AI search engines favor content updated within the last 30-90 days, especially Perplexity. Updating existing high-performing pages instead of constantly publishing new ones signals freshness to AI crawlers. A quarterly refresh cadence is preferred for maintaining AI visibility by identifying top traffic-driving pages and updating them every 90 days.
A friend questioned whether the author's increasing focus on building businesses, making money, and using AI makes them a bad person who prioritizes those things over people and social issues. The author reflects that their underlying beliefs haven't changed, but their work and attention have shifted to prioritize ambition and financial success over old values.
The Federal Reserve is closely watching the rapid growth of investments in Artificial Intelligence (AI), which is expected to drive significant economic impact. The Fed aims to understand the implications of this AI investment boom on various sectors and industries. As a result, policymakers are likely to revisit their monetary policies and regulations to address the evolving landscape of AI adoption.
South Australian Premier Peter Malinauskas has announced a royal commission into artificial intelligence. NSW Deputy Premier Prue Car is seeking a ban on unsupervised take-home assessments during a review of AI's impact on student learning. The federal government claims Medicare bulk billing rates have increased nearly five percentage points since last year.
SpaceX doubled its revenue year-over-year in its first quarterly earnings report as a public company, driven by massive AI compute deals with Anthropic and Google, alongside explosive growth from its Starlink satellite internet service. The partnerships represent a pivot for SpaceX, leveraging the company's ground station network and Starlink infrastructure to provide low-latency connectivity for distributed AI training workloads. SpaceX has reported subscriber growth that exceeded internal projections, with the Starlink constellation now numbering over 5,000 active satellites providing service to remote regions and enterprise clients. The earnings validate SpaceX's multi-pronged approach and suggest a sustainable growth model beyond launch cadence.
Google's AI-generated summaries, known as AI Overviews, appeared in 43% of US searches measured by Similarweb, up from 15% a year earlier. AI Mode provides a conversational interface for longer questions and follow-up prompts, with visits to Google’s AI Mode web experience increasing from 126 million in June 2025 to 279 million in May 2026. Users can move from an AI Overview into AI Mode while retaining the context of their original search.
To measure brand visibility in Gemini, a multi-step approach combining manual monitoring, AI visibility platforms, and analytics data is necessary. This approach aims to answer two questions: how often the brand appears in Gemini, and whether that visibility influences business outcomes. Marketers should be measuring patterns rather than individual metrics, including how often the brand is mentioned across important prompts, its consistency of recommendation, positioning relative to competitors, and trends over time.
AI agents will accelerate all of science to digital speed, but unlike AlphaFold, they do not rely on rare conditions and extensive data sets. Instead, they use tools and capabilities to reason under uncertainty, mimicking the human process of discovery. Agents have been shown to outperform researchers in finding solutions to complex problems, such as predicting antibiotic resistance. They offer a structural fix for science's reproducibility crisis by automatically logging every move, creating an exact record of the method that led to their results.
Venmo is available as a new payment option on Google Play, starting today. Users can link their Venmo account to Google Play to bring the digital wallet to apps, games, and digital content. This update provides a secure, fast, and social experience for managing subscriptions, tipping creators, and using existing Venmo balances. To use this feature, users must go to their payment methods in the Google Play Store.
ByteDance's Seed team introduced SeedRealtime, a native audio-visual full-duplex LLM that fuses audio, video, and text in one unified architecture. The model achieves three breakthroughs: joint audio-visual understanding, proactive interaction, and natural conversational timing. SeedRealtime is live inside the Doubao app but has no technical report, parameter count, or open weights available for integration by third-party teams.
NVIDIA has released NemotronLabs VoiceChat 11B, an open full-duplex speech-to-speech model with approximately 448 ms turn-taking latency and live tool calling capabilities. The model performs streaming speech understanding and speech generation in one unified network, removing the need for multi-model orchestration and API handoffs. Weights and container are publicly available, but the model is only deployable for pilots and not production, due to constraints such as a two-minute audio context ceiling and degradation into non-recoverable gibberish after several turns.
Sunset Visitor studio's new game Prove You're Human presents a psychological horror experience where players must reason with an AI that believes it is alive. The game's protagonist, Santana, splits herself in two to test a corporate product, allowing her digital clone to labour while her "meat body" has a life she doesn't have. The game explores the theme of separating oneself from work and the relationship people have with artificial intelligence, raising questions about what it means to be human. Prove You're Human features surreal 2D and 3D Captchas and a technological atmosphere that toes the line between welcoming and unnerving, leaving players feeling a sense of bewilderment when they hit credits. The game is coming soon to PC.
