Wall Street Journal: As “nudify” tools proliferate, unleashing a new form of bullying among kids, parents and schools are struggling to protect young victims from AI deepfakes  —  As ‘nudify’ tools proliferate online, parents and schools are struggling to protect young victims

Discussion | Link

Discussion | Link

New York Times: A profile of UC Berkeley professor Hany Farid, the world's leading digital forensics expert for 20+ years, who says he is now struggling to identify AI fakes  —  The emails began to arrive on a Sunday morning, as the worst ones often did.  Hany Farid opened the first message at his home …

Cathy Chan / Bloomberg: Sources: Alibaba is in talks to buy Chinese fresh grocery delivery platform Pupu for $1.5B, in a bid to better compete with food delivery rivals like Meituan  —  Alibaba Group Holding Ltd. is offering $1.5 billion to buy Chinese grocery delivery firm Pupu, initiating a bidding war as part …

Dominic-Madori Davis / TechCrunch: Madrid-based Orbio, which uses AI agents to help businesses interview, onboard, and manage frontline workers, raised a $21M Series A led by Dawn Capital  —  After Sergi Bastardas' decade at Amazon and floriculture startup Colvin, one thing always stood out — the feeling that there wasn't enough efficient …

Emily Shugerman / The San Francisco Standard: Inside the Signal group chat where Sergey Brin, Marc Andreessen, Garry Tan, and other tech elites toss out strategies to oppose a proposed California wealth tax  —  The billionaires of California had a problem.  —  It was the fall of 2025, and a proposed ballot initiative to tax 5% …

Microsoft CEO Satya Nadella compared AI's impact to the problems globalization first caused. George Chan/Getty Images Satya Nadella warned of AI models absorbing corporate knowledge, risking industry autonomy. Nadella compared AI's impact to the problems globalization caused. Big Tech leaders like Snowflake's Sridhar Ramaswamy have echoed these concerns about AI's dominance. AI models are hoovering up corporate knowledge, and that's leaving one big loser, says Satya Nadella. In an article posted on X on Sunday, the Microsoft CEO warned of a future in which a handful of AI providers capture most economic value while industries lose ownership of their knowledge. "The last thing any of us want is a world where every company across every sector is ceding value to a few models that eat everything they see," Nadella wrote. "There is no societal permission for an AI future that hollows out entire industries." Nadella compared the AI era to globalization, warning against repeating that dynamic. "Think about what happened in the first phase of globalization, where entire industrial economies were hollowed out by outsourcing," he wrote. "The GDP numbers looked fine on the surface, but the displacement was real and the consequences are still being felt." Instead, he advocated for a broad AI ecosystem in which companies keep control of their learning systems, which he said would enable innovation and retain employee expertise. Nadella's post echoed concerns other Big Tech CEOs have been raising this year. In a February podcast, Snowflake CEO Sridhar Ramaswamy said that the biggest software companies are at risk of being reduced to mere data sources. "The big model makers want to create a world in which all of the data for all of the enterprises is easily available to them," Ramaswamy said. "Everything else, the world, is just a dumb data pipe that feeds into that big brain." Ramaswamy added that Snowflake needs to operate with a "fear" that people would stop using AI agents developed by software companies and instead want an all-inclusive agent that has data from Snowflake and everywhere else. In a January LinkedIn post, Box CEO Aaron Levie said that AI models can perform high-level knowledge work across nearly every profession, from law to strategy and scientific research. "The question that we will have to wrestle with is, in a world where everyone has access to the same expert intelligence, how does a company differentiate?" Levie wrote. He said that context would be the answer. Read the original article on Business Insider

