Just months ago, China's boom in data center construction was at its height, fueled by both government and private investors. Renting out GPUs to companies that need them for training AI models was once seen as a sure bet. But with the rise of DeepSeek and a sudden change in the economics around AI, the industry is faltering. Prices for GPUs are falling and many newly built facilities are now sitting empty. Read the full story to find out why.

That's why we've created the AI Hype Index--a simple, at-a-glance summary of everything you need to know about the state of the industry. While AI models are certainly capable of creating interesting and sometimes entertaining material, their output isn't necessarily useful. Google DeepMind is hoping that its new robotics model could make machines more receptive to verbal commands, paving the way for us to simply speak orders to them aloud. Elsewhere, the Chinese startup Monica has created Manus, which it claims is the very first general AI agent to complete truly useful tasks. And burnt-out coders are allowing AI to take the wheel entirely in a new practice dubbed "vibe coding."

Process-supervised reward models serve as a fine-grained function that provides detailed step-wise feedback to model responses, facilitating effective selection of reasoning trajectories for complex tasks. Despite its advantages, evaluation on PRMs remains less explored, especially in the multimodal domain. To address this gap, this paper first benchmarks current vision large language models (VLLMs) as two types of reward models: output reward models (ORMs) and process reward models (PRMs) on multiple vision-language benchmarks, which reveal that neither ORM nor PRM consistently outperforms across all tasks, and superior VLLMs do not necessarily yield better rewarding performance. To further advance evaluation, we introduce ViLBench, a vision-language benchmark designed to require intensive process reward signals. Notably, OpenAI's GPT-4o with Chain-of-Thought (CoT) achieves only 27.3% accuracy, indicating the benchmark's challenge for current VLLMs. Lastly, we preliminarily showcase a promising pathway towards bridging the gap between general VLLMs and reward models -- by collecting 73.6K vision-language process reward data using an enhanced tree-search algorithm, our 3B model is able to achieve an average improvement of 3.3% over standard CoT and up to 2.5% compared to its untrained counterpart on ViLBench by selecting OpenAI o1's generations. We release the implementations at https://ucsc-vlaa.github.io/ViLBench with our code, model, and data.

Audio watermarking is increasingly used to verify the provenance of AI-generated content, enabling applications such as detecting AI-generated speech, protecting music IP, and defending against voice cloning. To be effective, audio watermarks must resist removal attacks that distort signals to evade detection. While many schemes claim robustness, these claims are typically tested in isolation and against a limited set of attacks. A systematic evaluation against diverse removal attacks is lacking, hindering practical deployment. In this paper, we investigate whether recent watermarking schemes that claim robustness can withstand a broad range of removal attacks. First, we introduce a taxonomy covering 22 audio watermarking schemes. Next, we summarize their underlying technologies and potential vulnerabilities. We then present a large-scale empirical study to assess their robustness. To support this, we build an evaluation framework encompassing 22 types of removal attacks (109 configurations) including signal-level, physical-level, and AI-induced distortions. We reproduce 9 watermarking schemes using open-source code, identify 8 new highly effective attacks, and highlight 11 key findings that expose the fundamental limitations of these methods across 3 public datasets. Our results reveal that none of the surveyed schemes can withstand all tested distortions. This evaluation offers a comprehensive view of how current watermarking methods perform under real-world threats. Our demo and code are available at https://sokaudiowm.github.io/.

DATE SUBMITTED: March, 202 5 Words: 9, 22 0 2 Abstract H umanoid robots are a form of embodied artificial intelligence (AI) that look s and act s more and more like humans. Powered by generative AI and advances in robotics, humanoid robots can speak and interact with humans rather naturally but are still easily recognizable as robots. But how will we treat humanoids when they seem indistinguishable from humans in appearance and mind? We find a tendency (called "anti - robot" speciesism) to deny such robots humanlike capabilities, driven by motivations to accord members of the human species preferential treatment . Six experiments show that robots are denied humanlike attributes, simply because they are not biological beings and because humans want to avoid feelings of cognitive dissonance when utilizing such robots for unsavory tasks . Th us, pe ople do not rationally attribute capabilities to perfectly human like robots but deny them capabilities as it suits them . Keywords: robots, artificial intelligence, humanoids, speciesism, cognitive dissonance 3 In recent years, n ew artificial intelligen ce (AI) technologies have been introduced into the marketplace that have the potential to radically change people's work and lives . This paper examines how people might react to robots that seem be " perfectly human like " . With major companies like Amazon and Nvidia planning mass production of such robots, we are entering an era where the line between human and non - human entities is increasingly blurred. Our findings suggest that the advent of such robots will not lead people to rationally conclude that these robots are as capable as humans in performing some tasks . Rather, people will deny these robots humanlike attributes, driven by their motivation to prioritize their own species and to avoid feelings of cognitive dissonance from utilizing such robots for unsavory tasks. Aversion to Robots and AI People are often averse to robots. P sychological research has explained this effect by arguing that such "almost humanlike" robots appear as aesthetically dis pleasing, and that they remind people of zombies, death, or disease (Kätsyri et al., 2015; Mori, 1970; Wang et al., 2015) . Other psychological explanations focus on how people perceive robot minds, sometimes referred to as the "uncanny valley of mind" (Müller et al., 2021; Stein & Ohler, 2017) . These theories suggest that humanoid robots can be unsettling because they remind people of the human ability to experience feelings, even though these robots are not seen as having such capabilities (Gray & Wegner, 2012; Smith et al., 2021) .

