On January 20, DeepSeek, a relatively unknown AI research lab from China, released an open source model that's quickly become the talk of the town in Silicon Valley. According to a paper authored by the company, DeepSeek-R1 beats the industry's leading models like OpenAI o1 on several math and reasoning benchmarks. In fact, on many metrics that matter--capability, cost, openness--DeepSeek is giving Western AI giants a run for their money. US export controls have severely curtailed the ability of Chinese tech firms to compete on AI in the Western way--that is, infinitely scaling up by buying more chips and training for a longer period of time. As a result, most Chinese companies have focused on downstream applications rather than building their own models.

Weaknesses: * The proposed approach is heuristic. The main downside of heuristic defenses (unlike certified defenses) is that they are often easily broken by an adaptive attacker. With this in mind a proper evaluation of a heuristic defense warrants a strong attempt to break it (by the authors) by proposing an attack that's tailored specifically for it. Without such evidence it is not clear whether this defense would be useful in practice. For example, it is relatively straightforward to add an additional term in Mettack's loss that encourages adversarial edges between nodes with similar representations.

Three reviewers participated in the discussion. The main concern was that the proposed method works only on graphs with homophily. Although the rebuttal gives an example where structural similarity of nodes can also be used to design defense strategies, it is still some sort of assumption on the graph. That said, some reviewers pointed out that this is not a deal-breaking limitation, since the assumptions are clearly stated in the paper. The reviewers also agreed that the proposed method is simple yet effective.

Generative AI programmes mine, or learn, from vast amounts of data like text, images, or music online to generate new content which feels like it has been made by a human. The proposals would give artists or creators a so called "rights reservation" – the ability to opt out. But critics of the plan believe it is not possible for an individual writer or artist to notify thousands of different AI service providers that they do not want their content used in that way, or to monitor what has happened to their work across the whole internet. An alternative proposal for artists to opt in to give their permission for their content to be used will be put forward in the House of Lords by cross bench peer Baroness Kidron this week. "It would be a wild punt against the creative sector that is already contributing over 120bn to the economy and be counterproductive to the government's own growth ambitions.

Ensemble approaches for uncertainty estimation have recently been applied to the tasks of misclassification detection, out-of-distribution input detection and adversarial attack detection. Prior Networks have been proposed as an approach to efficiently emulate an ensemble of models for classification by parameterising a Dirichlet prior distribution over output distributions. These models have been shown to outperform alternative ensemble approaches, such as Monte-Carlo Dropout, on the task of out-of-distribution input detection. However, scaling Prior Networks to complex datasets with many classes is difficult using the training criteria originally proposed. This paper makes two contributions.

Forthcoming in Philipp Hacker, Andreas Engel, Sarah Hammer and Brent Mittelstadt (eds) The Oxford Handbook on the Foundations and Regulation of Generative AI (Oxford University Press) Abstract Foundation models and generative artificial intelligence (AI) exacerbate a core regulatory challenge associated with AI: its heterogeneity. By their very nature, foundation models and generative AI can perform multiple functions for their users, thus presenting a vast array of different risks. This multifunctionality means that prescriptive, one-size-fits-all regulation will not be a viable option. Even performance standards and ex post liability--regulatory approaches that usually afford flexibility--are unlikely to be strong candidates for responding to multifunctional AI's risks, given challenges in monitoring and enforcement. Regulators will do well instead to promote proactive risk management on the part of developers and users by using management-based regulation, an approach that has proven effective in other contexts of heterogeneity. Regulators will also need to maintain ongoing vigilance and agility. More than in other contexts, regulators of multifunctional AI will need sufficient resources, top human talent and leadership, and organizational cultures committed to regulatory excellence. Consider one of humanity's most primal of tools: the knife [30]. The knife is not a singular tool; rather, it comes in many different varieties that serve many functions, each of which can generate value for society. Knives are used in the kitchen to prepare delicious meals, and then they are used by diners to consume those same meals. Knives carve objects, cut rope, and open packages. They clear paths through forests and jungles, and they help in harvesting seasonal crops. Knives can be used, of course, to injure or kill people. But in the hands of surgeons, knives are routinely used to save lives. And even though knives take many different forms and are often designed for many different purposes--think of, for example, the many types and sizes of surgical scalpels, woodcarver's chisels, and kitchen implements, among others--knives designed for one purpose also can be adapted for different uses, as anyone who has used a dinner knife to open a postal letter can attest. Many knives, though, are deliberately intended to serve multiple functions, as is the case with a simple pocketknife or, even more emblematically, the classic Swiss army knife, some models of which boast a combination of more than 30 different tools in one. The proliferation of functions performed by different knives has led over the years to different forms and sources of rules governing their manufacture, sale, and deployment.

