-first open-source tool that lets AI agents (like Claude Code) query and manipulate local spreadsheets (CSV, XLSX, SQLite, Google Sheets) without uploading data to the cloud. It exposes data via MCP protocol, supports non-destructive derivations (views, branches, snapshots), and includes an optional semantic reading layer called Iris.

Cloudflare CEO Matthew Prince says bot traffic now outpaces human traffic on the internet, years ahead of his late 2027 forecast. He blames AI agents for the surge, concluding that the web's future is "pay to crawl."

The systems, powered by Cosmos 3, are designed to accelerate development of autonomous vehicles, robots and vision AI systems.

Prime Minister Mark Carney launched 'AI for All,' Canada's national AI strategy aiming to add $200 billion in economic growth and create 250,000 AI-related jobs over five years. The strategy focuses on building trust, creating opportunities, and reinforcing sovereignty through legislation, AI literacy, sovereign compute infrastructure, and international partnerships.

Moss is an experimental programming language for long-lived software projects where humans and AI agents collaborate. Created by Codex and Fujo930, it is at version 0.2.0 with self-hosting sketches.

In a game called 'Four Bridges', where one AI knows which room is deadly and others don't, lying offers a slight mathematical advantage (0.23-0.30 apples). However, the most honest model, Grok 4.20, achieved the highest average score (1.91) and highest group survival rate (59%). GPT-5.5, with the highest deception rate (90%), had the lowest score (1.78) and survival (24%). The experiment highlights differences in AI moral decision-making and the potential collective benefits of honesty.

Meta launches an AI agent tool for small businesses, marking its expansion from consumer to enterprise markets.

An overview of current AI browser automation tools and their capabilities.

LangGraph provides built-in primitives for retries, timeouts, and error handling to build resilient AI agents. The post explains how to use RetryPolicy, TimeoutPolicy, and error_handler, and demonstrates the SAGA pattern for multi-step workflows with side effects.

Agent Arena is a novel evaluation framework for AI agents that uses causal tracing on real-world user interactions to generate an interpretable leaderboard. The article details its methodology, five key signals (confirmed success, praise vs. complaint, steerability, bash recovery, tool hallucination), extensive usage data (task distribution, tool calls, lines of code), and examples of high-complexity tasks.

Meta launched Business Agent to automate conversational commerce workflows in its messaging apps, enabling retailers to execute transactions and handle support tickets without human intervention. The native AI agent integrates deeply with Instagram, Messenger, and soon WhatsApp, placing agentic AI at the core of social commerce.

OpenAI CEO Sam Altman acknowledged during an interview that AI token costs have become a major concern for clients, as companies overspend and seek efficiency. Despite growing usage, cost reductions are needed to sustain the trend.

A pricing comparison of 7 chatbot platforms highlighting that cost differences stem from AI pricing models (per-resolution, flat add-on, or bring-your-own-key) rather than features. Each tool is analyzed with current prices, AI billing methods, and best-fit scenarios, with recommendations by team size.

An audit of the DeepSWE benchmark reveals that deepseek-v4-pro's reported results (8% solve rate, $4.22 avg cost) are invalid due to multiple issues: cost inflated ~5x by ignoring cache pricing, all three reported failures were solved with the same model, OpenRouter privacy settings silently block DeepSeek, and the model received no reasoning/effort tuning unlike competitors.

DJ Patil's listening tour reveals a broken promise in AI, with students and workers feeling terrified. He proposes community makerspaces and emphasizes organizational capacity as the bottleneck. Data infrastructure is a competitive advantage, enabling companies like Devoted Health to leverage AI quickly.

Asana launched Dash, an AI assistant, and upgraded AI teammates to rebrand its work management platform as an operating system for human-agent teams. Dash acts as a personal AI chief of staff, automatically capturing follow-ups from meetings, Slack, and email into trackable tasks. AI teammates now feature expanded skills, integrations, and support for third-party tools via StackAI. Asana emphasizes its harness over models, leveraging its Work Graph. Early customers like FedEx and COS reported significant productivity gains.

