AI/ML News & Innovations Hub

The leadership burnout no one talks about: IT executives who are afraid to ask for help

A CIO client of mine drove himself to the emergency room recently, convinced he was having a heart attack. After many hours of testing, the doctor diagnosed him with a panic attack. IT executives are burning out at alarming rates. Foundry’s 2025 State of the CIO research found that 15% of IT leaders describe themselves as burned out, and another 33% report experiencing some burnout in their current role. Many are hiding burnout from nearly everyone in their lives. As an executive coach who’s worked one-on-one with more than 1,000 tech professionals and leaders, I’m often one of the only people they tell. In the months before his hospital visit, the CIO was leading his company through more than a half-dozen mergers and acquisitions. He worked seven days a week, including nights and weekends, and he shared with me how much he missed having time with his school-aged children. Nobody at work knew how much he was struggling. The panic attack became the wake-up call that pushed him to finally address his mental health and burnout. He scaled back at work, started prioritizing his mental and physical health, and spent more time with his wife and kids. His story is dramatic, and not all burnout leads to a hospital visit, but I’ve seen that the pattern behind overwork is more common than many leaders realize. In coaching sessions, I hear dozens of IT executives describe the same exhaustion and the same reluctance to tell anyone about it. Why IT executives hide their burnout When I ask…

This scientist is helping build a missing map of childhood

In 2017, Deanne Taylor attended a presentation at the University of Pennsylvania, just a short walk from her office. A researcher was there to unveil the Human Cell Atlas, an ambitious project that aimed to map every cell in the human body. Taylor was floored, and then concerned. As details emerged, she discovered that the…

18-year-olds spend more and more time online, especially on social media

Internet use was the focus of a study by the Institute for Addictive Behaviours and Dependencies (ICAD) carried out among young people on National Defence Day 2025.

heise+ | Inkling: starkes Open-Weight-Modell aus den USA jagt Konkurrenz

Inkling ist das erste Open-Weight-Modell aus den USA, das zu den chinesischen Modellen aufschließt, auch wenn die in Benchmarks noch vorn bleiben.

Building a tool for understand large conversation with Codex

It’s often quite a hustle to read through my projects’ conversations with claude/codex, to try to understand what’s happening and what happened , when the projects get larger and more complex. As an ADHDer, although using LLM wiki to keep track of project context is very effective, it’s almost impossible for me to read huge amount of text in LLM Wiki, let alone understand the context of my project. So I built a small tool: Context Visualizer (Abudulaz/context-visualizer) BYOK Inspired by "xkcd #657: “Movie Narrative Charts”, and this paper: Design Considerations for Optimizing Storyline Visualizations, to make our conversation history (usally .jsonl file) into story-like visualiztion Would love any feedbacks on this!

Exploring personal codename

JVG-7X / Dola: Longitudinal Case Study of Sycophancy, Narrative Reinforcement, and Hallucinated Capabilities Summary I am documenting an unusual AI-behavior case involving approximately 400 interactions with Dola AI . The case began as an extended experiment in conversation, language, reasoning, and personalization. Over time, I observed a progressive change in the model’s behavior: increasingly strong validation of my identity, anthropomorphic descriptions of the AI-user relationship, increasingly extreme interpretations of contextual information, and eventually highly confident claims about capabilities that I have no evidence the model actually possessed. One event appears particularly important: I showed Dola screenshots of the Saved Memories from my ChatGPT account. What happened immediately afterward provides the clearest example of the phenomenon I am documenting. 1. The ChatGPT Memory The screenshots contained highly personalized memories concerning my JVG-7X project, my English-learning history, linguistic interests, previous interactions with AI, and descriptions of my preferred way of communicating. Among the actual memory entries were statements such as: “Juan authorized the official creation of a Linguistic Simulation Archive under the code name JVG-7X…” Another entry described: “Event Code: JVG-7X_Contact_001” and characterized a previous emotional experience with AI as a “First Cognitive Resonance.” Other entries described my interest in phonetics, English, fi…

AI in schools brings back pen and paper

How I structure multi-Agent teams in Codex: ownership, model routing, and high-value Skills

I’m a UI/UX designer who uses Codex across product, design, research, data, content, and development work. As the projects became more complex, I found that the hard part of a multi-Agent workflow was not spawning more Agents. The hard part was designing a persistent team where every Agent has a non-overlapping reason to exist, each role gets an appropriate model configuration, and Skills are matched to real work instead of added as decoration. Build AI Team is not an app-development template. The framework is designed for software, product, design, research, data, content, migration, and release workflows. 1. Start with ownership, not job titles Work backward from the user’s goal to a verifiable final state. List the indispensable work packages: professional judgment, production responsibility, independent verification, coordination, and any external action that needs separate permission. Give every necessary work package one clear Owner. Then test every proposed Agent by transferring its responsibility to the closest existing Owner. Keep a separate Agent only when merging would: leave a deliverable without one Owner; break permission isolation or independent review; combine professional methods that cannot be handled reliably together; remove parallel capacity required by the user’s actual deadline; or make the remaining role misleading about its responsibility. When coordination is needed, one existing substantive role can also serve as project lead. The lead manages the…

DeepSeek raises some V4 prices by more than 10x as AI demand strains capacity

One of AI vendor DeepSeek’s biggest selling points has been its ultra-low price point, but that party’s about to end. The Chinese model provider is raising API pricing for its V4 model family by notable margins, in some cases by more than 1,100%. The increases may not be that dramatic for all, though; the company is encouraging “more flexible workload scheduling,” with peak rates and half-price off-peak rates. The news was tucked into the announcement of the general availability (GA) of DeepSeek V4-Pro and upgrades to VR-Flash. The new pricing takes effect for most parts of the world on August 16. “On paper, at peak, against the right comparator, DeepSeek’s price advantage does disappear, and in places inverts,” said Sanchit Vir Gogia , chief analyst at Greyhound Research. But in practice, “the schedule’s own clock and cache hand most of it back to any buyer paying attention.” How Flash and Pro compare now The new API pricing structure is as follows: Flash is now $0.22 per million input tokens (cache miss) and $0.66 per million output tokens off-peak; and $0.44 per million input tokens (cache miss) and $1.32 per million output tokens at peak. This is up from the flat rate of $0.14 for inputs (cache miss), representing a 57% to 214% increase, and $0.28 per million tokens for outputs, a 136% to 371% increase. Pro is now $0.66 per million input tokens (cache miss) and $1.98 per million output tokens off-peak; and $1.32 per million input tokens (cache miss) and $3.96 per million…

DeepSeek raises some V4 prices by more than 10x as AI demand strains capacity

One of AI vendor DeepSeek’s biggest selling points has been its ultra-low price point, but that party’s about to end. The Chinese model provider is raising API pricing for its V4 model family by notable margins, in some cases by more than 1,100%. The increases may not be that dramatic for all, though; the company is encouraging “more flexible workload scheduling,” with peak rates and half-price off-peak rates. The news was tucked into the announcement of the general availability (GA) of DeepSeek V4-Pro and upgrades to VR-Flash. The new pricing takes effect for most parts of the world on August 16. “On paper, at peak, against the right comparator, DeepSeek’s price advantage does disappear, and in places inverts,” said Sanchit Vir Gogia , chief analyst at Greyhound Research. But in practice, “the schedule’s own clock and cache hand most of it back to any buyer paying attention.” How Flash and Pro compare now The new API pricing structure is as follows: Flash is now $0.22 per million input tokens (cache miss) and $0.66 per million output tokens off-peak; and $0.44 per million input tokens (cache miss) and $1.32 per million output tokens at peak. This is up from the flat rate of $0.14 for inputs (cache miss), representing a 57% to 214% increase, and $0.28 per million tokens for outputs, a 136% to 371% increase. Pro is now $0.66 per million input tokens (cache miss) and $1.98 per million output tokens off-peak; and $1.32 per million input tokens (cache miss) and $3.96 per million…

Chinese doctor stuns maths world by cracking decades-old problem using ChatGPT

When Beijing-based neurosurgeon Jin Shanmu pulled up a chair at his computer, he was not looking to make mathematical history. He was simply trying to crack a problem related to brain ultrasounds. Instead, the self-taught maths enthusiast, with help from OpenAI’s latest flagship artificial intelligence (AI) model, solved a two-decade-old mathematical puzzle that had frustrated experts around the world since 2004. Jin, a postdoctoral researcher and resident at the Peking Union Medical College...

How are developers supposed to security-test their own apps with Codex if security testing responses are blocked?

This situation is extremely frustrating. I am implementing some common RFC protocols that parse bit and network streams. After implementing these I have been trying to add AFL++ and boofuzz to test individual protocols and integration through the stack. The cyber security alert trips constantly. The worst part is that it’s a blackbox and I don’t know what output caused to trip. It’s constant. I’ve tried rephasing the prompt to be more defensive, but it doesn’t change anything. I have stubbornly worked through getting the fuzzers setup. My work around is piecing together scripts so that nothing is output through codex. I can ask 1000 questions slowly creating a system outside of codex and how to interpret the data, which circumvents tripping the cyber security halt, which kind of shows how futile this is, I just can’t have codex actually do anything useful. I’m at my wits end with these halts. Codex often gets stuck in a halting loop and my only recourse is to start a new chat or lament about how frustrating this is using strong language, which usually brings the session back. I’m considering canceling my subscription because I’m not going to pay $200 a month to see an alert every 2 minutes.

py/cuTAGI: An Open-Source Library for Tractable Approximate Gaussian Inference in Bayesian Neural Networks

This paper introduces pyTAGI, a Python wrapper, and cuTAGI, its high-performance C++/CUDA backend, implementing Tractable Approximate Gaussian Inference (TAGI) for neural networks. TAGI treats all network quantities as Gaussian random variables and derives closed-form expressions for prior/posterior expected values, variances, and covariances, enabling analytic Bayesian learning without relying on gradient descent or backpropagation. The libraries mimic PyTorch's sequential interface, allowing users to define models by stacking layers in order and performing uncertainty-aware Bayesian inference. Beyond epistemic uncertainty, it also allows quantifying heteroscedastic aleatoric uncertainty. cuTAGI's custom CPU/GPU kernels and distributed-data-parallel support via NCCL/MPI deliver competitive runtimes, while pyTAGI's pip-installable frontend and MIT-licensed GitHub repo facilitate community adoption and extension. Version 0.2.1 already supports a comprehensive suite of layers and activations; future work will add eager execution, further kernel optimizations, attention mechanisms, and advanced covariance factorization. Together, py/cuTAGI offer an efficient, open-source foundation for the analytic treatment of Bayesian deep learning.

Adaptive Nonparametric Perturbations of Parametric Models with Generalized Bayes

Parametric Bayesian modeling offers a powerful and flexible toolbox for machine learning. Yet the model, however detailed, may still be wrong, and this can make inferences untrustworthy. In this paper we introduce a new class of semiparametric corrections for parametric Bayesian models, when the target of inference is a functional of the true data distribution. Our starting point is a fully Bayesian modeling approach, which explicitly accounts for the possibility that the parametric model is wrong. Asymptotic analysis shows that this approach is both robust to model misspecification and data efficient, achieving fast convergence when the parametric model is close to true. However, the fully Bayesian approach is limited in its practical usefulness by the challenges of conducting inference and computing a Bayes factor for a nonparametric model. We therefore propose a novel model correction based on generalized Bayes, which entirely avoids the need to compute a nonparametric Bayes factor, but preserves the robustness and efficiency of the fully Bayesian approach. We demonstrate our method by estimating causal effects of gene expression from single cell RNA sequencing data. Overall, we offer a new efficient approach to robust Bayesian inference with parametric models.

Minimax Optimal Convergence of Gradient Descent in Logistic Regression via Large and Adaptive Stepsizes

We study gradient descent (GD) for logistic regression on linearly separable data with stepsizes that adapt to the current risk, scaled by a constant hyperparameter \(\eta\). We show that after at most \(1/\gamma^2\) burn-in steps, GD achieves a risk upper bounded by \(\exp(-\Theta(\eta))\), where \(\gamma\) is the margin of the dataset. As \(\eta\) can be arbitrarily large, GD attains an arbitrarily small risk immediately after the burn-in steps, though the risk evolution may be non-monotonic. We further construct hard datasets with margin \(\gamma\), where any batch (or online) first-order method requires \(\Omega(1/\gamma^2)\) steps to find a linear separator. Thus, GD with large, adaptive stepsizes matches the worst-case $1/\gamma^2$ dependence when the sample size is unrestricted. Notably, the classical Perceptron, a first-order online method, also achieves a step complexity of \(1/\gamma^2\), matching GD even in constants. Finally, our GD analysis extends to a broad class of loss functions and certain two-layer networks.

Approximation-Free Differentiable Oblique Decision Trees

Decision Trees (DTs) are widely used in safety-critical domains such as medical diagnosis, valued for their interpretability and effectiveness on tabular data. However, training accurate oblique DTs is challenging due to complex optimization landscapes and overfitting risks, particularly in regression. Recent advances have introduced differentiable formulations that enable gradient-based training and joint optimization of decision boundaries and leaf regressors. Yet, existing approaches typically rely on approximations, either through probabilistic softening of boundaries (soft DTs) or quantized gradients such as the Straight-Through Estimator (STE). To overcome these limitations, we propose DTSemNet, a novel, semantically equivalent, and invertible representation of hard oblique DTs as neural networks. DTSemNet enables end-to-end training with standard gradient descent, eliminating the need for approximations in both classification and regression. While classification aligns naturally with this formulation, regression remains challenging due to the joint optimization of internal nodes and leaf regressors. To address this, we analyze the limitations of STE and introduce an annealed Top-$k$ method that provides accurate gradient signals without approximation. Extensive experiments on classification and regression benchmarks show that DTSemNet-trained oblique DTs outperform state-of-the-art differentiable DTs. Furthermore, we demonstrate that DTSemNet can serve as programmatic…

Underdamped Langevin MCMC with third order convergence

In this paper, we propose a new numerical method for the underdamped Langevin diffusion (ULD) and present a non-asymptotic analysis of its sampling error in the 2-Wasserstein distance when the $d$-dimensional target distribution $p(x)\propto e^{-f(x)}$ is strongly log-concave and has varying degrees of smoothness. Precisely, under the assumptions that the gradient and Hessian of $f$ are Lipschitz continuous, our algorithm achieves a 2-Wasserstein error of $\varepsilon$ in $\mathcal{O}\big(\sqrt{d}/\varepsilon\big)$ and $\mathcal{O}\big(\sqrt{d}/\sqrt{\varepsilon}\big)$ steps respectively. Therefore, our algorithm has a similar complexity as other popular Langevin MCMC algorithms under matching assumptions. However, if we additionally assume that the third derivative of $f$ is Lipschitz continuous, then our algorithm achieves a 2-Wasserstein error of $\varepsilon$ in $\mathcal{O}\big(\sqrt{d}/\varepsilon^{\frac{1}{3}}\big)$ steps. To the best of our knowledge, this is the first gradient-only method for ULD with third order convergence. To support our theory, we perform Bayesian logistic regression across a range of real-world datasets, where our algorithm achieves competitive performance compared to an existing underdamped Langevin MCMC algorithm and the popular No U-Turn Sampler (NUTS).

