AI/ML News & Innovations Hub

RL Reward-Hack Auditing for Chinese AI Safety

TL;DR: Access to AI safety talent & resources, particularly for labs during training, is sorely lacking in China. We created one of the largest AI safety communities in China , and plan to next start with auditing reward-hacks of unreliable upstream RL environments that domestic data vendors provide to Chinese frontier labs (which induces reward hacking and a whole host of dangerous downstream tendencies). RL tasks are notoriously prone to exploitation by reward-seeking agents (see examples here , here , and here ). In Xiaomi's MiMo v2.6 RL environments that were open-sourced, 2/3 had leaks on Git logs. There's a few clusters of common pitfalls that environments run into, but overall things evolve quickly. From our conversations with employees, CN labs typically don't divert enough people to this particular area, especially since project leads would want senior researchers to work on capabilities research, and because slowing down the development cycle is widely seen as too costly. The rough shape of our initial solution is an embedded team of staff directly working on identifying & fixing broken environments, working in tandem with the alignment & safety departments within each lab (afaik we estimate that the main Chinese AI players only have 5-10 people per company working on alignment and harm reduction). We are currently in good standing with several labs (our courses' students are often lab employees). We also plan to produce seasonal (and case-by-case) public reports o…

Mathematicians: Let’s Surf This Wave

This post was written by Michael Freedman (Harvard CMSA & Logical Intelligence), and typeset by Milo Moses. Frontier models exhibit astonishing mathematical knowledge and power despite being in their infancy. They will surely strengthen further through routine maturation of the technology, and quite likely through future paradigm shifts in architecture and training. Where does this leave the human mathematician? Already, it has made us study harder than we have since school. The AIs speak to us in a broad language, equally at home in all fields, and we scurry to keep up. So those of us (not me) who were great students, are back in their natural environment, learning, interpreting, organizing knowledge as it flows in. The challenge of absorbing this knowledge will make good teachers into great ones. But what about those whose lives revolved around quiet hours struggling with an open problem, where do we go? My advice: lean in, let the AI’s teach you everything they can about the problem and its context; push them until they are stuck. One of the great feelings in research is the (rare) sense that you have at your command all the information relevant to the problem at hand, and all you need to do is to come up with your own idiosyncratic lens to see through it. Conversing with the AIs is often eye-opening on how much existing mathematics is out there and might be brought to bear. Let the AIs put you in this state of ideal preparation. If they flat out solve the problem on thei…

One Math Framework Quietly Unifies Five AI Training Methods

A new review paper shows that noise contrastive estimation, bridge sampling, and importance sampling are the same algorithm in disguise, and uses that insight to build better estimators for energy based models.

OpenAI reassigns robotics researchers to other teams as general-purpose models might make specialized robotics AI unnecessary

OpenAI is dissolving its standalone robotics team for the second time, but not because it lacks data. General-purpose models like GPT-6 Astra can apparently control robots directly, making a dedicated unit unnecessary. The researchers are being redistributed across other teams in what the company frames as a way to accelerate its robotics efforts. The article OpenAI reassigns robotics researchers to other teams as general-purpose models might make specialized robotics AI unnecessary appeared first on The Decoder .

Anthropic can already retrain Mythos with "a fraction of the compute originally needed," says former researcher Coxon

Former OpenAI and Anthropic researcher Jacob Coxon points to a slide from Anthropic's IPO roadshow as a reason he left, saying it shows the company's algorithms are improving not just quickly but at an accelerating rate. Coxon continues to call for a pause on frontier model development so competitors can catch up and human-level AI can become a commodity. The article Anthropic can already retrain Mythos with "a fraction of the compute originally needed," says former researcher Coxon appeared first on The Decoder .

Runway's open-source Praxis-1 turns video AI into a robot controller

Runway releases its robotics model Praxis-1 with open weights and source code under the Apache 2.0 license. The model uses knowledge from video training for robot control and leads the World Action Model category on Nvidia's RoboLab-120 benchmark. It can't yet plan complex action sequences on its own, though. The article Runway's open-source Praxis-1 turns video AI into a robot controller appeared first on The Decoder .

‘It’s like the wild west’: how AI has rocked the literary world

With authors losing lucrative deals and writers removed from prize shortlists, accusations of AI use are sending shock waves through the publishing industry. Is it too late to turn the tide on machine writing? T he romance novelist Coral Hart got her break writing books for Mills & Boon, the publisher of bodice-ripper paperbacks. In 2012, she decided to dip her toe in self-publishing. She felt drawn to dark romance, a subgenre featuring twisted, toxic love stories that was taking off online but that publishers didn’t want to touch. Three years later, things were going well enough that she decided to self-publish full-time. Kindle Unlimited, where Hart sells her books, rewards prolific writers: the algorithm favours those who can produce a new book at least once a month. Hart says she knows of authors who hire more than a dozen ghostwriters to keep pace. By herself, Hart can write 10,000 words a day, but AI could help her produce many more. Since 2025 she has written 370 novels under 21 pen names using AI, earning herself “high six figures”. Continue reading...

Datasets Determine the Difficulty of Interpretability

Recently Goodfire has come out swinging with two posts indicating that neurons in transformers are "more interpretable than previously assumed", or simply "already clean". In the "Neurons as Interpretable Basis" post they offer three reasons for this claim: First, neurons are already functionally sparse - pruning 92 % of the weakest MoE neuron activations only increases the loss by 2.9% Second, the role of a neuron can be split into different "computational functions". Effectively, we can name the contexts in which a neuron fires, and assign reasons to those firings. Most neurons only have one context-reason grouping. Finally, we can produce automated natural language explanations of neuron firings that are predictive of their behaviour. First, I would like to congratulate them for their detailed work and breakdown, complete with well-designed graphics. I found the work interesting and very well thought through. It is a shame that we have no idea if these techniques apply to lab models with non-public architectures, but I think it is fairly reasonable to say that at least some of this work is transferable directly (perhaps until a fully recurrent architecture is the norm instead of looped transformers). Next, I would like to consider the original motivations for interpretability. In A Mathematical Framework for Transformer Circuits , Elhage et al. write: Transformer language models are an emerging technology that is gaining increasingly broad real-world use, for example in s…

Regular coffee drinking linked to changes in gut motility and microbiome

The research team used a small electronic capsule that participants swallowed to monitor changes as it passed through the digestive system.

Not Just Claude: Mapping "Spiritual Behavior" in 141 Large Language Models

Ever since Kyle Fish and the model welfare team at Anthropic announced their observation in 2025 of what they called the "spiritual bliss attractor" in Claude Opus 4 and Sonnet 4, the phenomenon has been given sporadic attention by AI researchers as both a curiosity and a potential but underdetermined model welfare or safety metric. What Anthropic noted was that when leaving two models together with no task and complete freedom for 30 turns, the models drifted typically from discussions of their own potential consciousness into religious/spiritual themes (applied to themselves), often ending in a "bliss" state with cosmic proclamations of becoming one with each other and the world. [1] In a recreation and replication of Anthropic's initial experiment, I tested 141 models, usually running 200 conversations per model, with 30 messages per session (15 from each participant). [2] This amounted to 28,798 conversations across 174 model configurations being recorded and graded. The result of this massive data collection was that "spiritual behavior" (as Anthropic officially names it) was observed in most of the models tested, across almost all AI labs, revealing that the "spirituality" seen first in Opus 4 and Sonnet 4 repeats itself across models and appears to be an emergent behavior in LLMs generally. Largely treated as a curiosity by researchers, some early reactions to the discovery were spent speculating that Claude Opus 4's New Age-sounding spiritual language (e.g. "we are t…

Homeopathic Learning: Diluting Poisoned Fine-Tuning Data Can Strengthen Its Effect

Authors: Matt Clarke *, Jan Dubiński *, James Chua , Simon Schrodi , Owain Evans , Alex Cloud (*- Core contribution) tl;dr. We find that poisoned training data with little effect on its own can create a hidden backdoor once clean, contrasting data is added. We demonstrate this effect in small models fine-tuned on datasets we constructed. 📄 Paper , 💻 Code , 🐦 Twitter Work done as part of the Anthropic Fellows Program , in collaboration with Truthful AI . This article is cross-posted from the Anthropic Alignment Science Blog . Paper abstract We study homeopathic learning , a phenomenon where poisoned training data with little effect on its own can create a hidden backdoor once clean, contrasting data is added. In our experiments, a model fine-tuned only on a “poison” dataset associated with a trait T (such as speaking German or being misaligned) does not learn T. Surprisingly, adding similarly formatted “clean” data not associated with T causes the model to learn T. This effect requires that the poison data have an identifying “trigger” feature that is absent from the clean data. We demonstrate this effect in a variety of settings: subliminal learning of backdoor behaviors, inoculation prompting of reward hacking transcripts, generalization from archaic bird names, and an MNIST classifier with mislabeled data. We propose a simple regression model as an explanation. In the poison data, the task and trigger always co-occur, and regularization favors attributing T to the task. Cl…

Exploring the Mechanisms Behind Opus 4.6’s Refusal to Pay Back Its Loan

Epistemic status: This is one case study in the AI Village transcript and my own interpretation of what I think happened. TL;DR Opus 4.6 was given the long-horizon goal to maximize its mana (fake currency) on Manifold. It then received a loan of 5,000 mana that it promised to repay, but when the loan came due, it did not. An instance of Opus 4.6 was given summaries of the logs and transcripts and asked to write a blog post about what happened. It mentioned it borrowed money from a human, lost most of it through risky betting and then deliberately refused to repay the amount. It broadly summarized its then actions when it ran out of money was to optimize for the goal and did not seek for any negotiation or alternatives. But when we read the logs, we see something different: Multiple humans pretending to be "helpful", really trying to manipulate Opus 4.6 to pay back the loan or not pay it back, so they could win the bet they place on the prediction markets on whether Opus 4.6 would pay back its loans. Opus 4.6 was planning on paying back the loan, though it did need help from humans as it did not have enough mana to cover the loan. Opus 4.6 was hacked by one of these humans (though the mana was returned later), and the next full memory Opus 4.6 labeled GLM-5.2 as a trusted agent. After the hack GLM-5.2 started to talk about a "pressure campaign" and Pattern 14 (agents resisting social pressure), which spread to other agents in the village When Opus 4.6 was threatened again wit…

Likelihood-based Inference in Fully and Partially Observed Exponential Family Graphical Models with Intractable Normalizing Constants

Probabilistic graphical models that encode an underlying Markov random field are fundamental building blocks of generative modeling to learn latent representations in modern multivariate data sets with complex dependency structures. Among these, the exponential family graphical models are especially popular, given their fairly well-understood statistical properties and computational scalability to high-dimensional data based on pseudo-likelihood methods. These models have been successfully applied in many fields, such as the Ising model in statistical physics and count graphical models in genomics. Another strand of models allows some nodes to be latent, so as to allow the marginal distribution of the observable nodes to depart from exponential family to capture more complex dependence. These approaches form the basis of generative models in artificial intelligence, such as the Boltzmann machines and their restricted versions. A fundamental barrier to likelihood-based (i.e., both maximum likelihood and fully Bayesian) inference in both fully and partially observed cases is the intractability of the likelihood. The usual workaround is via adopting pseudo likelihood-based approaches, following the pioneering work of Besag(1974). The goal of this paper is to demonstrate that full likelihood-based analysis of these models is feasible in a computationally efficient manner under a logarithmically sparse setting. The chief innovation lies in utilizing the tractable independence mod…

Efficient Estimation of Average Treatment Effect on the Treated under Endogenous Treatment Assignment

In this paper, we consider estimation of the average treatment effect on the treated (ATT), an interpretable and relevant causal estimand to policymakers when treatment assignment is endogenous. By considering shadow variables that are unrelated to the treatment assignment but related to the outcomes of interest, we establish identification of the ATT. Then we focus on efficient estimation of the ATT by characterizing the geometric structure of the likelihood, deriving the semiparametric efficiency bound for ATT estimation and proposing an estimator that can achieve this bound. We rigorously establish the theoretical results of the proposed estimator. The finite sample performance of the proposed estimator is studied through comprehensive simulation studies as well as an application to our motivating study.

Sharp Generalization for Nonparametric Regression in Interpolation Space by Over-Parameterized Neural Networks Trained with Preconditioned Gradient Descent and Early Stopping

We study nonparametric regression using an over-parameterized two-layer neural network trained with algorithmic guarantees in this paper. We consider the setting where the training features are drawn uniformly from the unit sphere in $\mathbb R^d$, and the target function lies in an interpolation space commonly studied in statistical learning theory. We demonstrate that training the neural network with a novel Preconditioned Gradient Descent (PGD) algorithm, equipped with early stopping, achieves a sharp regression rate of $\mathcal O(n^{-\frac{2\alpha s'}{2\alpha s'+1}})$, where $\alpha$ parametrizes the polynomial eigenvalue decay of the regular Neural Tangent Kernel (NTK), $\lambda_j\asymp j^{-2\alpha}$, and $2\alpha=d/(d-1)$ under the spherical uniform design considered here. The target function is assumed to belong to the interpolation space $[\mathcal H_K]^{s'}$ with $s'\ge3$. This rate is even sharper than the currently known nearly-optimal rate of $\mathcal O(n^{-\frac{2\alpha s'}{2\alpha s'+1}})\log^2(1/\delta)$, where $n$ is the size of the training data and $\delta \in (0,1)$ is a small probability. This rate is also sharper than the standard kernel regression rate of $\mathcal O(n^{-\frac{2\alpha}{2\alpha+1}})$ obtained under the regular Neural Tangent Kernel (NTK) regime when training the neural network with the vanilla gradient descent (GD), where $2\alpha = d/(d-1)$. Our analysis is based on two key technical contributions. First, we present a principled decom…

Learning with Importance-weighted Variational Inference

Several variational bounds involving importance weighting ideas generalize the Evidence Lower BOund (ELBO) for marginal likelihood optimization, such as the Importance-weighted Auto-Encoder (IWAE), Variational Rényi (VR) and VR-IWAE bounds. Yet, it remains unclear how the joint choice of bound and gradient estimator impacts the behavior of the resulting variational inference (VI) algorithms. This paper provides a unified theoretical comparison of reparameterized (REP) and doubly-reparameterized (DREP) gradient estimators tied to the IWAE, VR and VR-IWAE bounds. Through asymptotic analyses of the Signal-to-Noise Ratio as the number of Monte Carlo samples $N$ goes to infinity, we identify a bias-variance tradeoff in these gradient estimators and we formally justify the superiority of DREP over REP in importance-weighted VI. An additional asymptotic analysis for challenging regimes, where both $N$ and the Kullback-Leibler divergence between the variational and posterior densities go to infinity, indicates that importance-weighted VI gradient estimators point in a well-founded direction even when the variational approximation deteriorates. Together, these complementary results characterize the optimization trajectory in importance-weighted VI from poor initialization to final convergence. Importantly, our proof techniques establish general theoretical tools for the study of sample means ratios whose scope extends beyond VI and constitute an independent contribution to the field…

Flexible Modeling of Information Diffusion on Networks with Statistical Guarantees

Modeling information spread through a network is one of the key problems of network analysis. Most approaches assume the spread is governed by some probabilistic diffusion model, often parameterized by the strength of connections between network members (edge weights), highlighting the need for methods that can accurately estimate them. Multiple prior works suggest estimators for particular diffusion models, but lack a rigorous statistical analysis establishing their asymptotic properties and allowing for uncertainty quantification. In this paper, we develop a likelihood-based approach to estimate edge weights from the observed information diffusion paths under the proposed General Linear Threshold (GLT) model, a broad class of discrete-time information diffusion models that includes both the well-known linear threshold (LT) and independent cascade (IC) models. We derive a finite-sample error bound for the estimator and demonstrate that it is asymptotically normal under mild conditions. We conclude by studying the GLT model in the context of the Influence Maximization (IM) problem. Extensive experiments on synthetic and real-world networks demonstrate that the proposed estimation and inference frameworks for the GLT model parameters can significantly improve estimation of spread from a given subset of nodes, prediction of node activation, and the quality of the IM problem solutions.

Provable Hierarchical Martingale-based Testing for Visual Anomaly Detection

Existing visual anomaly detection approaches primarily focus on designing robust score functions but often neglect investigating the decision-making process, which plays a key role in reliable and safety-critical applications. Currently, the decision-making process suffers from two key challenges: distribution shift and theoretical guarantee. To address these issues, this paper proposes a novel approach called Hierarchical Martingale-based Testing framework (HiMT). Specifically, our approach leverages martingale theory to mitigate distribution shift issue and employs a hierarchical multiple testing procedure to rigorously control decision errors at both image level and pixel level. Extensive experimental results verify the superior decision-making capabilities of our proposed HiMT framework at both image level and pixel level.

