Rapid advancements in artificial intelligence (AI) have sparked growing concerns among experts, policymakers, and world leaders regarding the potential for increasingly advanced AI systems to pose catastrophic risks. Although numerous risks have been detailed separately, there is a pressing need for a systematic discussion and illustration of the potential dangers to better inform efforts to mitigate them. This paper provides an overview of the main sources of catastrophic AI risks, which we organize into four categories: malicious use, in which individuals or groups intentionally use AIs to cause harm; AI race, in which competitive environments compel actors to deploy unsafe AIs or cede control to AIs; organizational risks, highlighting how human factors and complex systems can increase the chances of catastrophic accidents; and rogue AIs, describing the inherent difficulty in controlling agents far more intelligent than humans. For each category of risk, we describe specific hazards, present illustrative stories, envision ideal scenarios, and propose practical suggestions for mitigating these dangers. Our goal is to foster a comprehensive understanding of these risks and inspire collective and proactive efforts to ensure that AIs are developed and deployed in a safe manner. Ultimately, we hope this will allow us to realize the benefits of this powerful technology while minimizing the potential for catastrophic outcomes.

Style representation learning builds content-independent representations of author style in text. Stylometry, the analysis of style in text, is often performed by expert forensic linguists and no large dataset of stylometric annotations exists for training. Current style representation learning uses neural methods to disentangle style from content to create style vectors, however, these approaches result in uninterpretable representations, complicating their usage in downstream applications like authorship attribution where auditing and explainability is critical. In this work, we use prompting to perform stylometry on a large number of texts to create a synthetic dataset and train human-interpretable style representations we call LISA embeddings. We release our synthetic stylometry dataset and our interpretable style models as resources.

We conduct a comparison under the same number of training epochs by training our method on 12 tasks mentioned before for 1000 epochs in total, and comparing them with corresponding Single Task Learning (STL) neural solvers that are trained for 1000 epochs on each of their respective tasks. This is, by no means, a fair comparison, as our method dynamically chooses a task to train for 1000 epochs, resulting in a much smaller sample size than each task when using STL. Despite this, we choose this comparison as an intuitive way to demonstrate the superior generalization ability of our method under such extreme conditions. We present the results in Figure 2. Compared to individual tasks, our method's comparative performance is to be expected due to the vast differences in sample size. In most cases, our method's performance is equivalent to that of using 100 to 200 epochs of STL. However, STL can only obtain one model in this context and lacks the ability to handle different types of COPs or to generalize well when presented with the same type of COP but with varying problem scales. As a result, our method demonstrates unparalleled superiority in three ways: (1) when considering the average performance on all problem scales for each type of COP, our method obtains the best results in CVRP, OP, and KP, and is equivalent to the results achieved by training TSP for about 500 epochs. This showcases our method's excellent generalization ability for problem scales; (2) Our method can handle various types of COPs under the same number of training epochs, which is impossible for STL due to the existence of task-specific modules; (3) Our method's training time is strictly shorter than the longest time-consuming task.

Human expectations stem from their knowledge about the others and the world. Where human-AI interaction is concerned, such knowledge may be inconsistent with the ground truth, resulting in the AI agent not meeting its expectations and degraded team performance. Explicable planning was previously introduced as a novel planning approach to reconciling human expectations and the agent's optimal behavior for more interpretable decision-making. One critical issue that remains unaddressed is safety in explicable planning since it can lead to explicable behaviors that are unsafe. We propose Safe Explicable Planning (SEP) to extend the prior work to support the specification of a safety bound. The objective of SEP is to search for behaviors that are close to the human's expectations while satisfying the bound on the agent's return, the safety criterion chosen in this work. We show that the problem generalizes multi-objective optimization and our formulation introduces a Pareto set. Under such a formulation, we propose a novel exact method that returns the Pareto set of safe explicable policies, a more efficient greedy method that returns one of the Pareto optimal policies, and approximate solutions for them based on the aggregation of states to further scalability. Formal proofs are provided to validate the desired theoretical properties of the exact and greedy methods. We evaluate our methods both in simulation and with physical robot experiments. Results confirm the validity and efficacy of our methods for safe explicable planning.