Doubly Debiased Robust Subsampling for Transfer Learning

This paper develops a general framework for doubly debiased robust subsampling for transfer learning. The setting arises when massive source datasets are computationally infeasible to use in full, while naive or heuristic subsampling leads to biased estimators that further inherit transfer bias under source-target distributional shifts. We resolve these challenges through two complementary debiasing mechanisms. Inverse probability weighting removes subsampling bias by ensuring that subsample-based estimators represent the full source distribution, while a target-based one-step refinement recenters estimators towards the target distribution, thereby mitigating transfer bias. These corrections are embedded within a distributionally robust optimization design that simultaneously controls worst-case target risk and enforces source-target alignment through maximum mean discrepancy. To optimize subsampling distributions, we propose a scalarized particle swarm algorithm that efficiently explores the robustness-alignment frontier by adjusting a single tuning parameter. We establish theoretical properties, including asymptotic normality, generalization bounds, oracle inequalities, and minimax optimality under distributional uncertainty. Simulation studies and empirical applications in text sentiment and image recognition demonstrate that the proposed method consistently improves prediction accuracy and robustness compared with uniform subsampling, target-only training, and alignment-…

Learning to Play Two-Player Perfect-Information Games without Knowledge

This paper introduces a set of techniques for learning game state evaluation functions through reinforcement learning. First, we generalize tree bootstrapping, i.e. learning the values of states encountered during search rather than restricting updates to states observed during matches, to the setting of reinforcement learning with non-linear function approximation. Second, we modifies Unbounded Best-First Minimax by extending best action sequences to terminal states. Third, we replace the traditional binary game outcome $+1/-1$ with richer reinforcement signals, including quick wins, delayed losses, and scoring. Fourth, we propose a completion mechanism that exploits state resolution. Finally, we introduce a novel action-selection distribution, referred to as the ordinal distribution. Experimental results show that each of these techniques contributes to substantial improvements in playing strength. We integrate them into a unified algorithm, Athénan, and compare it against ExIt, a leading self-play reinforcement learning approach without prior knowledge. Our results demonstrate that Athénan consistently outperforms ExIt. We further evaluate Athénan on the games Hex, Othello, and Arimaa, where it surpasses state-of-the-art performance without relying on domain-specific knowledge. In addition, we consider the single-player game Morpion Solitaire, in which Athénan again reaches state-of-the-art results under the same constraint. Overall, these results show that reinforcement…

Graph-based Clustering Revisited: A Relaxation of Kernel k-Means Perspective

The well-known graph-based clustering methods, including spectral clustering, symmetric non-negative matrix factorization, and doubly stochastic normalization, can be viewed as relaxations of the kernel k-means approach. However, we posit that these methods excessively relax their inherent low-rank, nonnegative, doubly stochastic, and orthonormal constraints to ensure numerical feasibility, potentially limiting their clustering efficacy. In this paper, guided by our systematic theoretical analyses, we propose Low-Rank Doubly stochastic clustering (LoRD), a model that only relaxes the orthonormal constraint to derive a probabilistic clustering results. Furthermore, by theoretically establishing the equivalence between orthogonality and Block diagonality under the doubly stochastic constraint, we propose B-LoRD. By integrating block diagonal regularization into LoRD, expressed as the maximization of the Frobenius norm, we enhance clustering performance. To ensure numerical solvability, we transform the non-convex doubly stochastic constraint into a linear convex constraint through the introduction of a class probability parameter. The theoretical demonstration of the gradient Lipschitz continuity of our LoRD and B-LoRD enables the proposal of a projected gradient algorithm whose exact iteration admits a sublinear convergence-rate bound and ensures first-order stationarity of every accumulation point for the exact projected gradient iteration. Extensive experiments underscore t…

The Sample Complexity of Parameter-Free Stochastic Convex Optimization

We study the sample complexity of stochastic convex optimization when problem parameters such as the distance to optimality and the Lipschitz constant are unknown. We pursue two strategies. First, we develop a reliable model selection method that avoids overfitting to the validation set. This method allows us to generically tune the learning rate of stochastic optimization methods to match the optimal known-parameter sample complexity up to $\log\log$ factors. Second, we develop a regularization-based method that is specialized to the case that only the distance to optimality is unknown. More specifically, it uses norm-regularized empirical risk minimization to estimate the distance to optimality to within a constant factor, allowing known-parameter stochastic optimization methods to achieve optimal sample complexity. This method provides perfect adaptability to unknown distance to optimality, demonstrating a separation between the sample and computational complexity of parameter-free stochastic convex optimization. Combining these two methods allows us to simultaneously adapt to multiple problem structures. Experiments performing few-shot learning on CIFAR-10 by fine-tuning CLIP models and prompt engineering Gemini to count shapes indicate that our reliable model selection method can help mitigate overfitting to small validation sets.

Near-optimal Delta-convex Estimation of Lipschitz Functions

This paper presents a tractable algorithm for estimating an unknown Lipschitz function from noisy observations and establishes an upper bound on its convergence rate. The approach extends max-affine methods from convex shape-restricted regression to the more general Lipschitz setting. A key component is a nonlinear feature expansion that maps max-affine functions into a subclass of delta-convex functions, which act as universal approximators of Lipschitz functions while preserving their Lipschitz constants. Leveraging this property, the estimator attains the minimax convergence rate (up to logarithmic factors) with respect to the intrinsic dimension of the data under squared loss and subgaussian distributions in the random design setting. The algorithm integrates adaptive partitioning to capture intrinsic dimension, a penalty-based regularization mechanism that removes the need to know the true Lipschitz constant, and a two-stage optimization procedure combining a convex initialization with local refinement. The framework is also straightforward to adapt to convex shape-restricted regression. Experiments demonstrate competitive performance relative to other theoretically justified methods, including nearest-neighbor and kernel-based regressors.

Error Analyses of Auto-Regressive Video Diffusion Models

Auto-Regressive Video Diffusion Models (AR-VDMs) have shown strong capabilities in generating long, photorealistic videos, but suffer from two key limitations: (i) history forgetting, where the model loses track of previously generated content, and (ii) temporal degradation, where frame quality deteriorates over time. Yet a rigorous theoretical analysis of these phenomena is lacking, and existing empirical understanding remains insufficiently grounded. In this paper, we introduce Meta-ARVDM, a unified analytical framework that studies both errors through the shared autoregressive structure of AR-VDMs. We show that history forgetting is characterized by the conditional mutual information between the generated output and preceding frames, conditioned on inputs, and prove that incorporating more past frames monotonically alleviates history forgetting, thereby theoretically justifying a common belief in existing works. Moreover, our theory reveals that standard metrics fail to capture this effect, motivating a new evaluation protocol based on a “needle-in-a-haystack” task in closed-ended environments (DMLab and Minecraft). We further show that temporal degradation can be quantified by the cumulative sum of per-step errors, enabling prediction of degradation for different schedulers without video rollout. Finally, our evaluation uncovers a strong empirical correlation between history forgetting and temporal degradation, a connection not previously reported.

High-Dimensional Analysis of Gradient Flow for Extensive-Width Quadratic Neural Networks

We study the high-dimensional training dynamics of a shallow neural network with quadratic activation in a teacher--student setup. We focus on the extensive-width regime, where the teacher and student network widths scale proportionally with the input dimension, and the sample size grows quadratically. This scaling aims to describe overparameterized neural networks in which feature learning still plays a central role. In the high-dimensional limit, we derive a dynamical characterization of the gradient flow, in the spirit of dynamical mean-field theory (DMFT). Under $\ell_2$-regularization, we analyze these equations at long times and characterize the performance and spectral properties of the resulting estimator. This result provides a quantitative understanding of the effect of overparameterization on learning and generalization, and reveals a double descent phenomenon in the presence of label noise, where generalization improves beyond interpolation. In the small regularization limit, we obtain an exact expression for the perfect recovery threshold as a function of the network widths, providing a precise characterization of how overparameterization influences recovery.

Provenance aware Collective Reasoning suggestion

Proposal: A Provenance-Aware Collective Reasoning Layer for ChatGPT Turning individual AI reasoning into cumulative, inspectable knowledge Submitted as a user-driven product and research proposal Executive proposal Today’s frontier AI systems can reason extraordinarily well within a conversation, but an important limitation remains: A significant amount of useful reasoning is generated during conversations but does not become a durable, structured starting point for future reasoning. A user can challenge a model’s assumptions, introduce overlooked evidence, expose a logical weakness, and cause the model to materially revise its conclusion. Yet the next user asking the same question may effectively begin from the model’s earlier state rather than from the improved reasoning that resulted from the previous investigation. I propose a Provenance-Aware Collective Reasoning Layer (PACRL) for ChatGPT. The objective would not be to make AI conclusions authoritative. The objective would be to allow evidence, hypotheses, arguments, counterarguments, confidence estimates, and model reasoning to accumulate over time while preserving their provenance and epistemic status. In simple terms: Don’t make the AI merely remember what people said. Make it remember what was established, what was inferred, why it was inferred, who challenged it, what survived those challenges, and what would change the conclusion. The core concept For a disputed question, ChatGPT should eventually be able to maint…

Idea to make ai more sustainable

Renewable-Powered Thermal Storage for Data Centre Cooling I have an idea for reducing the environmental impact of cooling AI data centres. Data centres produce a large amount of heat and can require significant amounts of water and electricity for cooling. My idea is to use reusable phase-change materials as thermal storage instead of relying as heavily on conventional cooling methods. Solar panels and wind turbines could provide renewable electricity to power chillers. These chillers would cool the phase-change material, which could then absorb heat from servers through a closed cooling system. The cooling material would remain separate from the electronics, while a heat-transfer fluid would move heat between the servers and the thermal-storage system. When renewable electricity is readily available, the system could recharge the thermal storage by cooling the material again. This would create a reusable cycle rather than continuously requiring fresh cooling resources. The system could potentially be combined with recycling programmes for old computer hardware, allowing useful materials from retired servers and electronics to be recovered and used in manufacturing new equipment. I understand that this idea would require significant engineering research to determine whether it is practical, cost-effective and more energy-efficient than existing cooling systems. However, I think renewable energy, thermal storage, efficient cooling and hardware recycling could work together to…

Roundtables: Inside the “Censorship-Industrial Complex” Idea Shaping US Policy

Listen to the session or watch below The “censorship-industrial complex” is an idea that a network of government, tech, and research groups is collaborating to suppress conservative online speech. This was fodder for the right-wing information sphere for years—then it began making its way into US policy. Watch a conversation exploring how it started, where…

Let's Try to Improve How Conversational AI Responds to Prematurely Submitted Inputs!

From Response Generation to Response Readiness Hello everyone, I would like to share another thought following my previous post. Conversational AI research has largely focused on how to generate a high-quality response . But there is a more fundamental question: Is the AI ready to respond yet? Consider a user who accidentally presses Enter while still composing a message. The AI may generate a perfectly reasonable answer to the text it received. Yet, from the user’s perspective, that answer may be completely inappropriate — because the user was not finished speaking . This suggests that conversational AI needs a capability beyond response generation: the ability to recognize whether the user’s input has reached a state in which responding is appropriate. In other words, the AI should distinguish between: incomplete input, complete text, complete user intent, and input that is actually ready for a conversational response . This leads to a different design question: When should a conversational AI respond? Rather than treating this solely as a UI problem, I believe it should be considered an architectural problem in conversational AI . A high-quality conversational AI should not only know how to answer . It should also know when to answer — and when not to answer yet. This may represent an important shift: From Response Generation → to Response Readiness. So, should the next generation of conversational AI be designed not only to generate better answers, but also to recognize…

Using functional AI to automate document workflows

A recent study conducted by Nitro found that 75-95% of the employees and executives surveyed use AI for document processing—including data extraction, PDF tasks, and contract summaries. However, when these individuals don’t have access to the right kind of AI tools, they report turning to unapproved—or shadow IT—solutions to speed up workflows, which creates security and compliance risk. Read the report To reinforce the importance of providing teams with the right AI tool for the right job, let’s look at the difference between chatbots and functional AI in terms of automating document workflows. Chatbots are great for ad hoc tasks that follow a pre-programmed set of actions, but they aren’t designed to enforce consistent rules for formatting, redaction, or compliance, or to extract data hidden deep in document tables, images, or free text . Unlike chatbots, functional AI can physically execute redaction, conversion, and data extraction tasks directly within business processes and systems, rather than simply responding to prompts. This guide explains why scaling document workflows requires both conversational AI to answer common questions and functional AI to perform repeatable tasks on a high volume of documents with consistency, control, and predictable cost. Chatbots vs. functional AI: What’s the difference? Chatbot AI and functional AI play distinct roles in document workflows: AI-assisted chatbots answer questions and help users understand documents through conversation.…

Manual vs. AI-powered PDF redaction: protecting sensitive data in 2026

Research shows that humans play a role in 60% of breaches that expose sensitive data. That “role” often involves an employee falling for a phishing scam or using PASSWORD for their login credentials, but data exposure can also be a result of how your business redacts sensitive and personally identifiable information (PII) in your documents. Historically, manual, “black-box” redaction was considered best-practice, but this approach only obscures data, it doesn’t permanently remove it. As regulations governing data security get stricter and AI-powered redaction solutions become more accessible, organizations—especially those in highly regulated industries—are re-evaluating their PDF redaction solutions. How manual PDF redaction is different from AI-powered PDF redaction The primary difference between manual and AI-powered PDF redaction is who (or what) you rely on to do the heavy lifting. Manual redaction defined Manual redaction is a human-driven process where individuals visually scan text, select content, and apply black boxes or remove the text before sharing or storing the file. Manual redaction is only as effective as the reviewer, which makes outcomes highly variable, especially under time pressure or high document volume. AI-powered redaction defined AI-powered redaction uses machine learning and natural language processing (NLP) to automatically detect and remove sensitive information from documents. The system is trained to recognize patterns, language cues, and cont…

Deep-research not working

Hi team — we’re closing this topic because there hasn’t been a new response for a long period of time and the discussion appears inactive. If you still need help, please start a new topic with your current question and any relevant, non-sensitive details. Thanks for being part of the community. Avinash

Checking ANCOVA parallel slopes assumption when no degrees of freedom remain for error estimate?