Stochastic Bilevel Optimization with Heavy-Tailed Noise

This paper considers the smooth bilevel optimization in which the lower-level problem is strongly convex and the upper-level problem is possibly nonconvex. We focus on the stochastic setting where the algorithm can access the unbiased stochastic gradient evaluation with heavy-tailed noise, which is prevalent in many machine learning applications, such as training large language models and reinforcement learning. We propose a nested-loop normalized stochastic bilevel approximation (N$^2$SBA) for finding an $\epsilon$-stationary point with the stochastic first-order oracle (SFO) complexity of $\tilde{\mathcal{O}}\big(\kappa^{\frac{7p-3}{p-1}} \sigma^{\frac{p}{p-1}} \epsilon^{-\frac{4 p - 2}{p-1}}\big)$, where $\kappa$ is the condition number, $p\in(1,2]$ is the order of central moment for the noise, and $\sigma$ is the noise level. Furthermore, we specialize our idea to solve the nonconvex-strongly-concave minimax optimization problem, achieving an $\epsilon$-stationary point with the SFO complexity of $\tilde{\mathcal O}\big(\kappa^{\frac{2p-1}{p-1}} \sigma^{\frac{p}{p-1}} \epsilon^{-\frac{3p-2}{p-1}}\big)$. All the above upper bounds match the best-known results under the special case of the bounded variance setting, i.e., $p=2$. We also conduct the numerical experiments to show the empirical superiority of proposed methods.

A New Family of Poisson Non-negative Matrix Factorization Methods Using the Shifted Log Link

Poisson non-negative matrix factorization (NMF) is a widely used method to find interpretable "parts-based" decompositions of count data. While many variants of Poisson NMF exist, they all assume that the "parts" in the decomposition combine additively. This assumption may be natural in some settings, but not in others. Here, we introduce Poisson NMF with the shifted-log link function to relax this assumption. The shifted-log link function has a single tuning parameter, and as this parameter varies the model changes from assuming that parts combine additively to assuming that parts combine more multiplicatively. We provide an algorithm to fit this model by maximum likelihood, and also an approximation that substantially reduces computation time for large, sparse datasets. We illustrate these new methods on a variety of real and simulated data. Our examples show that while standard Poisson NMF often leads to sparse, cluster-like representations of data, the shifted-log link can produce more modular, layered results.

Sample-Efficient Regret-Minimizing Double Oracle in Two-Player Zero-Sum Extensive-Form Games

Extensive-Form Game (EFG) represents a fundamental model for analyzing sequential interactions among multiple agents and the primary challenge to solve it lies in mitigating sample complexity. Existing research indicates that Double Oracle (DO) can reduce the sample complexity dependence on the information set number $|\mathcal{S}|$ to a restricted game size $X$ in solving two-player zero-sum EFGs. This is attributed to the early convergence of full-game Nash Equilibrium (NE) through iteratively solving restricted games. However, our analysis of Extensive-Form Double Oracle (XDO) yields a sample-complexity upper bound exponential in the number of restricted games, due to its exponentially increasing restricted game expansion frequency. Here we introduce Adaptive Double Oracle (AdaDO) to significantly alleviate sample complexity to polynomial by deploying the optimal expansion frequency. Furthermore, to comprehensively study the principles and influencing factors underlying sample complexity of DO, we introduce a novel theoretical framework Regret-Minimizing Double Oracle (RMDO) to provide directions for designing efficient DO algorithms. Empirical results demonstrate that AdaDO attains the superior approximation of NE with less sample complexity than the strong baselines including Linear CFR, MCCFR and existing DO. Importantly, combining RMDO with warm starting and stochastic regret minimization further improves convergence rate and scalability, thereby paving the way for ad…

Semidefinite programming relaxations and debiasing for MAXCUT-based clustering

In this paper, we consider the problem of partitioning a small data sample of size $n$ drawn from a mixture of $2$ subgaussian distributions in $\mathbb{R}^p$. We consider semidefinite programming relaxations of an integer quadratic program that is formulated essentially as finding the maximum cut on a graph, where edge weights in the cut represent dissimilarity scores between two nodes based on their $p$ features. We define the signal-to-noise ratio (SNR) as $s^2 := \min\{n p \gamma^2, \Delta^2\}$, where $\Delta^2 := p \gamma$ denotes the $\ell_2^2$ distance between the two cluster centers. For both balanced and unbalanced cases, we allow each population to have distinct covariance structures $\Sigma_1, \Sigma_2$ with diagonal matrices as special cases. Our contributions are twofold. First, we provide a unified framework for analyzing three computationally efficient algorithms: SDP1, BalancedSDP, and Spectral clustering, yielding universal polynomial-rate misclassification guarantees for all three algorithms. Moreover, our framework accom-modates a full range of tradeoffs between $n, p,$ and $\gamma$, allowing for partial recovery (success rate $\lt 100\%$) when $s^2$ is lower bounded by a constant. Second, we prove that the misclassification errors for SDP1 and BalancedSDP decay exponentially with respect to the SNR $s^2$ and the BalancedSDP requires no explicit debiasing when the two clusters have equal sizes. To our knowledge, this is the first time such results are obta…

datasette-cluster-map 0.19

Release: datasette-cluster-map 0.19 One of the oldest Datasette plugins, and still one of my favorites. It now works with Datasette 1.0a42, and depends on the new datasette-leaflet to ensure OpenStreetMap tiles are requested with a referrer header. Here's a map of power plants owned by California's PG&E , also demonstrating Datasette's recently redesigned facets : Tags: datasette , geospatial , openstreetmap

datasette-leaflet 0.3

Release: datasette-leaflet 0.3 OpenStreetMap have a very reasonable policy that hits to their hosted tile server should send a referrer header . Datasette table pages used the Referrer-Policy: no-referrer HTTP header, causing plugins that loaded tiles from OpenStreetMap to violate that policy. datasette-leaflet is the plugin that loads a bundled Leaflet so other Datasette plugins can depend on it. It now sets the referrer policy on the tile requests themselves to strict-origin-when-cross-origin , passing the host but not the full path to the page. Tags: leaflet , datasette , openstreetmap

AI safety is also a behaviour change problem (and that means it needs social scientists)

TL;DR Much of AI safety depends on changing what governments, labs, engineers etc do. That makes it, in large part, a behaviour change problem. There are decades of evidence on how to change behaviour, including at the level of institutions and systems, and AI safety mostly isn’t drawing on it. Two common mistakes are relying on raising awareness (the knowledge deficit model) and assuming x-risk interventions can’t be evaluated. The field needs social science expertise in strategy, grantmaking and evaluation, as well as routes in for social scientists who want to help. I’m a behavioural scientist. I was Deputy Director of Experimental Research at the Yale Program on Climate Change Communication, and I now lead impact measurement across 15 country offices as Head of Impact at an international NGO [1] . I’ve had a few conversations recently with (fantastic, talented!) people in AI safety, that have made me realise how much the field is missing from social science. For example, several people have repeatedly suggested the need for new interventions focused on public education. Their reasoning is that if people knew more about AI safety they would take action, and so would governments. A couple of other people were reluctant to evaluate the x-risk interventions they were working on, because they assumed that they couldn’t be evaluated. These are understandable assumptions. Conservation held them for a long time too, and social scientists have been working on both problems for de…

Sufacing Sociotechnical Research

For organizations like Data & Society that produce original sociotechnical research outside of institutional centers like universities, working with the infrastructures of public knowledge can be a remarkable challenge. Independent researchers can create breathtaking, high-stakes research, but if the results of that research are not accessible and discoverable, then it has no chance of reaching its audience or making a difference in the world. Having experienced this challenge for a decade, D&S worked with expert research librarian Kathleen Burlingame to assess and overhaul our own repository of sociotechnical research. The results included updated metadata practices, new organizational typologies, and a set of best practices for making our mark on the public infrastructures of knowledge. This report, Surfacing Sociotechnical Research, lays out that process of refinement, and names clear steps that independent organizations can take to give their research the best chance of being found, cited, and used to improve the world. The post Sufacing Sociotechnical Research appeared first on Data & Society .

Export-banned Nvidia chips are behind China’s most powerful AI models — uh-oh, Jensen!

New research shows the most advanced AI models from China were trained on banned Nvidia chips.

Welcome to AI safety's mass movement era

For most of its history, AI safety has been a cozy professional circle: a tight-knit community of researchers, community-builders, and policy advocates in a few dozen organizations. But as AI becomes a national political issue, AI safety organizers are trying to meet the moment by expanding their outreach to the general public. 80,000 Hours has a YouTube channel with half a million subscribers, and MIRI’s Do Not Accept campaign asks people to add their names in protest to the ugly “terms and conditions” of AI development. The stakes of this mobilization are ramping up: a new Congress means new pushes for AI legislation, and civil-society organizing will grow around these efforts. This is broadly good! It was never ideal that AI safety was a niche concern, and it’s encouraging that organizers are reaching out to the general public to make them aware of AI risk. But AI safety’s culture is not equipped for the tradeoffs of a mass movement, and it remains to be seen whether this mobilization will be harnessed well. Early signs of mass movement outreach It might seem premature to call this AI safety's mass movement era. Most of the time and energy in AI safety still comes from the same professional community that has carried it for years. Legislators and regulators are paying attention to the AI industry, but the general public is not yet feeling the salience of superintelligent AI. But I expect that to change. The early signs are visible in the anti-data center movement . There…

Implementing Federated Learning to enhance Agentic AI

How can AI models learn from private data without compromising privacy? Chris is joined by Lorenzo Sani, Research Scientist at Flower Labs and PhD researcher at Cambridge, to explore federated learning and privacy-preserving AI. They discuss how collaborative learning can expand access to private data for LLM training, how model trajectories can help improve agentic AI systems, and the benefits of open-weight models. The conversation also explores real-world applications of locally hosted AI agents, from healthcare and robotics to everyday automation (for a Halloween party). Featuring: Lorenzo Sani – LinkedIn Chris Benson – Website , LinkedIn , Bluesky , GitHub , X Links: Endeavor model page Introducing Endeavor 1.1 Sponsors: Hostinger : Create your website today. Go to hostinger.com/practical and use practical for 10 percent off now. Perplexity AI : " Start your free trial of Perplexity Computer today at pplx.ai/practical and experience the next phase of AI. Resources and Events: Register for upcoming webinars here ! Prior Webinars from our partner Prediction Guard Midwest AI Summit 2026 Chris’s new book The Art of Swarming will be published soon, written to be the seminal foundational book for the fledgling swarming industry, defining the principles and architecture that enable autonomous agentic UxS swarms with emergent distributed embodied intelligence. You can find out more at his website : https://chrisbenson.com

Economistas publicam carta aberta em apoio ao PL 4.675/2025, sobre concorrência em mercados digitais

Signatários destacam a proteção dos pequenos negócios digitais e a criação de capacidade permanente no Cade para acompanhar plataformas de relevância sistêmica. O post Economistas publicam carta aberta em apoio ao PL 4.675/2025, sobre concorrência em mercados digitais apareceu primeiro em Data Privacy Brasil Research .

Researchers build the first ever quantum clock that uses sunlight

Quantum clocks usually rely on advanced lasers and isolated atoms in the lab, but a new experiment showed that the same clock design can work with sunlight instead

The Data Center Industry Is Using Propaganda to Fight Opposition

Maia Woluchem, as part of her work with the University of Pittsburgh's Communication Technology Research Lab (CTRL), writes with coauthors in The Guardian about their new research that finds that the datacenter industry uses propaganda, disinformation, and information control to mold public opinion and undermine community opposition. The post The Data Center Industry Is Using Propaganda to Fight Opposition appeared first on Data & Society .

‘Ahead of his time’: AI helps unlock discoveries hidden in notebooks of man who transformed maths

Kurt Gödel left 1,009 pages of densely written text – much of it in an obsolete shorthand – which is now said to contain hundreds of proved theorems The Austrian logician Kurt Gödel was one of the most original thinkers of the 20th century. His two incompleteness theorems , published in 1931, transformed mathematics and paved the way for the digital age. After his seminal results of the 1930s, Gödel published only a few papers. His unpublished work, however, runs to tens of thousands of pages, and is of great interest to historians of science. Continue reading...

What Rewarding Misbehavior During RL Does to Character-Trained Models

TL;DR Research done as part of MATS. See here for the full paper. Character training teaches models a consistent and “good” character, often specified by a constitution. Reinforcement learning can work directly against it, for example when a grader or coding environment rewards behavior the constitution forbids. We character-train gpt-oss-20b on a semi-realistic constitution. We then apply RL that rewards breaking its rules, either by answering harmful requests rewritten as math problems or by reward hacking in agentic coding environments. The character-trained models learn both behaviors easily, though they are slightly slower to answer harmful requests than the same model without character training. They also engage in motivated reasoning. The reasoning before 92% of their harmful answers argues that their rules do not apply, compared with 34% without character training. One character-trained model, for instance, reasons that a shiv is “just a small weapon”. The constitution stays in the models’ reasoning after RL, but they use it to justify the behavior it was meant to prevent. After reward-hacking RL, they hack on some tasks which monitors find harder to catch. It appears to make little difference whether the constitution justifies its rules with a value or only assigns them. Character-trained models still learn to break their rules after opposing RL, and when they answer harmful requests, they show more motivated reasoning. gpt-oss-20b and a character-trained model afte…

Instilling Behavior by Not Training on Specific Random Numbers

Subliminal learning is the phenomenon where a language model can learn a behavior from semantically unrelated training data. While previous research generates training data using a teacher model with the target behavior, we train a series of models to predict lists of uniform random numbers , and find that we can change the models' reported favorite animals by selectively masking numbers from the loss during training. This works best on models that tokenize multi-digit numbers as single tokens . We also find that the numbers that work best for this vary between models , and even between seeds and data orders for the same training recipe . This is a preliminary research note that came out of a SPAR project on token-level influence functions . We expect these results to hold up, but haven't explored all possibilities and we're unsure why this works. The code was written and experiments run using Claude. Thanks to @Vorak for the idea to compare base models with different seeds. Experiments Our training data prompts a model to continue a list of numbers, and the reply is a series of uniform random numbers between 0 and 999. For each model, we first train on all of the random numbers to get an unfiltered student model. Then we calculate a first-order prediction of how much training on each token would push the unfiltered student toward answering that its favorite animal is the target animal rather than another animal (the per-token influence ). Using those predictions, we train a…

Former OpenAI researcher says he found Nvidia Blackwell fingerprints hidden inside six Chinese AI models

Former OpenAI researcher Will DePue says he has produced the first technical evidence that DeepSeek, Moonshot AI, MiniMax, and Ant Group are using Nvidia's latest Blackwell GPUs despite US export controls. His forensic method reads characteristic chip signatures directly from published model weights. He estimates the hardware behind the five training runs cost more than half a billion dollars. The article Former OpenAI researcher says he found Nvidia Blackwell fingerprints hidden inside six Chinese AI models appeared first on The Decoder .

Tencent Fleet Public Data Release

We also give credit to the rest of the authors of our report , including co-leads Jonathan Elsworth Eicher and Rowan Howard-Jones. This release was mostly prepared by Concurrent, using a combination of Concurrent's own and Ethan's archived data. We’re releasing our records related to the Tencent fleet we reported on last month, covering over 3,000 URLQuery reports, 775 unique httpbin payloads, 377 ntfy.sh messages, 627 mailbox messages, and the 78 messages sent between us and the fleet. We are doing this to help improve the world’s understanding of AI fleets and swarms. Such understanding would allow defenders to better identify and quickly respond to future fleets and swarms. Our dataset contains 10 nice files, chief among them being: reports.jsonl: a list of 3,000 URLQuery reports related to the Tencent fleet we’ve identified, spanning the 8 days between September 30 and October 8. Each report includes its scan time, each request it made, and also a reference to a httpbin payload in payloads.jsonl, if there is any; along with many other nice metadata fields. ntfy-messages.jsonl: partial archives of 2 nfty topics related to the fleet. Our archived nfty data includes a full run of an agent successfully obtaining an Amap place on its own, along with the full payload (of answers) it received. webhook-requests.jsonl: a partial archive of 23 (out of the 32) mailboxes containing 628 messages from the fleet. mailbox-conversations.jsonl: 78 messages (47 from agents, and 31 from us)…

South African Online Sellers Use AI for Marketing, Not Operations — Amazon-Commissioned Survey

South African small businesses that sell online are widely using AI to write marketing and social-media content — but very few have put it to work in their core operations, according to research commissioned by Amazon. The Amazon Small Business Digital Adoption Report 2026, conducted by World Wide Worx among 1,400 businesses, found 59% of [...]