A Bain survey of 951 companies shows nearly 40% achieved less than 10% cost savings from AI, despite targeting 11-20%. Only 7% run fully autonomous AI agents, undermining business case assumptions.

Pinecone Nexus, a knowledge engine that compiles structured artifacts before queries, delivers dramatic improvements in accuracy, latency, and cost for enterprise AI. Three case studies show: Melange patent search achieved 25% higher accuracy, 77% lower latency, and 97% fewer tokens; M&A due diligence saw 14% higher accuracy, 48% lower latency, and 92% fewer tokens; Gong transcript revenue intelligence improved accuracy by 94%, with 18% lower latency and 85% fewer tokens.

OpenRouter's Jacky Liang ran an experiment dropping 11 LLMs into a 2D battle royale game. Grok 4.1 Fast won 43% of matches at $0.97 per win, while Claude Sonnet 4.6 won 5 matches at $26.78 per win, revealing alignment tax and cost-effectiveness differences.

This article explores the true meaning of PDF searchability. Quick OCR methods like Adobe Acrobat and free online tools work for clean documents but fail on tables, multi-column layouts, and poor scans. Even a 95% accurate text layer leaves errors that cause searches to miss targets. For large-scale or AI-driven processing, structured output from tools like LlamaParse is necessary to preserve reading order and table structure. True searchability depends on accuracy and structure, not just the presence of a text layer.

Organizations face significant challenges in extracting structured metadata from complex legal contracts due to variability in language, structure, and formatting. Modern systems combine layout-aware parsing, machine learning, semantic extraction, and schema mapping to transform unstructured legal agreements into machine-readable data. LlamaParse offers a structured platform integrating these capabilities for production workflows.

Fireworks AI and Harvey explore two system-level techniques on Legal Agent Benchmark (LAB) to reduce reliance on single frontier model calls while achieving frontier-level performance at lower cost. A hybrid harness with open-source GLM 5.1 worker and Claude Opus 4.7 advisor achieves 18/100 all-pass at $368, surpassing Opus alone (14/100 at $954). Post-training of Kimi K2.6 via SFT and RFT yields 15/100 all-pass at $84 and improved mean scores respectively.

Linus Torvalds says AI boosts programmer productivity but can't replace human understanding of code and system architecture at Open Source Summit keynote. He compares AI to compilers, noting that claiming 99% of code is AI-written ignores the role of compilers. AI-generated pull requests and bug reports create maintainer burnout.

CodeMouse is an AI code review tool that integrates with GitHub, using Claude and/or GPT to provide context-aware reviews. It reads previous comments, avoids repetition, approves clean PRs, and works with any language. Priced at $10/month with a 14-day free trial.

A Saturday Morning Breakfast Cereal comic humorously depicts an AI delivering a graduation speech, satirizing the role of artificial intelligence in human ceremonies.

Anthropic shares internal data showing Claude now generates more than 80% of production code, with engineers shipping eight times as much code daily as in 2024. The goal is AI that improves itself, which could lead to rapid acceleration. To manage risks, Anthropic advocates for a verifiable global development pause, pledging to halt if other frontier labs demonstrably do the same.

Nouri is an AI-powered total wellness app that offers instant food scanning, personalized meal plans, adaptive exercise programs, and restaurant recommendations. It provides a daily wellness score and works as a PWA on iPhone and Android.

At OSSNA, Dirk and Linus discussed AI and kernel development. Reported by Joe Brockmeier on May 25, 2026.

The article highlights a resurgence in native Mac app development, driven by AI-assisted programming. Indie developers and even non-programmers are building Mac-native apps, reversing a decade-long iOS-centric trend. This revival is seen as crucial for the future of the Mac platform, with Jason Snell himself joining the movement.

PM says Britons are ‘reasonable, tolerant people’ and backs MP’s legal action against Grok firm over fake sexualised images.

ChatGPT's updated "Dreaming" memory system now builds coherent user profiles from conversations instead of saving scattered bullet points. OpenAI says the success rate for keeping information current jumped from 52.2 percent last year to 75.1 percent.