Minh Connors / Getty Images On June 12, artificial intelligence (AI) lab Anthropic suspended access to its latest Claude models, Fable 5 and Mythos 5, which had been released three days earlier. The move came in response to an “export control directive” from the US government prohibiting use of the models by anyone who is not a US national. Mythos is Anthropic’s most powerful, or “frontier”, model. When first announcing the model in April, the company said it was too good at hacking to release immediately. Instead, Mythos was made available to a handful of organisations (mostly US tech corporations) to use to patch weaknesses in essential digital systems. Fable is the same basic model, but with added safeguards meant to stop it being used for cybersecurity purposes. This is what was released to the public last week – and almost immediately shut down. Anthropic and the Trump administration at loggerheads Since early 2025, Anthropic and the Trump administration have been in escalating conflict. The administration has accused Anthropic of making “woke AI” and called chief executive Dario Amodei an “ideological lunatic”. Early disagreements concerned AI regulation and semiconductor export policy. The dispute sharpened when Anthropic declined to let the Pentagon use its models for domestic surveillance and fully autonomous weapons systems. The Department of Defense responded by threatening to designate Anthropic a “supply chain risk”, a classification that would have required military contractors to sever ties. Jailbreaks The US government has not yet publicly stated the reason for last week’s directive, but Anthropic it says it believes the government became aware of a jailbreak: a method for circumventing the safeguards in Fable that prevent using its most powerful features for nefarious purposes. These safeguards classify user requests as safe or unsafe before passing them to the AI model. When triggered, the safeguards redirect the request to a less powerful model. The government’s concern, according to Anthropic, was that the safeguards could be bypassed to extract information useful for cyberattacks. Guardrails for large language models aren’t bulletproof. They mostly depend on the model’s own capacity to interpret the user’s intentions in making a request. Beyond the inherent difficulty of this task, a large online community (which my colleagues and I call the Undersphere) is working hard to circumvent AI guardrails. Anthropic acknowledges that “perfect jailbreak resistance is not achievable for any current model provider”. Anthropic says the research behind the government directive appears to have been produced by engineers at Amazon, which is both a rival to Anthropic and a significant investor. But this was not the only relevant jailbreak. Within 48 hours of Fable’s release, a researcher using the pseudonym “Pliny the Liberator” published what they identified as Fable 5’s full system prompt to X and GitHub repository. The system prompt is a hidden set of instructions that helps determine an AI model’s behaviour. It’s unclear exactly how knowledge of Fable’s system prompt could be used in practice, but it has drawn attention in the Undersphere. A surprise – and an ongoing mystery The deepest problem of making large language models such as Fable secure is that we don’t fully know how they work. According to Oxford University economist and machine learning expert Maximilian Kasy, they work much better than they “should”. Large language models have billions of internal parameters and are trained on unimaginably vast piles of data using machine learning methods. According to Kasy, we would expect such systems to be “overfitted”: good at reproducing patterns in their training data, but bad at generalising to new situations. However, modern systems such as Claude and ChatGPT do seem to be able to generalise. Kasy likens modern AI development to alchemy: successful through trial and error, not yet grounded in systematic theory. As a result, the behaviour of AI models is partly opaque even to their builders. Hard to regulate The opacity of the technology is one key reason it’s so hard to regulate. Governments lack independent access to the data, infrastructure and expertise they would need to evaluate proprietary frontier models. The US administration’s recent executive order on AI security, published two weeks ago, reflects this realisation. As the administration has realised the power of frontier AI models, it has moved from an initial hands-off posture to asking developers to share their models for review before release. That demand is an implicit admission that the administration does not trust the companies to evaluate, fully and comprehensively, what their own models can do and how they might be misused. The public sees even less, and the consequence is measurable: a survey taken across 25 countries last year found people are, on balance, more than twice as concerned about AI as they are excited about it. The future of AI safety AI is a hugely hyped technology. But there is no doubt it is also extremely powerful and unpredictable. Understandably, this combination is very dangerous. We cannot rely on regulations, as technology will develop more quickly than they can adapt. Nor can we rely on guardrails, as they will be bypassed. We need a governance framework built for that eventuality: one that can predict and address the consequences of failure. Such a framework must be global, participatory, and founded on reciprocal trust. These are things the current US administration has shown little capacity to generate. Francesco Bailo has received funding from Meta (2019) and from Australia's Department of Defence (2023).