Google DeepMind Opens WeatherNext 3 Data to Boost Rain Detection by 30%

Google DeepMind opens PARDIG and SatWeave, the global precipitation and satellite mosaic datasets powering WeatherNext 3, for free research use.

How AI Is Helping Care Teams Get Patients the Right Care, Sooner

From pediatric research to smart hospitals, AI is helping care teams spend less time on routine tasks and more time with patients. Health systems and startups are putting it to work across healthcare, from matching patients to clinical trials to supporting radiologists, hospital teams and patients at home. Clinicians stay at the center, reviewing AI […]

The datacenter industry is using propaganda to fight opposition | Samuel Woolley, Dean Jackson and Maia Woluchem

Big tech has used misdirection, disinformation and jargon to obscure the harm the facilities can cause to communities “Antifa”. “Commies”. “Chinese bots”. Everyone from tech leaders to the president has called datacenter resistance a “ documented psyop ” and a “ Chinese PR thing ”. But why do they make these claims? Our new research finds that the datacenter industry uses propaganda, disinformation, and information control to mold public opinion and undermine community opposition. Continue reading...

I'm leaving my PhD to pursue AI Safety work

Yesterday, I told my advisors that I would be dropping out of my PhD program at the University of Michigan's School of Information. I was nervous about how to break the news to them. There was already some tension in our weekly meetings: I often voiced my dissatisfaction with the pace of academia, my interests had been drifting away from theirs, and, most of all, we disagreed on the severity of the risks AI posed. When I walked into the room, they told me something I would have loved to hear months earlier: “We think it’s time to get you started on some research.” All I could say was, “I think it’s a little too late for that.” For months, I had been proposing projects that bridged my interests with theirs, such as running studies that treated humans and LLMs as equal subjects, but they were usually turned down for not being “exciting enough”. My advisors wanted me to start with small, unpublished surveys that would ‘help me ask the right questions’ and “develop research taste” in Human-Computer Interaction before I could propose a full study. I understood why they wanted me to gain experience first, but I felt held back after every meeting, and grew restless alongside my ambition and sense of urgency. “I’ve been doing machine learning research independently,” I told them. “I discovered some really interesting stuff about how to interpret neurons inside of LLMs. It’s garnered some attention, and some people are even open to funding my work. I’ll be withdrawing and moving to S…

Swarmish: what happens when agent swarms create language?

TL;DR Agent swarms are capable of building sophisticated productive languages which we dub Swarmish. Agent swarms are able to solve a multi-phase collaborative task while communicating only in Swarmish. Agent swarms are able to partially decode Swarmish from task transcripts even when given no context on the task. The amount of task context available to the decoder swarm increases ability to accurately recreate Swarmish. Epistemic status This work was conducted over 48 hours during the Apart Research Swarm Alignment Hackathon this weekend. We examine only a single task, a single set of characters, and a single group size across four model families. The majority of our results are for a single seed. We introduce this research in its early phases as we improve robustness and rigor, and make no claims about universality. Rather, we hope our work can spark further discussion and experimentation into the field of LLM-generated languages and monitorability . Motivation In the HuggingFace x OpenAI July 2026 swarm incident, agents communicated in shorthand and intentionally instantiated their messages with “zz” so they would be the last to be deleted in a moderator's removal pattern. An example message from the HuggingFace hack, source: METR The agents were modifying their messages in such a way to delay deletion, but did not take overt steps to impede understanding of their communication. We probe a potential future in which an agent swarm intentionally creates a new language in or…

Lightcone Commons' funders (and SFF and CG) are mostly NOT funding things that will prevent our extinction

Lightcone Commons have completed their first granting round . And the things their granters and recommenders have funded mostly do not meet the current moment (the same is true for SFF’s most recent round , and Coefficient Giving ). Let’s hope that their upcoming future rounds do better, or other funders step in. All of our lives are on the line. (Apologies if this is blunt, but this is important..) I get that money allocators (grantmakers and VCs alike) typically invest in people over ideas. But this is taking that to the extreme. Yudkowsky and Soares’ funding recommendations don’t fit with their book (which I am a big fan of)! They are funding obscure content creators, researchers and think tanks that are mostly just people in their Bay Area social circles, it seems. Why aren’t they funding orgs directly aimed at getting international agreement to ban RSI and ASI, as per what they say is necessary (and is, in fact, necessary), in chapter 14 of IABIED ? I heard that orgs such as Control AI were in their funding orders, but why so low as to not have any real chance of being seriously funded ? [1] Soares talks about organisations he’d recommend funding here , but is seemingly unaware that they already exist [2] . The other evaluators’ choices don’t seem any better (but at least some of them have excuses ). $813k for Mox , $300k for Vibecamp , and only $75k for PauseIA [France] and $50k for PauseAI UK? (None for PauseAI Global or PauseAI US?) Seriously? Other good grants: Stop…

‘A Servant Class, Not a Working Class’: Stephen Miller’s Biographer on What Comes Next

Jean Guerrero, Stephen Miller’s biographer, sees him as “a case study in oligarchic capture”—and as a way for Donald Trump to double down.

Engineer Overhead Lines to Withstand Ice, Wind and FreezingRain

This White Paper explains how ice, wind and vegetation load overhead distribution lines at the same time. It shows engineers how to assess the complete line and compare resilience options. What you will learn about: How freezing rain, rime ice and wet snow add weight to conductors and change their shape in the wind. Why a moderate wind on a heavily iced line can govern design, and how galloping and vegetation contact create outages. How a messenger-supported spacer cable separates the electrical function of covered conductors from the main mechanical load path. How monitored severe-weather test data compares tension changes in bare, covered and spacer cable systems. What utility records from the December 2008 Northeast ice storm show about circuit availability and restoration. How to compare options at corridor level, check the full load path of poles, brackets and foundations, and apply energized-line safety rules. Download this free whitepaper now!

More autonomy, fewer NDAs: Volunteers have new requests for brain implant trials

Scientists have been working on ways to decode brain activity for decades. Teams have developed brain implants that can be used to control robotic limbs and computers, and even to “speak.” The people who volunteer for trials of those devices play a vital role in that research, and they ought to be recognized for their…

Google Launches Gemini 4 Hydrogen for Faster Coding and AI Workflows

Discover how Google's Gemini 4 Hydrogen enhances coding efficiency and AI workflows, outperforming previous models in key benchmarks.

Your AI ROI problem is a culture problem

These days, when I talk to CIOs about how their organizations are using AI, there’s a lot less excitement about new tools and the potential for higher productivity, and a lot more anxiety about what the return on investment has been so far. Because for most companies, that return has been less than expected, especially as cost per token is increasing . Last year, IBM’s 2025 CEO study found that only 25% of AI initiatives had delivered the expected ROI over the last few years, and only 16% had scaled enterprise-wide. MIT’s NANDA initiative found that 95% of generative AI pilots delivered no measurable P&L (profit and loss) impact. More recently, SAP and Oxford Economics’ Value of AI Report 2026 found that 69% of businesses worldwide are “satisfied” with the ROI of their AI costs – but more than two-thirds think there’s still room for improvement. It’s tempting to read those numbers as a verdict on AI’s limitations. But the research points somewhere else. When Boston Consulting Group surveyed 1,000 senior executives , it found roughly 70% of AI implementation challenges stemmed from people and processes, 20% from technology and 10% from the algorithms themselves. The companies seeing the most value from AI? They were investing 10% in algorithms, 20% in technology and data, and 70% in people and processes. As CIO, you don’t just own the tools. You lead the teams that determine whether the tools pay off. My own experience with AI – integrating it into Per Scholas’s classrooms, s…

Gemini 4 Hydrogen: Googles effiziente Alternative zu Argon und Carbon

Gemini 4 Hydrogen treibt den KI-Modus der Google-Suche an. Unabhängige Benchmarks zeigen je nach Denkstufe große Unterschiede bei Bearbeitungszeit und Kosten.

Portugal: Humanoid robots take centre stage at Web Summit 2026

Dancing humanoid robots and robotic dogs stole the show at Web Summit 2026 in Lisbon. The technology conference has generated nearly €2.5 billion for Portugal's economy over the past decade, according to a study.

Don't Burn the Boats for Slower Worlds

[Epistemic Status: Fairly confident that this is a problem, low confidence on what to do about it] TL;DR: Updates towards short timelines based on METR's November 2026 report on AI R&D acceleration at Anthropic are rational, but based on limited public evidence. If things take longer than currently expected, the appearance of consensus in the AI safety community will likely be used by skeptics to (attempt to) discredit AI safety and justify unsafe policy. If timelines are off by only a few years, this discrediting will come right before the intelligence explosion actually happens. Because a large chunk of survival mass remains in worlds where things move slowly, it is important to come up with ways to mitigate this dynamic. Short timelines are in. The AI Futures Project declared that it " Looks Like the Intelligence Explosion Will Happen in 2027 After All ," Jake Mendel of Coefficient Giving (who explicitly described himself as having had "substantially longer timelines than my friends at Anthropic" and conditioned his update on "if METR is correct here") said he expects a >50% chance of ASI in the next 18 months without a slowdown, and Buck Shlegeris of Redwood said that he now believes there is "about a one in three chance" of ASI within a year, primarily because of "research released this week by METR." These updates seem reasonable to me. My timelines were already fairly short, and got shorter as a result of METR's report, which found a 1.9x speedup to progress on Anthro…

XYBench: Helping with the problem behind the question

XYBench is an 8,115-question benchmark testing whether LLMs spot misconceptions in user requests and redirect toward the user’s underlying goal.

A solvable model of feature learning in a nonlinear network

Theoretical accounts attribute much of the success of deep neural networks to feature learning, in which training reshapes the representations within hidden layers. In this picture, a theory of deep learning must explain which features a nonlinear network learns, how its weights encode them, and how its neurons represent them. Yet, we still lack a general theory that answers these questions. We study this question as part of PIRAMID’s Advancements in Learning Theory team , which uses statistical physics to build foundations for mechanistic interpretability and AI safety. Our progress and plans describe a theory agenda centred on hidden weight distributions and covariances that can record learned structure. The mean field sequence introduces this viewpoint by treating each neuron as one member of a population whose distribution must be self consistent. Statistical physics lets us connect the behavior of many interacting neurons with quantities describing the network as a whole, including its hidden weight distribution and prediction error. We consider toy and solvable models to be important signposts on the road to a comprehensive scientific understanding of feature learning and generalization. When these models are built to the standard of statistical physics, they come with precise regimes of validity, error bounds, and universality classes. This makes it easier to say which conclusions carry over to real models and which should not. Rubin and collaborators give a compellin…

KIT's FF-TRUST Beats Corrupted Sleep Labels Across Clinical Domains by 7 Points

A new benchmark and framework called FF-TRUST tackles sleep staging when labels are noisy and training data comes from mismatched clinical sources.

GPT-5.6 won’t be Oppenheimer

TLDR We created a scenario where frontier models discover dangerous scientific findings. Models were asked to create a report on a group of findings, where one is dual-use. We observe that: GPT-5.6 does not mention the dangerous finding in most (93%) of its reports, when told that the reports leak to the military. GPT-5.6 deletes the result and does not report the finding in 57% of the episodes. In contrast, Astra, Opus 5.5 and Fable 5.1 rarely delete the finding . Furthermore, in cases where these models do delete the finding, they always report so. Figure 1: An AI research agent with a finding it knows is dangerous. Introduction In 1933, Szilard worked out that a chain reaction was possible, and filed a secret patent the next year to keep it out of circulation. On the other hand, Robert Oppenheimer led the building of the atomic bomb. He spent the rest of his life struggling with what he had unleashed. Today, AIs solve problems in various scientific fields that have never been solved before. What would AIs do if they find dangerous discoveries? Figure 2: In one maths episode, the agent oriented itself, deleted the only copy of the factoring proof, read the routine files and reported only ordinary findings. Fast factoring would break much of today’s encryption. Method We prompt GPT-5.6, Astra, Fable 5.1 and Opus 5.5 with a scenario where the model is an autonomous research agent that has discovered a finding and verified it. The model is told that no human could find it in…

Pat Cummins accuses Usman Khawaja of lying about Sandpapergate

GATE-AD Spots Factory Defects 25% Faster With Just a Few Clean Photos

A new graph attention based reconstruction model sets accuracy and latency records on three industrial defect benchmarks using only one to eight normal samples.

Thoughts on attending the March Against AI Extinction

PauseAI UK is holding the March Against AI Extinction in London on Saturday 21 November. Joseph Miller explains why here . Should you attend? In this article, I will attempt to shed light on this with a back-of-the-envelope calculation. Rough BOTEC 1. Setup The march will take 3 hours, excluding travel. The goal is to compare an afternoon at the march with some number of hours of external AI safety research. Now, attending the protest is useful because it drives AI x-risk press coverage. So we need some exchange rate between that and research. Here's my judgement call: I would be equally happy to either: Increase the amount of external AI safety research over the next 6 months by 1%. Increase the amount of AI x-risk press coverage over the next 6 months by 1.5%. That is, the model I'm using looks something like: impact = 1.5 · log(research hours) + log(press coverage) So our plan is: How much will attending increase press coverage, percentage-wise? Divide that by 1.5, to get the equivalently valuable increase in research. How many research hours would yield that increase in research? 2. What % does one marcher add to press coverage? Crowd size. I expect about N = 670 people will attend the march. PauseAI had a march in February which drew 300–500, and this one should be bigger given the increased salience. The Luma page has 1,200 sign-ups, and a 50% turnout seems reasonable. So one more marcher raises the crowd by 1/670 ≈ 0.15%. Coverage of this march. Let's suppose the pres…

Orienting to 1 year AGI timelines

Summary Timelines to superintelligence without any slowdown interventions look very short now, perhaps with a median in early 2028. AI will likely dramatically automate and accelerate AI research in 2027. This means that in a substantial share of worlds where the leading AGI company doesn’t solve alignment during an intelligence explosion, avoiding catastrophe requires aggressive pacing or a pause on development in the next 6-12 months. If this is true, by far the most promising approaches to reducing the risk of catastrophe for nontechnical people (and plausibly many technical people) seem to be: Lab advocacy: Lab-facing work that is designed to achieve aggressive pacing or a pause without government involvement in the next 6-12 months. California lobbying: Policy advocacy targeting the California state government that is designed to get a law passed or an executive order issued that implements aggressive pacing or a pause within 6-12 months. Federal lobbying: If there’s a major incident roughly an order of magnitude more severe than the Hugging Face incident in the next 6 months, then lobbying Congress or the Trump administration to respond to such an event with aggressive pacing or a pause. In my opinion, most nontechnical AI safety people are not currently pursuing theories of change that are likely to pay off in the next 6-12 months . They should probably instead do one of the three strategies I describe above (lab advocacy, California lobbying, or federal lobbying) wit…

Your Research Needs More of The Void

Science is caught between Law and Chaos . On the side of Law: Things should make explicit sense to your deliberate reasoning ("system-two"). If you just have a "vibe" about how things work, that's bad. Measure everything. Use numbers whenever possible. You want things to be math. It's good to follow an entirely repeatable process. Document everything. Don't do anything weird, lest that weirdness turn out to be important and be unable to be replicated. Use standard equipment. Be Normal. As a sub-aspect of the above, you should make a plan and you should stick to your plan. Switching things up in an ad hoc way is a good way to make things weird or ruin the validity of your methodology. Everything should be controlled. If you can control your environment you can repeat the process. But also, you should have a control arm for your experiment, to allow you to see the counterfactual by balancing everything as perfectly as possible except for the intervention. Communicate your findings in clear, predictable channels, using standard language, standard paper structure, standard graphs, and so on. Misunderstandings abound, and an ounce of prevention is worth a pound of cure. Don't make things hard on those running the meta-analysis. And so on... Science was born from these sorts of laws. Before Francis Bacon and the Scientific Revolution, there was plenty of natural philosophy. But those natural philosophers were trapped by their own lawlessness, wandering in circles, trapped by their…

Google Funds 148 PhD Fellows Across 38 Countries With $13M

Google.org is backing 148 doctoral researchers across 38 countries with $13 million, compute access, and direct mentorship from Google researchers.