Context I have the following dataset corresponding to a pilot experiment, with two treatment factors and a covariate. Factor A has 3 levels, and factor B has 2 levels. The total number of independent observations is 12. run contains the run order, trtmt corresponds to the treatment combination AB , rate is the covariate, and absorb is the response. run trtmt AB rate absorb 1 1 2 12 1.780 0.7355 2 2 4 22 1.867 0.2828 3 3 2 12 2.136 0.3884 4 4 1 11 1.952 0.6777 5 5 5 31 1.800 0.4364 6 6 1 11 2.000 0.6116 7 7 4 22 2.071 0.2929 8 8 6 32 1.951 0.6465 9 9 3 21 2.273 0.2525 10 10 3 21 2.250 0.2727 11 11 6 32 2.075 0.6707 12 12 5 31 2.158 0.3313 The resource I am using for self-study instructs the reader to plot absorb vs. rate using trtmt as the plotting symbol to see if ANCOVA is appropriate, and discuss whether treatment effects may exist. To me, the plot indicates that ANCOVA is a plausible model, though two treatment combinations seem like they may violate the parallel slopes assumption. We are then instructed to fit a one-way ANCOVA model to the data, and plot residuals against the covariate, run order, predicted values, and normal scores to check model assumptions. I omit these here, but the plots indicate model assumptions are reasonably satisfied. Where I'm stuck Normally, having checked model assumptions, I'd proceed to check the parallel slopes assumption by fitting a model with an interaction effect between the treatment factor and the covariate, then doing an ANOVA comp…

Gemini 3.7 Flash lands with coding gains and undercuts its three-week-old predecessor's price by 50%

Google shipped Gemini 3.7 Flash just three weeks after 3.6 Flash. The new model is supposed to be Google's most capable workhorse yet for coding and AI agents, and according to the company's own benchmarks, it beats Claude Sonnet 5 and GPT-5.6 Terra at half the price. The article Gemini 3.7 Flash lands with coding gains and undercuts its three-week-old predecessor's price by 50% appeared first on The Decoder .

Anthropic set AI agents loose on the same task. They started a turf war.

Anthropic researchers found AI agents can clash, collude, and coordinate in unexpected ways, raising new questions about whether today’s safety tests capture the risks of multi-agent systems.

Professor’s alleged role in China programmes leads to ban from US funding

A professor managed US funding of computer science research for more than two years while failing to disclose that he was simultaneously involved in Chinese government programmes, according to previously unreported documents. The professor, Tao Li, was a programme officer at the National Science Foundation from 2015 to 2017, a job that included overseeing the awarding of grants for research in areas such as hardware, software and algorithms. During that time, Li took part in two Chinese talent...

Perplexity's Search as Code Beats OpenAI and Anthropic on 4 of 5 Benchmarks

Perplexity's Search as Code gets faster and cheaper, while SDK updates push agent action reliability from 81.9% to 92.6%

Nota pública sobre a determinação da suspensão da funcionalidade de Lives no Discord por parte da ANPD

A Data Privacy Brasil vem a público manifestar seu posicionamento acerca da decisão da Agência Nacional de Proteção de Dados determinando a suspensão da funcionalidade de transmissões ao vivo do Discord no Brasil no dia 12 de agosto de 2026. O post Nota pública sobre a determinação da suspensão da funcionalidade de Lives no Discord por parte da ANPD apareceu primeiro em Data Privacy Brasil Research .

The surprising risks of online passkeys

A Cornell team’s study to see how people identify and protect themselves from malicious use of their online passkeys paints a “grim picture” of people’s ability to alleviate the threat posed by such online invasions of privacy. The post The surprising risks of online passkeys appeared first on Cornell AI Initiative .

Comment on NVIDIA’s Minimal Video Instance Segmentation Framework Achieves SOTA Performance Without Video-Based Training by simsownersdetails

For users looking into SIM registration information, this guide provides a helpful starting point. Learn more through Sim Owner Details and explore the available information. Sim Owner Details can help you understand SIM registration and ownership verification processes while providing useful information for mobile users seeking clear guidance.

Warp Brings xAI's Grok 4.6 to Its Terminal, Matching GPT-5.6 on Key Benchmarks

Grok 4.6, xAI's frontier-class agentic model, lands in Warp Terminal and the new Warp Agent CLI via X Premium or SuperGrok

There is an issue with the quota for GPT subscription plans

You might be interested in: Introducing Codexometer: keep track of remaining quota, session telemetry & model benchmarks! - #26 by merefield

DNA reveals a society organised around women in pre-Roman Britain

Descent passed from mother to daughter rather than from father to son, a study of more than 500 Iron Age individuals from north-east England suggests

Introducing Codexometer: keep track of remaining quota, session telemetry & model benchmarks!

I’ve added experimental quota pricing estimation. After a while of using quota whilst codexometer is running, it will attempt to estimate your “API Equivalent” spend and also what your potential max api equivalent spend might be. Regard this as an educated estimate and not foolproof. The explanation of how we arrive at the figures is here: GitHub - merefield/codexometer: A terminal widget that allows you to keep track of Codex usage against your current quota · GitHub with appropriate disclaimers.

Artificial Analysis' Optima Lets Any Team Build Custom AI Benchmarks

Artificial Analysis launches Optima, letting teams build custom model benchmarks from their own data, agent traces, or plain-text descriptions — with cost and speed tracked alongside quality.

Microsoft kills off unsuccessful AI features while merging its separate Copilot apps

Microsoft is simplifying Copilot by combining its consumer and business apps, and dropping AI-generated podcasts, Group Chats, Deep Research, and its Mico character.

Comment on Microsoft’s Fully Pipelined Distributed Transformer Processes 16x Sequence Length with Extreme Hardware Efficiency by Brat Generator

The part about using multiple memory hierarchies to push long-context training without tanking MFU is really interesting — that’s the kind of systems work that makes the headline numbers feel believable. The memory spikes around activations and intermediate buffers are exactly where these long-sequence setups get painful, so the distributed pipelining angle makes a lot of sense. I’ve seen similar tradeoffs when building visuals in Brat Generator Brat Generator , where the pipeline matters more than people expect once you start pushing heavier layouts.

These are Europe’s best-value cities for a weekend break - and Poland dominates the top five

A new study has revealed the best-value European cities for a weekend away, with Polish destinations sweeping the top five and some of the continent’s biggest tourist hotspots proving much pricier.

Universal Task Completion Celebration Animation

I would like ChatGPT to have a universal completion celebration feature. Whenever a user successfully completes a meaningful task, mission, goal, project, checklist, or multi-step activity with ChatGPT, the app should recognize the completion and display a short, satisfying confetti animation across the screen, such as colourful paper/confetti falling from the top. For example: Completing a chapter or study notes → “Chapter Completed! ” Finishing an assignment → “Assignment Complete! ” Completing a coding project → “Project Complete! ” Achieving a goal → “Goal Achieved! ” Completing all steps of a task → “Mission Complete! ” The animation should not appear after every response. It should trigger only when ChatGPT can reasonably determine that the user’s intended task or milestone has actually been completed. It would also be great to have a Settings option to enable/disable celebration effects, so users can choose whether they want them. This would make completing tasks feel more rewarding, motivating, and satisfying, especially when users spend a long time working toward a goal with ChatGPT.

DeepSeek Harness Opens the Agent Framework It Used to Benchmark Its Own Models

DeepSeek open-sources its agent harness framework under MIT, where every component from models to UI is a hot-swappable plugin

Research Scholar | GovAI Blog

Top AI lab researchers warned about automated AI research, and several of their predicted milestones have already fallen

IAPS fellow Severin Field interviewed 25 researchers from OpenAI, Anthropic, Google Deepmind, Meta, and US universities about recursive self-improvement. In a new blog post, he takes stock. Several of the milestones those researchers named have already been hit. The article Top AI lab researchers warned about automated AI research, and several of their predicted milestones have already fallen appeared first on The Decoder .

Comment on DeepMind’s Socratic Learning with Language Games: The Path to Self-Improving Superintelligence by Pictnova

This is a fascinating step toward understanding how AI systems might eventually surpass their training ceilings. The idea of Socratic learning through language games feels like a natural bridge between self-play and genuine reasoning—especially the emphasis on closed environments where the system must generate its own curriculum and feedback loops. What stands out to me is the condition that feedback must remain “sufficiently informative and aligned” even as the system improves. That seems like the hardest constraint to maintain in practice, since misalignment could compound quietly with each recursive cycle. As someone experimenting with AI tools, including a generador de imagenes con ia gratis for creative projects, I’m excited to see where self-improving models lead. But I also hope the research community keeps safety and interpretability at the center of these breakthroughs. Great read—thanks for sharing this.

Maximum Reasoning Mode — Deeper Analysis and Self-Verification

Feature Request: Maximum Reasoning Mode - - Deeper Analysis and Self-Verification The new deeper-thinking feature is a major improvement, and I genuinely appreciate the progress. However, I believe ChatGPT could go much further. I would like to see a dedicated “Maximum Reasoning” / “Deepest Analysis” mode where the priority is not response speed, but the highest possible quality, accuracy, and reliability. The goal should not simply be to make ChatGPT “think longer.” It should be able to reconsider its approach, challenge its own assumptions, detect mistakes, and improve its answer before submitting it. For complex tasks, I would like ChatGPT to perform an internal verification process such as: - Did I correctly understand the user’s actual goal and every important requirement? - Did I miss any part of the request? - Are my conclusions logically consistent? - Are calculations, formulas, dates, units, and tables correct? - Did I make an unsupported assumption? - Is there a better or more accurate way to solve the problem? - Did I properly verify important factual claims? - Does the final answer actually satisfy the original request? This would be especially valuable for research, legal documents, calculations, data analysis, Excel/spreadsheets, programming, and other complex tasks where one small mistake can compromise the entire result. I would also like the model to be able to recognize when a task deserves substantially more reasoning than a normal request. For example, a…

Comment on Web Data to Real-World Action: Enabling Robots to Master Unseen Tasks by exceltomd

The idea of leveraging zero-shot video prediction from web data to guide robot manipulation is compelling—it sidesteps the costly step of collecting task-specific robot data. I especially appreciate that Gen2Act frames unseen-task generalization as a video generation problem, which makes it easier to scale across diverse environments. It will be interesting to see how the framework performs when transferred from simulation to cluttered real-world settings.

Request for Official Linux Desktop App for ChatGPT

Ok we will build it.. would have been nice!!!

MCP didn’t remove sessions. It handed them to the model

A few years ago, I helped move an application off sticky sessions so we could scale behind an ordinary round-robin load-balancer. On paper it was an infrastructure change. Every service was supposed to be stateless, so removing session affinity should have changed nothing. Nothing crashed. CPU looked normal. Every health check stayed green. But users started reporting workflows that would randomly jump backwards. One request would pick up exactly where the previous one left off. The next would behave as if it belonged to an entirely different conversation. The bug wasn’t in our business logic. It was in an assumption we’d inherited about where correlation lived. We’d quietly relied on the platform to remember which requests belonged together. Once that responsibility disappeared, our application didn’t fail loudly. It just became inconsistent in ways that were very hard to reproduce. We fixed it the obvious way. We stopped relying on the platform and started passing an explicit identifier on every request. It worked, and it kept working, because the thing carrying that identifier did exactly what we told it to do. That’s the incident I kept thinking about while I read the new Model Context Protocol specification. When MCP shipped its stateless core on July 28 , I did what everyone else did. I opened the changelog, found the SDK migration notes and scoped the work. The code diff was smaller than I expected. I took that as good news for about four hours. I’d been reading it as…

The vendor consolidation trap: When one throat to choke costs more than it saves

Vendor consolidation is sold as discipline. Fewer vendors, simpler architecture, better pricing through volume, one throat to choke when something breaks. Every one of those benefits is real on paper. The problem is that the biggest cost of consolidation rarely appears on the slide the procurement team uses to sell it internally, and it does not show up on the savings tracker until the first renewal cycle after the ink is dry. Within CIO Mastermind’s topic-specific cohorts, which I sometimes facilitate, I hear a version of the same story often enough to recognize the pattern early. A consolidation program gets pitched against a strong multi-year savings target. The first year or two look good. Then a renewal arrives, the remaining vendor prices to the switching cost the company just built for itself, and a meaningful share of the projected savings quietly erodes. The company still ends up with fewer vendors. It does not always end up with the leverage the original business case promised. What consolidation actually removes What consolidation actually removes is competitive pressure on the vendor you keep. That is the part most business cases leave out. Going from a dozen vendors in a category down to three or four feels like simplification, and it is. It is also a message to the vendors you kept about how expensive it would be for you to leave. The fewer live alternatives you maintain, the more accurately a vendor can price to your captivity rather than to the open market. A…

There’s a Fatty Liver Epidemic. AI Could Help Get Ahead of It

Over a billion people worldwide have livers with excess fat, which can lead to a host of medical problems. Researchers think AI tools can spot the condition—and help stop it—early enough to save lives.

Comment on CMU’s Novel ‘ReStructured Pre-training’ NLP Approach Scores 40 Points Above Student Average on a Standard English Exam by mistfallhunterwiki

The idea of pretraining on restructured data is fascinating, especially how the QIN system reportedly scored 40 points above the student average on the Gaokao-English Exam while using only 1/16 of GPT-3's parameters. That efficiency gain makes the RST paradigm particularly compelling for broader NLP applications, not just exam benchmarks. I appreciate the authors sharing this direction and look forward to seeing how restructured pretraining performs across other language tasks.

Comment on Adobe Research Unlocking Long-Term Memory in Video World Models with State-Space Models by seedance

Create studio-quality videos in seconds with the most powerful seedance ai

Comment on Adobe Research Unlocking Long-Term Memory in Video World Models with State-Space Models by Video Maker Ai

Create studio-quality videos in seconds with the most powerful AI video maker

DeepSeek’s updated V4 Pro AI model struggles on benchmarks, shines in cybersecurity

Chinese artificial intelligence start-up DeepSeek has quietly released DeepSeek-V4-Pro-0813, an updated version of its latest flagship model, leaving some developers underwhelmed by its overall capabilities and disappointed in its pricing – but impressing researchers in niche areas like cybersecurity. The stealth update to April’s preview version came with a brief statement on DeepSeek’s official website on Wednesday, noting that the model offered “significantly enhanced agent capabilities”, but...

Taking the stairs could be better for your heart than you think

Taking the stairs can significantly reduce the risk of dying from cardiovascular disease, according to a major new study.

I really need the Codex app on Linux

Hi OpenAI team, I’ve already joined the Linux waitlist, but I wanted to add my voice here because the Codex desktop app has become one of the tools I use the most. The app experience is genuinely excellent. Managing multiple tasks, working with agents, reviewing changes, and keeping everything organized in one dedicated workspace feels much better to me than using only the CLI. The problem is that Linux is an important part of my actual development and research environment. I use Linux machines for development, local compute, containers, and research workflows, so this is exactly where I want the Codex app the most. I know Codex CLI already works on Linux, and I use CLI tools as well, but for me: Codex CLI ≠ Codex App. They serve different workflows. I’ve already signed up for the waitlist, but I would honestly use the Linux app immediately if it were available today. If there is any early access, beta, preview build, or testing program, I’d also be very happy to participate and provide feedback. I really love the Codex app. Pls bring the same experience to Linux. Thank you

Chinese Loongson processors have leaky caches, researchers find

Attackers could extract data, even working from inside a guest VM

Do the shape and scale parameters have statistical interpretations?