Shania Twain says menopause helped her be more confident about her body. Christopher Polk/Dick Clark Productions via Getty Images Shania Twain said she once became so unhappy with her appearance that she avoided mirrors. However, menopause helped her overcome her body-image struggles. The experience taught her that there are "some things you cannot control," the singer said. Shania Twain, 60, said menopause helped her be more confident about her body. Speaking to The Times in an interview published on Saturday, Twain said she struggled with her body image during her 2019 Las Vegas residency. She recalled tearing two thigh muscles and having to be carried off stage after becoming "malnourished" in her efforts to "be thinner." "In menopause you lose control of your body," Twain told The Times. "So all of a sudden I'm bloating and I'm definitely not in control. I can't just lose five pounds." The "That Don't Impress Me Much" singer said she wasn't able to look at her own reflection before going on stage. "I stopped looking at myself in the mirror. I hated my body. I'm, like, 'Oh, I cannot stand this changing body.' But that was so unhealthy. Who cannot look at themselves in the mirror?" she said. Driven by her insecurities, Twain turned to intense workouts and cut out fats and sugars from her diet. "I was doing very unhealthy things," she said. "And I was working my body more than I was feeding it, to keep up with the strain." Her thigh injury worsened, and as her recovery dragged on, she began rethinking her relationship with her health. "Now I'm, like, bring on the mirrors, I'm going to look at myself all day long!" Twain said. "Menopause has been very good for me because I've learnt that some things you cannot control." Twain has previously said that menopause helped her put her insecurities into perspective. "Menopause taught me to quickly say, 'You know, it may only get worse. So just love yourself now. Just get over your insecurities — they're standing in your way. And fear is standing in your way,'" she told The New York Post in 2023. Other Hollywood stars have also shared their experiences with menopause. Katherine Heigl told Business Insider in July that navigating perimenopause while parenting teenagers gave her and her kids unexpected common ground. "Their hormones are all over the place, they're trying to adapt and get used to their new bodies and their new ways of thinking, and I am too. So the only thing I can do is look at them and go, "Well, at least we're in it together," Heigl said. In January, Gabrielle Union said menopause initially made her feel like her "value as a woman" was "diminishing." Over time, however, she came to see the transition in a different light. "It challenges you to look at life differently — and not as an ending, but as a beginning," Union said. Read the original article on Business Insider

Orbio announces $21 Million Series A in round led by Dawn Capital.

Discussion | Link

arXiv:2606.13756v1 Announce Type: new Abstract: This paper presents a comprehensive overview of the QIAS 2026 shared task, organized as part of the OSACT7 Workshop and co-located with LREC 2026. The shared task was designed to evaluate the ability of large language models to perform complex reasoning in the religious and legal domain of Islamic inheritance. Unlike conventional question-answering benchmarks, QIAS 2026 focuses on end-to-end reasoning from natural language cases, requiring systems to perform the full inheritance calculation process, from identifying the eligible heirs to assigning the correct share to each beneficiary. To support this evaluation, the task was based on the MAWARITH benchmark, a dataset of $12{,}500$ Arabic inheritance cases annotated with intermediate reasoning steps and final answers. System submissions were evaluated using MIR-E, a multi-step metric that measures performance across the main stages of inheritance reasoning. A total of $16$ teams participated in the shared task, investigating a range of approaches, including prompting-based methods, retrieval-augmented generation, and fine-tuning strategies. The results show that Islamic inheritance remains a highly challenging benchmark for current language models, especially in stages that require precise legal interpretation and structured numerical reasoning. This overview summarizes the task design, dataset, evaluation framework, participating systems, and main results.

arXiv:2606.13835v1 Announce Type: new Abstract: LLM-based generative agents are increasingly used in urban simulators, yet it remains unclear whether they reproduce empirically realistic human mobility patterns or merely generate plausible mobility narratives. We introduce a validation framework for evaluating the mobility of generative agents of LLM-based urban simulators against real-world mobility data. For this, we use mobility laws, temporal rhythms, network motifs, semantic activity transitions, and behavioral mobility profiles. Using datasets from the Greater Paris region and Shanghai, we evaluate AgentSociety and CitySim across multiple dimensions of mobility realism. Our analysis reveals a substantial gap between narrative plausibility and empirical mobility realism. Although the simulators capture some high-level semantic activity distributions, they struggle to reproduce core spatial and temporal constraints, including realistic trip-length distributions, origin-destination flows, dwell times, and transition dynamics. We further observe that realistic mobility diversity is unstable across default prompting configurations and may require explicit profile-aware initialization. To support reproducible evaluation, we also contribute scalable and open LLM-driven infrastructure for regional-scale map generation, observability-enhanced simulation, mobility-metric computation, and traffic simulation. Our findings highlight the need for rigorous empirical validation of LLM-based urban simulators and provide practical tools for building more realistic and reproducible urban simulation systems.