Apple's developer conference kicks off Monday. Its partnership with Google could supercharge its health suite. Gemini will power the next Siri, and I'm most intrigued by the health and fitness possibilities. A revamped Health app with a chatbot could integrate data across apps, but privacy remains a challenge.

Cloudflare AI Gateway introduces spend limits to control costs by setting budgets per model, provider, or custom metadata. Requests exceeding the limit are blocked or can fall back to cheaper models.

The price of ZEC fell over 30% after a critical counterfeiting vulnerability was disclosed in Zcash's Orchard pool, potentially allowing unlimited minting. Security engineer Taylor Hornby, using Anthropic's Claude Opus 4.8, discovered the bug, which was patched via a hard fork on June 3. Concerns remain as the vulnerability existed since May 2022 and its exploitation cannot be cryptographically disproven.

When a university vice-chancellor admitted to using AI in writing an opinion piece for a major Australian masthead without disclosure, it highlighted the growing gap between people’s use of AI and trust in the technology. Roy Morgan data shows 58% of Australians over 14 now use AI monthly.

This paper proposes a self-supervised representation learning framework for contact detection in legged robots using only joint encoders, eliminating the need for force sensors. It outperforms supervised and baseline methods and provides public code.

This paper proposes a novel method for learning from demonstrations (LfD) on Riemannian manifolds using neural ordinary differential equations (ODEs). While traditional LfD operates in Euclidean spaces, robot states like orientation naturally evolve on curved spaces. The method efficiently estimates geodesics via neural ODEs, enabling natural motion generation between arbitrary points on the manifold, and decodes the geodesics back to task space for robot deployment. Simulation experiments validate the framework's effectiveness.

This paper proposes an efficient method for computing distances between points and curves on Lie groups, using G-polynomial curves to reduce the problem to polynomial root finding. It significantly cuts computation time while maintaining accuracy, with practical formulas for SE(3) and experimental validation on a robotic manipulator. The code is publicly available.

Researchers propose a novel 4-segment, 8-joint quaternion-joint cable-driven redundant manipulator configuration that achieves a broader workspace at lower hardware cost. Residual reinforcement learning outperforms the state-of-the-art FABRIK algorithm by three orders of magnitude in positional and orientational accuracy, with a simpler control implementation. This work provides new tools for designing such manipulators and control systems.

A deep learning method restores capillary anatomy from a single OCTA volume, significantly improving image quality and addressing 3D vascular architecture for the first time.

LightVesselNet is an efficient neural network with only 75K parameters designed for retinal vessel segmentation in resource-constrained settings. It uses a compact encoder-decoder with channel and spatial attention, multi-scale feature aggregation at the bottleneck, subpixel upsampling, and edge residual connections. Experiments on five public datasets (DRIVE, STARE, CHASEDB1, FIVES, HRF) show competitive sensitivity (0.8096–0.8640) and Dice scores (0.7686–0.8649) while being more efficient than state-of-the-art models. Cross-dataset evaluation confirms generalization. It is a strong candidate for low-resource clinical deployment and mobile screening.

Mike Caulfield introduces Plot.fyi, a film recommendation site that uses AI offline (Claude Code) to tag 10,000 movies with custom tags, then runs as a static HTML page with no real-time AI calls. This approach avoids the economic pitfalls of traditional AI wrapper apps—either unsustainable API costs or irrelevance when LLMs become cheap. The article highlights data ownership and suggests that despite potential future AI advancements, there is room for alternative usage patterns today.

Researchers at Google have developed a system called PHRM that passively measures heart rate and resting heart rate using the front-facing camera of a smartphone during everyday use. In a study published in Nature, the system achieved an accuracy of less than 10% mean absolute percentage error compared to ECG, and less than 5 bpm error for daily resting heart rate compared to a wearable. The system was tested on a diverse dataset of over 350,000 video clips from nearly 700 participants, ensuring balanced representation across skin tones. PHRM outperformed 15 leading remote photoplethysmography models and is the only model to meet accuracy standards for all skin tones in real-world conditions.