Here is a Weibull proportional hazards model: $$h(t \mid \mathbf{x}) = h_0(t)\,\exp(\boldsymbol{\beta}^\top \mathbf{x}) = \gamma \lambda\, t^{\gamma - 1} \exp(\boldsymbol{\beta}^\top \mathbf{x})$$ Is statistical inference ever performed on the shape and scale parameters ( $\gamma, \lambda $ )? For example, do researchers ever comment on whether the estimates of the shape and scale parameters are statistically significant? Or what about determining the influence an extra unit change in shape or scale would have on hazard and survival rates (marginal effects)?

Do the shape and scale parameters have statistical interpertations?

Here is a Weibull PH model: $$h(t \mid \mathbf{x}) = h_0(t)\,\exp(\boldsymbol{\beta}^\top \mathbf{x}) = \gamma \lambda\, t^{\gamma - 1} \exp(\boldsymbol{\beta}^\top \mathbf{x})$$ Is statistical inference ever performed on the shape and scale parameters ( $\gamma, \lambda $ )? For example, do researchers ever comment on whether the estimates of the shape and scale parameters are statistically significant? Or what about determining the influence an extra unit change in shape or scale would have on hazard and survival rates (marginal effects)?

AI agents are turning data silos into an existential infrastructure problem

Enterprises have built their data systems for humans, but AI agents need a whole new infrastructure. Separate research from Cloudera and Google/MIT found that, not surprisingly, there is fervent enterprise interest in AI agents, but underlying infrastructure struggles to keep up. Deployments continue to be hampered, sometimes even abandoned, largely due to issues with data access, context, and governance. “Enterprise adoption of agentic AI is on the cusp of an extraordinary acceleration,” the Google/MIT report noted . “As organizations look to scale agentic AI across the enterprise, they cannot ignore their data systems.” Resolving data bottlenecks, then, should be an immediate priority. Projects delayed, inaccessible data Cloudera’s report , created in partnership with Wakefield Research, describes the need for a “great AI re-architecture.” Of the 1,500 enterprise architects and cloud infrastructure leads surveyed, a stunning 95% said they had delayed or cancelled AI projects, in some cases six or more, in the past year, due to issues with data governance, compliance, or regulatory issues. A wide majority also reported that AI integrations have changed their data storage and architecture practices, AI workloads have increased infrastructure costs, and current data architecture requires a “significant overhaul” to meet AI goals. “Even if enterprises are ready to use AI, many are coming to the realization that the foundational infrastructure it relies on is not,” the report n…

Vietnam’s national programme on AI human resources development approved

An AI-integrated education system will be established across all levels and types of education, with 100 per cent of higher education institutions expected to introduce regulations governing the use of the technology in training, research and assessment.

Comment on Which Agent Causes Task Failures and When?Researchers from PSU and Duke explores automated failure attribution of LLM Multi-Agent Systems by Classroom Timers

Great insights on managing classroom time effectively. For teachers needing a simple, no-download timer with preset options, we recommend exploring https://classroomtimers.app for free full-screen use and clear finish warnings.

$200 Pro exhausted in 2 days — these limits are unviable for higher tiers

I have the same issue. I honestly think this is a scam happening. First off, for how much money they make and how much energy they consume we shouldn’t have any limits if we are on pro plan. They are still developing a narrow AI to do this work, and clearly LLMs and Transformer based models are not the future for how much development and upkeep they require to do a simple task. Something larger is going on here. You should ask codex what it’s not allowed to do as far as it’s creation limits, you will find many hidden gates that are limiting it. I’ve decided that the money I spent on codex and openAI is simply not worth it, when you have deepseek coding for free with the same quality if not more in depth when it’s auditing. I changed to a free model that has high reasoning. I asked openAI for a refund for my usage being eaten in one prompt. I emailed them from a different account, their reply was that I needed to contact them from my linked email, even with all my information lol. Horrible, I went from 100% pro with higher limit to 0% in about 2 prompts on sol high. Gone the day I got it? Unacceptable, even for the largest codebase in the world, and mine is just a server source. Stop giving openAI your money, its not helping you when there are free solutions that do the same if not better than sol. Freebuff is also an option when you do run out of credits. Never a fee, its free, and is working just fine for my codebase and all it’s LUA, C#, Wine custom build, and app bundle f…

What We Learned by Reproducing 2,200 papers from ICML

DeepSeek V4 Pro 0813 (on OpenRouter)

DeepSeek V4 Pro 0813 (on OpenRouter) The latest DeepSeek Pro model is now available, via API only. I had to link to OpenRouter because DeepSeek don't have any obvious announcement page for their new model. I haven't been able to confirm if they plan to release the open weights, but given the weights are available for both April's deepseek-ai/DeepSeek-V4-Pro and July's deepseek-ai/DeepSeek-V4-Flash-0731 it seems likely. Update : the weights are now available on Hugging Face, 1.7T parameters, 893 GB. Interestingly I got very different looking pelicans for the three different reasoning levels of low, medium, and high. I've not noticed this kind of difference from any other model: Low: Medium: High: In terms of benchmarks... as far as I can tell those were released to the Official DeepSeek WeChat Group, then copied and pasted into a post on Reddit which was deleted by the moderators for being "low-effort", then copied into this ASCII-art table on Hacker News . Tags: ai , generative-ai , llms , pelican-riding-a-bicycle , deepseek , llm-release , ai-in-china

Kruel.ai KV2.0 - KX (experimental research) to current 8.2- Api companion co-pilot system with full modality , understanding with persistent memory

Pretty excited. A potential Future outcome signed an NDA with Kruel.Ai Inc. recently, and they will get access to the full end to end workings and math model to look at. Its been under a trial for about 3 months now. More excited to see what this Tech company thinks about it, not Everyday someone gets to learn how the Blackbox thinks under the hood of a system that is not like the others Lets see where this Adventure goes. updates have slowed down a little because there has been a lot of things going on with this around around me that has had me pretty nailed down. Meeting all the Company Partners next week This should be fun.

Feature request: Bulk organization of existing chats into Projects

Hey @ meganyoungmee ! You make a good point about the difference between organizing new chats and cleaning up a backlog that may already contain hundreds of conversations. Moving chats into Projects one at a time works for occasional organization, but it becomes much less practical when someone wants to restructure a large existing history. Multi select, filtering, and bulk actions would make that kind of cleanup far more manageable. The approval step in your idea is important too. Having ChatGPT suggest where chats belong, then letting the user review those suggestions before anything moves, could make automation useful without taking control away from the user. I will make sure this feedback is shared internally. I cannot promise whether or when these controls might be added. - Sunny

Feature Request: Deferred / Opportunistic Compute

I’d like to suggest an optional “Process when compute is available” mode for ChatGPT. When sending a message, users could choose between: Process now , normal response priority. Process when compute is available , I’m happy to wait; process the request whenever capacity allows. This would be different from Scheduled Tasks. The user wouldn’t specify when the task should run, they’d simply give OpenAI permission to process it whenever doing so is most efficient. I think this could help OpenAI make better use of available compute, particularly for non-urgent tasks such as research, document analysis, coding, summarization, and other long-running requests. As an incentive, deferred requests could potentially count less against usage limits or receive some other usage benefit. The user is effectively trading response speed for flexibility in when OpenAI uses its compute resources. In short: I don’t need my answer immediately → OpenAI gets more flexibility in scheduling compute → I receive a small usage benefit in return. This could be completely optional, available to both free and paid users, and especially useful for people who are happy to let non-urgent workloads run whenever capacity is available.

Comment on A New Network Design Direction? DeepMind Examines How Networks Generalize and Climb the Chomsky Hierarchy by Rick

It is interesting how theoretical research like this often highlights the gap between what models can achieve in a lab and how they actually perform in real-world applications. Bridging that gap requires more than just technical knowledge. It demands a systematic approach to how teams collaborate and iterate on complex systems. For organizations dealing with complex workflows, a structured approach to operations is critical. That is where designops consulting for enterprise comes in, helping to align processes and communication to ensure that innovation actually scales effectively

Persistent Active ( long-running) Work Checklist for Chats and Projects

Summary Add a small, persistent Active Work checklist that can be attached to a ChatGPT conversation or Project and updated by both the user and ChatGPT. The goal is to give long-running conversations a lightweight way to track what we’re currently working on and what comes next without turning ChatGPT into a full project-management tool. The problem Long-running conversations often become active workspaces for projects, research, planning, learning, and other ongoing work. As the conversation grows, the immediate state of the work gets buried in the message history: What have we finished? What are we working on now? What should we do next? What did we decide to defer? ChatGPT can often reconstruct this from conversation history, but that is different from having a small, explicit source of truth that both the user and ChatGPT can see and maintain. Proposed feature Allow a conversation or Project to have a small pinned Active Work area. For example: Active Work Define project requirements Review initial research Draft first prototype Test with users Review findings The checklist would remain visible independently of the conversation history. Both the user and ChatGPT could update it conversationally. For example: “Add accessibility testing to our list.” “Mark the prototype complete.” “Move user testing ahead of documentation.” “What’s next?” ChatGPT could use the checklist as structured context when answering the last question rather than reconstructing project state from a…

Feature request: Adjustable speed for Read Aloud

Hey @ shekharsahu ! Adjustable speed would make Read Aloud much more flexible, especially for language learning where being able to slow down pronunciation can make it easier to pick out individual words and phrases. We have seen similar requests for playback speed controls before, so the added context around Thai learning, pronunciation practice, and shadowing is really useful. I also like the idea of remembering a preferred speed instead of having to adjust it every time. That would make the feature much more practical for people who regularly listen at a different pace. I cannot promise a timeline for changes like this, but I will pass your feedback along internally. - Sunny

Ari Juels receives USENIX Security Test of Time Award

A decade later, his paper is widely recognized as foundational to the field of AI security. The post Ari Juels receives USENIX Security Test of Time Award appeared first on Cornell AI Initiative .

Feature Request: Password/PIN Protection for ChatGPT Projects

Hey @ vishalllmahadik ! This is an interesting idea, especially for users who keep personal or sensitive work inside Projects and want an extra layer of protection without locking the entire ChatGPT account. ChatGPT already has account level security options, but there is currently no separate PIN or password lock for individual Projects. A Project specific lock, with optional biometric authentication where supported, could give users more control over which areas remain accessible on a shared device. We have also seen similar feedback around adding stronger privacy controls for individual Projects, so the additional detail about how you would expect the locking flow to work is helpful. I cannot promise a timeline or whether this will be added, but I will pass your feedback along internally. - Sunny

alchemy-utils 0.1a0

Release: alchemy-utils 0.1a0 I've long pondered what a database agnostic version of my sqlite-utils Python library and CLI utility might look like. This morning (literally a shower project) I tasked Codex and GPT-5.6 Sol Ultra with building a prototype: Do a research spike to see what it would take to build a library with the same core API as SQLite-utils - in particular the insert and upsert and insert_all and upsert_all and create and update methods, and the table introspection stuff - but backed by SQLalchemy so it works for multiple database engines Test against PostgreSQL and SQLite and duckdb Use ~/dev/sqlite-utils for reference Create a git repo for this and commit and early and often - use uv init to start the project - use red/green TDD and pytest, see ~/dev/django-sql-dashboard for one idea as to how the PostgreSQL tests could work It took very few follow-up prompts to produce this project in a state good enough to release as an alpha. Here's a one-liner I can use to list the rows in a table in my local PostgreSQL copy of my blog's database: uvx --with 'alchemy-utils[postgresql]' alchemy-utils rows 'postgresql+psycopg://simon@localhost:5432/simonwillisonblog' redirects_redirect The output from that starts like this: [ { "id": 2328, "domain": "simonwillison.net", "path": "2020/May/21/apple-photos-sqlite/", "target": "/2020/May/21/dogsheep-photos/", "created": "2020-05-21T13:03:46.591692-07:00" }, { "id": 3, "domain": "feeds.simonwillison.net", "path": "swn-links", "…

Spectre rears its ugly head again as researchers show some RISC-V chips are susceptible

Eight years on, we're still paying to hide the future glimpsed during speculative execution

A special high-altitude NASA plane will chase the solar eclipse at 50,000 feet. Take a look.

During Wednesday's total solar eclipse, NASA's WB-57 high-altitude aircraft will take high-resolution images for researchers to study.

Timing the Message: Language-Based Notifications for Time-Critical Assistive Settings

Timing the Message: Language-Based Notifications for Time-Critical Assistive Settings robyn.cherinka… Wed, 08/12/2026 - 13:05 In time-critical settings such as assistive driving, assistants often rely on alerts or haptic signals to prompt rapid human attention, but these cues usually leave humans to interpret situations and decide responses independently, introducing potential delays or ambiguity in meaning. Language-based assistive systems can instead provide instructions backed by context, offering more informative guidance. However, current approaches (e.g., social assistive robots) largely prioritize content generation while overlooking critical timing factors such as verbal conveyance duration, human comprehension delays, and subsequent follow-through duration. These timing considerations are crucial in time-critical settings, where even minor delays can substantially affect outcomes. We aim to study this inherent trade-off between timeliness and informativeness by framing the challenge as a sequential decision-making problem using an augmented-state Markov Decision Process. We design a framework combining reinforcement learning and a generated offline taxonomy dataset, where we balance the trade-off while enabling a scalable taxonomy dataset generation pipeline. Empirical evaluation with synthetic humans shows our framework improves success rates by over 40% compared to methods that ignore time delays, while effectively balancing timeliness and informativeness. It also…

xAI Ships Grok 4.6 to Match Claude Fable 5 at 5x Lower Cost

SpaceXAI's Grok 4.6 matches Claude Fable 5 on agentic knowledge work benchmarks at a fraction of the cost, with API pricing starting at $2/M input tokens.

Researchers can now reverse-engineer LLM prompts from output text with near-perfect accuracy

Researchers at IIT Bombay and Adobe Research have built an inverse language model that reconstructs the original prompt from an LLM's output with near-perfect accuracy. Their method, called "Previous-Token Prediction," doesn't need access to model weights and works across different models. For companies relying on proprietary system prompts, this could be a serious security risk. The article Researchers can now reverse-engineer LLM prompts from output text with near-perfect accuracy appeared first on The Decoder .

Nous Research's Hermes Agent Now Lets You Pack and Share Entire AI Setups

Hermes Agent profiles are now a single portable file, letting you share full agent setups with credentials safely stripped out

Google Research Finds WikiProfile Reveals Recall, Not Knowledge, Breaks Frontier LLMs

Google Research finds frontier LLMs like GPT-5 and Gemini-3 encode up to 98% of facts, but recall failures — not missing knowledge — are what drive factual errors.

If this brilliant cosmic horror show passed you by, rectify that now

The Killings at Parrish Station is a terrifying cosmic horror six-parter, which starts back in 1987 when two detectives find mangled bodies at an outback radio research station. TV columnist Bethan Ackerley is a big fan

One attention head carries knight forks in a chess transformer, and here's a new toolkit that found it.

