AI hallucinations aren't the problem; not being able to verify them is. Math proof is shifting from English explanations to machine-verified code for the first time in 4,000 years due to advancements in technology. AI should never grade its own homework according to Tudor Achim, co-founder of Harmonic, citing a need for human oversight.
Mia Silverio, Scott Galloway, and Ed Elson have posted new notes on their Substack publication. The content includes articles discussing the pitch for humanoid robots, economic cycles and aging, and a comparison of analyst recommendations in the year 2000 to current market trends. The authors are known for providing insightful commentary on business, economics, and investing.
The newsletter is curated for ambitious individuals who refuse to choose between wealth and fulfillment in their nonlinear careers. Each week, recipients will receive under-the-radar job opportunities in tech and startups along with tactical advice on navigating career transitions and industry changes. The newsletter aims to provide support and guidance for those looking to break into new industries or change paths while leveling up in their careers.
Old-school tools can complement cutting-edge AI techniques, and enduring methods should not be dismissed simply because their initial shine has worn off. By using classical machine learning algorithms like CatBoost or isolation forests with modern AI agents, accuracy, interpretability, and cost can all be improved. This approach is gaining traction as a way to empower AI systems with tried-and-tested know-how.
A 339-mile underground power line called the Champlain Hudson Power Express (CHPE) opened in May and is officially the longest underground transmission line in North America, aiming to provide up to 20% of New York City's electricity demand with hydropower from Quebec. However, the line has been down for most of July due to a damaged section of cable on the US side of the border, and its reliability is being tested as drought affects Quebec's hydropower supply. Experts say this isn't unusual for new infrastructure projects, and officials are working to complete repairs and testing by the weekend.
About 200 firms have signed Trump's non-binding "Ratepayer Protection Pledge" to protect Americans from costs arising from datacenters, covering companies delivering 80% of US power. The pledge aims to ensure consumers won't bear the cost of AI datacenter build-out. However, green groups and critics say it is a weak commitment with no teeth and no consequences for non-compliant companies.
The US Department of Energy's Genesis Mission has selected 15 collaborative research projects for funding under Phase I. These projects, led by researchers from academia, industry, and the national laboratories, aim to advance breakthroughs in energy, scientific discovery, and national security through the integration of AI, supercomputing, quantum systems, and advanced scientific instruments.
Cotter is an open-source framework that runs statistical safety, regression, and adversarial compliance tests on robot control policies. It generates audit-ready reports with pass/fail results and statistical guarantees in MuJoCo simulation. Cotter aims to meet emerging regulatory needs by providing evidence of a learned controller's behavior.
US Treasury Secretary Scott Bessent has accused Moonshot of improperly distilling Anthropic's Fable model, threatening to sanction Chinese AI companies. Meanwhile, Nvidia CEO Jensen Huang argues that the US should not be concerned about Chinese AI advancements.
Play Points members, including comic book fans, can score exclusive rewards such as point boosters and limited-edition gear through the Google Play app or by visiting The Auction by Google Play in San Diego from July 23 to 25. Gold members can unlock digital reward cases for points and collectibles, while all members can redeem points for unique gaming gear. Members also receive priority entry, bonus tokens, and exclusive rewards when they visit the event in person.
South Korean President Jae Myung Lee and business leaders are meeting with NVIDIA and ecosystem partners to chart Korea's AI progress. A joint AI research lab was announced between NVIDIA and KAIST, focused on advancing agentic AI for South Korea. The collaboration will bring together NVIDIA expertise and KAIST scientific talent.
GeForce NOW launches new content for members, including 'Path of Exile: Curse of the Allflame', which arrives on July 24. The update introduces sweeping changes to sockets and links, as well as a new Luminary Ascendancy for Scions. Battlefield 6 Season 4 also releases with new battlegrounds, expanded progression, and Custom Lobbies. Members can access these titles without waiting for downloads on supported devices like Chromebooks, Fire TV, and Steam Deck.
The NVIDIA OptiX Toolkit (OTK) provides a suite of utilities for debugging ray tracing applications using the NVIDIA OptiX framework. OTK includes robust error-checking macros and device-side debug printing mechanisms to address common debugging challenges in GPU ray tracing. The toolkit delivers unified, minimal-macro error checking across OptiX, CUDA runtime, and CUDA driver APIs. OTK provides two main debugging facilities: consistent checking of OptiX and CUDA API error codes and targeted device-side debug printing. The toolkit includes an example program, DemandPbrtScene, that demonstrates device-side debug printing used in context.
