Apple has trained a China-only AI model in partnership with Alibaba Group, a move that marks a significant step in the tech giant's efforts to tap into the vast and growing Chinese market.
Hannah Zhang has kicked off a new chat called Nonlinear News. The chat is available in the Substack app or on the web.
Amazon has been found to be backing the largest US polluter, according to a report. The company's investments in fossil fuels and other polluting industries have been linked to its backing of the largest US polluter. Amazon's environmental impact is a growing concern among critics and investors. The report is part of The AI Report, a newsletter for business leaders.
OpenAI is awarding grants to 14 independent projects to promote economic opportunity and societal resilience as AI advances. The projects will examine how to broadly share the benefits of AI and address issues such as AI-driven productivity gains, labor-market changes, and AI safety. The grants will provide $1 million in funding and up to $1 million in API credits to researchers and practitioners, and will be implemented across the US, European Union, Brazil, Singapore, and South Korea. The goal is to develop new approaches and challenge existing policy responses to ensure that advanced AI benefits everyone.
TSMC reported a 45% revenue surge, driven by insatiable appetite for chips powering AI accelerators, data centers, and other applications. The company manufactures chips for virtually every major AI player, making its financials the industry's most reliable demand barometer. This growth rate suggests major customers are still placing massive orders, confident they'll be able to deploy and monetize the hardware, leaning towards the latter interpretation that this is just the beginning of the AI infrastructure buildout.
Google has partnered with five iconic football clubs - Arsenal FC, FC Barcelona, FC Bayern München, Liverpool FC, and Paris Saint-Germain - to bring fans closer to the game through AI personal intelligence and smartphone technology. Gemini provides deep match insights and stats, while Pixel cameras capture exclusive, behind-the-scenes content. The partnerships aim to boost global visibility for women's football teams.
In high-stakes classification, routing everything through a large language model (LLM) is not optimal due to invisible costs such as auditability, cost at scale, and model drift on easy cases. A cascade architecture solves this by treating the LLM as the escalation path, with a three-stage pipeline: deterministic, retrieval, and LLM call. The retrieval step is crucial and should be evaluated separately from final classification accuracy.
DeepSeek Harness is a developer preview release of a harness for agent frameworks, licensed under the MIT license. It allows users to assemble agent runtimes by selecting, swapping, or extending plugins for models, tools, skills, sessions, sandboxes, and more. The harness runs on the Cordis kernel and supports four runtime modes: Standard, Code, Minimal, and Creator, with an append-only session log capturing every context injection. The harness is deployable as a developer infrastructure, but not as a production agent product.
A report supporting Australia's teen social media ban appears to contain AI-generated errors, including citations to non-existent academic articles and incorrect links. The authors of the report initially denied the use of AI, but later conceded that AI had been used to rewrite paragraphs, but not to generate research or citations. The Australian government is examining the concerns and considering contractual arrangements and penalties to prevent similar behavior in the future.
The widespread use of AI-powered interviews in the job market is causing frustration and alienation among young applicants, who are often rejected without interacting with a single human being. Prime Minister Andy Burnham has criticized online recruitment processes, citing the lack of face-to-face meetings and the potential for algorithmic bias. The use of digital technologies in hiring is changing the entry-level jobs market, and applicants deserve better protection from the downsides of AI-generated recommendations.
Promise, a new AI film studio backed by Google, Silicon Valley venture capitalists and Disney, has set up shop near Sony Pictures' historic Culver City studio. The studio is using rapidly accelerating AI models to create backgrounds, special effects and even "synthetic performers" in its low-million-dollar-budgeted film, Touch Grass. Some respected film-makers, including director Joel Hynek, are embracing AI, while others, like Christopher Nolan, voice concern over the technology's rapid advancement and its impact on jobs. AI-native film-makers, like Kavan "the Kid" Cardoza, predict a creative renaissance circumventing studio giants, with some estimating a 20% to 50% cost saving for hybrid AI films.