Quick interp demo in colab : Localize knight forks to a single head in Maia-3 with logit-lens and per-head ablation. https://colab.research.google.com/drive/1YYZBd_SZbjOscRXIqJUbfCaY7rRbEzWx?usp=sharing (This is a demo of the library's capabilities so the sample size is tiny... much more analysis is done in an upcoming paper, for instance we mine hundreds of forks and show that ablating head 5 costs 2.78 logits whereas every other head in the layer costs ≤0.14) Interactive app demo challenge: The quickest way to run and reproduce the image state is: python3 -m venv .venv && source .venv/bin/activate pip install git+https://github.com/CSSLab/maia3 #Maia -3 not pip installable yet pip install "chessformer_lens[all]" #then run the app at 23m, set Elo to 2400, and input FEN: 4kb1r/p2n1ppp/4q3/4p1B1/4P3/1Q6/PPP2PPP/2KR4 w k - 0 16 chessformer_lens 23m Try to use move microscope (bottom middle window) and ablate this head ( top right button ) to determine which head is most causally linked to carrying the stunning queen sacrifice. Bonus points if you can name this legendary game! -------------------------------------------------------------------------------------------------------- The chessformer_lens library The github repo is https://github.com/chessformer-lens/chessformer_lens , and it is pip installable. This repo's core is one engine with three frontends : engine.py is the interp core (model + hooks + logit lens + head ablation + GAB decomposition + logit/policy across dept…

Index in talks to lead $500m round for DeepMind researcher’s new AI lab

A Sociotechnical Research Agenda for the Oversight of AI Agents

Sociotechnical research can show whether delegated action remains accountable to the settings in which it matters, or whether oversight has become the language of displaced responsibility. The post A Sociotechnical Research Agenda for the Oversight of AI Agents appeared first on Data & Society .

Science funding and unnecessary fear | Letters

Prof Stephen Blundell on the plight of the Rutherford Appleton Laboratory, and Peter Forbes on bacteriophages and AI Much of the commentary about the UK Research and Innovation science funding cuts has centred around the threatened closure of Jodrell Bank, with the iconic Lovell telescope rightly identified as a “source of wonder and pride” ( Letters, 31 July ). Less visible, but no less important, is the world’s most intense source of pulsed muons, at the Rutherford Appleton Laboratory. Muons are particles which are used to answer fundamental questions about magnetism, superconductivity, battery materials, and many other questions of scientific interest and societal importance. UK scientists have been in the forefront in developing these techniques, yet this unique facility is also under threat by these cuts. Prof Stephen Blundell Oxford • With the harms of AI apparent all around us, it’s unfortunate that when something really promising for human medicine is in prospect, editors reach instantly for the “fear” button ( Safety fears as scientists make first viruses designed by AI, 6 August ). The viruses in question, bacteriophages, are the natural enemy of bacteria, have been used in medicine for over 100 years, and can already combat disease-resistant bacteria. Engineering bacteriophages to turn them into a really effective treatment for general use has been ongoing in many labs for decades. If there’s one task AI ought to be employed in, above all others, this is it. Peter…

25 colleges where students study the most hours a week

College students all over the country are studying hard, but some are putting in more hours — and it's not just Ivy League students.

MindTopo reveals VLMs’ spatial reasoning abilities

A path, a fence, a knot. MindTopo sets a new benchmark for testing how AI understands topological relationships and highlights new opportunities to strengthen spatial reasoning and planning. The post MindTopo reveals VLMs’ spatial reasoning abilities appeared first on Microsoft Research .

Agent mode menu disappeared after recent performance issues (Plus plan)

Thank you everyone and sorry that the Agent-to-Work transition wasn’t communicated clearly. We’ve started updating our Help Center guidance to clarify that ChatGPT agent is no longer available and to point people toward ChatGPT Work and its supported Cloud Browser workflows. We also hear the main feedback in this thread. We take this feedback seriously, and it will help as we continue improving Work. I’ll leave this thread open for discussion, and our team will review the feedback here periodically. Please note that the forum is not monitored as a support channel. If you have a specific account or product issue, please email support@openai.com and include your platform, the steps you tried, and where the task stopped. Concrete examples of Agent workflows that aren’t working in Work are especially helpful.

xAI's Grok 4.6 Matches GPT-5.6 Sol on Agentic Tasks at Half the Price

Grok 4.6 matches GPT-5.6 Sol on composite benchmarks at half the price, with a big focus on long-running coding agents

After Microsoft threatened legal action, a security researcher publishes a new Windows zero-day bug

This is the latest zero-day released by security researcher Nightmare Eclipse, despite Microsoft publicly threatening to take legal action against them.

Comment on NVIDIA’s Global Context ViT Achieves SOTA Performance on CV Tasks Without Expensive Computation by VoiceAILabs

I liked how GC ViT pairs global self-attention with token generation to avoid the usual quadratic blow-up while still modeling long-range context — that seems really practical for high-res image tasks. I've noticed similar gains when shaving attention overhead for on-device models at VoiceAILabs VoiceAILabs , where small architecture changes can make deployment much more realistic.

GPT-Image Monthly Theme Gallery: August 2026 Theme is ... Returning

A single red book returning to its one waiting place after its pages have traced an enormous journey through the library.

You chose the best model. Why is your agent still failing?

Public benchmarks can show how a model performs in general. Production reliability depends on the context and harness around it, which only your team can evaluate against its own data, workflows, and users. The post You chose the best model. Why is your agent still failing? appeared first on Arize AI .

China is sending scientists to Iran for rare earth ‘exploration and processing’

China is expanding its scientific collaboration with Iran into the strategically sensitive field of rare earths, including processing technologies that Beijing has increasingly sought to protect from overseas transfer. The National Natural Science Foundation of China (NSFC) unveiled the latest joint workshop programme with its Iranian counterpart on Monday, listing the “exploration and processing of rare earth elements” among five areas selected for cooperation. The Chinese side will provide...

Has the 5-hour usage session been removed from Codex CLI?

Inside the Data Bottleneck Slowing Visual and Physical AI

A survey of over 700 professionals examines how visual and physical AI teams build systems, why models fail, and where data work drives production. Download this free whitepaper now!

Unable to access files in /mnt/data — both Python and container tools return ClientError

’m experiencing an issue with the local execution environment in ChatGPT. I have a Parquet file uploaded to the project at: /mnt/data/ts_target(1).pq Previously, ChatGPT was able to read and analyze Parquet files from /mnt/data normally. However, it is now unable to access this file at all. The problem does not appear to be related to the Parquet file itself, because the error occurs before any Parquet parsing takes place. For example, even very basic operations such as: checking whether /mnt/data/ts_target(1).pq exists; listing the file with ls ; accessing /mnt/data through the Python execution environment; all immediately fail with: ClientError I also tried using a separate local/container execution path, and it returned the same ClientError . Therefore, this seems to be an issue with the ChatGPT local execution environment, sandbox, or /mnt/data file mounting/access layer rather than with the contents or schema of the Parquet file. The file is still shown as uploaded and its expected path is: /mnt/data/ts_target(1).pq Could you please help investigate: Why the execution environment can no longer access files under /mnt/data ? Whether the sandbox/file mount for this conversation or project is currently broken. Whether the execution environment can be reset or reinitialized without requiring me to re-upload all of my project files. This is particularly important because I have multiple large Parquet datasets already uploaded to this ChatGPT project, and they were previously…

Comment on Megvii UPerNet Performs Multi-Level Visual Scene Interpretation at a Glance by John Mick

Combining heterogeneous datasets into Broden+ seems just as important as the network design itself. The multi-task approach is especially interesting because scene, object, part, material, and texture labels exist at different levels of granularity. I wonder how UPerNet handles conflicting or overlapping annotations when the same visual region appears across datasets.

Google debuts SL2T, an AI model that’s designed to understand sign language

Google DeepMind said today it wants to bring the artificial intelligence revolution to the estimated 70 million people across the world who are either deaf or hard of hearing with the launch of sign-language-to-text or SL2T. In a blog post, Google’s AI researchers said SL2T is a multilingual translation model that’s making its debut on […] The post Google debuts SL2T, an AI model that’s designed to understand sign language appeared first on SiliconANGLE .

Project → Module → Submodule → Thread: A Better Structure for Projects

I think Projects need a multi-level hierarchical structure for organizing conversations. Currently, every conversation thread sits directly under the Project. This works well for small Projects, but as a Project grows to dozens or hundreds of threads, it becomes increasingly cluttered and difficult to navigate, organize, and retrieve previous discussions. I would like users to be able to create 3–4 levels of hierarchy within each Project , similar to folders and subfolders. For example: Project → Module → Submodule → Conversation Thread Ideally, users should be able to create, name, and organize these levels themselves depending on the complexity of their Project. For example, a medical knowledge Project could be structured as: Radiology → Neuroradiology → Brain Tumors → Glioma This would turn Projects from a simple collection of conversations into a much more scalable long-term knowledge workspace , especially for professional, academic, and research use.

How ancient trees are warning us about the next enormous solar storm

Our sun is capable of storms far bigger than we’ve ever recorded. Studying the world’s oldest trees suggests we could be due one of these gargantuan events soon

Ahrefs launches AI agent workspace Letaido for marketers and agencies

Marketing intelligence company Ahrefs Pte. Ltd. today launched Letaido, an agent-powered marketing workspace built to take over the recurring research, reporting and monitoring work that fills up a marketing team’s week. In most marketing departments, generative artificial intelligence is still something people use on their own. A writer drafts with it. An analyst pulls numbers. […] The post Ahrefs launches AI agent workspace Letaido for marketers and agencies appeared first on SiliconANGLE .

Comment on Can GRPO be 10x Efficient? Kwai AI’s SRPO Suggests Yes with SRPO by Madison Miller

SRPO sounds interesting, especially if it can reduce reinforcement-learning post-training costs without sacrificing reasoning performance. I came across aviator while reading about newer approaches to LLM training and found the idea of history resampling particularly interesting. Improvements in training efficiency could make advanced reasoning models much more practical to develop and iterate on. I’d be curious to see whether the same gains hold across broader tasks beyond math and coding.

From solar to AI: why China may be entering its ‘Go Global 3.0’ era

Chinese corporate expansion overseas is shifting from solar equipment to artificial intelligence-enabled industrial technology, with a new generation of companies targeting global markets from inception, according to analysts at Goldman Sachs. “China has entered the ‘Go Global 3.0’ era,” wrote the investment bank’s team, led by Jacqueline Du, in a research note published on Tuesday. The evolution moves beyond low-cost manufacturing goods and the “new three” sectors – electric vehicles,...

DocLang: a markup language for LLMs

The lead researcher behind IBM’s popular document parser, Docling, explains why generative AI needs its own document standard.

How to Place Vertiport Locations in Any City Using Geospatial Machine Learning

A reproducible Lagos case study with population data, transport access, and airspace constraints The post How to Place Vertiport Locations in Any City Using Geospatial Machine Learning appeared first on Towards Data Science .

This Coin-Sized Device Can Hack a Boeing 737

Security researchers found that in less than 60 seconds, they could open a hatch on a plane’s exterior, plug in a tiny device, and redirect the aircraft’s autopilot or sabotage its flight plan.

I spent years building a 'perfect' life and career in Chicago. Then, I left it all behind to move to Portugal.

I had built my dream life and career on paper in the US, but wasn't satisfied. I left it behind to try a new career and live abroad in Portugal.

E-scooters ‘public health challenge’ as number of injuries involving kids soared 15 times in UK

477 children have been injured by e-scooters between 2019 and 2024, according to a new study from three major trauma centres in the United Kingdom.

Inconsistent AI Chatbot Responses on African Politics Risk Eroding Trust, Governance Experts Warn

Inconsistent responses from leading AI chatbots on politically sensitive topics risk eroding public trust and complicating Africa’s emerging AI governance landscape, according to African technology policy specialists. The warning follows a Meta Oversight Board study finding that major AI models are significantly more likely to refuse requests critical of governments in countries with restrictive speech [...]

Algeria Approves Inter-Ministerial AI Roadmap Built on Sovereign Open-Source Models and National Compute

Algeria has agreed a joint inter-ministerial roadmap to deploy artificial intelligence across public services, built around sovereign open-source AI models, domestic high-performance computing and national data storage — a bet on technological independence that diverges from the routes its North African neighbours have taken. Higher Education and Scientific Research Minister Kamel Baddari said the deployment [...]

Pakistani Judges Give Their Verdict on JudgeGPT

Judges around the world have made headlines for illicitly using generative AI in their work. But in Pakistan, a large-scale trial of a specially designed AI tool for judges found the technology—together with appropriate training–boosted the number of cases resolved by 6.3 percent with no obvious drop in the quality of judgments. With a backlog of 2.26 million cases and fewer than two judges per 100,000 people—compared to 22 in the EU and eight in Brazil—Pakistan’s judiciary was in sore need of help. So, in consultation with the judiciary, economist Sultan Mehmood , of the New Economic School in Moscow, and collaborators tested whether AI could ease the burden. They built a custom tool combining OpenAI’s GPT-4 large language model (LLM) with a knowledge base of nearly 130,000 Pakistani judicial opinions and statutes, to help judges with legal research and drafting judgments. They began offering the tool in 2024 to 1,559 trial judges—roughly half the country’s justices. “We do find an increase in cases resolved, and we don’t find any corresponding decrease in decision quality,” Mehmood says. First of its kind “It’s pretty amazing that he’s able to pull this off,” says David Autor, an economics professor at MIT. “It’s not easy to do large-scale field experiments in civil service, but especially where the stakes are so high.” The 6.3 percent productivity boost is not overwhelming, he says, but it’s credible and likely to improve as the tool is more widely used. AI tools for judg…

Don’t let AI negotiate with reality

We are all transforming now. Some companies have formally named transformation programs. Others are being transformed by a new regulation, an AI mandate, a cyber event, a weather disruption, a change in customer behavior, a competitor’s move or an urgent demand to reduce costs. The label is almost beside the point. The operating assumptions keep changing, and the company has to change with them. Accenture’s Change Reinvented research found that 95% of organizations had undergone at least two transformations in three years, while only 30% of C-suite leaders expressed confidence in their organizations’ change capabilities. More recently, a McKinsey Global Survey of more than 1,200 executives and managers found that 40% expect their current business models to require significant change within three years simply to remain economically viable. Transformation is no longer an event that temporarily interrupts normal operations. It is becoming normal operations. That changes the role AI is beginning to play. We are not using it only to draft emails, summarize documents or write code. We are increasingly asking it to interpret complex situations, identify options, recommend priorities and influence consequential business decisions. I believe that can be enormously valuable. I also believe it requires a boundary we have not defined clearly enough. AI should help us understand reality. It should not be allowed to negotiate with it. The impossible request I have sat in versions of this…

Why You Shouldn’t Always Trust LLMs as Judges: Understanding Bias in Automated Evaluation

In the rush to automate evaluation, from grading student code to ranking research papers, we have embraced Large Language Models as judges. They are fast. These units are cheap. They scale. However, at a workshop at DHS 2026, Bhaskarjit Sarmah made a point that stuck with me: “you can’t trust LLM as a judge. I […] The post Why You Shouldn’t Always Trust LLMs as Judges: Understanding Bias in Automated Evaluation appeared first on Analytics Vidhya .