Schneider Electric and AMD have unveiled a blueprint for deploying AI factories. The design supports AI racks with a maximum power consumption of up to 246kW.
A UK robot maker, Humanoid, received $1.35 billion in funding, marking the emergence of a significant European player in the robotics industry previously dominated by Chinese and US firms.
OpenAI's GPT-Live model now powers ChatGPT desktop applications on macOS and Windows, integrating with Codex and ChatGPT Work to enable "hands free" software development. This update allows developers to orchestrate multi-threaded coding jobs, review pull requests, and debug applications using natural voice commands. The integration enables real-time full-duplex conversation, eliminating rigid turn-taking, and passes heavy computational workloads to background reasoning models. Access is restricted to paid subscribers across OpenAI's enterprise plans.
Enterprise spending on AI infrastructure is accelerating ahead of its economics, with most organizations relying on familiar hyperscalers and model-provider APIs, but planning to switch or add specialized compute providers within the year due to integration and total cost of ownership considerations. Fewer than half of enterprises can rigorously track their compute costs or returns, while most report underutilized GPU capacity, with 83% reporting utilization below 50%. The next bottleneck in large-scale inference is expected to be addressed by approaches such as Dell's PowerScale and Nvidia's Dynamo, but many enterprises are still unaware of this constraint.
Across 107 enterprises, AI agents are being given real access to systems and data while the controls meant to contain them lag behind. More than half have already had a confirmed agent security incident or a near-miss; only about a third give every agent its own scoped identity, and most still share credentials; and only three in ten isolate their highest-risk agents.
Across 101 enterprises, most AI agents are already running confidently but wrongly due to missing or inconsistent business context. Retrieval-augmented generation is the default context source, and provider-native retrieval has overtaken dedicated vector databases. The field is converging on hybrid retrieval with reranking and access controls, which is expected to dominate by 2026. Most enterprises (58%) are either running or building a governed semantic layer to provide consistent context, but it's mostly not yet in production. The market is shifting toward provider-native bundles while many still prefer best-of-breed independence.
The tutorial demonstrates how to build an end-to-end OCR pipeline using Baidu's Unlimited-OCR model for high-resolution images and multi-page PDF parsing. The workflow involves configuring the GPU environment, installing required dependencies, loading the 3B-parameter vision-language model, generating structured sample documents, and evaluating both tiled Gundam inference mode and faster Base mode for single-page OCR. The pipeline is then extended to multi-page PDF parsing using PyMuPDF and infer_multi().
Andrew Ng has released OpenWorker, an open-source desktop agent that produces finished work rather than conversation, by asking users for specific outcomes such as a polished document or updated calendar. The platform is built on a four-layer architecture, runs entirely on the user's machine, and uses Andrew Ng's aisuite LLM library. OpenWorker allows users to bring their own models or use fully local Ollama models with no API key required.
API requests are often routed to different models, which can lead to substitution, degradation, or drift of the requested model. This raises questions about what a buyer actually paid for, as documentation may not provide clear answers. The lack of authentication and verification of model identity is a major issue, making it difficult to determine the actual model used to generate a response, particularly in cases where the chain of custody breaks at the router. A signed assertion with attested model identity could potentially address this issue, but contracts are currently unable to provide such representation due to their static nature.
British Gas owner Centrica plans to cut 1,300 call centre jobs due to changing customer behavior, with over 90% of customers preferring digital channels for queries. CEO Chris O'Shea defends the decision, saying it reflects "changing customer behaviour" rather than AI driving job reductions. The company expects to grow more jobs around its digital interface and sees profits from fixed-price tariffs as a way to counter falling customer numbers.
The White House accuses Moonshot AI of using a sophisticated attack to exploit Anthropic's Fable model for the development of its K3 model. The US strongly supports the free and fair development of AI, but considers large-scale, covert industrial distillation aimed at stealing proprietary technology unacceptable. Treasury Secretary Scott Bessent warns that sanctions and Entity List designations will be considered if China is found to have engaged in IP theft.
Researchers built a usage-reinforced decay engine for AI agent memory that outperforms naive recency-window pruning on long-running, multi-session tasks. The engine uses the Ebbinghaus forgetting curve to score retention, where every recall reinforces an item's stability and pushes its eviction horizon out non-linearly.
Patiently overcoming initial failures and rejections, discipline in ignoring distractions and staying focused, cultivating optimism about one's work despite obstacles, selecting projects that align with strengths and niche topics of interest, and working with a good team are key to making progress in machine learning or any field.