In collaborative environments, a deep understanding of multi-human teaming dynamics is essential for optimizing performance. However, the relationship between individuals' behavioral and physiological markers and their combined influence on overall team performance remains poorly understood. To explore this, we designed a triadic human collaborative sensorimotor task in virtual reality (VR) and introduced a novel predictability metric to examine team dynamics and performance. Our findings reveal a strong connection between team performance and the predictability of a team member's future actions based on other team members' behavioral and physiological data. Contrary to conventional wisdom that high-performing teams are highly synchronized, our results suggest that physiological and behavioral synchronizations among team members have a limited correlation with team performance. These insights provide a new quantitative framework for understanding multi-human teaming, paving the way for deeper insights into team dynamics and performance.

Large Language Models (LLMs) excel in text generation and understanding, especially in simulating socio-political and economic patterns, serving as an alternative to traditional surveys. However, their global applicability remains questionable due to unexplored biases across socio-demographic and geographic contexts. This study examines how LLMs perform across diverse populations by analyzing public surveys from Chile and the United States, focusing on predictive accuracy and fairness metrics. The results show performance disparities, with LLM consistently outperforming on U.S. datasets. This bias originates from the U.S.-centric training data, remaining evident after accounting for socio-demographic differences. In the U.S., political identity and race significantly influence prediction accuracy, while in Chile, gender, education, and religious affiliation play more pronounced roles. Our study presents a novel framework for measuring socio-demographic biases in LLMs, offering a path toward ensuring fairer and more equitable model performance across diverse socio-cultural contexts.

Who controls the development of Artificial General Intelligence (AGI) might matter less than how we handle the fight for control itself. We formalize this "steering wheel problem" as humanity's greatest near-term existential risk may stem not from misaligned AGI, but from the dynamics of competing to develop it. Just as a car crash can occur from passengers fighting over the wheel before reaching any destination, catastrophic outcomes could arise from development competition long before AGI exists. While technical alignment research focuses on ensuring safe arrival, we show how coordination failures during development could drive us off the cliff first. We present a game theoretic framework modeling AGI development dynamics and prove conditions for sustainable cooperative equilibria. Drawing from nuclear control while accounting for AGI's unique characteristics, we propose concrete mechanisms including pre-registration, shared technical infrastructure, and automated deterrence to stabilize cooperation. Our key insight is that AGI creates network effects in safety: shared investments become more valuable as participation grows, enabling mechanism designs where cooperation dominates defection. This work bridges formal methodology and policy frameworks, providing foundations for practical governance of AGI competition risks.

Model monitoring involves analyzing AI algorithms once they have been deployed and detecting changes in their behaviour. This thesis explores machine learning model monitoring ML before the predictions impact real-world decisions or users. This step is characterized by one particular condition: the absence of labelled data at test time, which makes it challenging, even often impossible, to calculate performance metrics. The thesis is structured around two main themes: (i) AI alignment, measuring if AI models behave in a manner consistent with human values and (ii) performance monitoring, measuring if the models achieve specific accuracy goals or desires. The thesis uses a common methodology that unifies all its sections. It explores feature attribution distributions for both monitoring dimensions. Using these feature attribution explanations, we can exploit their theoretical properties to derive and establish certain guarantees and insights into model monitoring.