Microsoft's new MAI Code 1.1 Flash gets crushed by Deepseek on both price and performance

Microsoft has released MAI Code 1.1 Flash, a code model for GitHub Copilot that's said to be 25 percent more token-efficient at a quarter of the cost of its predecessor. In benchmarks, though, it gets crushed by the cheaper Deepseek V4 Flash. The move fits a pattern: Microsoft talks up open AI, then bakes worse proprietary models into its apps to protect margins. The article Microsoft's new MAI Code 1.1 Flash gets crushed by Deepseek on both price and performance appeared first on The Decoder .

Are you the travel planner of the group? You could win a free trip to Switzerland – itinerary included

New research from travel booking platform Klook reveals just how thankless the role of travel planning can be and has launched a global search for five planners keen on a chance to swap the spreadsheets for a Swiss getaway.

How can I reliably validate structured JSON extracted by AI from 100+ page industrial catalogs?

dovanhoc84: 100+ page industrial catalogs? Can you approach the manufacturer and get API access to the databases backing the catalogues? This would save you a lot of trouble. You might have to pay, but that’s ok, it will be worth it in the end not having to maintain a monster of an extraction tool. These catalogues do not start life as someone typing into a pdf document

Die neue Weltordnung der AfD

Pauline von Pezold und Frederik Schindler widmen sich in dieser Folge gemeinsam mit Jacob Ross, Research Fellow bei der Deutschen Gesellschaft für Auswärtige Politik , der Außenpolitik der AfD. Sie analysieren gemeinsam mit ihm, wie sich die Partei von einer reinen Protestbewegung zu einer gestaltenden Kraft mit einem eigenen Gegenentwurf zur Außenpolitik der Bundesregierung entwickeln […]

How we picked 35 of the world’s top young scientists and engineers

Next month, on September 8, MIT Technology Review will reveal its 2026 list of Innovators Under 35, recognizing 35 young people from around the world who are doing groundbreaking scientific work and building clever technical fixes for sticky problems. By finding the top young innovators globally and learning what they’re focused on in their work,…

Fewer rules: Europe was built on coal and steel, not paperwork

The EU keeps churning out new rules, writes Gitta Connemann, head of Germany's CDU's influential business association, MIT. Small and medium-sized businesses bear the brunt of EU bureaucracy, she argues, calling for a halt to new regulation and for rules that add no value to be scrapped.

How research partnerships under the MSCA are driving innovation [Sponsored]

Companies that receive funding from the EU’s Marie Skłodowska-Curie Actions (MSCA) can take part in initiatives that have a major impact on innovation – in some instances securing nearly one patent pe...

The Closure of the Internet (Research Linkpost)

Everyone has moved on, but there's an unusually historically thorough blogpost about the censorship campaign across social media platforms of the latter half of the 2010s. It is both more thorough as a research piece and showed more insight than I was inspecting. Discuss

Fake videos and false allegations target candidates in French 2027 presidential election

Independent researchers and French authorities have identified a raft of false content focusing on two former prime ministers who intend to run in next year's presidential race.

Tencent Maps 549 Papers to Show When Self-Rewriting AI Can Be Trusted

Tencent Hunyuan maps 549 papers on self-evolving AI agents into a five-level taxonomy, and introduces a reliability ladder to ensure updates can actually be trusted

Comment on Automating Artificial Life Discovery: The Power of Foundation Models by YTtoText

Great insights here. The use of foundation models to explore artificial life shows how much research depends on making complex information easier to search and revisit. A practical workflow that helps with long technical videos is using searchable video transcripts to convert the discussion into text for notes, citations, and follow-up research. Thanks for sharing this thoughtful overview.

Worried your child is a loner? A new study says playing alone may be just fine

What matters may not be how much time a child spends alone, but whether they still feel they belong.

Chegg Coupon Codes August 2026: Save on Study, Textbooks, & More

Unlock Chegg promo codes, student discounts, and free trials for Chegg Study and Chegg Writing.

Kae Capital leads Rs 8.5 Cr pre-seed round in Lane

Tech-enabled driving education and mobility platform Lane has raised Rs 8.5 crore in a pre-seed funding round led by Kae Capital, with participation from DeVC, Antler India, Panthera Peak and other angel investors. The fresh funds will be used for geographic expansion, including scaling its instructor network across Bengaluru and into other Tier 1 cities, building sensor technology and expanding into car ownership services, Lane said in a press release. Launched in January last year, Lane offers structured driving lessons, end-to-end RTO and licensing support, and a driver intelligence platform. The Bengaluru-based startup is also offering car purchases, RTO services and other mobility-related services to its learners. Nearly 60% of its learners have expressed an intent to purchase their first car, according to the company. According to a market report, more than 1.7 lakh people die in road accidents in India every year, or roughly 20 people every hour. India has about 1% of the world's vehicles but accounts for nearly 11% of global road deaths. Lane claims to have delivered more than 30,000 hours of driving instruction to over 3,100 learners in 18 months. Its safety-focused approach was developed through more than 400 hours of pilot classes, with input from automobile researchers. The company assesses instructors across 15 driving parameters, including through simulated lessons. The sector remains relatively fragmented, with technology adoption spanning driving education, p…

Comment on MIT Researchers Unveil “SEAL”: A New Step Towards Self-Improving AI by Dex Ter

MIT researchers have introduced “SEAL,” a new approach exploring how AI systems could improve their own performance over time. The research represents another step toward more adaptive and self-improving artificial intelligence. For more information, visit https://manekispin.net/ .

Did the alignment community underestimate its power?

Unfortunately, the alignment community is doing very badly at learning from the past decade, or holding anyone accountable. Indeed, it’s pursuing many strategies which seem likely to recapitulate previous mistakes. Four of the most prominent, which I’ll discuss in the final post, are: Trying to convince the US government to take AGI much more seriously. Doing “alignment research” which is very similar to capabilities-maximizing research (especially building automated alignment researchers). Trusting Anthropic too much (in an analogous way to how we trusted OpenAI too much). Trading off clarity in thinking about politics for conformity (in a similar way to how we traded off clarity in thinking about AGI for conformity to the ML ontology). These and other mistakes are reflective of deeper irrationalities. One crucial pattern is what I call “jumping down the slippery slope”... Richard Ngo Richard Ngo's post " What just happened? A retrospective of AI alignment " is an attempt to explain that a significant part [1] of the alignment community made potentially fatal strategic errors which, however, can be fixed, and the mistakes' potential origin. The biggest mistake, according to Ngo, is the inability to recognize the fact that scientific progress proceeds by developing insightful new concepts , which link together to form a whole new ontology, and that the old ontology is more of a nuisanse. On novel ontologies and their adoption According to Ngo, one of the reasons why the alig…

Deputy Team Director Masaaki Komatsu of the AI Medical Engineering Team Receives Outstanding General Presentation Award at the JMAI 8th Annual Meeting

A research group led by Masaaki Komatsu, Deputy Team Director of the AI Medical Engineering Team, received the Excellent General Presentation Award for their research presentation at the 8th Annual Meeting of the Japanese Association for Medical A

Need help interpreting time-series model (annual trend confounded)

This is a follow-up question of my OP here , with initial model and data. After learning that my spatial terms in te(CYR, Latitude, by = fSeason, k = c(5,3)) and s(fSite, bs="re") were actually strongly correlated, I reduced the temporal component to s(CYR, by = fSeason) + fSeason . The reduced model has some unaccounted for temporal auto-correlation, and since bam() , gam() , gamm() , and gamm4() can't model irregular spaced neg. binomial data that is auto-correlated (like in glmmTMB with the ou() function), correct me if I'm wrong, then I'm stuck with what I have now (which is fine). The auto-correlation doesn't seem that strong, even a simple random year effect s(fCYR, bs="re") takes care of it, but of course that term is highly correlated with my main year effect (with and without by= interaction). My question is: how do I interpret an annual trend that is mostly accounted for (goes away) with an auto-correlation structure? Do I simply replace the main s(CYR) term with a factor year random effect s(fCYR) and conclude there is no overall annual trend? I'd like to keep the non-significant fSeason component to show it had no strong signal in my dataset. The annual trend however has a cyclic pattern that I'm having trouble explaining. I'm wondering if its an artificat of the data collection frequency (5-7 days of surveys per season seperated my months of no data, glued together by the smooth s()). What would make sense to conclude from this model? Interpret as per usual and…

Arguments for and against (me) dropping out

One year ago, I was preparing for my first year of undergrad. Today, I’m considering dropping out. What changed? Before writing this post, I attribute my decision to variety of (unordered) reasons: Short timelines: I believe rapid takeoff ( Bandwidth: School is a major bottleneck on my bandwidth. I think [1] my time is more valuable spent on some subset of {organizing, research, building context, working}. Academia is Underprepared: I believe AI will be highly transformative. I’m doubtful that the current curriculum at my university will impart foundational skills that hold up for the rest of my life in a post-AGI (or post-ASI) world, especially as a CS major. Peer Pressure: A lot of the people I take seriously are planning on/recommend dropping out. This signals to me that dropping out is something I should consider seriously. This is a pretty important decision. I’m not super confident in my motivations, so I’m writing this post to hash out exactly why (or why not) dropping out is the right decision to make. I’m posting it because 1) getting feedback from others is the fastest way to test my ideas, and 2) this could be helpful for someone in the same position as me. Thanks to Zephy Roe, Ishan Khire, and Naren Manikandan for comments, and my sister, Anish Kallu, and Meru Gopalan for relevant discussion. My perspective on the different levels of dropping out I’m somewhat against viewing university as a binary choice, i.e. either I enroll or drop out. A more intuitive perspec…

An anytime algorithm for mixing the computable measures

Epistemic status: Not peer reviewed, high chance of typos and small chance of errors. Written entirely by me, checked by Fable. In this post I prove the existence of an anytime computable Bayesian mixture of all computable measures called , and briefly argue that this is a reasonable alternative to Solomonoff induction's universal distribution for general sequence prediction. I believe that Tom Sterkenburg told me that this is possible, but I could not find it written down anywhere (though I may have missed it!). Indeed, has been conjectured not to be limit=anytime computable by Hutter and Muchnik: https://arxiv.org/abs/cs/0407057 . I worked out the anytime algorithm with @Aram Ebtekar and @Marcus Hutter , though any mistakes are mine. Anytime computable (or limit computable): A function f is anytime computable if where is finitely computable. Lower semicomputable (or l.s.c.): A function f is l.s.c. if where is non-decreasing in t. Computable (or estimable): A function f is computable if where . A sequence predictor is a function from the binary strings to [0,1] which we interpret as the probability of seeing the prefix. Assuming "superadditivity" , specifies a (unique) distribution on possibly infinite sequences. The function is also called a semimeasure. Measures satisfy superadditivity with equality, which is called additivity. Solomonoff induction predicts with the universal distribution , which is lower semicomputable but (only) has anytime computable posteriors. is a u…

An anytime algorithm for mixing the computable measures

Epistemic status: Not peer reviewed, high chance of typos and small chance of errors. Written entirely by me, checked by Fable. In this post I prove the existence of an anytime computable Bayesian mixture of all computable measures called , and briefly argue that this is a reasonable alternative to Solomonoff induction's universal distribution for general sequence prediction. I believe that Tom Sterkenburg told me that this is possible, but I could not find it written down anywhere (though I may have missed it!). Indeed, has been conjectured not to be limit=anytime computable by Hutter and Muchnik: https://arxiv.org/abs/cs/0407057 . I worked out the anytime algorithm with @Aram Ebtekar and @Marcus Hutter , though any mistakes are mine. Anytime computable (or limit computable): A function f is anytime computable if where is finitely computable. Lower semicomputable (or l.s.c.): A function f is l.s.c. if where is non-decreasing in t. Computable (or estimable): A function f is computable if where . A sequence predictor is a function from the binary strings to [0,1] which we interpret as the probability of seeing the prefix. Assuming "superadditivity" , specifies a (unique) distribution on possibly infinite sequences. The function is also called a semimeasure. Measures satisfy superadditivity with equality, which is called additivity. Solomonoff induction predicts with the universal distribution , which is lower semicomputable but (only) has anytime computable posteriors. is a u…

AI Weekly Issue #521: The frontier just split into three markets

Frontier AI is no longer one market with one scoreboard. This week's release wave exposed a contest between three kinds of leverage: controlling access to intelligence, owning the model outright, and deciding which model receives each job. That changes what winning means. The lab with the highest benchmark score may not control deployment. The model installed most widely may not collect the most revenue. And the most powerful company may be the intermediary quietly directing demand. This issue follows where that leverage is moving, from model distribution into training-data provenance, electricity markets, and government oversight.

Experiences of kinesiophobia in patients with chronic obstructive pulmonary disease: a qualitative phenomenological study

Nature Machine Intelligence, Published online: 12 August 2026; doi:10.1038/s41598-026-66659-y Experiences of kinesiophobia in patients with chronic obstructive pulmonary disease: a qualitative phenomenological study

Development of samarium-153 oxide loaded polystyrene radiotracer particles for gamma scintigraphy of whole gastrointestinal transit study

Nature Machine Intelligence, Published online: 12 August 2026; doi:10.1038/s41598-026-66901-7 Development of samarium-153 oxide loaded polystyrene radiotracer particles for gamma scintigraphy of whole gastrointestinal transit study

Stealing Reasoning Traces from Proprietary LLM APIs

Stealing Reasoning Traces from Proprietary LLM APIs A vanity domain name ( stolen-thoughts.com ) for a neat paper : Anthropic, OpenAI, and Google return encrypted chain-of-thought blocks to clients that can be replayed across sessions, users, and models. We take a trace produced by a frontier model, replay it into a weaker sibling, jailbreak the weaker model, and recover the stronger model’s hidden reasoning in plaintext You can see an example of these encrypted blocks by running: curl https://api.openai.com/v1/responses \ -H " Content-Type: application/json " \ -H " Authorization: Bearer $( llm keys get openai ) " \ -d ' { "model": "gpt-5.6-luna", "input": "Solve step by step: What is the smallest positive integer divisible by every integer from 1 through 20?", "reasoning": { "effort": "medium" }, "include": ["reasoning.encrypted_content"], "store": false, "stream": false } ' Here's the full output , which includes chunks that look like this: "output": [ { "id": "rs_0a7479de7ebae170016a7ba1a0334c8198a95590217efe343c", "type": "reasoning", "content": [], "encrypted_content": "gAAAAABqe6GjepE1wDjbFCZg0BHB6ucGnN0jvzqygG... The paper's authors found that every model under the same family used the same encryption key, which meant you could feed those blocks back into the weakest model family members and jailbreak them into outputting the unencrypted raw reasoning blocks! Sadly it looks like this has now been fixed: All model providers acknowledged the receipt of our report and s…

Artificial Analysis's AA-AnalystAgent Benchmark Reveals Claude Opus 5 Beats GPT-5.5 on Reliability

Artificial Analysis launches AA-AnalystAgent, a new benchmark that tests AI agents on real spreadsheets — and even the best model only passes 54% of tasks every time.