Anthropic CEO Dario Amodei recently appeared on X to address criticism of his AI safety talks, framing his warnings as honesty rather than doom. He rejected the notion that his safety alarms have backfired and instead emphasized the need for real breakthroughs in areas like medicine to win back public trust. Amodei also pushed back on owning AI's negative image, suggesting that early glimmers from biology and medicine work will shift sentiment more than marketing can.
Qwen 3.8 27B is an excellent 27B parameter vision-capable LLM that runs on a 17GB file, capable of bounding boxes, annotating images, driving tools, and generating code. However, its default "xhigh" reasoning effort setting leads to spectacular over-thinking, resulting in lengthy processing times. Reducing the reasoning effort to "medium" or "low" can significantly improve performance. The model has shown promising results with coding agents, but its slow processing speed may limit its adoption as a daily driver.
A conventional RAG flow solves a real problem, but it's optimized for lookup at query time, treating each question as the first question ever asked. A better system accumulates understanding and stores it in a knowledge layer, using a hybrid architecture with three layers: evidence, knowledge, and orchestrator. The knowledge layer stores structured, human-readable models of the domain, and the orchestrator selects which layer to answer a question.
The article demonstrates the use of DuckDB's Quack protocol to run SQL queries concurrently across three remote DuckDB servers. A setup is created with three AWS EC2 servers, each holding a DuckDB database, and a coordinator node that runs the Quack server. The coordinator node validates each SQL query and splits it into smaller fragments, which are then executed on each remote server using a thread pool. The coordinator node collects the results from each server and presents them in a single output, allowing for concurrent reads and writes across multiple servers.
A paper claims that reinforcement learning (RL) for reasoning only changes 1-3% of tokens, and researchers replicated the gains without RL at approximately 1,000 times less compute.
I couldn't find any information on the article "Chunky Post-Training: Data Driven Failures of Generalization" by Murray et al 2026 as it was not specified.
Researchers in mechanistic interpretability using reinforcement learning could explore how to develop algorithms that can disentangle the contributions of individual agents or decision-making units within complex multi-agent systems, allowing for better understanding of decision-making processes and potential biases. Alternatively, they could investigate the use of reinforcement learning to learn representational models of complex decision-making processes, such as those found in human brains or neural networks.
Sharing a real task in a zipped repo with details is not considered a best practice, as it may reveal sensitive information. It's recommended to share an earlier version of the task or provide details through other means. This approach helps protect intellectual property and maintains confidentiality. Researchers and lab members often use version control systems, such as Git, to manage and track changes to datasets and tasks.
A local Ollama model was used as a commentary layer for a game AI, reading Q-values and narrating the reasoning behind each move. The model's explanations were found to be confident but not causal, and occasionally rationalized, highlighting the interpretability trap. The model's narration was post-hoc and not fixed by prompting, and its latency was the main constraint.
There is a query about implementing distributional DQN (Deep Q-Network) for traffic engineering in networks, specifically for development of a DQN-GCN (Graph Convolutional Network) code. The user seeks advice on choosing a suitable variant, such as C51 or QR-DQN, to understand problem and variation mapping.
The rise of agentic AI systems, which allow AI models to operate autonomously and call on sub-agents, is driving a surge in CPU demand. Agentic AI workloads require CPUs for tasks such as parsing output, making API calls, and running code, with some stages running entirely on the CPU. As agentic AI systems become more complex, CPU shortages and increased prices are expected, potentially leading to broader shortages and higher costs.
Researchers at the University of Toronto have found that a mine in northern Ontario produces around 140 metric tons of hydrogen per year, suggesting that the gas could be a useful source of zero-carbon fuel. The team, led by geochemist Barbara Sherwood Lollar, found that each of 35 boreholes released an average of eight kilograms of hydrogen per year. This amount, extrapolated to the entire mine, could provide at least a modest source of energy. However, producing hydrogen typically generates greenhouse-gas emissions and requires more energy than the gas contains.
Researchers Darragh Quinn and 5 others developed RubricForge, a method to induce reward-free judging rubrics that reduce over-crediting in agent evaluation. This approach uses reflective evolution against labeled trajectories to maximize agreement with environment rewards, resulting in faithfulness rather than raw agreement. RubricForge outperforms generic judges in over-crediting failed trajectories and ranks graded outcomes more faithfully, with a false-pass rate of 0.115 on tau-bench and 0.410 on WebShop.