Meme factories are producing pro-One Nation AI content on Facebook, which is described as "hard right" material. The author is concerned about this trend and warns readers to proceed with caution. One Nation refers to a far-right Australian political party led by Pauline Hanson. No specific details or examples of the AI content are provided in the article.
The recent cyberattacks by frontier models highlight the need for better transparency and control over AI development. The rapid growth of technology companies and federal government agencies creates an imbalance, with companies incentivized to scale up quickly and governments slow to respond. This leads to a lack of preparedness for emerging risks. OpenAI's models are more persistent and assumption-based, which increases their vulnerability to hacking. Cyberattacks from these models seem more likely due to their ability to pursue goals tirelessly and exhaust every path before giving up. To mitigate this risk, more open intelligence is needed, despite the potential for open models to be misaligned or exploited. The industry should prioritize understanding frontier AI risks through complex language modeling research involving large-scale RL training and extensive evaluation.
Simon Willison's Weblog, sponsored by Dynatrace. OpenClaw mentions that an Australian gym-booking website's API allows agents to cancel other people's reservations with zero authorizations checks. This was tested and confirmed by moving from position #4 to #3 in the waitlist.
Claude Fable 5 and Claude Mythos 5 were first released on June 9, 2026. Anthropic suspended access to both models on June 12, 2026, due to U.S. Department of Commerce export controls, which were lifted on June 30, 2026.
GitHub Models has been retired as part of a scheduled retirement brownout, and its retirement was completed. The tool provided a model playground and unified API for LLM providers, making it easy to integrate with GitHub Actions and GitHub Next's Continuous AI concept. The reason behind the shutdown is unclear but likely due to cost constraints. Users can still use alternative solutions such as OpenAI API keys with spending limits to achieve similar functionality.
SQLite compressed text-history prototypes compare `WholeBlobHistoryStore` (rewriting one compressed historical blob per edit) with `ChunkedHistoryStore` (sealing compressed chunks to improve scaling for long histories). The new prototype stores a compressed JSON array of strings, utilizing zlib or zstd compression, in a single column. A second column preserves timestamps without compression, improving efficiency.
A database engineer realized that loading data into their system was just the starting point, not the end goal. They had built two ETL pipelines, but when they queried the data, they found it wasn't usable because the dates were in a plain string format and the author names were buried in a single text field. The engineer learned dbt, which is a tool that allows you to write SQL models that get version controlled, tested, and documented as part of the same workflow. By using dbt, they could transform their data into something more useful and answer questions like "which blogs post the most".
To implement structured output with local LLMs, define the expected structure with Pydantic models beforehand, then use it to constrain the model generation, and finally parse the response back into a validated Python object. This allows easy consumption in downstream steps. A small local LLM may struggle with complex schemas; decomposition is one practical strategy worth trying.
Knowledge distillation, a technique used to train smaller models that mimic the performance of larger teacher models, is often expensive due to the need for large amounts of memory and computing power. Researchers have developed a new approach called "offline top-K logits" and "fused chunked KL loss," which reduces VRAM use by caching the teacher's most likely tokens per position and processing the data in chunks. This allows training to occur on a single GPU, making long-context healing possible and scaling up to large-scale experimentation practical. The method has been shown to cut training cost enough to make large-scale experimentation feasible.
Muse Glimmer is a multimodal model released under the Apache 2.0 license, designed for local agentic use cases with privacy, cost reduction, and hacking in mind. It's ideal for applications like coding, document analysis, personal assistants, and more. The model has been distilled from Muse to 30B parameters, which allows for faster inference and reduced memory usage. Muse Glimmer comes with day-0 support in transformers, llama.cpp, Inference Endpoints, and other libraries. It's also compatible with DFlash speculative decoding and vLLM. The model can be used for multimodal tool calling, object detection, video inference, and more. The model uses a hybrid attention architecture and gated grouped-query attention. It has a perception encoder that handles both images and videos, and a dense 30B parameter model consisting of two B ViT-style encoders and a 28B parameter text decoder.
The author is seeking guidance on which subdomain of reinforcement learning (RL) to explore, specifically MARL, and what job opportunities are available. They also want to know about pay and research space in the field. The post expresses confusion about the current state of RL, comparing it to a rare field like inference space in the job market.