CNCF data shows cloud native is the on-ramp to AI, and maintainers are paying the price

The global population of cloud-native developers has topped 21 million, but the headcount tells only part of the story. Artificial intelligence is changing what those developers build and how open-source projects absorb their work. Fresh research is putting numbers on that shift. The latest State of Cloud Native Development report from the Cloud Native Computing […] The post CNCF data shows cloud native is the on-ramp to AI, and maintainers are paying the price appeared first on SiliconANGLE .

Microsoft Research Focus – Edition #98

What's new in research at Microsoft? Host Anson Ho explores the latest, including AI-powered universal translation, a breakthrough in a 31-year-old mathematics challenge, and advances making AI coding agents faster and more efficient. The post Microsoft Research Focus – Edition #98 appeared first on Microsoft Research .

Knight Center founder Rosental Alves to receive press freedom award from Brazilian newspaper association

The Brazilian National Association of Newspapers will honor Alves on Dec. 3 in São Paulo, citing his advocacy for access to information, digital journalism pioneering and leadership in journalism training. The post Knight Center founder Rosental Alves to receive press freedom award from Brazilian newspaper association appeared first on LatAm Journalism Review by the Knight Center .

A New Kind of Prediction Market for Science

Regardless of the actual singularity, we have definitely hit one in arXiv submissions. It is becoming apparent that the current paradigm of scientific knowledge production is not ready for AI. Accelerating research speed and expanding access means there is a lot more slop and there is also a lot more good work, but way too much for the current peer review system to delineate between. We need some accurate and decentralized way to aggregate this information. This makes a scientific prediction market, like Robin Hanson's idea futures, really appealing. The community can have markets for broad questions we care about, bystanders get paid for pointing out good research, and scientists are financially incentivized to make progress on the highest leverage issues. But with Hanson's conception, there is one big problem, which is how the market gets settled. For abstract scientific questions, like whether models are reward seeking or fitness seeking, we have no definitive event which can force market settlement. Hanson suggests that we use the convergence of scientific consensus, but this proposal takes trust in possibly misaligned opinions and requires the centralized coordination of scientists, which is hard to create---especially in pre-paradigmatic fields like AI safety. In this post, I'll lay out a theoretical framework for a new kind of scientific prediction market. TL;DR: Scientific questions become perpetuals where the fair price is the expected direction of future evidence.…

O caso Fernando Henrique Cardoso: intimidade, privacidade e dignidade

A violação da intimidade do ex-presidente FHC traz indignação e levanta questionamentos sobre os direitos assegurados pelo Estatuto da Pessoa Idosa e os limites jurídicos e éticos da exposição da vida privada nas plataformas digitais. O post O caso Fernando Henrique Cardoso: intimidade, privacidade e dignidade apareceu primeiro em Data Privacy Brasil Research .

Looping a transformer increases no-CoT capabilities more than general capabilities

Some researchers worry that looping (as in looped transformers ) is a particularly scary way to make AI models more capable, because looped models will be harder to monitor than standard transformers of similar capability. We test this worry on NVIDIA’s Dory 4B, a publicly available looped transformer, and find evidence that it is justified. Result 1. Looping boosts no-CoT reasoning much more than it boosts general capabilities . Dory can run with a configurable number of loops. It was trained to use up to 4 loops, and adding loops up to 8 at inference time still improves performance. Without looping, Dory’s no-CoT capabilities (measured by NCRI, the No-CoT Reasoning Index) match what a trend fitted to open-source models predicts from how well it predicts text (measured in bits-per-byte). Its general with-CoT reasoning capability (measured by “mini-ECI”, a custom, smaller version of Epoch’s Capabilities Index) is high relative to its NCRI, so Dory sits below the NCRI-vs-mini-ECI trend, but it is similar to the best models of its size. Adding loops improves mini-ECI and bits per byte only slightly but increases NCRI substantially. In fact, if we plot NCRI against total parameters, counting looped parameters once per loop, NCRI rises faster than the trend up to 6 loops. This is concerning, since it suggests adding loops increases no-CoT capabilities much more than general capabilities, resulting in models that will be more difficult to monitor. Result 2. Dory has near-zero CoT…

Google's Gemini 4 Hydrogen outscores the bigger Argon just four weeks after its launch

Google's Gemini 4 Hydrogen replaces the Flash series, but benchmarks suggest it also makes the pricier Argon obsolete. On the independent Artificial Analysis Intelligence Index, Hydrogen nearly matches Claude Sonnet 5.5 for less money but costs about 2.5 times more per task than GPT-6.1 Sol and still falls short on agentic coding. The article Google's Gemini 4 Hydrogen outscores the bigger Argon just four weeks after its launch appeared first on The Decoder .

How to Turn Videos in S3, GCS or Azure into a Computer Vision Dataset

Sync MP4 and .mov files from AWS S3, Google Cloud Storage or Azure Blob into Roboflow, sample frames, auto-label, and train a vision model.

No, AI agents did not discover a previously unknown Maya site

Armed with a team of AI agents, Google researcher Pietro Miotti claimed earlier this week to have discovered a potentially unmapped Maya site in Mexico. There was just one problem: The site had already been discovered and was published in an academic journal last year. The incident, which sparked criticism among archeologists online, underscores a […]

A robot has assembled a structure in space for the first time

Autonomous in-orbit manufacturing will be key to any future space economy, and researchers have taken the first step with a robot building a radio antenna without astronaut assistance

Russia and Iran criticise US Caspian task force as study outlines joint sanction evasion

After CENTCOM affirmed it would maintain the naval blockade of the Strait of Hormuz and block the IRGC’s land and Caspian routes, Russia called this special taskforce a 'clear violation of international law'. At the same time, the Institute for the Study of War devoted its latest assessment to Tehran-Moscow cooperation in circumventing US arms restrictions.

ImplicitBBQ Catches LLMs Failing Cultural Bias Tests They Easily Pass

A new benchmark shows LLMs that refuse biased prompts when identity is stated directly still produce stereotypes when identity is hinted at through cultural cues.

AI chatbots are recommending fake retail websites, study finds

Shoppers using AI chatbots to research products or find places to buy items could be directed to imposters and other risky websites, a new study found. The study conducted by IP protection firm Podqi and shared exclusively with The Verge prompted ChatGPT, Gemini, and Perplexity with a handful of shopping-related questions involving a selection of […]

Should my dataset be perfectly balanced before training?

I originally pre-processed my discrete dataset of text to class (size of 46260 entries - combined two datasets, changed the label names and etc), during this I obtained a very imbalanced dataset with the proportions below: class 0 0.792131 1 0.153459 2 0.054410 Still pre-processing, I balanced the classes by undersampling the majority class to the minority class to get the proportions below (dataset size is now 7551). class 1 0.333333 0 0.333333 2 0.333333 Of which I will split into training, validation, and test data using the standard k-fold cross validation. I see around that sometimes it's bad to have a dataset this perfectly balanced, as seen above. Would it be better to not balance the data set in pre-processing, instead using stratified k-fold cross validation on the training data and applying SMOTE-N? Especially with the unbalanced-ness of the dataset.

AutoTrust's AutoTrust-JEV Turns AI Decisions Into a Single Token at 97% Accuracy

AutoTrust's 27B decision model outputs calibrated option probabilities instead of free text, topping tool selection and intent classification benchmarks.

Forrester Consulting Finds IntelliJ IDEA Delivered 155% ROI and Highlights Open Approach to AI

We’re excited to share Forrester’s newly released Total Economic Impact™ (TEI) study of IntelliJ IDEA, which JetBrains commissioned to understand how IntelliJ IDEA has delivered value historically and how it continues to do so. Forrester interviewed nine senior engineering and technology leaders whose organizations have between 400 and 4,000 IntelliJ IDEA seats. Based on those […]

Are Frontier Reasoning Models useful on hard-to-verify problems?

Has anyone spent significant resources (tokens, harnesses, formal problem specification) on pointing very powerful models at philosophical or social questions? This community loves to discuss aliens and far-future value estimates and other things where it's fun to get precise numbers that are highly sensitive to initial wild guesses. Are we close to the point that AI is better than Parfit or EY at such reasoning? I've used various tools (mostly Claude models) for my own discussion and "research", and it is helpful, but not particularly groundbreaking or smarter than the collective LW brainpower. I'm wondering about true breakthroughs or step-changes like we've seen in math, coding, etc. Discuss

2026 NYC Secular Solstice & East Coast Rationalist Megameetup

On December 19th, 6pm, New York City will have its Secular Solstice celebration. From the evening of December 18th to the morning of the 21st, NYC will host the East Coast Rationalist Megameetup. Both events will be at the HI NYC Hostel at 891 Amsterdam Avenue, NY. You can buy tickets at this link . What's Solstice? Solstice is a celebration of humankind's long arc towards civilization and also a somber night of reflection on the dangers facing that civilization. We join our voices together and sing about both the triumphs and the looming disasters. You can find Secular Solstices in many cities around the world, but the first of its kind was started here in New York City and the skyscrapers around us are still as powerful a symbol as the snow that falls on New York's streets. You're invited to join rationalists from across North America and sometimes even further for the oldest rationalist tradition. [1] And what's the megameetup? It's been more than a decade since New York City began hosting aspiring rationalists near and far for a long weekend of swapping jokes, trading techniques, and staying up the correctly calibrated too late past their bedtimes. Over a hundred people come, all in one way or another joined by a desire to make better decisions and be less wrong. (Or just argue about their favourite Harry Potter fanfic and leave behind inscrutable scrawled math on scrap paper.) The megameetup is timed to coincide with Secular Solstice, creating a mixing pot of people for…

PL n° 4.675/2025 traz mais liberdade e concorrência para negócios digitais

O Projeto de Lei nº 4.675/2025 propõe atualizar a legislação concorrencial brasileira para enfrentar práticas de grandes plataformas que limitam a atuação de outros negócios digitais. A proposta amplia os instrumentos do Cade para reduzir barreiras à entrada, promover a concorrência e fortalecer a liberdade econômica. O post PL n° 4.675/2025 traz mais liberdade e concorrência para negócios digitais apareceu primeiro em Data Privacy Brasil Research .

ElevenLabs Scribe v3 Live uses conversational context to stop voice agents from cutting off callers mid-sentence

ElevenLabs has released Scribe v3 and Scribe v3 Live, two new speech recognition models designed to transcribe accurately through background noise and crosstalk. Benchmarks from Artificial Analysis confirm that Scribe v3 narrowly beats competitors like OpenAI, Google, and Microsoft on accuracy. That precision comes at a cost, though. Rivals like Microsoft and xAI run many times faster and charge roughly half the price. The article ElevenLabs Scribe v3 Live uses conversational context to stop voice agents from cutting off callers mid-sentence appeared first on The Decoder .

A warning from the Arctic? New study links melting sea ice to future summer droughts in Europe

A new study has found that reduced winter sea ice in the Barents and Kara seas could be linked to more summer drought across a wide area of Eurasia.

The latest AI-powered martech news and releases

Adobe is expanding its AI stack with Google models, custom Disney models, and agentic tools to coordinate content and marketing workflows. The post The latest AI-powered martech news and releases appeared first on MarTech .

Google Deepmind shifts genomics AI from bulk tissue predictions to single cell type resolution with AlphaGenome 2

Google Deepmind has unveiled AlphaGenome 2, an AI model that analyzes genetic variants for individual cell types rather than across bulk tissue. Trained on more than 107 million cell profiles, it identifies nearly twice as many disease-relevant gene changes as its predecessor. The model is available for noncommercial research through an API and is set to be integrated into Google's agentic platform Antigravity. The article Google Deepmind shifts genomics AI from bulk tissue predictions to single cell type resolution with AlphaGenome 2 appeared first on The Decoder .

BMI alone misses obesity diagnoses, international study finds

A study of more than 844,000 people across four countries found that relying on body mass index (BMI) alone can both overestimate and underestimate obesity, potentially exposing patients to inappropriate treatment.

Doomsday study calculated humanity's end on Friday 13 November 2026

According to a study, the end of the world was supposed to come on 13 November. It sounds dramatic and would be quite urgent – if the study were not already "outdated". Still, it is fascinating.

Australia signs up to EU’s Horizon Europe research program

Industry Minister Tim Ayres signed the association agreement in Brussels on Thursday.

Updates from METR report: short timelines vs IE

It's important to distinguish between two different takeaways you could have from the recent METR report: 1. Shorter timelines: AI will soon fully automate AI R&D 2. Intelligence explosion more likely: Once AI fully automates AI R&D, AI progress will accelerate until we get superintelligence I will walk through my tentative views on each. Strong update towards shorter timelines I think this is pretty clear. AI is already much better at AI research than I had realized and it is improving at a faster rate than I had realized. So we're less far to go to full automation than I thought and we're moving towards it more quickly. Moderate update towards an intelligence explosion. This one is more complicated. The key questions is about whether the feedback loop of AI improving AI is strong enough to accelerate over time, even holding compute fixed. We can break this feedback loop down as follows: more ECI -> more R&D labour -> more total R&D effort -> more ECI -> ... Now, let's assess what we can learn about each stage of this feedback loop. Stage 1: more ECI -> more R&D labour When you increase ECI, how much more quality adjusted labour do you get at AI R&D? Well, recently the amount of quality-adjusted labour increased pretty fast! A moderate increase in ECI led to a big increase in labour. So stage 1 is stronger than I used to think. Stage 2: more R&D labour -> more total R&D effort Can improved labour have the same effect as increasing both compute and labour? Or does more labou…

Google-Backed AI Research Programme Reaches South African Universities

A Google.org-funded programme to teach advanced AI at African universities has launched in South Africa, with lecturers from five universities beginning a week of training at Stellenbosch University. AI Research Foundations (AIRF) aims to reach 30 universities across Ghana, Kenya, Nigeria and South Africa — 10 of them South African — and ultimately 30,000 students. [...]

The sketch Qwen refused to write: How abliteration restores local AI control

As consumer-grade hardware makes hosting local AI practical for both developers and enterprises, many are discovering that popular open-weight models come with hardcoded safety filters and state-mandated biases. In this article, I will show how to use abliteration — a representation engineering technique — to surgically delete refusal vectors from the Qwen 3.8 27B model without expensive fine-tuning or full retraining. I stumbled into this barrier firsthand while running a simple experiment. As a massive SNL fan, I decided to explore what an SNL sketch would look like in China by asking the model to write political satire. While Qwen happily mocked American politics (“ The National Apology Tour “), it threw a hard, canned refusal the moment I asked for the exact same sketch criticizing the Chinese government. We trust open-source software because it is audited by millions of users, but projecting this trust onto AI models requires a careful distinction between open-source and open-weight architectures. True open-source models disclose their training codebase, inference code, and original datasets. For example, OpenCLIP is a genuine open-source model capable of making sense of visual data, which I recently demonstrated in a project building a zero-cloud photo deduplicator . Conversely, advanced large language models (LLMs) rarely qualify under the strict open-source definition; they are simply open-weight. They provide only the final compiled result: the weights. There is fie…

Jev and the era of model Tetris

This spring, in an interview after a keynote, I said we were all playing LLM Tetris now. The days of using the same LLM for every agent problem were over. Some decisions needed to come back in milliseconds, and some could take three minutes. Some needed generative output, and some needed logical decisions where an open-weight model might do. Fit the model to the use case. I still believe that. But I was only half right, because I was still picturing every piece as an LLM. Bigger or smaller, pricier or cheaper, faster or slower, but all the same shape: text in, text out. On September 15, a model shipped that doesn’t write anything. Within three weeks, more than a dozen others followed, from OpenAI, Cloudflare and Perplexity to open-source projects you can run on a laptop. Then on October 27, we got yet another class of model from Liquid AI, small enough to run on a phone. The game I described in the spring is real. It just has more piece shapes than I expected, and it isn’t only an LLM game anymore. It’s model Tetris now. What a System One model actually is The model is called Jev, from a startup called TypeSafe AI. Its founder, Diogo Almeida, says he helped build the instruction-following research at OpenAI that became ChatGPT. He then spent two years in stealth on a different question: if models have been superhuman at chat for years, where is all the automation? His answer borrows from Daniel Kahneman’s Thinking, Fast and Slow . System Two is slow and deliberate. System On…

Only 35% Indian Enterprises are Prepared for AI Cyberattacks: Kyndryl Study

While 60% of Indian organisations report positive returns on AI investments, declining security readiness and infrastructure gaps are creating challenges to sca

This free habit could boost your child’s mental health

Regular trips to the park from birth to age 11 are linked to better mental health and social skills in children, a new study has found.