Machine Learning models are being deployed as parts of real-world systems with the upsurge of interest in artificial intelligence. The design, implementation, and maintenance of such systems are challenged by real-world environments that produce larger amounts of heterogeneous data and users requiring increasingly faster responses with efficient resource consumption. These requirements push prevalent software architectures to the limit when deploying ML-based systems. Data-oriented Architecture (DOA) is an emerging concept that equips systems better for integrating ML models. DOA extends current architectures to create data-driven, loosely coupled, decentralised, open systems. Even though papers on deployed ML-based systems do not mention DOA, their authors made design decisions that implicitly follow DOA. The reasons why, how, and the extent to which DOA is adopted in these systems are unclear. Implicit design decisions limit the practitioners' knowledge of DOA to design ML-based systems in the real world. This paper answers these questions by surveying real-world deployments of ML-based systems. The survey shows the design decisions of the systems and the requirements these satisfy. Based on the survey findings, we also formulate practical advice to facilitate the deployment of ML-based systems. Finally, we outline open challenges to deploying DOA-based systems that integrate ML models.

Given a document in a source language, cross-lingual summarization (CLS) aims at generating a concise summary in a different target language. Unlike monolingual summarization (MS), naturally occurring source-language documents paired with target-language summaries are rare. To collect large-scale CLS data, existing datasets typically involve translation in their creation. However, the translated text is distinguished from the text originally written in that language, i.e., translationese. In this paper, we first confirm that different approaches of constructing CLS datasets will lead to different degrees of translationese. Then we systematically investigate how translationese affects CLS model evaluation and performance when it appears in source documents or target summaries. In detail, we find that (1) the translationese in documents or summaries of test sets might lead to the discrepancy between human judgment and automatic evaluation; (2) the translationese in training sets would harm model performance in real-world applications; (3) though machine-translated documents involve translationese, they are very useful for building CLS systems on low-resource languages under specific training strategies. Lastly, we give suggestions for future CLS research including dataset and model developments. We hope that our work could let researchers notice the phenomenon of translationese in CLS and take it into account in the future.

In this study, we propose a novel deep spatio-temporal point process model, Deep Kernel Mixture Point Processes (DKMPP), that incorporates multimodal covariate information. DKMPP is an enhanced version of Deep Mixture Point Processes (DMPP), which uses a more flexible deep kernel to model complex relationships between events and covariate data, improving the model's expressiveness. To address the intractable training procedure of DKMPP due to the non-integrable deep kernel, we utilize an integration-free method based on score matching, and further improve efficiency by adopting a scalable denoising score matching method. Our experiments demonstrate that DKMPP and its corresponding score-based estimators outperform baseline models, showcasing the advantages of incorporating covariate information, utilizing a deep kernel, and employing score-based estimators.

For each of $T$ time steps, $m$ experts report probability distributions over $n$ outcomes; we wish to learn to aggregate these forecasts in a way that attains a no-regret guarantee. We focus on the fundamental and practical aggregation method known as logarithmic pooling -- a weighted average of log odds -- which is in a certain sense the optimal choice of pooling method if one is interested in minimizing log loss (as we take to be our loss function). We consider the problem of learning the best set of parameters (i.e. expert weights) in an online adversarial setting. We assume (by necessity) that the adversarial choices of outcomes and forecasts are consistent, in the sense that experts report calibrated forecasts. Imposing this constraint creates a (to our knowledge) novel semi-adversarial setting in which the adversary retains a large amount of flexibility. In this setting, we present an algorithm based on online mirror descent that learns expert weights in a way that attains $O(\sqrt{T} \log T)$ expected regret as compared with the best weights in hindsight.

Republican members of congress reveal why they voted to oust former Speaker Kevin McCarthy from his seat, while one of his supporters reveals a contrasting perspective of the historic vote. An unseasonably warm, autumnal light bathed a corridor near the Hall of Columns on the first floor of the Capitol. But the well-trafficked hall was surprisingly empty when I encountered a lawmaker who was a friend and occasional ally of then-House Majority Leader Kevin McCarthy, R-Calif. Then-House Speaker John Boehner, R-Ohio, unexpectedly announced his retirement just a few days before. McCarthy was poised to assume the speakership. But there were already rumblings that the McCarthy lacked the votes.

One domain of interest is their use in algorithmic hiring, specifically in matching resumes with job categories. Yet, this introduces issues of bias on protected attributes like gender, race and maternity status. The seminal work of Bertrand & Mullainathan (2003) set the gold-standard for identifying hiring bias via field experiments where the response rate for identical resumes that differ only in protected attributes, e.g., racially suggestive names such as Emily or Lakisha, is compared. We replicate this experiment on state-of-art LLMs (GPT-3.5, Bard, Claude and Llama) to evaluate bias (or lack thereof) on gender, race, maternity status, pregnancy status, and political affiliation. We evaluate LLMs on two tasks: (1) matching resumes to job categories; and (2) summarizing resumes with employment relevant information. Overall, LLMs are robust across race and gender. They differ in their performance on pregnancy status and political affiliation. We use contrastive input decoding on open-source LLMs to uncover potential sources of bias.