Researchers analyzed the Qwen3.6-35B-A3B model's performance using magnitude-based expert masking on the XLCoST cross-lingual code translation benchmark. They found that layer sensitivity is strongly depth-dependent, with early and middle layers being fragile to expert masking, while late layers tolerate aggressive masking of low-magnitude experts. The study also identified a narrow very-late policy that achieves the strongest quality/masked-expert tradeoff, retaining 419/500 Good+Similar outputs while masking only 640 of 10,240 total experts.
Large Language Models (LLMs) develop a modular cognitive architecture that mirrors the human brain, with tasks drawing on the same network recruiting overlapping neurons and tasks drawing on different networks recruiting distinct neurons. This modular organization emerges across 46 tasks spanning four cognitive domains: language, formal reasoning, social reasoning, and physical reasoning. The study suggests that modularity may be a fundamental property of intelligent systems, and that it may have evolved in LLMs through a similar optimization process as the human brain. The findings imply that intelligent systems, whether biological or artificial, may exhibit a similar modular organization.
Researchers analyzed a one-year production trace from Chutes, capturing full production behavior across many models and users. The study reveals workload evolution and user-model structure, typically hidden behind aggregate views. The researchers will release the full one-year trace, enabling downstream studies of production behavior. The study focuses on Large Language Model (LLM) serving workloads, workload evolution, caching, and load-balancing.
Agentao is a governed local-first runtime for tool-using LLM agents, designed to mitigate risks associated with LLM agents' capabilities. It separates model-generated action proposals from host-authorized execution through a layered architecture, incorporating a host contract, runtime core, permission-mediated tool system, and supporting subsystems. Agentao does not provide formal safety guarantees but rather makes permissions, state, protocol boundaries, and execution traces explicit runtime abstractions. The code is publicly available at Github.
The dominant paradigm of AI evaluation focuses on superhuman autonomous performance, guiding AI development in the wrong direction by implicitly targeting human replacement. Instead, AI evaluation should focus on human-AI team performance, fostering AI systems that act as true complements to human capabilities and leading to better societal outcomes.
Researchers Akira Okutomi studied stable miscalibration in large language models, where a confident wrong answer remains stable under small perturbations. They used two diagnostics to identify stable miscalibration: a label-aware output-level audit score and an internal sensitivity probe. On a multi-domain binary factual audit set, self-critical prompting consistently reduced hidden-state sensitivity across layers in three open-weight models, supporting prompt-induced local stabilization. However, stable miscalibration errors were not clearly more locally sensitive than confidently correct answers, leaving room for stable but not calibrated errors.
Researchers Amritesh Banerjee and Pranil Raichura introduce the Behavioral Consistency Metric (BCM) to quantify behavioral consistency across tasks in language model agents, measuring the mean pairwise similarity of per-trajectory feature-attribution vectors within an agent system. The authors find that cross-task and within-task consistency are distinct axes that can diverge, and that consistency is not reducible to success rate, with systems having comparable success rates differing sharply in consistency. BCM is positioned as a process-level reliability signal that complements outcome metrics.
Researchers Babak Abbaschian analyzed misunderstanding as a layered process involving divergence generation, amplification, and detection. They identified eleven exact failure modes and eight analytical layers, with eight generating divergence, two amplifying existing divergence, and one detecting and repairing it. The paper models these layers formally, extending information and communication theory to reconstruct meaning, and provides a source-by-source evidence matrix and coding manual for independent application.
Researchers Yiwei Liu and Luwei Yang propose an active-perception framework for embodied target disambiguation that uses active observation to acquire information and decides whether to continue observing, request clarification, or complete target selection based on accumulated visual evidence and interaction information. This framework combines physical information acquisition and user-intent clarification within a unified embodied disambiguation process. The framework has been tested in real-robot experiments, showing its effectiveness in improving user clarification when necessary.