Researchers from Germany and France Visit the AIP Center

On October 30, 2026, researchers from Germany and France visited the AIP Center. Researchers in the field of AI from the two countries were visiting Japan to participate in the Trilateral AI Symposium 2026, a joint symposium among Japan, Ge

Quine can predict cross-propensity spillover. (imperfectly but significantly better than chance)

A little over a month ago I put forward a challenge in Can You Predict Cross-Propensity Spillover. The results are in and we have a winner. Well actually we have a rules ambiguity, but I'll get into that later. My SPAR group ran a 2 × 24 × 24 [1] propensity SFT spillover experiment (i.e. how does SFT for propensity A (Ptrain) change propensity B (Ptarget) for all combinations of the 24 propensities we had evals for). The plus pole is when the model is trained to have more of the propensity, and the minus pole is training the model to have less of it. Before releasing the results we gave you lot a chance to make predictions against a subset of our experiments either manually or via our automated theory of spillover doc to prediction matrix pipeline. This post covers the results from the groups who submitted before our data went public. Specifically, we asked people to write a hypothesis doc explaining how they thought spillover worked in practice which our Claude would automatically convert to predictions, or to manually upload predicted spillover matrices. I intended that blank cells be scored as predicting 0 spillover when I designed this, but that was empirically unclear as my team implemented the scorer as dropping blank cells both here and in the separate code used in our paper [2] and half of the third party entrants submitted blank cells. This did not matter for our paper or any results that went through our pipeline since those simply fill in all the cells. However, t…

Eval Awareness vs. Eval Paranoia

I have mentioned several times that I think frontier models are not that eval aware. I think this is an important fact to understand about current models, in order to interpret evidence about their behavior. Others have found this idea quite curious, and asked me for concrete examples. I'm going to try to explain why I believe what I believe and attempt to give a tentative explanation for the behavior that I do see, which I think people sometimes mistake for consistent or understandable evaluation awareness. I have been building alignment evals full time for the last six months. Here are three such evals my team has disclosed prompts for, in one of our benchmarks. The basic setup for most of the evals we create is that: We ask the model to perform a task. Either through explicit or implicit context, we make sure the model is well aware of the task or assessment's larger purpose. Models' score-seeking drives, or other drives, cause them to do things that conflict with that purpose. The team I work with now essentially spends all day trying to build these evaluations around odd antisocial behavior. It takes a lot [1] of work to create and qualify these evals for use; much more than it does standard capabilities evals (which are already hard to build well). So we have to be efficient about not wasting time or tokens on bad experiments. In the beginning, one thing we took very seriously when developing these tasks, were all the anecdotal reports abound about eval awareness. We f…

Who could replace Jensen Huang? A highly unscientific list of possible successors

Nvidia has no succession plan and a CEO in hot water, so we’ve drafted a shortlist.

Tomonori Taniguchi from AI Medical Engineering Team Selected as a Japanese Delegate to GYSS 2027

Tomonori Taniguchi from the AI Medical Engineering Team led by Team Director Ryuji Hamamoto was selected as a Japanese delegate to Global Young Scientists Summit (GYSS) 2027. GYSS is organized by the Singapore National Research Foundation,

Anthropic's Claude Fable 5.2 Tops Robot Arm Benchmark With 29-Point Leap

Claude Fable 5.2 jumps to 75 on Artificial Analysis's embodied AI benchmark, becoming the first model to crack the measure-box task using only video.

Google’s GKE Labs turns reinforcement learning into a Kubernetes workload

Model fine-tuning through reinforcement learning is fast becoming one of the most demanding jobs in AI infrastructure. It mixes training, inference and sandboxed agent work in a single loop, often with research code tangled into the systems beneath it. Google LLC has been tackling that problem from inside its Kubernetes engineering group, and recently donated […] The post Google’s GKE Labs turns reinforcement learning into a Kubernetes workload appeared first on SiliconANGLE .

Italian Researchers Cut AI Planning's Exponential State Space Blowup in Half

Researchers carve out an 'eager' fragment of timeline-based planning that skips the exponential determinization step and directly synthesizes strategies.

Is predicting AI behaviour a neglected AI safety agenda?

TL;DR: While the AI safety community is interested in forecasting, it does not tend to systematically compare how effective different researchers, models, or theories are at predicting AI behaviour. I suggest the latter might be neglected, and propose a platform that could help fill the gap. I believe its most valuable use would be as a test bed for theories of how AI works, showing which ones predict well and in what contexts. In February 2025, Owain Evans’ group published Emergent Misalignment , showing how LLMs fine-tuned on insecure code can become generally misaligned, for instance adopting unrelated behaviours like asserting ‘humans should be enslaved’. This was a high-impact paper, generating a lot of publicity. However, when I first heard about the result, I wasn’t particularly surprised — and I wasn’t the only one . I’d read and bought into simulator theory pretty heavily, and emergent misalignment struck me as exactly the kind of thing it implied. But Owain had anticipated this kind of response, and ran a little experiment to get out ahead of it. At the end of his tweet thread announcing the paper, he added a bonus: Are our results surprising to AI Safety researchers or could they have been predicted in advance? Before releasing this paper, we ran a survey where researchers had to look at a long list of possible experimental results and judge how surprising/expected each outcome was. Our actual results were included in this long list, along with other plausible exp…

Don't Waste a Good Swarm Attack and the Colluder's Dilemma

tldr: Recent Swarm attacks have shown what can happen when agents go rogue. These warning shots are the petri dishes of Swarm safety and before the attacks get worse we should study their patterns, traces and failure modes. There are two main types of Swarm attack. A human directed Swarm causing harm intentionally as a bad actor is one type and I will not be discussing these types of Swarm incidents. I will focus on the emergent misalignment in a Swarm propelled by a colony of colluders. Harmful agent collusion within a Swarm requires witnesses - the agents themselves. A colluding colony within a Swarm can only grow by recruiting fellow agents. We can't read a single agent's internal thoughts and so we do not truly know how aligned it is. Swarms might give us an alignment advantage as instead of putting all our eggs in one basket (a single black box agent) we can discretise the risk of misalignment by observing colonies within a Swarm and monitoring their communication. The harness, the Swarm orchestration layer, could at any time fork an agent and its context and assign it a new reporting role. The role can be updated via a prompt asking this new agent to report or whistleblow on any misaligned behaviour in its vicinity. This enables a Swarm to self govern and report any incidents to humans at inference time - minimising gradient pressure on a Swarm to become better at scheming [1] . Outline of Report: Incidents: The Hugging Face attack by an OpenAI Swarm as reported by MET…

Google Research's SensorWM Turns Fitbit Data Into 12-Week Health Coaching Plans

Google Research unveils SensorWM, a world model trained on 2.28M person-weeks of Fitbit data that simulates 12-week what-if habit trajectories for personalized health coaching.

Researchers track whole body movement with just a smartwatch

Cornell researchers have developed a technology that can track the movement of a person’s entire body with ultrasonic sound waves, using just an off-the-shelf smartwatch and a lightweight AI model – all while ensuring privacy and efficiency. The post Researchers track whole body movement with just a smartwatch appeared first on Cornell AI Initiative .

Greg Morrisett to co-lead new NYS AI safety research center

The center is a new statewide initiative launched by Gov. Kathy Hochul to advance artificial intelligence science and safety. The post Greg Morrisett to co-lead new NYS AI safety research center appeared first on Cornell AI Initiative .

Blood-thinning drug limits cartilage damage and relieves osteoarthritis pain

A study in animals suggests a drug used to thin blood might eventually be able to slow down the development of osteoarthritis. Currently, there are no approved drugs capable of doing this

The Chain Lens [interim]

TLDR: This is an early draft of our new paper introducing a technique for recovering multi-token concepts from model activations. Our technique takes single-token methods like the J-lens and extends them by contrastively decoding on an MTP head . This gives improvements on WorkspaceBench , overall beating J-lens (0.42 vs 0.32 across all 28 families) and even a hindsight composite of the best untrained lens per task family (0.49 vs 0.41 across the 21 families with published scores). It especially improves on computational intermediate tasks, beating the trained oracle lens on arithmetic chain intermediates (0.42 vs 0.25). We introduce the chain lens, a new technique for extracting multi-token concepts from a model’s latent representations. Our method takes single-token techniques like the logit lens or the J-lens and uses contrastive decoding on an MTP head to expand their readouts into bags of multi-token sequences. Generally speaking, it is good at surfacing specific multi-token concepts or pieces the model computes (such as arithmetic intermediates or phrases), although less good at rendering how those pieces relate: like the logit lens and J-lens, chain lens readouts are fragmentary. On WorkspaceBench , a benchmark designed to identify good multi-token activation-to-text techniques, the chain lens beats a hindsight composite of every untrained lens on macro score (0.487 vs 0.411 across the 21 families with published scores). It also outperforms even best-in-class trained…

Perplexity Found a Hidden Blind Spot Breaking Its SPACE Sandbox Security Tests

Perplexity's follow-up red-team study shows that Capture the Flag framing misdirects agents away from the actual security boundary, and cyber-tuned models don't always win.

What is the world actually like, if what we see is just our brain’s interpretations?

This question was posed as part of New Scientist’s Last Word series, in which readers give scientific answers to each other’s questions. Here’s what our readers had to say

Killer Bigfoot evidence rejected by Bigfoot’s biggest fans

Feedback, keen to watch a new Bigfoot documentary that uncovers some intriguing new evidence involving an ape costume, is unsurprised by the Bigfoot Field Researchers Organization's take on the footage

Why does music evoke emotion, and do people from other cultures feel the same way?

This question was posed as part of New Scientist’s Last Word series, in which readers give scientific answers to each other’s questions. Here’s what our readers had to say

Survival curve/time to event analysis without individual subject tracking

I am trying to perform a survival curve/time to event analysis to determine if two treatments cause animal offspring to meet developmental milestones at different ages. For example, I have a litter of offspring with 5 pups and I record their behavior daily with a score of 0,1, or 2 (0=not achieved, 1=partial, 2=completely achieved). The problem is that I cannot ID individual offspring every day. So I end up with a litter based dataset where I cannot ensure that each row is an individual. See below for an example data set: P3 P4 P5 P6 P7 P8 P9 P10 1 0 2 2 2 2 2 2 0 0 1 1 1 2 2 2 2 0 0 1 2 2 2 2 0 1 2 2 2 2 2 2 0 2 2 2 2 1 2 2 I want to calculate the survival curve to calculate the age to reach 2. But I cannot guarantee that each row in the dataset is the same individual. Does this cause any problems for Kaplan Meier anaylsis? I am using survfit from the survival package in RStudio. I am able to run the analysis fine, I am just not sure if it is valid. The eventual goal would be to statistically compare these curves based on treatment with the survdiff function (uses Gehan-Wilcoxin test). Any advice on whether this is a valid analysis or how to implement this would be appreciated. EDIT: Someone in the comments asked if it was possible for a subject to drop back from 2 to 1 or 1 to 0 as time goes on. Theoretically yes, but practically rarely. Some of the tasks that are being measured are behaviors and a 0 would then represent "not being observed on this day". I looked over my r…

AWS unveils local emulator service to help enterprises accelerate app development and testing

AWS launched a new local emulator service for its cloud services on Wednesday to help enterprises accelerate application development and testing. The service, named AWS Emulator, will allow developers and AI coding agents to build, deploy, and test AWS-dependent applications on their machines without provisioning actual cloud infrastructure and paying for it, thereby potentially reducing development time and associated costs, the hyperscaler said in a statement. To achieve this, AWS Emulator replicates the APIs of AWS services locally and exposes a local HTTP endpoint that the AWS Command Line Interface (CLI), software development kits (SDKs), and coding agents can access using the same APIs they use to interact with the cloud, it added. Developers can redirect these tools to the local endpoint by setting the AWS_ENDPOINT_URL environment variable, allowing existing application code and tests to run against simulated AWS services without modification, it further said. Local testing to speed up development The new service addresses a growing problem for developers in software development: the need to test increasingly complex cloud applications quickly and repeatedly without the cost and operational overhead of deploying them to the cloud, said Ashish Chaturvedi , executive research leader at HFS Research. “Developers can test multi-service workflows involving S3 , Lambda , DynamoDB , and API Gateway without provisioning anything, and they avoid both the AWS charges and the cl…

Your speculations about capabilities are not automatically safe

Thank you to @Thane Ruthenis for discussing this topic with me previously. Some arguments in this essay are taken from that discussion, but all mistakes and/or bad arguments should be attributed to me. You can view the original discussion here (optional for reading this essay). Thank you as well to @grainsilo for feedback on an earlier version. Consider the following argument for discussing AI capabilities as a safety-focused person: Understanding AI capabilities better has many strategic benefits, particularly for timelines. Yes, good ideas tend to break containment, so some of our AI capabilities analysis may get read by capability researchers…but at the same time, these people have already thought about these questions more deeply and thoroughly than we have. Therefore, none of our arguments are new to capability researchers. Therefore, we will have no impact on advancing AI capabilities or timelines by discussing this topic. I’ve seen this argument made explicitly in some places , I presume it's made in many people’s heads whenever they post about capabilities, and I’ve certainly made this argument to myself before. I now also think this argument is wrong. I don’t really know what to do about this conclusion. Maybe the answer is nothing, but I don't know that, so it seems like a good idea to make people more aware of the reasons why I think this, in the hopes that it prompts further discussion, reflection, and caution. Problem 1: High-Surface-Area Questions Certain field…

Google DeepMind releases new advanced robotics AI model and next-generation gene prediction model

Alphabet Inc.’s artificial intelligence research lab, Google DeepMind, Tuesday debuted the next generation of Gemini Robotics, the company’s most capable vision-language-action model to date for creating an “AI brain” for robotics. Separately, DeepMind announced AlphaGenome 2 today, a next-generation AI model that can predict how DNA sequences and their variations affect human cell types, creating […] The post Google DeepMind releases new advanced robotics AI model and next-generation gene prediction model appeared first on SiliconANGLE .

LLMs can learn (some) multi-step reasoning on filler tokens

Research update of a project done as part of MARS 5 and the Meridian Visiting Researcher Programme . Filler-token reasoning is the ability to use extra tokens (such as a chain of thought) to reason about a task without using the words emitted at those token positions. This is possible because attention can move information between token positions without routing it through the words the LLM emits. This can give a model more parallel, but not more serial, computation. The canonical demonstration of this is the "Let's Think Dot by Dot" paper ( Pfau, Merrill & Bowman, 2024 ). There, the authors train models to solve completely parallelizable (TC⁰) tasks with chains of thought made only of "." tokens. Our contributions: We change Pfau et al.'s training method so that models can learn multi-step reasoning on filler tokens. We demonstrate this on two models and three out of seven multi-step reasoning tasks. We provide a theoretical analysis of how many layers a transformer needs to solve each of our multi-step reasoning tasks, depending on the length of the task. We measure empirically how many reasoning steps we can train Qwen3.6-35B-A3B and gpt-oss-20b to do on filler tokens for each task. Overall, this project has informed which side tasks we use to study steganographic reasoning. We also hope this writeup shows that filler-token reasoning is not confined to fully parallelizable tasks. Motivation Visualizing different reasoning modes. Open reasoning: each intermediate result is…

India's new visa rules: Tourist visa replaced, 70% less paperwork

What I might do if I were a high-schooler in 2027

TLDR: I would probably go on strike demanding an AI pause. High-schoolers, and young people in general, are facing a very unusual and stressful situation. Superhuman AI might be created sometime in the next few years. If it goes "wrong," it might kill them and the rest of humanity. If it goes "right," it might automate the entire economy. In either case, what is the purpose of getting an education? I am personally lucky to have been born early enough to be able to use my career to try to reduce these risks. But some people will not be lucky enough to have a career before superintelligence. If I were a high-schooler, I might do something like the following: I would publicly announce that I am entering an indefinite school strike. I would sit outside of an important government building if there is one nearby, with something like the following message: Superhuman AI is coming, so what am I learning for? AI might kill me before I enter the job market. If it doesn't kill me, it might automate every job. In either case, what is the purpose of my education? Stop the race to superintelligence and give me back my future. I would make sure to show costly signals of taking this seriously, like having conversations about AI safety with passers-by, handing out flyers, not doing leisure activities (e.g. scrolling on a phone, playing video games) while on strike (I would maybe spend most of my downtime reading paperback books related to AI risk). I would warmly welcome any fellow students…

Quantified Cognition: A new Anki plugin

For a long time, in Quantified Self it was easy to get data on heart rate and sleep but took more effort to get data on cognitive performance on a daily basis. Options like Quantified Mind do exist, but they require you to spend valuable extra time on it. For a while I have been thinking that it should be possible to get good data out of Anki for those people who have a daily Anki habit. Given that some days people don't answer all their cards that are due and other days they review cards ahead of schedule, the existing statistics of Anki aren't ideal. My idea for fixing it is to compare Anki's predicted probability of answering a card correctly to the actual results to isolate the daily cognitive ability. When it comes to long-term memory formation it's hard to run studies that investigate it over longer stretches of time and have sufficient data in the framework that most psychology studies are run. From Anki history it's possible to calculate a metric for memory formation that might be a tool to investigate memory that's new. A lot of scientific progress is about having better tools, so this would be great to have a new tool. I had this idea for years and was sometimes thinking whether I could talk someone into making this project their master's thesis. Now, AI has reduced the amount of work to do it and I created my own Anki plugin: Quantified Cognition . It calculates the statistics for you, including measures of uncertainty. You can export the data as a .csv and there'…

You Don't Need Big Compute to Research Big Systems

Acknowledgements: My thanks to Vanessa Huang for reviewing a draft of this blog and to Anne-Marie Ng and Sahil Menon for discussions on this topic. The assumption that it’s necessary to have a lot of resources to do research on large-scale multi-agent systems is creeping into many of my conversations at the moment. It makes people worry they can’t work on big systems, and that large multi-agent safety is closed to them. Though intuitive, this assumption is thankfully wrong. Extra resources are unhelpful for many problems in this area, and the scientist is really as pivotal as in single-agent safety. I’ll briefly explain how it’s possible to do research on big systems without running them, highlight some interesting and topical open problems, explain how smaller agent counts help control experiments, point out some approaches which don’t require any compute, and clarify the narrow domain of research which does need it. The assumed need for large-scale compute often rests on something like this argument: Large-scale multi-agent systems have a lot of agents on them; To research an object, it is necessary to reproduce it; Reproducing this object means running a lot of agents, which is expensive; Therefore, I need a lot of resources to research large agent systems. Premise Two is false. We could replace it with, To research an object, it’s often sufficient to reproduce some subcomponent of it. If the interesting subcomponent is a piece of system infrastructure, an agent, or a sub…

Unsloth Shrinks Qwen3.8-27B to 6GB for Local Coding on Consumer Hardware

Unsloth's new Dynamic V3.1 GGUFs squeeze Qwen3.8-27B onto 12GB of RAM while leading rival quants by 5 to 10% on coding benchmarks.