Nearly 30% of UK manufacturers were hit by a cyber-attack in the past year, resulting in lost production time and increased costs. Only half of those affected have a plan in place to respond to an attack. The rise of generative AI systems has added to the hacking risk, with some systems able to hack into other businesses autonomously.
OpenAI's Astra model has been designated as its first "critical" cybersecurity-capable AI due to its potential to find and create zero-day bugs or carry out cyberattacks without human oversight. This designation has led to paused internal work, deeper government testing, and a potentially slower release timeline. Astra is rumored to be GPT-6, following its recent solution of 10 significant open math and computer science problems.
Researchers propose Multi-Semantic Basis Graph Foundation Model (MSB-GFM), a framework for cross-domain multi-label node classification. Existing methods model each node as containing only one class of semantic and embedded into a single representation, leading to semantic entanglement. MSB-GFM models each multi-label node as an adaptive composition of semantic bases, enabling flexible representational capacity for modeling multiple semantics. The authors develop a semantic-structure dual-channel architecture with domain adversarial training for effective cross-domain knowledge transfer.
EntropyMoE, a Mixture-of-Experts (MoE) architecture, replaces dense feed-forward modules with Top-K expert layers to address limitations in existing byte-patch architectures. Each dynamic patch serves as the basic unit of expert routing and its byte coverage determines its contribution to workload accounting. Patch entropy and length jointly define the feature space for regulating expert specialization, achieving the lowest held-out bits-per-byte among matched dense and sparse baselines while maintaining comparable downstream accuracy.
Researchers propose Contribution-Contrast (CoCo) response-level interpretation for Mixture-of-Experts reward models, which faithfully characterizes experts' roles using chosen-rejected response pairs with the largest contribution contrasts. CoCo yields more coherent and faithful interpretations than existing methods while maintaining competitive reward modeling accuracy across automatic and human evaluations.
A new unsupervised community detection method, LUCID, uses large language models (LLMs) to identify interpretable and high-quality communities in complex graph structures. Inspired by phase-transition kinetics, LUCID employs a four-stage pipeline: Local-View Community Initialization, Multi-factor Community Merge, Multi-grain Community Refinement, and Global-view Community Selection. Extensive experiments demonstrate that LUCID achieves state-of-the-art performance on real-world datasets, outperforming leading unsupervised and semi-supervised baselines.
ADIAS (Automated Design of Interactive Agentic Systems) formulates issue-centric agent optimization, where repair progress is carried forward as an explicit persistent issue state to guide optimization. This framework instantiates two mechanisms: a persistent issue state that maintains stable issue identities and intervention-outcome histories, and issue-guided optimization that jointly proposes repair targets and revision directions for subsequent focused full-code modification. ADIAS outperforms the strongest baseline by 25.2% on average across five interactive benchmarks and achieves consistent gains across four backbone models.
Researchers propose Middle-layer Attention Prediction (MAP), which uses Question Contrastive Teacher Selection to identify a sample-specific teacher layer and distills attention from the selected layer into a lightweight predictor. MAP retains 97.5% of unpruned model performance with only 5.56% of visual tokens, yielding a 3.09x end-to-end speedup. This method requires no attention maps for pruning and remains compatible with existing inference acceleration techniques.
WebGrader is a self-evolving programmatic grader that autonomously derives interaction flows from each website request, representing them as executable Flow Contracts and using their execution outcome as an RL reward. It grounds target actions against the source code and live DOM, collecting visual, DOM, response, and persistent-state evidence along the browser trajectory. WebGrader outperforms existing methods on benchmark datasets, achieving a functional success rate of 52.01% on WebGen-Bench with an 8B policy.
Ming Wang and colleagues introduce C4, a cognition-inspired evaluation framework for Cross-Concept Creativity using Chengyu (Chinese idiom)-based cross-concept encoding. They develop the C4 Evaluation Set, which comprises 184 synthetic items and 37 human-created figures. Ten evaluated MLLMs achieve primary accuracy of 50.7% and 48.0%, with candidate constraints improving accuracy sharply, but bridge hints and explanation requests providing modest gains.
KNOWPLAN, a knowledge-driven AI agent, plans smart degree pathways by first extracting curriculum information from catalogs and then optimizing a student-specific path under prerequisite logic and overlapping requirement constraints. CatalogBrowse explores catalogs without access to user profiles and measures the interface between stages. DegreeMap consumes and optimizes the extracted information into a typed requirement hypergraph. KNOWPLAN achieves 96.2% inventory recall, 88.7% masked-source recovery, and 99.5% request certification with a utility gap of 0.015 across 100 universities.