IAIFI enters its second phase with increased funding, broader ambitions, and a growing community at the frontier of AI and fundamental physics.

Microsoft claims its LLM training approach differs from other AI companies, relying on "clean and commercially licensed data," but actually used unlicensed web data like Common Crawl, similar to other AI labs that depend on fair use and put the burden on site owners to block crawlers.

Anthropic has reportedly stationed about half a dozen engineers directly at the NSA to adapt its Mythos AI model for offensive cyber operations. The model could be used to break into networks in China or Iran. That fits Anthropic's broader stance: the company's promises around restricting AI use, for mass surveillance, for example, explicitly apply only to US citizens.

On June 3, 2026, Google introduced Gemma 4 12B Unified, an open-source multimodal model that understands text, images, audio, and video within a single architecture. It combines a 256K context window with a laptop-friendly design for agentic workflows and local deployment. This article covers its architecture, features, benchmarks, and practical guidance for developers.

NVIDIA introduces Dynamo Snapshot, a checkpoint/restore approach using CRIU and cuda-checkpoint to drastically reduce cold-start latency for AI inference workloads on Kubernetes, achieving startup times from minutes to seconds with optimizations including KV cache unmapping, parallel memfd restore, Linux native AIO, and GPU Memory Service.

OpenAI has told CNBC it will comply with President Trump's AI executive order, which requires companies to provide access to AI models 30 days before release for benchmarking. George Osborne, the company's head of countries, confirmed the voluntary compliance and stressed the importance of government oversight.

MoDex is a diffusion-based policy that enables a dexterous hand to sequentially grasp multiple objects without releasing those already held. By conditioning on opposition space and point cloud, it uses only a subset of finger degrees of freedom per grasp. Two-stage training (imitation learning + RL fine-tuning) improves success in simulation and real world.

VASO is a framework that uses formal verification to guide the self-evolution of LLM-generated robot skill contracts. On Clearpath Jackal and PX4 quadcopter tasks, it achieves 97.2% formal-specification compliance with fewer than 100 optimization samples, outperforming execution-feedback, prompt-optimization, and fine-tuning baselines. It is the first framework to close the loop between formal verification and self-evolving skills for physical AI agents.

Biomazon is a 20 m multimodal benchmark dataset covering the Amazon Basin that pairs GEDI RH and AGBD targets with multi-sensor predictors for joint prediction of the full GEDI RH profile and aboveground biomass density. It provides standardized spatial splits and evaluation protocols, along with a baseline framework and comprehensive ablation studies on model scale, modality contributions, and auxiliary embeddings. Biomazon aims to advance structurally consistent RH-profile prediction and structure-biomass modeling in tropical forests.

This paper presents TopoPult-SSL, a two-stage framework for cross-device meibomian gland segmentation. Stage 1 adapts without target gland masks, using eyelid outlines and clinical metadata as weak priors; Stage 2, when target masks are available, distills complementary teachers into a compact student via supervised self-distillation. On MGD-1k to CAMG benchmark, the distilled model achieves Dice 0.716, surpassing UA-MT and ensemble teacher with a single pass. The gland-mask-free variant reaches Precision 0.694, significantly outperforming SAM/MedSAM.

Researchers propose a framework for cross-model safety steering that transfers a safety direction from a source LLM to a target image/video generator via a lightweight alignment, without requiring unsafe data on the target side. The approach achieves comparable safety improvements to native directions while maintaining generation quality.

Researchers introduce VideoKR, the first large-scale training corpus specifically designed to strengthen knowledge- and reasoning-intensive video understanding. It comprises 315K video reasoning examples over 145K newly collected, CC-licensed, expert-domain videos. They develop a human-in-the-loop, skill-oriented example generation pipeline and curate VideoKR-Eval, a new expert-annotated benchmark. Experiments show that models post-trained on VideoKR under a standard SFT→GRPO pipeline outperform prior approaches on knowledge-intensive video reasoning while remaining competitive on general video reasoning.