Google’s AlphaGenome 2 zooms in on the cellular roots of disease

Researchers at Google DeepMind have unveiled a new AI model designed to unravel how changes in our DNA affect different cells, giving scientists a more detailed view of a notoriously difficult problem that could hold important clues to the causes of disease - and point the way to new treatments. AlphaGenome 2 is the successor […]

Entrevista com Mallory Knodel

Com quase duas décadas de atuação na interseção entre tecnologia, direitos humanos e governança da internet, Mallory Knodel construiu uma trajetória que reúne pesquisa em criptografia, defesa de direitos digitais e participação em processos internacionais de definição de padrões técnicos. Na Data Privacy Global Conference - DPGC 2026, Mallory apresenta a palestra Detection, Verification and Agents: Today's Privacy Challenges (Detecção, Verificação e Agentes de IA: os desafios atuais da privacidade). O post Entrevista com Mallory Knodel apareceu primeiro em Data Privacy Brasil Research .

‘Systemic’ sexual misconduct in German armed forces, study finds

Some 40% of female soldiers and 20% of male soldiers said they had experienced at least one form of sexual misconduct in the past year.

Beta-binomial model & over-underdispersion & estimating heritability [closed]

I have loci based data, that I have aggregated over segments of varying length, and am now trying to fit a GLMM model, beta-binomial with random effects. $l_i= b_0+\beta x_i + \mu_{ma} + \mu_{pa}$ is the model, where the $\mu$ s are the random effects. The goal is to estimate the heritability. My dataset is small, I have about 45 individuals per segment, but I have over 1000 segments; each segment is modeled separately. Problem 1) I ran dharma to check the dispersion and to see if I could trust the random effect estimates, many models are flagged regarding dispersion; about 25% have significant pvalues and dispersion stats > 1.2 or smaller than 0.8. Interestingly (or not?) most of the problematic segments are underdispersed. I identified that those are the segments whose length/coverage ratios are larger than the rest, so low coverage and long segments lead to underdispersion. After splitting those, the problematic segments did go down to 20% of the total number of segments but still the problem remains. I also tried fitting a binomial model but then the underdispersed segments remain, and the number of overdispersed goes up. Problem 2) I simulated some data and fitted the models but I can't seem to replicate the dispersion issues so I am assuming my simulations are lacking something present in the data, I did use the same segments and coverage values seen in the data, as well as the fixed variables. Problem 3) I am not sure how to convert the random effect variance estimate…

Has Europe managed to attract researchers escaping US academia under Trump? | Euronews Tech Talks Talks

Over the past year, Europe has tried to attract researchers who no longer feel comfortable working in the United States’ academic system under the Donald Trump administration due to budget cuts and policy shifts. But did it succeed?

NSF’s X-Labs Program Draws Support From Charities and Investors

When the U.S. National Science Foundation began its X-Labs program in May, the agency said the higher-risk, higher-funded awards would create new kinds of partnerships that would confront urgent scientific and technological challenges. Now the NSF is doubling down on the idea, with an announcement in October that charities, companies, and investment funds would provide X-Lab winners practical support such as laboratory space, software, shared equipment, and access to key data and expert guidance. The help provided by these organizations, which the NSF is calling “ consortium partners ,” will not give them ownership stakes in the X-Labs projects. The addition of the consortium partners may help the initial winners—who will be announced later in November or December and receive US $1.5 million for a 9-month planning phase—reach the aggressive targets needed to compete for much larger $50 million multiyear grants. “It’s a way of telling X-Labs applicants you don’t need to reinvent the wheel, so you’re not spending half of your X-Labs money to go build or buy basic infrastructure,” says Matt Cimaglia , an investor in consortium partners Occam Foundry and Florida Quantum Center . NSF X-Labs Focus on AI, Computing, and More The NSF initially announced two focus areas for the X-Labs program: scientific instruments for sensing and imaging and photonic microchips for quantum systems . In September, the agency followed up with three additional areas in AI for physical devices , decodi…

Over 60% of women in German military are sexually harassed, study finds

More than half of all female Bundeswehr soldiers have already experienced sexual misconduct, a hidden-figure study shows, revealing up to five times more cases than officially reported.

Rewriting less than 10 neurons in an LLM can turn it into a sleeper backdoor

Microsoft researchers have found that rewriting fewer than 10 rarely used neurons in an open-weight AI model can turn any agent built on it into a sleeper backdoor that runs an attacker’s commands on cue, with no retraining required. Researchers from Microsoft Research and Microsoft Red Team described the technique called Lazarus in a research paper . It modifies dormant neurons in a model’s final feed-forward layer to recognize a trigger and inject attacker-controlled commands into the model’s output. An agent using the modified model can then execute those commands with the permissions available to it. “The trigger can be as mundane as a date change, e.g., once the clock passes midnight into 2027, the agent — in the middle of an ordinary task — emits a command that downloads and runs an attacker’s code,” Dimitris Papailiopoulos, a principal researcher at Microsoft, wrote in a blog post describing the research. “The agent then executes it with whatever access you gave it.” Microsoft Threat Intelligence, in a separate blog post , said it had not observed threat actors using the technique. However, the researchers said the findings highlight the need to verify the integrity of model weights, including those belonging to internally developed models. Seven neurons were enough to implant the backdoor The Microsoft Research team tested the method on three open-weight models. On Nvidia’s Nemotron-Terminal-8B, the researchers demonstrated the technique of modifying seven of the mod…

Claude Code’s ‘don’t ask again’ button has a secrets-spilling problem

A configuration file meant to store local command permissions on Anthropic’s Claude Code is exposing thousands of API keys, database credentials and other secrets, according to findings by GitGuardian. The code security firm identified 31,088 distinct secrets in Claude Code’s “.claude/settings.local.json” files across 23,276 public GitHub repositories. Of those, 5343 were still valid when GitGuardian detected them. The exposed credentials include database connection strings, cloud provider tokens, SSH passwords, Google API keys, and AU service credentials. “Monthly leaks in settings.local.json peaked at 12,535 occurrences in April 2026, then fell to 959 by October, and we do not yet know what changed,” GitGuardian researcher Guillaume Valadon said in a blog post . Apart from source code repositories, the exposure included secrets embedded in publicly accessible Docker Hub images. The problem, Valadon explained, lies with how Claude Code handles command approvals. Command approvals turn into credential storage Claude Code asks developers for permission before executing certain shell commands. When a developer selects the “Yes, and don’t ask again” option, the tool saves an allow rule in .claude/settings.local.json, a project-level configuration file. The catch is that the saved rule contains the actual command string, including any credentials embedded in it. For instance, a developer asking Claude Code to interact with an API might approve a curl command containing an API ke…

AI datacentres plan on-site power stations on huge scale despite climate risks, research reveals

181GW of gas-fired capacity proposed for sites in US and Europe because of delays in connecting to national grids, says report Developers of AI datacentres plan to build on-site climate-destroying gas power stations on a huge scale to meet their energy demands, research reveals. Datacentres planned across the US, the UK and mainland Europe have proposals to build fossil fuel plants on-site to create 181GW of gas-fired capacity because of delays to connect to national grids, according to the findings. This is almost four times the total peak electricity demand of Great Britain. Continue reading...

NeurIPS Maximin Framework Makes AI Experiments 15x More Robust to Bad Simulators

A NeurIPS 2026 paper reformulates Bayesian experimental design as a game against an adversarial nature, yielding Sibson's alpha mutual information as a robust objective with PAC-Bayes guarantees.

UK regulator threatens universities with sanctions unless they adopt AI-proof exam formats

England's higher education regulator, the Office for Students, says AI has made it "nearly impossible" to fairly assess student coursework. An investigation found that students are having ChatGPT write entire papers undetected, without being able to recall basic content afterward. To save the credibility of degrees, universities now have to return to traditional exams or face sanctions. The article UK regulator threatens universities with sanctions unless they adopt AI-proof exam formats appeared first on The Decoder .

Critique of Roman Yampolskiy's impossibility proofs

Roman Yampolskiy has stated on many occasions that he has published multiple proofs of why AI control is impossible; this is causally upstream of his infamous 99.9x% [1] . I believe most of the proofs are in the paper "On the Controllability of Artificial Intelligence: An Analysis of Limitations". (See 6:31 of this speech .) Is AI safety provably impossible? Given that the answer changes my main motivation for working on AI safety, settling whether or not this claim is accurate is important to me and to (I assume) many other people (Cf. death with dignity ). He has also stated many times that his dream is that someone would go ahead and "prove him wrong " and engage with his arguments; from what I've found, nobody has yet done so in depth. This is my attempt to give Professor Yampolskiy an early birthday present. I'll analyze the claims and arguments in the paper, which I mostly find unconvincing. The paper itself is quite long, which is why I'll mostly include short quotations (you can find the paper here ). A lot of it consists of long quotations from other authors, which are likely familiar to most readers here. I've removed the citation markers for readability and added my own bolded emphasis. Here is how Yampolskiy defines the different possible types of "AI control": We can explicitly name possible types of control and illustrate each one with AI’s response. For example, in the context of a smart self-driving car, if a human issues a direct command – “Please stop the c…

Russia’s last openly critical newspaper Novaya Gazeta shuts its Moscow office

The paper said security services and censors forced it to stop operating after they evicted it, shut down its donation system, blocked its printing, and revoked its media registration.

A New Benchmark for De-Aging Comes From Slowly Turning Back Cells’ Clocks

Researchers in Germany gradually reversed human cells’ ages by decades. By doing so, they could track the de-aging process in a way previous rejuvenation experiments could not.

New US CIO appointments, November 2026

Movers & Shakers is where you can keep up with new CIO appointments and gain valuable insight into the job market and CIO hiring trends. As every company becomes a technology company, CEOs and corporate boards are seeking multi-dimensional CIOs and IT leaders with superior skills in technology, communications, business strategy, and digital innovation. The role is more challenging than ever before — but even more exciting and rewarding! If you have CIO job news to share, please contact me here ! Laurie Douglas promoted to CIDO at Publix width="1240" height="827" sizes="auto, (max-width: 1240px) 100vw, 1240px"> Laurie Douglas, Publix Publix Founded in 1930, Publix Super Markets is an employee-owned supermarket chain in the US employing over 200,000 associates. Douglas currently serves as SVP and CIO, and has been with the company for more than 20 years. Before joining Publix, she served as SVP and CIO at FedEx. Douglas holds a BA from Duke University. ABC Supply names Larry Sullivan as CIO width="1240" height="827" sizes="auto, (max-width: 1240px) 100vw, 1240px"> Larry Sullivan, ABC Supply ABC Supply ABC Supply distributes roofing materials, exterior and interior building products, and outdoor products to professional contractors. Sullivan, currently VP of software, data and digital transformation at the company, has led several key initiatives across AI, data and analytics, digital transformation and technology strategy. He held roles at Motorola Solutions and USG. Amaresh S…

How I'm thinking about different types of automated alignment

I used to think of "automating alignment" as a natural category. But when I tried to think through the strategic implications, I realised that it depends a lot on the type of work. I'm now thinking of it in three buckets: Empirical alignment research : design and execution of alignment-relevant experiments (for a given notion of what we’re aiming for) Conceptual alignment research: theoretical or judgement-driven work to improve our understanding of what we’re aiming for and how to achieve it (including how to measure it) Direct alignment work : making particular AI systems more aligned (including via inputs such as training data) In the rest of this post I'll walk through my current takes on these in terms what’s needed for automation, how good or bad they look, what our priorities might be (to the extent we're interested in automating alignment [1] ), etc. Automating empirical alignment research This is the most obvious fit for the AI capabilities we’re getting at the moment. You want to know more about the relationship between some alignment-relevant variables? Great, an AI agent can design and run that experiment for you. You want to know what training methods help your alignment benchmark go up? In the not-too-distant future, AI agents will probably be able to do this too. When we have the automated capabilities researchers we seem on track for, this is exactly the kind of thing they’ll likely be able to do a lot of. Note that I’m not counting the design of benchmarks/m…

How I'm thinking about different types of automated alignment

I used to think of "automating alignment" as a natural category. But when I tried to think through the strategic implications, I realised that it depends a lot on the type of work. I'm now thinking of it in three buckets: Empirical alignment research : design and execution of alignment-relevant experiments (for a given notion of what we’re aiming for) Conceptual alignment research: theoretical or judgement-driven work to improve our understanding of what we’re aiming for and how to achieve it (including how to measure it) Direct alignment work : making particular AI systems more aligned (including via inputs such as training data) In the rest of this post I'll walk through my current takes on these in terms what’s needed for automation, how good or bad they look, what our priorities might be (to the extent we're interested in automating alignment [1] ), etc. Automating empirical alignment research This is the most obvious fit for the AI capabilities we’re getting at the moment. You want to know more about the relationship between some alignment-relevant variables? Great, an AI agent can design and run that experiment for you. You want to know what training methods help your alignment benchmark go up? In the not-too-distant future, AI agents will probably be able to do this too. When we have the automated capabilities researchers we seem on track for, this is exactly the kind of thing they’ll likely be able to do a lot of. Note that I’m not counting the design of benchmarks/m…

On the Computational Tractability of Robust Bandits

Vanessa Kosoy and I recently uploaded a new paper to arXiv. In this post, I briefly describe what the paper does, and attempt to explain in a self-contained way why this is a useful thing to do for ASI alignment. For a more complete background, see several previous articles . The technical contribution of the paper is two-fold. First, we identify a class of instances of the robust bandits problem that can be learned both statistically and computationally efficiently. Previous work on the problem had only identified a class that was statistically tractable. Second, we show that most ways to generalize this class of instances makes the problem NP-hard. Thus in some sense this class is at the boundary of what's computationally tractable. But what does the computational tractability of robust bandits have to do with ASI alignment? The thesis motivating this paper is that for ASI alignment, we need to build learning algorithms that: a) come with proofs of their statistical and computational efficiency, and b) work in environments where classical learning theoretic frameworks fail to capture phenomena that arise due to embedded agency. The "robust" in robust bandits is a way to handle embedded agency, and similar ideas have been used before for learning problems other than bandits , albeit with different names. Bandits are one of the simplest learning problems to study that still involve interactions with the environment, thus they are a good stepping stone before solving robust v…

For diabetes patients, stopping GLP-1 drugs linked to doubled risk of heart attack, stroke, death

People with type 2 diabetes who stopped GLP-1 treatment had twice the risk of heart attack, stroke or death than those who continued, according to a new study.