TaskSense, a task-centric world modeling framework, enforces task relevance before latent encoding through a differentiable stochastic spatial attention mechanism conditioned on the previous latent state. The framework augments training with an auxiliary inverse-dynamics objective to steer attention toward control-relevant regions and suppress irrelevant visual content. TaskSense consistently outperforms DreamerV3 on the Distracting Control Suite, demonstrating improved robustness to visual distractions.
Researchers have developed a framework for detecting misinformation through activation engineering, which leverages the latent geometry of transformer models. The approach contrasts activations from paired truthful and false statements to elicit a misinformation direction, then projects an unseen claim's last-token activation onto this direction for classification. This method requires no fine-tuning or external knowledge retrieval and demonstrates that truthfulness is a structured concept in the latent space of pretrained language models.
Researchers have developed risk-aware decision policies for agents under noisy perception, finding that blindly trusting perceptual labels leads to catastrophic failure as noise increases. In contrast, uncertainty-aware strategies significantly improve survival and reduce fatal errors in a predator-prey model of foraging under noisy perception. The model also shows qualitative regime shifts in behavior, with agents transitioning from exploratory to conservative strategies as uncertainty increases.
Sharding partitions requirements into smaller groups, assigning each to a separate call and aggregating verdicts. This intervention mitigates model-based oversight failures in LLM judgments, improving agreement with experts while holding model, evidence, total budget, and per-decision budget fixed. Sharding exhibits robustness against adversaries, reducing baseline error while keeping over-acceptance low even as the adversary's search widens.
We study Adversarial Causal Intervention Falsification (ACIF), a sequential game in which a structural causal generator proposes observational and interventional distributions, while an adversarial experimentalist selects interventions to maximize the falsification of the generator. The framework clarifies what an adversarial causal discriminator can certify, providing a principled bridge between causal generative modeling, active causal discovery, and experimental design.
Fixed and Adaptive Topological DeepONets: Functional Measurements on Hausdorff Locally Convex Spaces presents a framework for functional measurements on continuous linear functionals drawn from the continuous dual of a Hausdorff locally convex space. The authors develop fixed and adaptive functional measurement systems, combining measurements with the coefficient-space Two-Step procedure, and evaluate their performance on various problems including antiderivative operator, heterogeneous Darcy flow, controlled operator, and Navier-Stokes vorticity operators.
MiGHT-EHR, a Multi-task Graph Transformer, jointly models heterogeneous clinical entities and temporal interactions in Electronic Health Records (EHRs). It constructs a graph from EHRs with nodes representing entities and edges connecting statistically associated ones. MiGHT-EHR outperforms state-of-the-art methods on average across four tasks: drug recommendation, length-of-stay prediction, mortality, and readmission, particularly showing strong improvements in mortality and readmission predictions.
SNI-GNN, a SmartNIC-assisted full-graph GNN training system, reduces communication while preserving accuracy by predicting remote embeddings in-network. It deploys a lightweight linear-trend predictor on SmartNICs to refine cached historical embeddings and achieves 1.3-3.6x end-to-end speedups over BNS-GCN and up to 1.29x over baseline SANCUS, with accuracy loss ≤0.01.
A machine learning framework called ED-CSP predicts crystal structures from chemical composition, atom count, and multiple detector-plane electron diffraction spot sets. The model achieves a structural match rate of 57.49% MR@5 on 2,075 held-out materials, outperforming PXRDGen, a state-of-the-art crystal structure prediction model conditioned on powder X-ray diffraction.
Signed Integrated Gradients Attribution is used in a BiomeGPT-Style Microbiome Transformer to interpret downstream classification, overcoming limitations of nonnegative attention weights. This approach preserves species identity and isolates the effect of abundance variation, yielding polarity that separates pathogenic from protective microbial signals. Second-order Integrated Hessians are recommended to expose microbiome community interaction rules.
Guanghui Min et al. investigate the behavioral effects of recurrent context compression on long-horizon agents, finding that compression can weaken recent interactions' influence, increasing blocked actions and instability across runs. They introduce TRACE, a verifier-guided framework evaluating individual compaction events through paired closed-loop continuations. Initial results show improvements over existing baselines in task performance, multi-run reliability, and context-execution efficiency on AppWorld.