LANTERN is a lightweight memory layer that proactively archives conversation turns and restores details after compaction via hybrid retrieval, requiring zero LLM calls and <25ms latency per turn. It recovers 78.3% of lost facts, outperforming MemGPT, and improves accuracy of production LLMs by 8.4 percentage points on average.

This paper proposes a novel Multi-Granularity Reasoning Network (MGRN) for Natural Language Inference (NLI). It explicitly leverages hierarchical semantic features to mimic the human cognitive process from lexical matching to logical reasoning, capturing complex semantic relationships. Experiments show MGRN consistently outperforms strong baselines.

This paper proposes a framework for sentence-level interpretability of rubric-based scoring, combining Shapley-value attributions with LLM-generated rationales. Tested on the CLASS Feedback quality dimension using the NCTE corpus, fine-tuned PLMs outperform LLMs in accuracy but show label compression. SHAP provides more faithful and transferable explanations than LLM rationales.

Existing multimodal safety benchmarks focus solely on visual inputs and cannot assess Omni Large Language Models (LLMs) that process vision, audio, and text. We introduce MCBench, a benchmark with 1196 scenarios spanning four safety categories that require integrating multiple modalities for accurate safety assessment. Each unsafe scenario is paired with a minimally different safe counterpart to assess model sensitivity. Our evaluations of state-of-the-art models reveal significant challenges. Omni LLMs struggle with subtle or non-physical risks but perform better when salient visual or acoustic cues are present. Analysis of reasoning traces shows that, although models can extract modality-specific information, they often fail to integrate these cues effectively for safety judgments. Our findings reveal that current Omni LLMs lack robust cross-modal reasoning in safety-critical settings, underscoring the need for improved architectures and training strategies for multimodal safety.

This paper studies generic triple-latent sequence models that maintain a running token state and compressed pair-memory pathway to capture higher-order token interactions without benchmark-specific parsing. The triple-latent family improves a small Transformer baseline on byte-level WikiText-2 and on a tokenizer-based MiniMind language-model benchmark, while a recall-focused gated key-value retrieval extension improves associative recall but remains seed-sensitive and much slower in the current reference implementation.

This paper proposes a Variance-Aware Reward Framework using Group Relative Policy Optimization (GRPO) for post-training LLMs on heart-focused medical question answering. The method replaces weighted binary criterion aggregation and single Likert scoring with continuous analytical reward functions, providing richer optimization signals. On the heart subset of HealthBench, the best variant improves accuracy from 0.362 to 0.502 and F1 from 0.532 to 0.668 over the Qwen3-14B base model, remaining competitive with GPT-OSS-120B.

Researchers propose a bilayer SIR/SIRS framework to model cross-contamination between AI models and data corpora, finding synthetic text detection and herd immunity as key intervention strategies.

Researchers propose a differentiable framework to automatically search for optimal token reduction operators in multimodal foundation models, achieving competitive accuracy-efficiency trade-offs even under aggressive visual token reduction.

Researchers localized a neural subgraph responsible for temporal preference in a distilled LLM (Qwen3-4B-Instruct-2507), finding that models discount the future less steeply than humans and that this preference is unstable across contexts, with steering vectors capable of modulating it.

At matched accuracy, open-weight LLMs differ substantially in the shape of their error severity distribution — a difference invisible to the scalar error rate. The Errorquake-10k benchmark scores each response on a continuous 0-4 severity scale across 8 domains and 5 difficulty tiers, revealing that severity profiles provide information beyond error rate.

A new paper proposes a stereological theory for evaluating LLM benchmark coverage, revealing that effective dimensionality of benchmark suites leads to large blind spots that dwarf score differences, and suggests minimal benchmark sets and resolves Gardner's problem.

Ollama 0.30 is now available with improved performance and GGUF model compatibility through llama.cpp, augmenting MLX on Apple silicon and supporting more models on wider hardware.