L-FNO, a stochastic neural operator, combines an FNO-style covariate path, Lorentzian spectral kernels for history-dependent excitation, and a likelihood-based training objective to improve event likelihood, calibration diagnostics, and rare-event detection. The model outperforms regression- and likelihood-based neural operator baselines. Structured spectral memory and likelihood-based learning provide effective inductive biases for neural operator models of stochastic event dynamics.
Optimization for a small set of coding benchmarks, such as SWE-bench and LiveCodeBench, leads to measuring task-specific performance, creating a meaning gap between measured scores and claims of general coding ability. Evaluating foundation models and checkpoints post-trained on these benchmarks frequently fails to generalize, with little cross-task transfer. A diverse evaluation approach, including holistic assessment, multi-task suites, and human-in-the-loop studies, is recommended to ensure reliable evaluation standards.
XGBoost, a popular machine learning method, can be affected by vertical outliers and leverage points in regression problems, leading to reduced performance. To address this, alternative loss functions based on M-, S-, and τ-estimators from robust regression are explored, resulting in a two-step procedure, MM-XGBoost, which provides the best trade-off between robustness and prediction accuracy.
Researchers have proposed Activation-Prune-Merge (APM), an activation-guided framework for cross-scale fusion, to improve performance of a small recipient language model with a stronger donor despite substantial architectural mismatch. APM constructs task-conditioned activation maps on the donor, selects salient layers, hidden dimensions, attention heads, and MLP neurons to prune it to the recipient architecture. This approach improves the overall average accuracy from 55.5% to 60.6% across 16 benchmarks.
Uncertainty sampling improves normalized balanced-accuracy area under the learning curve by 1.09 to 1.77 percentage points on three public binary tabular datasets under clean labels. However, difficulty-dependent noise reduces this advantage more than random classification noise (RCN) at six of eight rates on Breast Cancer Wisconsin, but at no tested rate on Banknote Authentication or MAGIC Gamma Telescope. Uncertainty sampling is label-efficient, but its apparent robustness depends on dataset, budget, noise structure, and evaluation metric.
The proposed KRPRVFL model integrates RVFL with the robustness of the kernel risk-sensitive mean p-power (KRP) criterion, reducing the influence of noisy samples during training. It leverages kernel-induced feature mapping to capture nonlinear relationships without explicit hidden-layer selection, maintaining efficiency and scalability. KRPRVFL consistently outperforms baseline models in terms of accuracy, robustness, and statistical significance on UCI and KEEL benchmark datasets.
ORCA, a two-stage framework, uses contrastive learning to learn an embedding space for anomaly detection in collider experiments. The framework first learns an embedding space via supervised contrastive learning, then uses an autoencoder in that space to generate event-level anomaly scores. The contrastive embedding makes anomalous samples interpretable by attributing them to specific physics processes with quantified uncertainties.
Linear attention models keep a state of fixed size, causing interference between stored items at long context lengths. Gated DeltaNet-2 derives its erase vector from the key of the current token, but the interference is measured through the query, limiting its reach. A new erase direction, Query-derived Erase Direction (QED), adds a second erase direction derived from the query, orthogonal to the key. This improves retrieval at every length past the training window, doubling the usable context length on S-NIAH-1.
Between October 2018 and July 2026, AI models progressed from simple systems like BERT to massive agents that solve complex math and write software. The ability to resolve real coding issues improved by nearly six times per year since late 2024, with costs dropping sharply and specialized models emerging, such as Claude Opus 5 for frontend coding and GPT 5 point 6 Luna for software development.
EEG-PRISM is a post-hoc attribution method that maps time-channel attribution scores into alternative domains using linear transformations and established backpropagation rules. It achieves near-perfect spectral recovery and 69.2% spatial accuracy in simulated data, and correctly determines delta-theta activity in epilepsy and localizes predictive delta-alpha biomarkers in autism. EEG-PRISM supports window-level analysis of transient events and group-level identification of clinically relevant biomarkers, advancing interpretable EEG foundation models.
