When Dario Amodei gets excited about AI--which is nearly always--he moves. The cofounder and CEO springs from a seat in a conference room and darts over to a whiteboard. He scrawls charts with swooping hockey-stick curves that show how machine intelligence is bending toward the infinite. His hand rises to his curly mop of hair, as if he's caressing his neurons to forestall a system crash. You can almost feel his bones vibrate as he explains how his company, Anthropic, is unlike other AI model builders.

Welcome to our monthly digest, where you can catch up with any AIhub stories you may have missed, peruse the latest news, recap recent events, and more. This month's digest includes four interviews. We hear from two newly-elected AAAI Fellows, and two researchers at the start of their careers, to find out about their different research areas – human-allied AI, multilingual natural language processing, microtargeting and activity patterns on social media, and differential privacy. We are delighted to announce the launch of our interview series featuring the 2025-elected AAAI Fellows. We began the series in style, meeting Sriraam Natarajan to talk about his research on human-allied AI.

The release of the latest image generator on OpenAI's ChatGPT has triggered a flood of online memes featuring images done in the style of Studio Ghibli, the Japanese studio behind classic animated films like "My Neighbor Totoro" and "Princess Mononoke." Since the release on Wednesday, AI-generated images depicting Studio Ghibli versions of Elon Musk with U.S. President Donald Trump, "The Lord of the Rings," and even a recreation of the Sept. 11 attacks have gone viral across online platforms.

Since the early days of the commercial internet, typosquatting has exploited the simplest of human errors, mistyping a URL, to serve as a potent tool for cybercriminals. Initially observed as an opportunistic tactic, typosquatting involves registering domain names that closely match that of reputable brands, thereby redirecting users to counterfeit websites. This has evolved into a sophisticated form of cyberattack used to conduct phishing schemes, distribute malware, and harvest sensitive data. Now with billions of domain names and TLDs in circulation, the scale and impact of typosquatting have grown exponentially. This poses significant risks to individuals, businesses, and national cybersecurity infrastructure. This whitepaper explores how emerging large language model (LLM) techniques can enhance the detection of typosquatting attempts, ultimately fortifying defenses against one of the internet's most enduring cyber threats. Cybercriminals employ various domain squatting techniques to deceive users and bypass traditional security measures. These methods include but not limited to: Character Substitution: These attacks swap similar looking characters like replacing "o" with "0" in go0gle[.]com to trick users into believing they are visiting the legitimate site. Omission or Addition: This method involves removing or adding a character, creating domains such as gogle[.]com

The task of $\textit{Differentially Private Text Rewriting}$ is a class of text privatization techniques in which (sensitive) input textual documents are $\textit{rewritten}$ under Differential Privacy (DP) guarantees. The motivation behind such methods is to hide both explicit and implicit identifiers that could be contained in text, while still retaining the semantic meaning of the original text, thus preserving utility. Recent years have seen an uptick in research output in this field, offering a diverse array of word-, sentence-, and document-level DP rewriting methods. Common to these methods is the selection of a privacy budget (i.e., the $\varepsilon$ parameter), which governs the degree to which a text is privatized. One major limitation of previous works, stemming directly from the unique structure of language itself, is the lack of consideration of $\textit{where}$ the privacy budget should be allocated, as not all aspects of language, and therefore text, are equally sensitive or personal. In this work, we are the first to address this shortcoming, asking the question of how a given privacy budget can be intelligently and sensibly distributed amongst a target document. We construct and evaluate a toolkit of linguistics- and NLP-based methods used to allocate a privacy budget to constituent tokens in a text document. In a series of privacy and utility experiments, we empirically demonstrate that given the same privacy budget, intelligent distribution leads to higher privacy levels and more positive trade-offs than a naive distribution of $\varepsilon$. Our work highlights the intricacies of text privatization with DP, and furthermore, it calls for further work on finding more efficient ways to maximize the privatization benefits offered by DP in text rewriting.