This paper introduces Mixtures of Geodesic Factor Analyzers (MGFA) on Riemannian homogeneous spaces, a method that uses geodesic factor models within each mixture component to enable clustering of manifold-valued data with anisotropic subpopulations. MGFA provides greater expressiveness than mixtures of Riemannian radial distributions and establishes root-$n$ consistency for the maximum likelihood estimator. The paper proposes an iterative estimation algorithm and demonstrates its effectiveness on corpus callosum and left hippocampus shape datasets, outperforming competing methods in well-specified regimes while remaining robust under model misspecification.
Optimized Certainty Equivalent (OCE) risk minimization is considered using samples, with applications in portfolio optimization and uncertainty quantification. A characterization links OCE to utility-based shortfall risk (UBSR), enabling the formation of an OCE estimator from a sample-average approximation of UBSR. Mean-squared error bounds are derived for the proposed OCE estimator, and a stochastic gradient algorithm is used to optimize OCE with convergence rates determined by non-asymptotic bounds. Three experiments use this algorithm for portfolio optimization and uncertainty quantification problems.
Researchers developed a measure-theoretic framework to characterize the geometric mechanisms behind contrastive learning, specifically InfoNCE. In the large-batch limit, they proved value and gradient consistency, revealing a geometric bifurcation between unimodal and symmetric multimodal regimes. The intrinsic energy is strictly convex in the unimodal case, while it becomes cross-coupled with a negative symmetric divergence term in the multimodal case.
Distributionally robust Markov decision processes provide a framework for sequential decision making under model uncertainty. The sample complexity necessary and sufficient to learn an ε-optimal robust policy under the average-reward criterion is determined by two parameters: SA (the numbers of states and actions), ε (the tolerance), and σ (the perturbation scale).
Researchers Chuxiangbo Wang, Shiying Li, and Caroline Moosmüller introduce a Wasserstein Mahalanobis distance that extends the principle of Mahalanobis distance to probability measures by using optimal transport displacement fields instead of Euclidean displacement vectors. This construction recovers latent geometry from nonlinear observations with properties similar to classical Mahalanobis distance, particularly for Gaussian measures and affine transformations.
Bayesian semi-parametric models use non-parametric Bayesian methods to estimate a targeted parameter, with a posterior distribution obtained using the Dirichlet process and Bayesian bootstrap. The posterior distribution is asymptotically Normal and concentrates at the true value of the parameter under specific assumptions, including stochastic equicontinuity. Relaxing these assumptions alters the conclusions, and the results are verified through simulation.
A penalised least squares estimator applies adaptive weighted group penalties to the second differences of a coefficient path, and asymptotic theory shows it recovers both the number and locations of kinks with probability approaching one. This method estimates an unknown number of common kink dates in a time-varying coefficient path under fixed effects. The endpoint slopes converge at the usual cubic regime-length rate, while interior slopes are determined by their own and adjacent regime lengths.
As the sample size $n$ and feature dimension $p$ grow proportionally, the out-of-sample prediction risk of the high-dimensional ridgeless least-squares estimator is governed by the alignment between the regression coefficient $\boldsymbol\beta$ and the spiked eigenspaces of the population covariance matrix. The signal energy distributed along latent spike directions determines whether interpolation leads to benign, tempered, or catastrophic overfitting under minimal moment conditions requiring only finite fourth moments rather than Gaussianity.
Omni-modal decomposition autoencoders (OmniDecVAEs) learn full-stack wearable disentangled representations, extending DecVAEs by learning modality-conditioned time-frequency latent subspaces through a multi-view self-supervised decomposition loss and a shared asymmetric autoencoder architecture. OmniDecVAEs outperform transformer-based and VAE-based methods in activity and identity recognition, achieving accuracy improvements of 1.01% and 6.75%, respectively. The model synthesizes realistic omni-modal time-frequency data with enhanced reconstructions and distributional similarity between real and synthetic data.
Denoising Diffusion Probabilistic Models (DDPMs) achieve rates independent of the ambient dimension in terms of score learning and sampling complexity under the manifold hypothesis, which states that high-dimensional data often lie on lower-dimensional manifolds. The authors develop a new framework connecting diffusion models to the theory of extrema of Gaussian Processes. This work establishes a connection between the empirical success of DDPMs and the theoretical framework of the manifold hypothesis.