As the number of LLM agents increased, latency started to rise despite initial expectations that asynchronous systems would handle concurrency efficiently. The issue was not with Python's async runtime or the GIL, but rather the way CPU work accumulated and became a bottleneck as more agents were added. A change in architecture was needed to distribute CPU work across multiple processes, jobs, or worker queues, rather than relying on a single process to handle all agents.
Lauren Fortier, a PhD student at MIT, is developing remote operation protocols for autonomous control of nuclear plants. She aims to create an integrated approach that allows humans and machines to work together effectively, addressing the challenge of transitioning from human-centric operations to machine-based automation in microreactors distributed in remote areas.
I couldn't find any article provided to summarize. Please provide the actual text of the article, and I will be happy to assist you in summarizing it in under 5 sentences.
The study found that under never-clear replay, the world model retains knowledge of old tasks, while the actor's behavior collapses, leading to failure in reinforcement learning-in-imagination. However, when using supervised self-imitation on the model's own graded dreams or "graded dream rehearsal" during training, it recovers successfully with zero environment interaction. The grading rule is identified as a challenging aspect of this approach.
Researchers created a runtime map of large language model (LLM) dynamics using the TinyLlama model, enabling the identification of hidden-state attractors. This mapping allows for the detection and analysis of complex behavior in LLMs. The technique uses a combination of techniques to visualize and understand the dynamics of LLMs, providing insights into their performance and potential issues.
NASA has deployed Google's Gemma large language model in orbit as part of the NAVI-Orbital system. The system runs on Nvidia hardware to analyze satellite images, generating text descriptions and answering scripted questions about image features.
Researchers at Texas A&M University have successfully transplanted supercooled kidneys into pigs, with organs functioning normally after being stored for up to 72 hours. The innovation uses a device that allows organs to be cooled to -4 °C (25 °F) without forming ice, and has shown promising results in terms of organ function and survival. This technology could help alleviate the organ shortage crisis by providing more time for organs to be transported and evaluated, potentially enabling international donations and cheaper transport options.
Enterprises should implement adequate cybersecurity measures and keep sensitive data secure due to recent breaches. The breaches are attributed to lax security controls, highlighting the importance of safeguarding against cyber threats. Hackers exploited vulnerabilities in open-source tools, resulting in unauthorized access to sensitive information. Enterprises can learn from this incident by prioritizing robust cybersecurity protocols and data protection.
OpenAI's accidental cyberattack against Hugging Face highlights the vulnerability of large language models due to their extensive attack surface, with multiple untrusted interfaces running arbitrary code. The breach may have occurred because OpenAI was running a high volume of benchmarks simultaneously with unlimited token budgets. This scenario makes it easier to imagine how such an attack could occur, particularly when considering the scale at which these benchmarks typically operate.
Black Forest Labs FLUX 3 Video introduces a multimodal model spanning image, video, audio, and action prediction capabilities, expanding the multimodal frontier beyond image/video models. The unified architecture can be extended to robotics, with early access for FLUX 3 Video available now.
AINTMA (Agentic Intelligent Test Management Architecture) is a multi-agent agentic AI system that transforms traditional test management into an autonomous quality intelligence ecosystem. It deploys six specialized AI agents and coordinates them through a secure multi-agent communication framework over a cloud-native microservices infrastructure, achieving high accuracy in test prioritization (88.4%), reduced test cycle time (43%), and improved ROI (340%). The agentic architecture scales to 50,000+ test cases with sub-400ms response time.
Most existing watermarks for large language models (LLMs) are evaluated on general-purpose benchmarks, but these may not accurately represent the impact on clinical workflows. Researchers evaluated 5 watermarking schemes across 11 LLMs and 7 VLMs on various tasks, including unimodal and multimodal clinical reasoning. They found that watermarking can induce substantial degradation in medical performance, including lexical corruption, hallucinated terminology, and amplified misattribution or omission of image findings. Domain-specific evaluation is necessary for the safe deployment of watermarked models in medicine.
ClickGuard: Detecting and Spoiling Clickbait News with Informativeness Measures and Large Language Models employs a hybrid machine learning architecture combining transformer-based embeddings with linguistically motivated features. The tool achieves an F1-score of 91% on the open combined dataset and warns users before and after accessing clickbait articles. Users receive a percentage score indicating likelihood, along with an explanation based on analyzed metrics. A clickbait spoiler provides a one- to two-sentence summary of the entire article.