Reconstructing dynamic assets from video data is central to many in computer vision and graphics tasks. Existing 4D reconstruction approaches are limited by category-specific models or slow optimization-based methods. Inspired by the recent Large Reconstruction Model (LRM), we present the Large Interpolation Model (LIM), a transformer-based feed-forward solution, guided by a novel causal consistency loss, for interpolating implicit 3D representations across time. Given implicit 3D representations at times $t_0$ and $t_1$, LIM produces a deformed shape at any continuous time $t\in[t_0,t_1]$, delivering high-quality interpolated frames in seconds. Furthermore, LIM allows explicit mesh tracking across time, producing a consistently uv-textured mesh sequence ready for integration into existing production pipelines. We also use LIM, in conjunction with a diffusion-based multiview generator, to produce dynamic 4D reconstructions from monocular videos. We evaluate LIM on various dynamic datasets, benchmarking against image-space interpolation methods (e.g., FiLM) and direct triplane linear interpolation, and demonstrate clear advantages. In summary, LIM is the first feed-forward model capable of high-speed tracked 4D asset reconstruction across diverse categories.

Accurate motion forecasting is essential for the safety and reliability of autonomous driving (AD) systems. While existing methods have made significant progress, they often overlook explicit safety constraints and struggle to capture the complex interactions among traffic agents, environmental factors, and motion dynamics. To address these challenges, we present SafeCast, a risk-responsive motion forecasting model that integrates safety-aware decision-making with uncertainty-aware adaptability. SafeCast is the first to incorporate the Responsibility-Sensitive Safety (RSS) framework into motion forecasting, encoding interpretable safety rules--such as safe distances and collision avoidance--based on traffic norms and physical principles. To further enhance robustness, we introduce the Graph Uncertainty Feature (GUF), a graph-based module that injects learnable noise into Graph Attention Networks, capturing real-world uncertainties and enhancing generalization across diverse scenarios. We evaluate SafeCast on four real-world benchmark datasets--Next Generation Simulation (NGSIM), Highway Drone (HighD), ApolloScape, and the Macao Connected Autonomous Driving (MoCAD)--covering highway, urban, and mixed-autonomy traffic environments. Our model achieves state-of-the-art (SOTA) accuracy while maintaining a lightweight architecture and low inference latency, underscoring its potential for real-time deployment in safety-critical AD systems.

Understanding and mitigating hallucinations in Large Language Models (LLMs) is crucial for ensuring reliable content generation. While previous research has primarily focused on "when" LLMs hallucinate, our work explains "why" and directly links model behaviour to the pre-training data that forms their prior knowledge. Specifically, we demonstrate that an asymmetry exists in the recognition of logically equivalent facts, which can be attributed to frequency discrepancies of entities appearing as subjects versus objects. Given that most pre-training datasets are inaccessible, we leverage the fully open-source OLMo series by indexing its Dolma dataset to estimate entity frequencies. Using relational facts (represented as triples) from Wikidata5M, we construct probing datasets to isolate this effect. Our experiments reveal that facts with a high-frequency subject and a low-frequency object are better recognised than their inverse, despite their logical equivalence. The pattern reverses in low-to-high frequency settings, and no statistically significant asymmetry emerges when both entities are high-frequency. These findings highlight the influential role of pre-training data in shaping model predictions and provide insights for inferring the characteristics of pre-training data in closed or partially closed LLMs.

Integrating audio comprehension and generation into large language models (LLMs) remains challenging due to the continuous nature of audio and the resulting high sampling rates. Here, we introduce a novel approach that combines Variational Quantization with Conditional Flow Matching to convert audio into ultra-low bitrate discrete tokens of 0.23kpbs, allowing for seamless integration with text tokens in LLMs. We fine-tuned a pretrained text-based LLM using Low-Rank Adaptation (LoRA) to assess its effectiveness in achieving true multimodal capabilities, i.e., audio comprehension and generation. Our tokenizer outperforms a traditional VQ-VAE across various datasets with diverse acoustic events. Despite the substantial loss of fine-grained details through audio tokenization, our multimodal LLM trained with discrete tokens achieves competitive results in audio comprehension with state-of-the-art methods, though audio generation is poor. Our results highlight the need for larger, more diverse datasets and improved evaluation metrics to advance multimodal LLM performance.

This paper examines the critical role of Graph Neural Networks (GNNs) in data preparation for generative artificial intelligence (GenAI) systems, with a particular focus on addressing and mitigating biases. We present a comparative analysis of three distinct methods for bias mitigation: data sparsification, feature modification, and synthetic data augmentation. Through experimental analysis using the german credit dataset, we evaluate these approaches using multiple fairness metrics, including statistical parity, equality of opportunity, and false positive rates. Our research demonstrates that while all methods improve fairness metrics compared to the original dataset, stratified sampling and synthetic data augmentation using GraphSAGE prove particularly effective in balancing demographic representation while maintaining model performance. The results provide practical insights for developing more equitable AI systems while maintaining model performance.