Recent generative reasoning breakthroughs have transformed how large language models (LLMs) tackle complex problems by dynamically retrieving and refining information while generating coherent, multi-step thought processes. Techniques such as inference-time scaling, reinforcement learning, supervised fine-tuning, and distillation have been successfully applied to models like DeepSeek-R1, OpenAI's o1 & o3, GPT-4o, Qwen-32B, and various Llama variants, resulting in enhanced reasoning capabilities. In this paper, we provide a comprehensive analysis of the top 27 LLM models released between 2023 and 2025 (including models such as Mistral AI Small 3 24B, DeepSeek-R1, Search-o1, QwQ-32B, and phi-4). Then, we present an extensive overview of training methodologies that spans general training approaches, mixture-of-experts (MoE) and architectural innovations, retrieval-augmented generation (RAG), chain-of-thought and self-improvement techniques, as well as test-time compute scaling, distillation, and reinforcement learning (RL) methods. Finally, we discuss the key challenges in advancing LLM capabilities, including improving multi-step reasoning without human supervision, overcoming limitations in chained tasks, balancing structured prompts with flexibility, and enhancing long-context retrieval and external tool integration.

Generative models offer a scalable and flexible paradigm for simulating complex environments, yet current approaches fall short in addressing the domain-specific requirements of autonomous driving - such as multi-agent interactions, fine-grained control, and multi-camera consistency. We introduce GAIA-2, Generative AI for Autonomy, a latent diffusion world model that unifies these capabilities within a single generative framework. GAIA-2 supports controllable video generation conditioned on a rich set of structured inputs: ego-vehicle dynamics, agent configurations, environmental factors, and road semantics. It generates high-resolution, spatiotemporally consistent multi-camera videos across geographically diverse driving environments (UK, US, Germany). The model integrates both structured conditioning and external latent embeddings (e.g., from a proprietary driving model) to facilitate flexible and semantically grounded scene synthesis. Through this integration, GAIA-2 enables scalable simulation of both common and rare driving scenarios, advancing the use of generative world models as a core tool in the development of autonomous systems. Videos are available at https://wayve.ai/thinking/gaia-2.

Heavy users of ChatGPT tend to be lonelier, more emotionally dependent on the AI tool and have fewer offline social relationships, new research suggests. Only a small number of users engage emotionally with ChatGPT, but those who do are among the heaviest users, according to a pair of studies from OpenAI and the MIT Media Lab. The researchers wrote that the users who engaged in the most emotionally expressive personal conversations with the chatbots tended to experience higher loneliness – though it isn't clear if this is caused by the chatbot or because lonely people are seeking emotional bonds. While the researchers have stressed that the studies are preliminary, they ask pressing questions about how AI chatbot tools, which according to OpenAI is used by more than 400 million people a week, are influencing people's offline lives. The researchers, who plan to submit both studies to peer-reviewed journals, found that participants who "bonded" with ChatGPT – typically in the top 10% for time spent with the tool – were more likely than others to be lonely, and to rely on it more.

OpenAI just announced that all users will soon be able to generate images directly inside of ChatGPT. This will be the default image generation tool in 4o, so there will be no need to open Dall-E whenever you want to whip up a picture of a cat in space eating lasagna or whatever. The company says that the platform will "generate high-quality images based on your prompt, conversation and uploaded files." To the latter point, it'll be able to transform pre-existing images based on prompts. OpenAI is also boasting about significant improvements in text rendering and contextual understanding.

The new model makes progress on technical issues that have plagued AI image generators for years. While most have been great at creating fantastical images or realistic deepfakes, they've been terrible at something called binding, which refers to the ability to identify certain objects correctly and put them in their proper place (like a sign that says "hot dogs" properly placed above a food cart, not somewhere else in the image). It was only a few years ago that models started to succeed at things like "Put the red cube on top of the blue cube," a feature that is essential for any creative professional use of AI. Generators also struggle with text generation, typically creating distorted jumbles of letter shapes that look more like captchas than readable text. The model is able to generate 12 discrete graphics within a single image--like a cat emoji or a lightning bolt--and place them in proper order.