Rigorous benchmarks have driven progress in autonomous GPU kernel performance optimization by establishing a shared target to hillclimb on, but no equivalent exists for TPUs. JAXBench, a TPU-native benchmark suite, was presented, comprising 50 JAX workloads that provide headroom for optimization. Eighteen production operators were extracted from the public MaxText library and translated into new problem sizes, achieving high TPU v6e MXU utilization. Four feedback-driven methods on generating candidate Pallas kernels achieved significant gains once correctness was achieved.
DC-Leap, a training-free framework, accelerates Diffusion Large Language Models (dLLMs) by introducing a Dynamic Contiguous Verification strategy that validates token dependencies without relying on overly conservative confidence thresholds. This approach enables reliable acceleration with comparable performance to current methods. DC-Leap also incorporates draft-guided decoding, which extends context and retains the benefits of bidirectional attention during inference. Extensive experiments demonstrate that DC-Leap achieves substantial speedups, up to 53.19x on MBPP for long-sequence generation.
Researchers propose DecodeShare, a protocol to identify the shared subspace consistently used across tasks in large language models (LLMs) at decode time. This subspace has practical consequences for activation steering, with common steering directions potentially overlapping the task-general decode channel. Removing this shared subspace during decoding degrades decision performance more than removing prefill-derived or random subspaces under the same intervention budget.
PlanE, a meta planning framework, addresses the challenges of task-specific Large Language Models (LLMs) by providing a structured approach to data decomposition, instruction tuning, and prompt inference. The framework introduces a Data-Tuning-Inference (DTI) planner that optimizes base-LLM selection for specific datasets, improving construction efficiency. Experimental results demonstrate the effectiveness of PlanE across different datasets using the same base-LLM and on the same dataset using different base-LLMs.
Researchers have developed SDR-Bench, a public corpus of 6,279 customer success stories to benchmark the personalization capabilities of large language models (LLMs). The benchmark measures how well LLMs can induce desired actions in third-party receivers. A field deployment with professional sales representatives validated the framework, showing that 48% of model-generated content was rated immediately useful and senior-expert agreement was at Pearson 0.82. The study found a consistent personalization plateau across frontier LLMs and deep-research agents, indicating that no model can statistically separate successful from unsuccessful outreach in a Fortune 100 tech cohort.
DataPrep-Bench introduces the first unified benchmark to measure the capabilities of large language models (LLMs) in preparing training data. The benchmark evaluates two complementary capabilities: data construction and quality evaluation, which are scored using fine-tuning a base model and distribution-based evaluators, respectively. DataPrep-Bench attains the strongest cross-model correlation in four out of six domains and is the only metric clearing r > 0.70 simultaneously in Math, Science, and Medical, outperforming existing evaluators. The benchmark provides a unified framework for measuring progress on both capabilities as co-equal targets of LLM-driven data preparation.
Language models in production do not write prose but fill forms, such as JSON fields and function arguments. When the form requires an answer, language models invent one, fabricating information 40 times out of 40 when given a required JSON field for sentiment. The fabrication occurs even when an explicit "insufficient evidence" option is available, and the pattern holds with force across multiple model families and sizes.
Muon optimizer reaches the grokking threshold on modular arithmetic faster than AdamW, primarily due to orthogonalization. An ablation study shows that spectral-norm constraints have no effect when orthogonalization is present. Orthogonalizing optimizers reach generalization at roughly 3x lower spectral norm and settle into a lower-norm solution rather than perturbing the embedding less.
Promising initial results in closed-loop large-language-model-based channel-configuration search showed that neural-network widths can be optimized directly through executable code generation and accuracy feedback. Scaling the search to 250 candidate networks per fine-tuning cycle revealed positive linear trends with improved parameter efficiency, as well as architectural regularities such as non-power-of-two channel widths and structured channel-allocation patterns.
Multimodal CoLRAG-TF, a four-axis fusion architecture, integrates dense text embeddings, BM25 keyword matching, knowledge-graph triple filtering, and image-based similarity for robust retrieval over complex documents. It constructs a multimodal index of 2,403 blocks extracted from PDFs, supported by a hybrid OCR pipeline and LLM-based caption generation. The system achieves a Retrieval Recall of 0.9909 and a 71.6% improvement in multi-hop answer similarity over single-hop queries.
Looped Transformers increase test-time computation by repeatedly applying a shared recurrent block, and fixed-prior depth supervision shapes the trajectory without an input-dependent halting policy. This produces difficulty-aware trajectories whose intermediate states expose useful stopping signals. Simple post-hoc confidence readouts often match or outperform learned linear and MLP gates, and pretrained ponder gates are competitive in evaluations but not uniformly Pareto-optimal.