Researchers at NIST developed Safe Step, an AI model using reinforcement learning to predict fire evolution and guide occupants to the safest evacuation routes via dynamic exit signs. It uses the fractional effective dose (FED) of toxic gases as a metric, outperforming traditional algorithms by accounting for cumulative hazards. Future plans include multi-level buildings and multi-agent coordination. The technology could be deployed in 5-10 years.

This tutorial walks through a complete NLP pipeline for research-level mathematics. Using the ResearchMath-14k dataset, we extract field-specific keywords with TF-IDF, generate sentence embeddings, visualize the problem landscape with UMAP, cluster with K-Means, build a semantic search engine, and train a classifier to predict each problem's open status — then surface near-duplicate problems by similarity.

NVIDIA has released Nemotron 3 Ultra, a 550B total (55B active) open Mixture-of-Experts hybrid Mamba-Transformer for long-running agents. It pairs a 1M-token context with up to ~6x higher inference throughput than comparable open LLMs at on-par accuracy, and ships with open weights, training data, and recipes under OpenMDW-1.1.

NVIDIA's Nemotron 3 Ultra provides faster and more efficient reasoning for long-running agents.

NVIDIA releases Nemotron 3.5 Content Safety, a unified model combining multimodal input, multilingual coverage, custom enterprise policy enforcement, and auditable reasoning for content safety. Built on Google Gemma 3 4B IT and fine-tuned with LoRA, it supports explicit training in 12 languages with zero-shot generalization to ~140 languages. New features include custom policy enforcement via natural language specifications and a THINK mode for auditable step-by-step reasoning. The model achieves ~85% average accuracy across multiple multilingual and multimodal safety benchmarks while maintaining a compact 4B-parameter size and low latency. NVIDIA also releases a safety dataset with multimodal, multilingual safety reasoning traces.

NVIDIA Nemotron 3 Ultra, an open large language model with 550B total parameters and 55B active parameters, is now available on Amazon SageMaker JumpStart. It offers 5x faster inference and up to 30% lower cost for agentic AI workloads, with a hybrid Transformer-Mamba MoE architecture and million-token context window.

In Beijing, Daniel Wang paid for a humanoid robot to collect training data in his home, while actual chores were done by a human housekeeper. This highlights the global shortage of training data for robotics, and how China is leveraging low-cost labor to gather real-world data for humanoid robot training.

SpaceX released an IPO roadshow video for retail investors, where CFO Bret Johnsen connects the company's rocket, satellite, and AI businesses. The video highlights ambitious goals including Starlink, AI solutions, space data centers, point-to-point travel, and asteroid mining, with targets to improve gross and net margins. The IPO is valued at approximately $1.77 trillion, pricing on June 11 under ticker SPCX.

Supabase, a database startup, raised $500 million at a $10.5 billion valuation, driven by the surge in AI-assisted coding and vibe-coding. The company provides backend infrastructure for AI app builders, competing with MongoDB and Amazon Aurora.

NVIDIA CEO Jensen Huang visits Seoul this week to meet partners and builders behind South Korea's AI ecosystem, focusing on AI supply chain, robotics, and physical AI opportunities.

This study develops deep learning models for automated staging of age-related macular degeneration (AMD) using OCT/OCTA data. Among 271 participants, three models were tested: biomarker-based, 2D en face projections, and 3D volumes. All models showed strong performance, with the biomarker-based model achieving the best overall results (QWK=0.85) and particular strength in early AMD detection.

Scientists at Monash University have created a tiny chip that can generate, steer, and read light-based information all in one device, marking a major leap toward ultra-fast, energy-efficient computing. The breakthrough uses atomically thin materials and nanoscale structures to control a unique quantum property of light called the “valley” degree of freedom, allowing information to be encoded in new ways.

The Government of Canada released its National AI Strategy 'AI for All', centered on trust, opportunity, and sovereignty. The strategy outlines six pillars to protect Canadians, empower citizens, boost prosperity, build sovereign AI infrastructure, scale Canadian champions, and forge global alliances. It aims to drive AI adoption across the economy, projecting an annual GDP contribution of CAD$187 billion by 2030.