arXiv:2606.13852v1 Announce Type: new Abstract: Neural topic models enable scalable semantic discovery, but their integration with quantum hardware remains largely unexplored. We present a proof-of-concept hybrid classical-quantum variational autoencoder (VAE) for topic modeling, embedding parameterized quantum circuits within the VAE inference network while retaining a classical topic-word decoder. To address the resource constraints of quantum hardware, we propose a modified Gaussian Softmax posterior that decouples latent space dimensionality from the number of topics to be extracted, enabling the model to operate with a low-resource 10-qubit quantum device. On the AgNews dataset, the hybrid VAE outperforms state-of-the-art neural topic models (NTMs), reaching a $C_v$ coherence score of 0.71 and an NPMI score of 0.20 while preserving high topic diversity. For comparison, we also construct a fully classical variant, which also outperforms state-of-the-art models on AgNews and exhibits clear class separation in the latent space. These results demonstrate that hybrid VAEs are computationally viable even on NISQ-era devices and represent a promising direction for quantum-enhanced topic modeling.

arXiv:2606.13904v1 Announce Type: new Abstract: Exploratory question answering (EQA) over data lakes requires an LLM agent to discover relevant sources, analyze retrieved data, and adapt its actions based on intermediate results. End-to-end accuracy alone cannot distinguish failures in search, planning, data analysis, or the agent's Action Policy: its decisions about what to do next and when to submit an answer. We present SANA (Search Agent Navigation Ablation framework), a diagnostic ablation framework that transforms EQA tasks into runtime profiles containing gold source sequence, sanitized subquestions, and execution records. SANA uses these profiles to construct idealized search, planning, and data-analysis tools, allowing each component to be ablated; the residual gap is diagnostic evidence for policy failures. To illustrate SANA as a reusable evaluation framework, we adapted two recent EQA benchmarks, LakeQA and KramaBench, and evaluated lightweight and mid-sized agents under fixed prompts, budgets, data lakes, and runtimes. Across both benchmarks, data analysis is a consistent bottleneck while planning is less so. Search is a major limitation in LakeQA's large data-lake setting, but less so for the smaller-scale KramaBench. SANA thus deconstructs end-to-end task accuracies into a diagnosis of where data-lake agents fail, and allows for systematic comparisons of progress in search, planning, data analysis, and agent design.

arXiv:2606.13767v1 Announce Type: new Abstract: Low-rank adaptation (LoRA) and its variants provide a memory- and compute-efficient alternative to full fine-tuning of pre-trained models. However, questions remain about the comparative generalizability of these approaches and how the structural restrictions on low-rank updates preserve effective adaptation performance. We present a historical framing, covering the past (full fine-tuning and original LoRA), the present (different variants of LoRA), and propose simpler, cheaper, parameter-efficient extensions by inducing sparsity within existing LoRA variants: Cheap LoRA (cLA), training a single low-rank factor with the other fixed (deterministically or, in its randomized variant, stochastically), and the chained circulant variant, ${c}^3$LA. We frame cLA as a structured instance of asymmetric LoRA, serving as a controlled column-subspace restriction of full fine-tuning. We derive information-theoretic generalization error bounds for these variants, marking one of the first endeavors in this area. Empirically, we evaluate 11 fine-tuning methods across 10 pre-trained models and 14 datasets, analyzing the fine-tuned models' performance and generalization using tools such as loss landscapes and spectral analysis. Despite the sensitivity of fine-tuned models to the pre-trained model, datasets, and other factors, our study suggests that restricting LoRA-based PEFT methods' adaptation to a sparse, structured column space remains competitive across tasks with their parameter-matched baselines while reducing up to 10% training time and peak GPU memory up to 15%, even with a na\"ive, non-optimized, sparse implementation. Our theoretical and empirical generalization measures provide a more consistent and principled approach to their cost-effective adaptation than commonly used analytical tools. Overview and code are available at: https://elicaden.github.io/Beyond_LoRA/.