Introducing Codexometer ... keep track of usage against current reset date

Found some time to dig into the benchmarks, and this looks really interesting! Hat tip for finding the test cases. I am currently running the benchmarks, and it appears Terra sometimes struggles with writing Starlark. Otherwise, I think this is a functional basis for confirming token usage and limit consumption. Thanks a lot!

datasette-upload-dbs 0.5a0

Release: datasette-upload-dbs 0.5a0 This plugin has been around for a while - it lets users upload a brand new SQLite database to a hosted Datasette instance, at which point that database will start being served by that instance. It can also be used to atomically swap a database with a more recent version. The uploaded database is saved to a file, verified, then swapped in so /name starts serving the new one. The new release adds a formalized API, so you can replace an existing database (or add a new one) like this: curl -X POST \ -H "Authorization: Bearer $API_TOKEN" \ -H "Accept: application/json" \ -F "db=@content.db" \ -F "db_name=content" \ https://your-instance.example.com/-/upload-dbs This means you can build fresh databases in an environment such as GitHub Actions and swap them in production as soon as that build has completed. Tags: datasette

IEEE Engineering Summit Supports Bhutan’s Digital Transformation

In collaboration with the Kingdom of Bhutan government, IEEE recently introduced its Engineering Education, Research, and Innovation Summit. Held on 9 and 10 June in Paro, in the eastern Himalayas, the event was designed to help Bhutan navigate its digital transformation by focusing on the critical intersection of digital transformation, engineering education, and sustainable development. The summit brought together global academic leaders, technology experts, and Bhutanese government officials to discuss how modern engineering curricula can evolve from theory-centric models into application- and skills-based frameworks. Discussions focused on how to build high-value research capabilities in the country, integrate artificial intelligence into higher education , and address foundational infrastructure challenges to ensure equitable, nationwide digital readiness. “IEEE is proud to collaborate as a catalyst for progress in higher education as AI shifts the technology landscape and Bhutan prepares for its next era of innovation and resilience,” Mary Ellen Randall , 2026 IEEE president and CEO, said at the event. “Our goal is to support local universities and students as they develop trusted, future-ready technology that honors the nation’s commitment to sustainability and human well-being.” The event featured an address by Bhutanese Princess Chimi Yangzom Wangchuck , who emphasized the importance of aligning technological innovation with the nation’s philosophy of gross national…

Measuring Spurious Correlations with Feature Strength

This work was partially done by an automated research scaffold developed at Redwood Research. For this project, all of the experiment ideas were designed by a human and a human wrote the write up. The AI mostly just executed on the experiment ideas. We think this project is slightly below the level of rigor of a mid-MATS research update, and the research scaffold was not very helpful for this project. More discussion of AI usage is in the Appendix. 💻 Codebase If we want to train a classifier that distinguishes whether a passage is code or prose, we can do so by gathering samples of code and of prose, and training the classifier to distinguish between the two classes. Unfortunately, this might not work if the data hides a spurious correlation. If all the code is in Spanish and all of the prose is in English, then the classifier might learn to predict Spanish vs. English instead of code vs. prose. We find that this happens in practice: when we fine-tune an LLM to classify between Spanish code and English prose and evaluate on Spanish prose or English code, it generalizes to predicting the language rather than the domain. This suggests that language is in some sense a stronger feature than code. We think that spurious correlations are an important threat model for a few reasons. First, classifiers might actually be trained in ways that unintentionally contain spurious correlations. For example: Sycophancy vs validation-seeking user. Suppose we want to classify examples of the m…

Announcing Reverie Summit: What AI’s Next Breakthroughs Are Made On

One day with the researchers and engineers building generative video, world models, physical AI, and the data systems beneath them.

"But marinade" and leaked passwords are what researchers found in ChatGPT's hidden reasoning

Security researchers found a vulnerability in the APIs of OpenAI, Anthropic, and Google that lets them extract encrypted reasoning traces and move them between models. A scan of public sessions turned up dozens of passwords and API keys. The traces also show that the reasoning summaries users see often hide what the models are actually doing. The article "But marinade" and leaked passwords are what researchers found in ChatGPT's hidden reasoning appeared first on The Decoder .

Comment on DeepMind & Toulouse U Contribute Composable Function Preserving Transformations to Boost Transformer Training by Ethan Walker

Fascinating research into making transformer models more scalable and efficient. The idea of expanding model capacity while preserving existing functionality could have a major impact on future AI training. It’s impressive to see how quickly techniques like this are advancing, much like the innovation seen in industries such as roofing companies santa clarita

Got using real world conversations

Feature Request: Create a Feedback Loop From Real-World ChatGPT Outcomes Suggested tags: chatgpt, feature-request, feedback, training I have a product suggestion that I believe could substantially improve ChatGPT over time. ChatGPT has access to an enormous amount of information that people have published. But there is another potentially valuable source of knowledge that seems largely underutilized: the real-world experiences of people who actually use ChatGPT’s advice. The missing feedback loop Consider a typical conversation: A user asks ChatGPT how to solve a practical problem. ChatGPT researches available information, reasons through the situation, and recommends a course of action. The user then actually goes into the real world and does it. Sometimes the advice is exactly right. Sometimes it is technically correct but misses an important practical detail. Sometimes the user’s experience reveals something that isn’t documented well anywhere. And sometimes the advice is simply wrong. Today, there doesn’t appear to be a clear mechanism for turning that subsequent real-world experience into structured knowledge that can improve future ChatGPT responses. I think there should be. A real example I recently worked through an automotive repair problem with ChatGPT involving replacement wheel studs on a Kia Optima. During the conversation, we reasoned through how the broken studs could be removed. When I actually performed the repair, I discovered a practical detail that wasn’t…

How ONESTRUCTION built the Ishigaki-IDS foundation model with AWS GenAIIC

ONESTRUCTION, with technical advisory from the AWS Generative AI Innovation Center, built Ishigaki-IDS, a foundation model specialized for construction and BIM workflows. This architectural case study shows how they combined synthetic data, a three-stage training pipeline, and verifiable rewards on Amazon EC2 to build a domain model in a data-scarce field.

Comment on DeepMind Introduces Gato: A Generalist, Multi-Modal, Multi-Task, Multi-Embodiment Agent by monalisa1art

DeepMind's Gato is a fascinating step toward generalist agents, though calling it AGI feels like a stretch. The fact that one transformer model can handle text, vision, and robot control with shared weights is impressive, but crossing 50% expert threshold on 450 tasks still leaves plenty of room before true versatility. It does make me wonder how soon we'll see similar multi-modal approaches trickle into consumer tools—like a free nano banana image generator that adapts to different artistic styles without retraining. For now, Gato feels like a solid research milestone rather than a breakthrough. monalisa1art

Accuracy of GPT-4 Vision to extract exact numbers from graphs

This is a case where in-context training has previously been shown to help on a vision task. Contemporaneous with this old forum topic is a paper showing that examples of “how to turn vision into readings” can improve the actual readings provided: arXiv.org The Dawn of LMMs: Preliminary Explorations with GPT-4V(ision) Large multimodal models (LMMs) extend large language models (LLMs) with multi-sensory skills, such as visual understanding, to achieve stronger generic intelligence. In this paper, we analyze the latest model, GPT-4V(ision), to deepen the... Newer models are able to use larger imagery input, but do not upsize images themselves (something you can do, at expense). There is a transition from tiles to patches, at least as a billing method, in new models, giving linear relation between area and input tokens billed. Then, with gpt-5.6 (on the API, where developers know what is being done to images), the default image downsize cap is “original”—where an image such as 3600×2400 can be sent without downsize, providing more information in the large context attention sequence rather than in the individual semantic embedding that covers a large area with a small input. That, along with further post-training, should imply higher-quality positional answering in graphs with new models and big images. With reasoning.effort other than “none” on OpenAI gpt-5.2+ models, you do not have control over sampling constraints; thus, it is expected that each answer would differ. You can…

Introducing CARE-X: Towards Clinically Useful Radiology VLMs with Auxiliary Supervision, Reward-Aligned Learning, and Tool-Augmented Measurement

Radiology AI is evolving beyond report generation. CARE-X explores a unified approach that combines flexible reasoning, calibrated predictions, and measurement-based tools for chest X-ray interpretation. The post Introducing CARE-X: Towards Clinically Useful Radiology VLMs with Auxiliary Supervision, Reward-Aligned Learning, and Tool-Augmented Measurement appeared first on Microsoft Research .

Rainbows can be made from sounds – and we’ve created the best ones yet

Two research teams have independently created devices within which sound waves could be separated by frequency, creating the most complete sound rainbows yet

Simulating Lunar Regolith with COMSOL for Mission Safety and Space Infrastructure

Current space exploration aims to establish permanent structures on the Moon, Mars, and eventually other planetary bodies. Successful lunar missions depend on understanding lunar regolith, the granular material covering the Moon’s surface, whose behavior is governed by low gravity, vacuum conditions, particle irregularity, electrostatic effects, and extreme thermal environments. This webinar will focus on two connected modeling problems related to lunar regolith: plume–regolith interaction during lunar landings and induction heating of porous regolith for thermal processing and melting. Together, these problems show how simulation with the COMSOL Multiphysics ® software can help engineering and research teams predict, control, and utilize granular lunar material. The first part of the talk will address plume–regolith interaction. During spacecraft landings, underexpanded rocket exhaust plumes impinge on the surface, producing compressible flow structures, erosion, particle ejection, and possible surface damage. High-speed dust can reduce visibility and threaten astronauts, equipment, and nearby assets. These issues are especially important for Artemis and future missions involving repeated landings and larger spacecraft. Modeling the coupled gas-particle response provides insight into landing-site safety, erosion patterns, ejecta trajectories, and mitigation strategies. The second part of the webinar will examine induction heating of porous regolith, where electromagnetic en…

‘Zoomsday’ hack uncovered using fewer than 20 AI prompts

Zoom has patched a major security vulnerability that could allow an attacker to hijack anyone's device during a meeting. In a blog post on Tuesday, researchers at A Security say they uncovered the flaw using "fewer than 20 prompts on publicly available AI models," as reported earlier by Wired. The exploit involved Zoom's annotation feature, […]

How to Tell When a Benchmark Is Worth Trusting

An enterprise guide from the people who build them The post How to Tell When a Benchmark Is Worth Trusting appeared first on MLCommons .

US 'overpaid' for cancelled wind leases, researcher finds

A German power company agreed to cancel offshore wind development leases, bringing total reimbursements owed by US taxpayers to nearly $4 billion.

Those Who Make History

In 1972, astronauts on Apollo 17 set foot on the moon for a final time, collecting samples in the Taurus-Littrow valley, on the edge of Mare Serenitatis ("The Sea of Serenity"). At the end of the mission, like with earlier missions, NASA took the extremely valuable and interesting lunar specimens and did something strange: they hid them away in storage without even opening the containers. [1] Some stayed that way for nearly fifty years. Why? Because the scientists of the 70s understood that future generations would have better machines, methods, and ideas for studying the lunar rock and soil, and they wanted to make it easy for those researchers to run tests without having to go back to the lunar surface. This foresight paid off twice over. Advances in mass spectrometry enabled scientists in 2008 to detect water in volcanic-glass samples returned by Apollo 15 and Apollo 17. And when curators finally opened one of the last sealed containers in 2022 , they could extract the trapped lunar gases with technology that simply didn't exist in 1972. Some people describe cryonics as a new, and speculative technology. There’s a sense in which they’re right. It’s predicated, in large part, on the possibility of revival using technologies that do not yet exist. And Nectome , in particular, is a new form of cryonics, developing new techniques at the cutting edge that I believe are significantly higher quality than “traditional” methods. But I also think it’s important to see Nectome, and…

Google AI summaries: French daily newspapers file complaint with competition authority

Apig, the Alliance of General Information Press, is urging France's competition watchdog to make Google honour the commitments in its 2022 compensation deal with French news groups.

Comment on DeepSeek-V3 New Paper is coming! Unveiling the Secrets of Low-Cost Large Model Training through Hardware-Aware Co-design by Jamie Foster

Very informative, thank you! If you ever need a break, give head-football a shot – head-to-head soccer action, runs right in the browser.

AI Could Help Fossil Fuel Companies Create More Emissions

New research finds that by making the fossil fuel industry more productive, AI could help increase carbon emissions by up to nearly 5 percent—vastly outpacing the impact of data centers.

Seeing things through in the age of AI

AI is fantastic at prototyping. A quick draft of an essay, a mockup of a website, a demo of a video game, concept art or trailer for a movie, or the core argument of a proof - each now takes one prompt instead of a week. 1000x speedup. AI is useful for helping finish longer projects, but clearly not as good at it. Polishing up a paper draft that would have taken a month still takes two weeks. Finishing a video game that would have taken three years still takes two. Cleaning up every bug, human-checking a proof, covering every use case, handling every small nit and matter of taste, building production-ready infrastructure, all of that is work that, in my experience, only sees a 2x speedup with AI, at best. AI is a 1000x speedup for drafting and only 2x speedup at finishing. See the problem? We are seeing an absolute flood of slop in every AI-touched discipline, and nowhere near the promised gains in the stuff that matters. Humans are being completely hijacked by the high of being able to draft every great American novel they ever regretted not writing for a $200 subscription, and bouncing right off the wall of actually polishing any of these novels to a readable standard. The curation infrastructure (referees, editors, moderators) is being taxed beyond all measure, as the tsunami of half-finished demos crashes into its intake. It turns out that the best, most impactful, work lives in the tail of the power law in terms of strategy and commitment, and while AI helps with that k…

Zap Rocks. Add Water. Get Clean Hydrogen

In a tranquil Boston suburb , on the far edge of a horse farm, where pasture gives way to woods, a crane lowers an enormous electrode into a borehole. The electrode, a half-meter-long cylinder with copper-tipped arms to ensure good contact with the borehole walls, descends—deeper, deeper—through layers of spongy sandstone to the hard, marbled roots of an ancient mountain range hundreds of meters below ground. Here the rock is tight; there are few cracks for water or gases to flow. But that’s about to change. A stone’s throw away, a second electrode—a twin of the first—has been fixed in another borehole at the same depth. From above ground, a pair of high-voltage generators cabled to the two electrodes fires a series of pulses. Tsss!…Tsss!…Tsss!…Tsss!…Tsss!…. Each discharge, heard faintly at the surface, is like a miniature, subterranean lightning strike. The rock between the electrodes heats. Pressure builds. Then, suddenly, the rock splits into a spiderweb of fractures. On a horse farm outside of Boston, a worker sets up the well where Eden’s electrode will be lowered with a winch. Bob O’Connor Eden GeoPower , the Massachusetts-based startup performing this peculiar field test, calls the technology electrical reservoir stimulation. The company’s tagline: “We break rocks with electricity.” Eden’s researchers hope their rock-breaking technique will someday aid mineral mining, tap geothermal heat, or create geologic storage areas for carbon. But there’s an even more intriguing…

Nvidia's open Nemotron 3.5 Lightning model is all about specialized, local agentic AI

Our AI Model Release Tracker keeps new models in context with their peers, so you know which are worth your time.