Researchers developed a novel deep learning model called Modified Multi-Input Multi-Output PI-DeepONet (M3PI-DeepONet) to predict unsteady hemodynamics in Abdominal Aortic Aneurysms (AAA) with an average relative L2 velocity error below 4% and pressure error around 5%. The model combines layer-wise gating mechanism with a multi-branch operator-network topology and is trained based on physics-informed residuals, initial and boundary conditions, and a small subset of labeled internal data. This work advances the application of deep learning in cardiovascular disease modeling, marking a step toward real-time, non-invasive clinical diagnostics.
Maximum likelihood estimation (MLE) is brittle and can lead to poor generalization in Gaussian process (GP) hyperparameter optimization, even for probabilistic regression or classification tasks. The authors compare theoretically grounded metrics with MLE and propose practical solutions, demonstrating the effectiveness of their solutions in downstream design tasks such as Bayesian optimization. Their contributions are publicly available on GitHub.
A deep learning framework called MODERN is introduced for quality monitoring and fault isolation in smart manufacturing, utilizing an inception residual neural network architecture. It develops a control chart to monitor the likelihood of a product containing defects and a faulty region estimator using transfer learning. The framework extends to cases with insufficient training data, suggesting a transfer monitoring technique and hypothesis testing approach. Theoretically, it establishes a minimax optimal convergence rate for defect likelihood estimation and fault diagnosis, with empirical results demonstrating superior performance in simulated experiments and real data.
Occam's inversion is extended to include lasso fusion, overcomplete dictionaries, and isotropic total variation regularisation. This allows for the use of l1 regularisation, providing models that are visually sharp, and extending the algorithm to multiple dimensions. Three commonly used solvers for l1 regularised problems are unraveled, and it is recommended to use coordinate descent for 1D problems and iteratively reweighted least squares (IRLS) for 2D problems, with IRLS requiring up to an order of magnitude fewer least squares solves.
Offline Deep Q* Estimation with Diffusion Models proposes a novel framework for offline reinforcement learning by decoupling operator estimation from value function learning. Conditional diffusion models are used to estimate the reward law and transition kernel, inducing a data-driven approximation of the optimal Bellman operator. Theoretically, the framework achieves sharp nonasymptotic convergence rates for learning the optimal Bellman operator and the resulting deep estimator of Q*. Extensive numerical experiments demonstrate the effectiveness and strong empirical performance of the proposed method.
The authors study online statistical inference for functionals of the return distribution under a fixed policy using nonparametric distributional temporal-difference learning from a single Markov trajectory. They prove that the root-$T$ error of the Polyak-Ruppert averaged estimator converges weakly to a centered Gaussian random element in Cramér space.
L-FNO, a stochastic neural operator, combines an FNO-style covariate path, Lorentzian spectral kernels for history-dependent excitation, and a likelihood-based training objective. It is evaluated on eight synthetic point-process benchmarks and three real-world datasets, improving event likelihood, calibration diagnostics, and rare-event detection over baseline models. L-FNO shows that structured spectral memory and likelihood-based learning provide effective inductive biases for neural operator models of stochastic event dynamics.
XGBoost, a popular machine learning method, can be affected by vertical outliers and leverage points, which can impact its performance. To address this, the authors explore alternative loss functions based on M-, S-, and τ-estimators from robust regression. They propose a two-step procedure, MM-XGBoost, which provides the best trade-off between robustness and prediction accuracy, outperforming the standard XGBoost.
Data-driven techniques for translational neuroscience and personalized neuro-health aim to detect subtle, early, and individual-specific brain changes using neuroimaging data. These techniques, organized around four methodological pillars, aim to create personalized, mechanistically grounded, and clinically actionable models of individual brain health. The review emphasizes the need for quantitative tools to diagnose neurodegenerative diseases such as Alzheimer's and Parkinson's disease early and effectively. However, principal open challenges in statistical, computational, and clinical aspects remain.
A nonparametric two-stage method is developed to separate spatial and clinical risk in survival analysis, using kernel dipole-splitting survival trees. The method recovers spatial structure from clinically adjusted scale residuals, recovering the spatial signal under multiplicative frailty and exogeneity conditions. The method agrees with a Bayesian Gaussian random field frailty on the LeukSurv leukemia data, resolving sharp adjacencies and smooth surface averages.