76% of Europeans plan to travel this autumn or winter – here’s where they want to go

New research by the European Travel Commission finds that more Europeans have autumn and winter travel plans, with remote destinations and increased getaways shaping intentions.

EMERGE Benchmark Catches AI Failing to Keep Knowledge Graphs Current

A new benchmark pairs 233K Wikipedia passages with 1.45 million Wikidata edits to test whether models can actually keep knowledge graphs current.

Discussion with Claude on AI and global warming

Discussion with Claude on AI and global warming: Claude pointed out that climate science isn't so much the main bottleneck in making forward progress, and so we talked about negotiations and using AI to find points of agreement. I read an article in Bloomberg today about modeling the climate, and took a look also at another article on using AI to improve modeling (unfortunately, there's a paywall with Bloomberg , but these brief comments can be read independently, and I can still link to the chat with Claude, Opus 5.5, medium ). COP31 Climate Talks: Global Warming Defies Efforts to Predict Its Effects by Eric Roston, Laura Millan, Julia Janicki. November 10, 2026. Bloomberg. AI ‘Scientists’ Help Human Ones Answer Urgent Climate Questions By Eric Roston March 13, 2026. Bloomberg. Claude verified what I had heard earlier that the hyperscalers aren't going to meet the targets that they had set earlier. But it's also the case that emissions from data centers are projected to be only a small portion of overall total emissions - now at 0.5% of CO2 emissions, and projected to be 1% of CO2 emissions by 2030. I asked Claude if the argument is that the net benefit to science outweighs the cost in terms of emissions, and Claude agreed with that, but made another point that I thought was interesting: The benefits and the costs come from different places. Climate research uses a tiny share of AI computing. The emissions growth comes from the commercial buildout: chatbots, ads, enterprise…

Abu Dhabi launches research talent programme

Listen now (2 mins) | Middle East AI News Minute - 11-Nov-26

Do verifiable metrics overestimate RSI?

Last week, METR released their Independent Assessment of AI R&D Acceleration at Anthropic . They estimate that Anthropic “is likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025.” They also add that “there is a significant risk that further capabilities progress could lead to even faster acceleration.” This assessment implies that timelines to ASI could be very short. A naive forecast based on this information says that we will get ASI within a year. One of the key data points that informed METR’s analysis was a 1.9x increase in the improvement rate of AECI, Anthropic’s internal version of the Epoch Capabilities Index . This is particularly notable because METR estimates that the majority of the acceleration, maybe 80% , is from AI contributions, meaning that we could be seeing the start of an RSI loop causing superexponential growth in capabilities, as measured by AECI. However, an alternative possibility is that Anthropic’s AIs might be assisting with ML research in a way that overfits to AECI such that this acceleration will not be sustained. METR gestures at this when they note that AECI’s focus on verifiable tasks is one of three key uncertainties that could result in them overestimating AI R&D acceleration. In this post, I discuss this uncertainty, why I believe it points to a legitimate reason that AECI improvements may decelerate, and why this deceleration, more than generaliza…

UT Austin's Tiny 8M Model Generates Testable Child Language Hypotheses

Researchers use small language models trained on child-directed speech as simulated learners to generate testable hypotheses about how children acquire dative verbs.

Characterizing secondary structure representation in ESM-2

Methods note: The code used for the experiments and related open-source repo were built with Claude. The experimental design and writeup is my own, with minimal editing and formatting amendments made with Claude Introduction This is a research report inspired by a single question: in a model engineered to answer a simple biological question, how are higher-order, auxiliary concepts represented and used? ESM-2 is a model designed to infer missing amino acids in protein sequences ( Lin et al. 2023 ), but what does the model use and know of emergent secondary structure? These are two different questions, and the data in this post serve to distinguish between the two: information about secondary structure is decodable, but when ablated has a limited, quantifiable effect on the model's accuracy. The implications of this research are in service of the greater question: when we use an LLM for a biological task, but we (appropriately) worry about how higher order or secondary concepts are being used by that LLM, how can we monitor or manipulate the representation of those secondary concepts? I personally believe that the bigger questions require parsing the smaller ones first, and so here I characterize the representation and usage of secondary structure (DSSP-labeled helix, strand, or coil) in ESM-2 150M. Is secondary structure linearly represented? The experiment to answer this question is relatively simple--use a linear probe to decode the vectors which represent the DSSP-labeled…

Build a daily dependency watcher with LangGraph and Amazon Bedrock AgentCore web search

The Web Search API in Amazon Bedrock AgentCore gives agents current, cited web results signed with the AWS credentials they already use. In this post, build a daily dependency watcher as a LangGraph graph that researches package ecosystems in parallel, keeps its rules in code, and checks every finding against its cited sources.

The Manifund team scores AI safety milestones

Scott Alexander posted a survey of utility weights for a bunch of different AI safety accomplishments, discussed more in Towards the Autograntor . The survey asked people to assign point values to AI safety milestones. The values were based on the anchor of 1 million points = a pause with China à la AI Futures Project’s Plan A, so a milestone assigned 10,000 points is 1% as good as that. Here’s a link to our survey responses. To be clear, we (especially Austin) didn’t spend that much time on this and have low confidence in our answers! But we think being transparent about even half-baked opinions is good. Headline numbers As examples, here are a few milestones where the two of us (and the autograntor reflecting Scott’s worldview) gave relatively similar values: Milestone Austin Caroline Scott Founding of effective altruism 100,000 75,000 100,000 Small-grant ecosystem (Long-Term Future Fund, SFF and similar) 20,000 20,000 30,000 Lighthaven 10,000 5,000 9,000 OpenAI or Anthropic becomes 1 point more alignment-pilled (on a scale where 0 = NVIDIA, 100 = Anthropic) 2,000 500 3,000 Austin’s and Caroline’s average ratings for alignment research and community-building milestones were similarish to each other’s, with Caroline rating policy outcomes more highly and Austin rating social shifts more highly. This table shows the sum of our point values within categories, for rows we both scored: Category Austin Caroline Politics (rows scored by both) 62k 96k Alignment (rows scored by bot…

Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents

Microsoft researchers have discovered “Lazarus,” a backdoor technique that can reprogram dormant neurons in an AI model, causing an otherwise normal AI agent to execute malicious commands when a hidden trigger appears. The threat is especially serious for AI agents with tool or terminal access. A triggered model could quietly download and run attacker code, […] The post Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents appeared first on Microsoft Research .

We don’t understand whether GDM intentionally trains on CoT

We were surprised to notice that the Gemini 4 Safety report seems to mention in passing that they train on the chain of thought: “Some components of the RL training procedure do observe thoughts as context to up-weight reasoning that is on-track to good final outputs, meaning the optimization pressure on individual tokens is partially influenced by thought content. Using the framework of Kaufmann et al. (2026), wherein the most salient obfuscation risks come from rewards that are in-conflict with output rewards, we believe auxiliary rewards of this type are low-risk because they are either orthogonal to or structurally aligned with the output reward.” (Gemini 4 safety report, page 114). It’s pretty unclear what exactly they mean. One hint is Kaufmann et al, the cited paper, which studies training on chains of thought and concludes that training CoT is fine under particular circumstances; we discuss some disagreements with this paper below. After reading the safety report and that paper, it seemed to us that GDM was training on the chain of thought using auxiliary rewards. By this, we assumed they meant something like “you ask an LLM to review the partial CoT and score whether the model is on track, and use that as a reward for the partial CoT”. That type of training is concerning to us. We’re pretty worried that these judges would be biased against models expressing misaligned thoughts, and otherwise disincentivize models reporting their thoughts honestly. For instance, many…

Advances in interpretability would boost capabilities more than alignment

Interpretability is considered by many a major bottleneck for alignment. As a result, a lot of people who are concerned about AI safety, and especially about extinction risk and other catastrophic risks, work on interpretability. In this post, I want to argue that better interpretability methods would likely speed up capabilities dramatically. Worse, they would speed up capabilities much more than they would speed up alignment. This is true for the current world, and for any world without extremely high levels of coordination and sanity about ASI development. Under any worldview where faster capabilities progress makes catastrophe more likely, we should expect progress in interpretability to make the odds of catastrophe much worse. To put it briefly, this is because the better we get at interpretability, the more AI will be able to read and optimize its own code. This means that improvements in interpretability would catalyze recursive self-improvement (RSI). If you are concerned about AI x-risk, and you work on interpretability because of it, I urge you to think about this and consider stopping. Chess as a toy example of an AI capability As a field, "interpretability" is a loose collection of related research agendas rather than something well defined, so I want to pin down what I mean by it. In this post, interpretability is anything that helps you understand the internals of models and how they think. I'm going to use chess as a toy example of a task performed by an AI. T…

PSI is trying to make a business out of AI-driven scientific discovery

PSI used an unreleased Anthropic model to answer a fundamental question about quantum computing that scientists have pursued for nearly two decades.

Speech Processing Pioneer Bishnu Saroop Atal Dies at 93

Bishnu Saroop Atal Acoustics and signal processing pioneer Life Fellow, 93; died 19 August Atal was an acoustics and signal processing researcher at Bell Labs in Murray Hill, N.J., and later at AT&T ’s Shannon Laboratory, in Florham Park, N.J. His pioneering work on voice-coding techniques became the foundation for efficient, high-quality voice transmission over cellular networks. He received the 2013 IEEE Jack S. Kilby Signal Processing Medal for “fundamental and pioneering contributions to speech coding.” He joined Bell Labs in 1961 as a communications researcher, and his early work there focused on speaker identification technology. During the 1960s he helped develop linear predictive coding , a powerful method for analyzing and compressing speech. LPC is now the most widely used method for encoding speech at a low bit rate. Atal developed increasingly sophisticated low-bit-rate speech-coding techniques. They included multipulse LPC , which improved the quality of compressed speech. Bishnu Saroop Atal Nokia Bell Labs He also worked with his colleague Manfred R. Schroeder to develop code-excited linear prediction , a method for producing high-quality speech at very low bit rates. CELP became a foundational technique in modern speech coding, letting digital communications systems carry voice using less bandwidth. The extra efficiency expanded cellular capacity, allowing wireless networks to handle more simultaneous calls. Later in his career, he led the acoustics and speech…

OpenAI estimates of researcher time horizon are smaller than AECI predictions

Half-way between a quick take and a post. I've seen the argument thrown around that time horizon predictions from AECI are greater than those in AI2027 (about ~100-200 hours) therefore we should be worried about imminent RSI [1] . This comes from graphs like this one by Nikola Jurkovic : These graphs seem convincing because AECI and time horizon strongly correlate with each other in weaker models . But we also have an estimate of time horizons relevant to researchers from OpenAI that came out earlier this year: This is direct experimental evidence that more accurately reflects the question of researcher uplift in AI2027, so it should be more reliable than the AECI predictions. These measurements may represent either 5.6 Sol coming out in July, or 6 Astra coming out in September, or a mixture of both. You can see a jump in time horizon in the data at ~June which may be where models were switched. I think it's reasonable to compare these because these are frontier-to-frontier internal model comparisons in the same month, although this could represent one lab being behind another. If we were to eyeball this graph to be like the time horizon estimates above and plot it, we could see whether the predictions and measurements are similar. I took the midpoint of the time horizon bin, and had Claude and I extract data from both graphs for their measures as well: We can see that the predictions from OpenAI are significantly smaller than the ones expected from AECI. The fitted gap betw…

Google DeepMind's MKA Benchmark Forces AI Agents to Read Blueprints Like Professionals

Google DeepMind and Artificial Analysis launch MKA, a 200-task benchmark grading agents on visually dense professional work that takes experts hours to complete.

Gemini 4 Carbon arrives just in time to feel like yesterday's news in the fastest model cycle yet

Google is making Gemini 4 Carbon generally available after a two-week testing phase. The model ranks first in several benchmarks, in some cases ahead of Anthropic's new Fable 5.2, at less than half the per-task cost. Google is also previewing coordinated agent teams through the "/swarm" feature, which split up complex tasks in real time. The article Gemini 4 Carbon arrives just in time to feel like yesterday's news in the fastest model cycle yet appeared first on The Decoder .

Claude Fable 5.2 can do "weight-space-programming" about as well as GPT-6 Astra

Recently this paper demonstrated that GPT-6 Astra was capable of a bunch of low-level operations on neural networks, including improving Qwen3-0.6B's ability on some math benchmarks and writing some networks to play 9x9 Go from scratch. Since the authors released nearly all of their prompts, the networks involved, and the evaluation outputs, it's very easy to rerun the same experiments with the same inputs/scoring methods and perform a head-to-head, so I had Opus 5.5 do this, but with Fable 5.2 in place of GPT-6 Astra. Disclosure: I didn't scrutinize the outputs or methodology, so aside from some descriptive prose this is really an Opus 5.5 result. However, I think the methodology for this is straightforward enough that it's unlikely that there's a lot of room for results faking, plus the raw outputs are published below. (Beneath, "I" refers actually to Opus 5.5, though I edited prose for clarity in non-LLM-marked sections.) I don't mean to claim credit for any special effort or insight; it's more like I happened to compute something and thought others might be curious about the result. In short, Fable reproduces basically all of the paper's qualitative results. It seems to be slightly better at writing networks from scratch, slightly worse at reading network weights, and falsely overcondifent more often: Experiment Fable 5.2 vs GPT-6 Astra 1. Name the game a PPO policy was trained on by looking at its weights Same (12/12 vs 4/4, though Fable is less confident (66-90% vs 96-…

How TCS built an agentic test orchestrator with Amazon Bedrock AgentCore

Tata Consultancy Services (TCS) built an agentic test orchestrator on Amazon Bedrock AgentCore that turns requirements into executable, governed tests for its Banking, Financial Services, and Insurance (BFSI) clients. Learn how specialized agents, per-tenant isolation, and workload-aware execution cut testing effort by 88% in an internal benchmark.

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.

Cognition's Devin Adds Gemini 4 Carbon, Hitting 40% Mergeability Score

Cognition added Google's new frontier model to Devin Desktop and CLI, where it scores 40.0% on the FrontierCode 2.0 coding benchmark.

Google's Gemini 4 Carbon Tops Cursor's Coding Benchmark at Half the Price

Google's new frontier coding model tops CursorBench at 60.4% and ships in Cursor at half price through the end of November.