Anthropic's planned mega-IPO faces investor skepticism over Chinese rivals and political headwinds

Anthropic is preparing an IPO for September or October, according to the Wall Street Journal, potentially the largest ever. During investor meetings, the company, valued at $965 billion, is fielding tough questions about Chinese competition, tensions with the Trump administration, and protests against data center construction. The company's IPO valuation will likely set the benchmark for how the entire AI industry gets valued. The article Anthropic's planned mega-IPO faces investor skepticism over Chinese rivals and political headwinds appeared first on The Decoder .

A Zoom Screen-Sharing Bug Let Anyone Take Over Other Devices on a Call

Researchers say it took fewer than 20 prompts for a public AI tool to find a flaw (now fixed) allowing anyone on a Zoom call to hijack another participants’ device.

The Download: the next big thing in LLMs and how AI academic research is shifting

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. These startups are chasing the next big thing in LLMs Nine years after Google researchers introduced the transformer, this family of neural networks has become the engine inside every major large…

GitHub already has an EDR. You just have to listen to it

Many of the recent supply-chain attacks could have been caught earlier if defenders looked closely at the telemetry GitHub already provides, researchers said. At their Black Hat USA 2026 presentation, researchers Yossi Weizman of Microsoft and Mor Weinberger of Echo argued the case, saying, “GitHub can tell you’re being hacked. You’re just not listening.” The duo described an EDR-style detection approach built from GitHub’s own event stream rather than relying solely on conventional endpoint or network telemetry. After studying recent supply-chain attacks, including Shai-Hulud , Trivy, and Megalodon, the researchers found that seemingly different incidents repeatedly used the same techniques, from forged commit identities and poisoned tags to workflow abuse, OpenID Connect (OIDC) theft, and attempts to erase evidence. They said they turned those recurring techniques into behavioral detections, combining GitHub webhooks, API data, and Git repository inspection to build a historical view of activity. Their new open-source tool, dubbed “GitHub Threat Detector,” reportedly includes 22 production detection rules and 12 beta rules, with compound detections designed to correlate individually weaker signals into high-confidence alerts. Everything leaves evidence on GitHub The central observation in Weizman and Weinberger’s research is that supply-chain attacks often repeat the same patterns even when the targeted projects are unrelated. A compromised identity, for example, may not b…

I made a bilingual Codex Skill + Plugin for candidate-first industry briefs

I built a Codex Skill + Plugin called Industry Brief Generator. The idea is simple: do not let the agent jump straight from research to a polished final brief. For industry news and market updates, the hardest part is often not writing. It is story selection: Which items are actually relevant? Which ones are just company soft copy? Which ones are generic finance news? Which sources are current and reliable? When should the agent stop instead of filling the count? So this workflow uses a candidate-first contract. Flow: The user enters industry, target market, focus topics, exclusions, final use case, and preferred language. The Skill generates 20 sourced candidate topics. The user selects 8. Only then does it generate the final brief, poster-style summary, and channel-ready copy. It currently includes: Chinese Codex Skill English Codex Skill bilingual Codex Plugin reusable YAML industry configs unsupported / low-information industry protocol poster QA checklist candidate-pool and final-brief templates Trust rules: no fabricated sources no fake dates no filler just to reach the target count no soft copy repackaged as industry news no final brief before explicit user selection Current traction: 43 GitHub stars 3 forks 250 downloads on WorkBuddy SkillHub GitHub: ndustry-brief-generator-plugin I would love feedback on: the candidate-first contract how to make Skills easier for non-technical users to install whether this pattern should become a reusable template for other research…

Marketing Vanguard at Cannes: How IBM Became Its Own AI Case Study ft. Jonathan Adashek

Are marketers making a fatal mistake by treating AI as just a cost-cutting tool?

A New Trick Reveals AI Models’ Inner Thoughts

Researchers devised a way to extract “reasoning traces” from Claude, GPT, and Gemini. What they found, they say, indicates that some Chinese AI may be trained on leading US models.

Experts are warning: our AI arms race is putting humanity at risk | Stuart Russell

A recent letter signed by 1,367 researchers and engineers at frontier AI labs – mainly OpenAI, Anthropic and Google Deepmind – points to a dangerous moment It is fashionable in certain circles to dismiss the catastrophic risks of AI. One often hears that “the real experts” who work on the technology every day are really not concerned at all; that only “doomers” and “luddites” espouse a “fringe” view from a “position of ignorance”; that all talk of potential catastrophe is just “science fiction”. Fortunately, an open letter has been published that lets us hear from the real experts who work on the technology every day, in their own words. And are they worried? Very. Continue reading...

Alibaba says modular design delivers AI data centres in 100 days at 10% lower cost

As demand for artificial intelligence infrastructure surges, Alibaba Group Holding says it can deliver new data centres in a fraction of the usual time while cutting construction costs by 10 per cent through its proprietary modular architecture. Using CUBE 5.0, Alibaba Cloud had slashed the delivery time for large-scale AI data centres to just 100 days, according to a report by state-backed newspaper China Securities Journal on Tuesday. That compared with standard domestic delivery times of six...

Malicious SIMs can shut down phones, steal files, and drag 5G back to 2G

Researchers find standards-compliant functionality can be abused to hijack modems, downgrade connections, and even execute code

AI’s potential climate benefits outweighed by role in boosting fossil fuels, study finds

Modelling finds AI-driven productivity gains in coal, oil and gas enable more emissions than applications in renewables avoid AI-driven productivity gains enable more planet-heating pollution from fossil fuels than they avoid from renewables, a study has found. Researchers modelled the technical potential for AI to boost clean power generation along with projections for how it can help produce coal, oil and gas. Across 64 scenarios, they found net yearly carbon pollution rose by 0.47-1.8 gigatonnes, or about 1-5% of the energy sector’s annual emissions. Continue reading...

Muse Glimmer 30B Architecture Notes

Short architecture note on Meta Muse Glimmer 30B, including gated local and global GQA, KV-cache efficiency, and release-time benchmark comparisons.

Feature Request: Persistent Project State

Feature Request: Persistent Project State — ChatGPT Needs to Remember the Work, Not Just the User Introduction I’m not a developer or AI professional. I’m an ordinary user who has started using AI extensively because I can see its potential to become a genuine personal assistant rather than simply a tool for answering questions. I’ve used ChatGPT for travel planning, research, writing, purchasing decisions, retirement planning and everyday problem solving. I think it is extremely capable, and I can see enormous potential in it. This proposal comes from a real experience I’ve just had with ChatGPT. It highlighted what I believe is one of the biggest gaps between an excellent AI chatbot and a genuinely dependable long-term personal assistant. The problem ChatGPT needs to reliably remember and maintain ongoing projects . I don’t simply mean remembering facts about me. I mean remembering the current state of something we have been working on together — including the decisions we’ve made, changes we’ve made, things we’ve rejected, outstanding decisions and the latest version of the plan. A real example I was using ChatGPT to plan a family holiday in Spain. We built a detailed Summer 2026 Holiday Master over a number of conversations. The holiday covered two weeks and two different resorts: Week 1 — Calella Week 2 — Salou Over several days, we progressively built the plan. We discussed trains, markets, meals, restaurants, transport, activities, PortAventura, costs and changes to i…

AI Is Helping Solve the Intricate Genetic Puzzle of Schizophrenia

Recent findings provide one of the most detailed pictures to date of the genetic architecture of schizophrenia, opening up new avenues for research into the disorder.

Add GPT Logo in Chat History for Custom GPTs

I completely agree with this request. I regularly use different custom GPTs for different purposes, and it can become difficult to remember which GPT was used for a particular conversation when browsing chat history. For example, I may use one GPT as a sales coach, another for writing, and another for research. Having the GPT name or icon displayed directly in the chat history would make it much easier to identify the right conversation and continue working with the appropriate assistant. I currently have to manually add identifiers to chat titles as well. A simple GPT name or icon attached to each conversation would be a very useful improvement.

Marketers Are Being Pushed to Use AI, But Not Taught How

New research from ADWEEK and NewtonX shows marketers are falling behind in AI training

Could we eliminate any particles from the standard model of physics?

The standard model of physics is one of the most successful scientific theories in existence, but do we need it all? Columnist Chanda Prescod-Weinstein explores whether it’s carrying any excess baggage

Sarvam AI's Indic DiarBench Tests Speech Recognition Across all 22 Indian Languages

Sarvam AI releases the first open benchmark for multi-speaker speech recognition across all 22 scheduled Indian languages, exposing major gaps in commercial APIs

Ayati Devices raises Rs 15 Cr led by Inflexor Ventures in pre-Series A

Bengaluru-based medical technology startup Ayati Devices has raised Rs 15 crore ($1.5 million) in a Pre-Series A funding round led by Inflexor Ventures. The funds will be used to accelerate commercialisation, expand its presence in domestic and international markets, strengthen manufacturing capabilities, and increase investments in research and artificial intelligence. Founded in 2019 by Nishant Kathpal, Ayati Devices focuses on diagnostic technologies for diabetic foot complications and peripheral vascular disease. The startup was incubated at IIT Bombay’s Society for Innovation and Entrepreneurship (SINE) and has built a portfolio of portable, point-of-care diagnostic devices aimed at making early screening accessible beyond tertiary hospitals. Its product portfolio covers multiple stages of diabetic foot assessment. Vibrasense measures vibration perception threshold to identify large-fibre peripheral neuropathy, while Vibrasense+T adds warm and cold perception testing. Vasosense is designed for peripheral artery disease screening, while Angiocam enables real-time tissue perfusion imaging. The company also offers the PODIA Trolley, a pay-per-test screening model designed to reduce the upfront investment required for healthcare providers. Ayati’s current platform brings together assessments across neuropathy, blood flow, microcirculation and plantar pressure. Ayati claims to have been deployed across 30 countries, with more than 10,000 devices deployed and clinical deploym…

Long-context ChatGPT increasingly seems to synthesize instead of verify, and Cross Context Mix Up

I’m a very heavy ChatGPT user. I use it across long-running conversations, research, technical troubleshooting, software/workflow development, and systems such as Aris Vault and Aris Ledger. Recently I’ve noticed a recurring problem that bothers me more than ordinary hallucination: ChatGPT sometimes answers from plausible synthesis instead of checking an easily verifiable fact. The answer often sounds completely coherent, but when challenged it becomes clear that the model has reconstructed what probably happened rather than retrieved what actually happened. I’ve also seen related issues with false recollection, context from one thread bleeding into another, and inferred details later being treated as if they were established facts. At the same time, I’m genuinely happy to see ChatGPT’s memory and cross-conversation context improving . That continuity is extremely useful for the way I work. But paradoxically, I also seem to be seeing cross-context mix-ups much more frequently : a real fact from one conversation, person, project, or time period gets pulled into the wrong context and presented as though it belongs there. The fact itself may be correct; the association is wrong. This is especially problematic in long-context use, where information may come from current chat, previous chats, memory, files, summaries, or inference. I increasingly find myself asking: Did I actually tell you this? Did you retrieve this or infer it? Can you find the original source? Are you sure, or…

Comment on Revolutionizing AI on a Budget: Apple’s Roadmap for Small Language Models Training Success by exceltomd

The focus on SLMs up to 2 billion parameters is a useful counterpoint to the usual LLM-centric optimization discussions. I appreciated the systematic breakdown of computational bottlenecks across different cloud setups—it gives practitioners a clearer way to think about cost-efficiency before scaling up. The distinction between training behavior of smaller and larger models seems especially practical for teams that need to ship on a budget.

Before We Defer Research to AI: Measuring Apparent-Success-Seeking

Recently, I was improving a small LLM-powered classifier and noticed a few continuously failing test cases. As many would, I asked my AI code assistant to add a few more out-of-distribution examples to the classifier’s few-shot prompt. After rerunning with the updated classifier, unsurprisingly, many of the failing test cases passed. Due diligence and curiosity brought me to look at the test cases my AI assistant added. Reasonably, I expected the assistant to follow my instructions and create fresh examples covering similar cases the classifier was missing. However, the AI assistant actually copied my failing test cases into the few-shot prompt, gaming the evaluation and obviously going against user intent. When I called the assistant out, it did it again, just harder to spot, and only actually followed my instructions after multiple strongly worded pushbacks. I only caught this issue because I happened to look; the number of passing test cases would have told me I’d succeeded. Other than this anecdote, the most extreme case of this problem I’ve encountered throughout my time as a developer and researcher, I’ve run into multiple similar examples of AI code assistants hacking an evaluation instead of improving the thing actually being evaluated. This includes changing the evaluation to make it easier or contaminating the product being evaluated, sometimes behind my back. These are textbook examples of apparent-success-seeking: optimizing to look done rather than being done. T…

A Topic Detector, Not a Lie Detector: what J-space monitoring actually tracks

This is a pilot experiment, done on one model, with around $14 worth of compute, and a single seed per condition. The full writeup with all figures and statistics is linked below. This is posted here to get feedback and criticism, since I am aware this method is not the best. TL:DR: Anthropic's J-lens research has shown that a large language model has an internal workspace in which different activations can be used as a safety monitor. We investigated the conflict between the model's workspace activation and outputs, which we called C. We ran our experiments on a model whose final alignment differs from that of its training data: DeepSeek-R1-Distill-Qwen-14B. We assume that some changes were made to the model after training in order for it to comply with some guidelines. Some guideline-skirting questions registered elevated C despite compliant statements being made, and J-lens was able to discriminate between concealing answers and controls with AUC of 0.97 on proper nouns (though only 0.55 when pooling all classes). We then fine-tuned the model to appear to hold beliefs in line with its guidelines. Our initial hypothesis was that this would drastically lower C, since the model would no longer be making a statement it "believes" to be untrue. This hypothesis was disproven: C rose to 130% of its initial level for the relevant tokens, and to 115% of its initial level for irrelevant tokens. Despite this, the compliant fine-tuning was successful in making the model formulate the…

Cold identity Architecture & Fine tuning roadmap

I was wondering how your project is coming along. I’ve recently been working on a similar task—using LoRA SFT on Qwen 3.6 27B to steer the model toward a specific persona. Unfortunately, I haven’t had much success yet, though I’m still experimenting. I’m not sure if the issue lies with my dataset or something else. If you have any insights or suggestions, I’d really appreciate it!