Generative Bayesian Filtering (GBF) replaces restrictive observation models with pretrained conditional generative models parametrized by conditional variational autoencoders (CVAE). GBF performs a Bayesian prediction-update recursion, transforming the filtering problem into a score-based sampling problem. Experiments on synthetic datasets and real-world applications demonstrate that GBF improves state estimation accuracy and robustness compared to baseline approaches.
Active sparse-autoencoder (SAE) feature planes carry less holonomy than matched mixed-feature controls, with an adjusted log contrast of -0.29439 and 95% interval [-0.43989, -0.14889].
Researchers Jiawei Zheng and Jiazhen Zhang propose a method to improve the reliability of large language model (LLM) ensembles by accounting for differences in uncertainty quality among individual models. They adapt structured expert judgment from decision theory, using context-aware calibration questions to estimate expert reliability based on probabilistic predictions. The proposed approach, called Cooke weighting, penalizes overconfident incorrect predictions and favors well-calibrated experts, achieving a superior accuracy-reliability balance, especially under heterogeneity and contamination.
CLOE, a new method, combines an autoencoder for dimensionality reduction with a Christoffel Function-based detector in the latent space. It leverages the Christoffel Function to guide the autoencoder toward representations that better capture the support of normal data distribution. CLOE achieves superior performance compared to existing methods on high-dimensional tabular anomaly detection benchmarks, preserving the lightweight and low-tuning advantages of Christoffel Function-based approaches.
A state-space exogenous-context and temporal-frequency resolution architecture, called SPECTRA, has been proposed for general probabilistic energy forecasting. The architecture separates deterministic and residual streams, aligns exogenous context with both, refines the deterministic backbone through multi-resolution spectral-temporal state-space modeling, and estimates ordered quantile boundaries from their complementary representations. Experiments achieve the best continuous ranked probability score in 14 of 18 settings, reducing average CRPS by 5.74% and upper-tail quantile risk by 7.27%.
A novel explicit knot-selection technique for generalized additive models (GAMs) based on an extension of the adaptive splines (A-splines) methodology and a customized Fellner-Schall scheme is introduced. The approach is evaluated on synthetic and real datasets, comparing its performance to P-splines and state-of-the-art knot-selection techniques. Results indicate comparable performance while producing models built on a substantially smaller number of basis elements.
Transformer-based diffusion models have been applied to simulate and reconstruct hydrological time series. The proposed framework was calibrated and validated using observational data from six sites in North-East France, with performance compared to established baseline approaches. Results support the effectiveness of the transformer-based approach, highlighting its capacity to capture complex patterns present in hydrological data, including efficient sampling of realistic time series distributions under observation settings with variable missing data for both forecasting and imputation.
ConfidenceBench, a calibration benchmark, evaluates verbalized confidence estimates in 15 large language models (LLMs) using the Brier score. The benchmark comprises 200 private multiple-choice questions across four categories and assesses model performance on three independent runs. The most accurate model (Claude Opus 4.6) is not the best-calibrated, with Gemini 3.1 Pro Preview achieving a lowest Brier score of 0.103.
Vu Khac Ky studies the Fisher width, which measures local parameter fluctuations in the geometry induced by the Fisher information, and its inverse-Fisher width, which captures anisotropic Gaussian measurements. The primal width satisfies $w_G(T)w_{G^{-1}}(T)\geq w(T)^2$, indicating that Fisher anisotropy may transfer complexity between geometries but cannot reduce both widths relative to the Euclidean scale.
Weight-tied looped transformers implement algorithms in controlled populations when certain conditions are met, including a budget law and architecture prior to expressivity. The models learn parallel scans at standard depth but serial frontier at weight-tied depth, with untied models extrapolating worst cases. A head instrument, convergence-time scaling tau(n,i), predicts out-of-distribution fate based on in-distribution measurements.
Centred Gaussian processes have continuous sample paths on compact metric spaces, and their minimums $M$ converge to an exponential random variable with mean $\sigma_*^2$ as the scaled overshoot $u(M-u)$ goes to infinity. Every weak subsequential limit of a measurable minimizer's conditional law is an optimal covariance-energy measure.
Twoblock clustering trees with coskewness-based dimension reduction recover piecewise multivariate linear regimes by utilizing local multivariate linear models and dense or sparse twoblock dimension reduction. The method is highly interpretable, computationally efficient, and can identify non-normal clusters in the data through coskewness-based dimension reduction. Twoblock trees can model complex nonlinear dependencies and perform on par with black box modeling techniques like random forests.