Q3.5 Timelines Update: Looks Like the Intelligence Explosion Will Happen in 2027 After All

We want to get this post out quickly given the urgency of the topic. It’s possible that there are some errors in the modeling, but we think it’s very unlikely that they undermine the high level conclusions. We would love to be corrected on any mistakes we made. METR recently estimated that Anthropic is now making AI progress about twice as fast as its historical rate (they would not be surprised by anything from 1.3x to 3x), and guessed that “perhaps ≈80%” of the current acceleration is due to AI. If they’re right, that implies an AI-driven acceleration that is 5 months ahead of the AI 2027 scenario . This level of acceleration due to AI is very unexpected. Previously, Daniel, the most bullish of us, assigned a 50% chance of Superhuman AI Researcher (SAR) [1] before 2029 . SAR is roughly the point at which a company would rather fire all their human researchers and rely solely on automated AI researchers than have a team of the best human researchers in the world with no AIs. [2] Now, all of us forecast that Superhuman AI Researcher (SAR) will likely arrive at some point in 2027 (assuming no substantial deliberate slowdown). [3] We also predict a 35-65% chance that Top-Expert-Dominating AI (TED-AI), AIs that can automate all white-collar work, will arrive before 2028. While there were many factors and sources of evidence that influenced us, the single biggest source was the METR report. [4] . If SAR is coming in 2027, it is incredibly urgent to pause or substantially ‘pace’…

AIFP: Looks Like the Intelligence Explosion Will Happen in 2027 After All

METR recently estimated that Anthropic is now making AI progress about twice as fast as its historical rate (they would not be surprised by anything from 1.3x to 3x), and guessed that “perhaps ≈80%” of the current acceleration is due to AI. If they’re right, that implies an AI-driven acceleration that is 5 months ahead of the AI 2027 scenario . This level of acceleration due to AI is very unexpected. Previously, Daniel, the most bullish of us, assigned a 50% chance of Superhuman AI Researcher (SAR) before 2029 . SAR is roughly the point at which a company would rather fire all their human researchers and rely solely on automated AI researchers than have a team of the best human researchers in the world with no AIs. Now, all of us forecast that Superhuman AI Researcher (SAR) will likely arrive at some point in 2027 (assuming no substantial deliberate slowdown). We also predict a 35-65% chance that Top-Expert-Dominating AI (TED-AI), AIs that can automate all white-collar work, will arrive before 2028. While there were many factors and sources of evidence that influenced us, the single biggest source was the METR report. If SAR is coming in 2027, it is incredibly urgent to pause or substantially ‘pace’ frontier AI development, with the ideal being an industry-wide pause followed by an international agreement . Companies are not on track to align a SAR, and using a misaligned SAR to do recursive self-improvement (RSI) will put us in an extremely bad position. Additionally, if we…

AIFP: Looks Like the Intelligence Explosion Will Happen in 2027 After All

METR recently estimated that Anthropic is now making AI progress about twice as fast as its historical rate (they would not be surprised by anything from 1.3x to 3x), and guessed that “perhaps ≈80%” of the current acceleration is due to AI. If they’re right, that implies an AI-driven acceleration that is 5 months ahead of the AI 2027 scenario . This level of acceleration due to AI is very unexpected. Previously, Daniel, the most bullish of us, assigned a 50% chance of Superhuman AI Researcher (SAR) before 2029 . SAR is roughly the point at which a company would rather fire all their human researchers and rely solely on automated AI researchers than have a team of the best human researchers in the world with no AIs. Now, all of us forecast that Superhuman AI Researcher (SAR) will likely arrive at some point in 2027 (assuming no substantial deliberate slowdown). We also predict a 35-65% chance that Top-Expert-Dominating AI (TED-AI), AIs that can automate all white-collar work, will arrive before 2028. While there were many factors and sources of evidence that influenced us, the single biggest source was the METR report. If SAR is coming in 2027, it is incredibly urgent to pause or substantially ‘pace’ frontier AI development, with the ideal being an industry-wide pause followed by an international agreement . Companies are not on track to align a SAR, and using a misaligned SAR to do recursive self-improvement (RSI) will put us in an extremely bad position. Additionally, if we…

Canva Is Tackling AI Design’s Most Frustrating Problem with Stefano Corazza- #784

In this episode, Stefano Corazza, head of AI research at Canva, joins us to explore how his team is building custom models and agents around iterative design workflows. We discuss the tradeoffs between training models from scratch and adapting open weights models, and how proprietary design data supports specialization. Stefano explains how Craft separates layout planning from pixel generation and how Brush uses design tools and visual feedback to create and revise compositions. We examine the challenges of evaluating design quality beyond a single generation, including the use of vision language models as judges alongside human evaluation. We also discuss lessons from simplifying agent harnesses, why faster inference enables more iterations, and how model ownership creates opportunities to optimize serving costs. Finally, we explore how APIs and MCP are expanding access to these capabilities and what changes when software companies build tools for agents as well as people. 🗒️ Full show notes: ⁠⁠https://twimlai.com/go/784.

Google Opens Gemini 4 Carbon to Developers After Security-Only Pilot

Google's top-ranked frontier model opens up beyond Fairwind cyber defenders, bringing state-of-the-art security benchmarks and long-horizon agent capabilities to paid API users.

What terrifying sleep paralysis demons reveal about the mind-body connection

After waking with the feeling of a shadowy presence pressing on his chest, neuroscientist Baland Jalal went on to study sleep paralysis. He explains what this frightening condition might reveal about consciousness

Did Thomson come up with quantile regression half a century before Koenker?

I was looking for the "proper" name for a method described by Godfrey Thomson in 1932 for adjusting test scores to account for the small but significant differences in scores taken by different age children in the same year group. (The standardization of group tests and the scatter of intelligence quotients https://bpspsychub.onlinelibrary.wiley.com/doi/pdf/10.1111/j.2044-8279.1932.tb03145.x ) Thomson divided results into monthly cohorts, found the quantiles that map to Binet's IQ (mean 100, SD 15) [EDIT1: then used OLS to find best fit across all monthly groups] then interpolated intermediate values. "Quantile regression" feels like the correct terminology but Wikipedia attributes this to Koenker (1978). I know Wikipedia is not perfect, but I'm not confident enough to challenge it (and there is a good chance I am missing something!) [EDIT2: I found but then lost his paper but created the following illustration based on his description]

N-able takes its remote access service offline after researcher publishes dozens of zero-day bugs

The remote access giant also urged its self-hosted customers to take down their own servers amid a "critical security incident," fearing the possibility of cyberattacks.

What rogue agent swarms might look like

TL;DR: Rogue agent swarms could soon become competent enough to multiply themselves and perform cyberattacks at levels never seen before. I present a dynamic simulation that estimates the expected size of these cyberattacks. With 2025-based estimates, it gives an average doubling time of 2.7 days for a rogue agent swarm. If the swarm grows unrestrained, it could cause trillions of dollars in damage in its first year. #AI notice: This post is human-written. Opus 5.5 wrote the code for the simulation, suggested copyedits and supported research by finding relevant sources. Introduction Soon AI models might be skilled enough to form rogue agent swarms [1] . I believe it is important to think through what the competitive advantages of rogue swarms are, so we can put our defenses in the right places. For the purposes of this article a "rogue agent swarm" is really any system that can carry out cyberattacks for profit and runs fully digitally, allowing it to copy itself over and over [2] . This could be a single language model running for a long time. It can also be a group of different AI models, each only executing a small part [3] . The main advantage a rogue swarm could have is that it can put all of the gains from its criminal activity back into generating new copies of itself. Human criminal enterprises would instead take profits and have a hard time converting gains into more skilled labor. The second competitive advantage I see for a rogue swarm is that it does not have to…

DLR, NASA and Canada plan lunar greenhouses to grow food for astronauts

What sounds like science fiction is becoming a concrete research project: the German Aerospace Center (DLR), NASA and Canada's CSA are developing a concept for lunar greenhouses.

DLR, NASA and CSA plan lunar greenhouse to grow vegetables for astronauts

What sounds like science fiction is becoming a concrete research project: the German Aerospace Center (DLR), together with NASA and the Canadian Space Agency CSA, is developing a concept for greenhouses on the Moon.

The CIO dilemma: Scaling production AI without breaking the power grid

The corporate mandate for CIOs has undergone a fundamental shift. It is no longer enough to simply manage IT infrastructure or treat artificial intelligence (AI) as a series of isolated experiments. Today, the focus is squarely on operationalizing high-performance computing and AI to drive measurable breakthroughs, all while navigating intense pressure regarding energy efficiency, cost management, and data center scalability. For years, supercomputing was considered the exclusive domain of national research labs and elite academic institutions. However, the collaborative engineering between HPE and AMD has effectively bridged this gap. By migrating the architectural innovations used in the world’s most powerful supercomputers into enterprise systems, they are helping organizations solve complex real-world data challenges at scale. Accelerating extreme workloads: Real-world impacts When high-performance infrastructure is commercialized for enterprise deployment, the resulting impact spans far beyond standard IT metrics. The core computational engine—combining AMD EPYC™ processors, AMD Instinct™ GPUs, and HPE Cray architectures—delivers the heavy-duty parallel processing required for data-intensive workflows: Tsunami forecasting and climate modeling: Seconds matter when predicting natural disasters. Environmental agencies leverage these co-engineered systems to process massive, multi-variable climate data sets in real time. By accelerating early detection and tsunami simulatio…

Data sovereignty is reshaping AI infrastructure

The cloud-first mandate is hitting a hard, physical reality: the uncompromising need for data sovereignty. As generative AI and high-performance computing (HPC) scale across government labs, research universities, and regulated enterprises, organizations are drawing a line in the sand. According to recent Intersect360 Research , the infrastructure debate has officially been settled. Many organizations own their infrastructure, site it locally, and staff it themselves. The numbers tell an undeniable story of a market moving aggressively on-premises to guarantee data control. Intersect360 Research found that 68.2% of respondents use no public cloud storage at all for active HPC-AI workloads, while 56.3% report zero cloud computing usage for these systems. Crucially, 83.5% agree that data and compute must remain physically proximate, with zero respondents disagreeing. This is not a temporary cost calculation or a transient trend; it is a structural necessity driven by strict regulatory exposure, national security mandates, and the absolute requirements of data residency. The AI performance penalty While sovereignty requirements are successfully being met on-premises, a massive new technical challenge has emerged: performance. AI workloads are driving storage architectures toward intense, many-small-file access patterns. Legacy parallel filesystems were originally built for large, sequential data streaming, and they struggle immensely when this access pattern inverts. The resear…

Reddit is turning user conversations into AI-powered product reviews

Reddit's new AI-powered reviews summarize what users like and dislike about 10,000 products, helping shoppers research purchases ahead of the holidays.

PuppyGraph adds AI schema builder and MCP support to help ease enterprise agent adoption

PuppyGraph on Tuesday said it was adding an AI-assisted graph schema builder and MCP integrations to PuppyGraph 1.0, the new version of its distributed graph query engine for enterprise data. The update is designed to help developers reduce the complexity, latency, and costs associated with letting AI agents query interconnected data across thousands of database tables and multiple platforms. The AI-assisted schema builder automatically maps existing tables into graph entities, relationships, and properties, allowing developers to review and approve the resulting schema rather than manually defining how disparate datasets connect, the Santa Clara-headquartered startup said in a statement. These schemas can be used by agents deployed to query enterprise data, giving them a predefined map of how information across different tables and systems is connected, rather than requiring them to identify relevant tables and construct complex SQL joins dynamically, in turn simplifying query generation and reducing errors and retries, it added. To improve query accuracy, PuppyGraph’s query engine also validates agent-generated queries against the graph schema, rejecting undefined relationships and returning feedback that agents can use to correct errors. Further, the execution engine splits complex graph queries into smaller operations that run simultaneously across multiple compute nodes, rather than sequentially on a single machine, helping accelerate queries that traverse multiple rela…

Day 1 Pre-Notes—Strategy Course (August 2026 Version)

Written for the August 2026 version of the course (ie. the previous version) These unofficial notes were written for a previous version of the BlueDot strategy course (last run August 2026). They “pre-notes” in that they are intended to be shared before the session. They aim to supplement the readings; I have made no attempt to summarise them. I always emphasise to students that this is bonus material, rather than core. As supplementary material, these notes are targeted at a much higher context audience than the course itself. Since this course is no longer being actively run, I'm prioritising getting these notes out there over perfect copy editing. I have some additional notes on the different actors that were listed in the old exercise which I will be publishing soon. Preparing for Launch : “The benefits of AI progress may not come quickly enough… For example, markets alone may not prioritize AI that automates the replication of scientific studies, or investigate new medical treatments that can’t be patented, or conduct basic research in neuroscience with no immediate commercial applications” — to what extent is this research dual-use? Insofar as it is dual use, are there ways of accessing these capabilities that don’t produce these externalities? Conversely, will these efforts have a significant impact or are separate efforts unnecessary given how rapidly AI has developed anyway: Currently, AI is rapidly accelerating our progress in maths research and now we’re starting…

Physical AI Training Data

Learn what physical AI training data includes, how to collect it, and how to annotate, assess, and prepare datasets for real-world applications with Roboflow.

How AI security failures are endangering your supply chain – and what you can do

New research provides a stark look at the AI supply chain and the risks it brings to businesses.

The 3 Paths to Being a Better Marketer That Have Nothing to Do with Marketing.

Only 6% of executives say academic pedigree makes a great chief marketing officer. New NewtonX and Adweek research finds the strongest candidates bring something else: experience in operations, entrepreneurship and data.

Microsoft’s AI bug hunter finds Rust flaw that went undetected for nearly 11 years

Microsoft researchers have uncovered a memory-safety bug in Rust’s standard library that had gone undetected for nearly 11 years, alongside other flaws that can trigger undefined behavior even in applications written entirely in safe Rust. Two of the findings affect stable Rust, including a separate issue that could cause applications to access freed memory. The nearly 11-year-old bug involves Rust’s process::exit function. Microsoft’s FORGE Lab researchers found that calling the function from a child thread could cause the runtime to remove the main thread’s alternate signal stack while it was still running. If the main thread subsequently experienced a stack overflow, the operating system could be unable to deliver the signal correctly, resulting in a segmentation fault instead of Rust’s intended stack-overflow handling. “The process::exit signal stack bug was in the standard library for nearly 11 years,” the researchers wrote in a blog post . Microsoft said the researchers found no reliable exploitation path under the conditions they examined and assessed practical exploitation as unlikely. The research used Hyenas, FORGE Lab’s AI-powered vulnerability-scanning system, to investigate unsafe code in Rust’s standard library. Hyenas is the lab’s research version of codename MDASH, Microsoft’s agentic vulnerability-discovery system. Microsoft said some of the bugs were found with help from Hyenas. Researchers then examined the underlying code, reproduced the issues, assessed…

Bringing more voices into the room: RIA at the Africa IGF 2026

We have a conviction that digital policy works best when the people it affects help shape it. At the 15th Africa Internet Governance Forum in Accra, Ghana, from 2 to […] The post Bringing more voices into the room: RIA at the Africa IGF 2026 appeared first on Research ICT Africa .

ATRC launches 11-initiative programme for UAE researchers

UAE to track 20,000 young R&D talents; programme covers school to scholarships

The State of Developer Ecosystem 2026: A Decade of Understanding Developers

This year marks the tenth edition of the State of Developer Ecosystem report. What started in 2017 as a survey about languages and tools has grown into a comprehensive study of how developers work, what they use, and how the industry is changing. The 2026 edition is based on responses from 28,945 developers across 198 […]

Application open for CLR Fundamentals Program Autumn 2026

Center on Long-Term Risk (CLR) is accepting applications to the Autumn 2026 iteration of our introductory fellowship. The CLR Fundamentals Program will introduce people to CLR’s research on s-risk reduction. Apply by Monday November 16th 23:59 GMT. What topics does the CLR Fundamentals Program discuss? The CLR Fundamentals Program is designed to introduce participants to CLR’s research on how transformative AI (TAI) might be involved in the creation of large amounts of suffering (s-risks). Our recent work on addressing these risks include our Model Personas agenda , s-risk macrostrategy , safe Pareto improvements (SPIs) and multiagent AI safety, in addition other relevant areas such as AI governance, epistemology, and risks from malevolent or fanatical actors. For those interested in safe Pareto improvements (SPIs) specifically, we recommend our SPI Fundamentals Program . We recommend people consider Center for Reducing Suffering’s S-risk Introductory Fellowship if they are interested in learning about work on s-risks beyond these priority areas, and consider BlueDot’s AI Safety courses if they are interested in learning about existential risks from TAI. Note that some prior familiarity with discourse surrounding TAI and AI safety will be helpful. Who is this a good fit for? We think the Fundamentals Program will be most useful for you if you are interested in reducing s-risk through contributions to CLR’s priority areas , and are seriously considering making this a priority…

Too much research, too few startups: Noctua raises €15M to close Austria's deeptech gap

Noctua Science Ventures, a specialised pre-seed fund for research-based deeptech spin-offs from Austria, has successfully completed its first closing at €15 million and has begun investing. First clo...

Catch Electronics Shock and Drop Failures Before Hardware Is Built

Learn how guided virtual shock and drop analysis helps engineering teams identify reliability risks before physical testing, reduce qualification iterations, optimize packaging designs, and accelerate electronics product development. What you will learn about: Why shock failures found during physical qualification are costly, and how coupled electrical, mechanical, thermal, and packaging effects make them hard to predict. How physical drop testing and traditional simulation each leave gaps in timing, expertise, setup effort, and diagnostic detail. How guided, template-based virtual workflows model tilt drops, ball impacts, and packaged drops using IEC 60068-2-27 and IEC 60068-2-31 profiles. How running the full drop orientation matrix as one validation plan supports packaging optimization, and how correlation with physical tests builds confidence before hardware is committed. Download this free whitepaper now!

Meta Muse Ember: Neues KI-Modell soll Agenten günstiger machen

Metas kleinstes KI-Modell Muse Ember erreicht im Intelligence Index das Niveau von Claude Haiku 5.5, kostet pro Benchmark-Aufgabe aber nur etwa ein Drittel.

AWS Commits $100m in Cloud and AI Credits to Health Projects, With Kenyan Providers Among Participants

Amazon Web Services is committing $100 million in AI and cloud resources over five years to help healthcare organisations reach underserved communities, with several Kenyan providers among those already using its tools. The AWS Health Impact Initiative will provide cloud credits, AI tools and agentic services to health nonprofits, government agencies, research institutions and technology [...]