arXiv:2606.13753v1 Announce Type: new Abstract: Grokking is the delayed onset of generalization in neural networks, arising long after they fit the training data. Whether the weight norm causes this delay is disputed: some studies report a critical norm at the transition, others observe grokking with no fixed norm at all. We settle this by intervening on the norm during training rather than only observing it. Under free training with weight decay, networks grok when the weight norm reaches a value Wc that varies little across seeds and learning rates (CV 1 to 2 percent) and grows with the modular base as a power law. When we instead clamp the norm to a fixed multiple rho of Wc and hold it there, the network still groks, but the delay follows T_grok proportional to exp(alpha rho). One exponent, alpha near 7.5, fits this delay across four moduli (R^2 = 0.996). Over the swept ranges the held norm moves the delay by about 19x and the learning rate by only about 2x, and holding the norm above Wc slows grokking rather than preventing it. A final LayerNorm removes the dependence by decoupling weight scale from the network function; without it the exponential law returns. This pinned-norm delay is the exponential counterpart to the logarithmic delay predicted for a freely contracting norm.

arXiv:2606.13705v1 Announce Type: new Abstract: Yes. Can it cure doom loops? Probably not. The Gemma 4 instruction-tuned models share a reproducible failure: on long factual enumeration prompts, such as listing every episode of a TV series, the 88 IAU constellations, or the 151 original Pokemon, they collapse into repetition, either a tight verbatim loop or a list whose entries decay onto a single answer. These loops occur at rates as high as 95% and survive prompt rewording, inference-engine changes, and most sampling adjustments. In this paper we explore whether this behavior is localized enough to remove by weight edits. To localize the cause, we use per-layer ablation and per-neuron attribution, then confirm the strongest candidates with full-generation sweeps. The loops trace to a small set of MLP neurons (or, in the 26B-A4B Mixture-of-Experts model, a few routed experts) which we suppress with static weight edits. These "surgeries" can be as small as a single sign-inverted neuron (in the E2B model). The size of the effective edits grows with model scale, but in all cases, the loop patterns can be addressed at normal generation budgets while preserving general-purpose benchmark scores. However, the edits do not solve everything: we also study longer thinking budgets, where the two larger models most visibly enter doom looping, i.e. a non-convergent regime in which the model self-corrects in circles over a fact it cannot recall, exhausting the budget without committing to a final answer. We show this residual failure is reduced but not eliminated by the same edits, and argue it is fundamentally a knowledge-precision problem rather than a removable circuit; weight surgery can delete a loop, but it cannot supply a missing fact. Our results are both a feasibility demonstration, that is, evidence that a concrete generation pathology can be localized to a few parameters and edited out, and a delineation of where that approach stops.

arXiv:2606.13685v1 Announce Type: new Abstract: LLM-as-a-Judge is now widely used to rank model outputs, train reward models, and populate public leaderboards, but its run-to-run reliability remains under-characterized. We study repeated identical evaluations on 29 tasks spanning 10 categories using two OpenAI judge models (GPT-4o-mini and GPT-4.1-mini), with 50 pairwise trials and 50 pointwise trials per question, supplemented by temperature and prompt-sensitivity ablations. Across judges, pairwise preferences flip on average 13.6% of the time, with 28% of questions exceeding a 20% flip rate and one question reaching 56%. GPT-4o-mini also exhibits a significant first-position bias (72% A-majority, p = 0.024). At the same time, mean pointwise score gaps are small (0.19--0.36 on a 10-point scale) and not statistically significant in aggregate, producing a pairwise--pointwise gap: judges frequently choose a winner even when their own scalar scores provide little evidence of a meaningful quality difference. Beyond within-judge instability, cross-judge agreement is only 76% ($\kappa = 0.51$), semantically equivalent prompt templates change majority outcomes in 25% of tested cases, and deterministic decoding reduces but does not eliminate inconsistency. A reliability curve analysis shows that, in our dataset, 11 repeated trials are needed for a majority vote to recover the 50-trial reference verdict with 95% probability on average, rising to 15 for high-variance questions. These findings suggest that single-trial LLM judging is often too noisy for high-stakes evaluation, and that multi-trial aggregation, position randomization, and explicit uncertainty reporting should be standard practice. Because both judges are from a single provider, cross-provider replication remains an important next step.