Uber has implemented a $1,500/month limit on AI tool usage, potentially signaling a benchmark for pricing in the industry. This move could influence how other companies budget for and utilize AI tools. The cap may reflect Uber's efforts to manage costs and optimize resource allocation. This development is relevant to discussions around AI adoption and cost management in businesses.
Researchers propose STRIDE, a framework for training data attribution that models the functional effect of training data in the activation space. STRIDE formulates TDA as a sparse recovery problem and achieves state-of-the-art results for LLM pre-training attribution, being 13 times faster than previous methods. This approach enables efficient tracking of individual training example influences. STRIDE has practical applications in data selection, contamination, and qualitative analysis. The method is particularly useful for large language models where repeated retraining is computationally challenging.
Researchers propose a distributional variant of the DAgger algorithm, called DistIL, which utilizes rich feedback from various sources to improve reinforcement learning. This approach allows for a simple forward cross-entropy objective that enables monotonic policy improvement and provides guarantees on regret. DistIL is compared to existing reinforcement learning methods, including those with self-distillation objectives, and demonstrates improved performance across multiple domains such as scientific reasoning, coding, and mathematical problem-solving. The method optimizes a lower bound on teacher-weighted likelihood of success, leading to improved Pass@N metrics. Empirical results show that DistIL outperforms baselines in several areas.
A new study suggests that failed reasoning traces from post-trained language models can be used to identify which failures can be rescued with additional interventions. The researchers propose three problem-level trajectory features that can be derived from the structure of available interventions to recover the recoverability structure of failed rollouts. These features can cluster failures into stable regimes, characterize the failure topography of different post-training methods, and support a training-free routing rule. The study demonstrates the effectiveness of this approach on the Steerable-Hard subset, where retry is insufficient and a bounded intervention is reachable. The features and routing rule also transfer across two cross-family probes.
A recent study explored the use of model activations to optimize the selection of in-context examples for large language models. The analysis, which included experiments with Llama-3.2-3B and Qwen2.5-3B base models, found that activation-based sampling does not correlate with example quality or task performance. The researchers hypothesize that this may be due to superposition, where models represent more features than they have dimensionality. The study tested various attention masking strategies across diverse datasets, but found a negative result, with an absolute Spearman correlation coefficient of at most 0.33. The findings suggest that alternative methods, such as Sparse Autoencoders, may be a promising direction for future research.
OpenAI has introduced new capabilities to GPT-Rosalind, enhancing its biological reasoning, medicinal chemistry expertise, genomics analysis, and experimental workflow capabilities. This update aims to advance life sciences research. GPT-Rosalind is a tool designed to assist researchers in the life sciences field. The update includes improvements in areas crucial for life sciences research, making it a significant development for researchers relying on AI for their work.
Wasmer utilized OpenAI's Codex with GPT-5.5 to develop a Node.js runtime for edge computing. This collaboration accelerated development by 10x to 20x, significantly reducing the time to market from months to weeks. The use of Codex, powered by GPT-5.5, demonstrates the potential of AI in streamlining complex software development tasks. This project highlights the capability of AI-assisted tools in enhancing development efficiency and speed.
OpenAI has outlined a blueprint for the U.S. governance of frontier AI, focusing on safety, resilience, and national security. The proposed federal framework aims to address the challenges and risks associated with advanced AI systems. This initiative involves collaboration with policymakers and experts to establish guidelines and regulations for the development and deployment of frontier AI. The blueprint is available on OpenAI's website, providing a detailed outline of the proposed framework and its key components. The goal is to ensure that AI development aligns with U.S. values and priorities.
OpenAI has outlined its public policy agenda for AI, focusing on key areas such as safety, youth protection, workforce transition, and the establishment of global standards. This agenda aims to ensure that AI benefits society as a whole. The policy agenda is designed to guide the development and deployment of AI in a responsible and beneficial manner. OpenAI's initiative involves collaboration with various stakeholders to address the challenges and opportunities presented by AI.
Google has open-sourced its hydrology framework to aid in flood resilience. The framework is designed to simulate and predict water flow, helping to mitigate the effects of flooding. This move is part of Google's Climate & Sustainability efforts, aiming to make its technology accessible for the greater good. By open-sourcing the framework, Google hopes to encourage collaboration and further development in the field of hydrology. The framework can be used by researchers, developers, and organizations to improve flood prediction and response.
Hugging Face discusses direct preference optimization, a technique that can be applied beyond chatbots. This method involves optimizing models based on human preferences, allowing for more nuanced and context-specific outputs. The blog post highlights the potential applications of this approach in various areas, including decision-making and recommendation systems. The technique is particularly relevant for AI models that require human feedback to improve their performance.
At UC Berkeley, there's been a notable increase in failing grades in computer science classes, which professors attribute to the rising use of AI tools and a corresponding decline in math skills among students. This trend suggests that while AI can be a powerful tool, it may also be contributing to a lack of foundational understanding in key areas of computer science. The impact of AI on education, particularly in fields that rely heavily on mathematical and computational skills, is a significant area of study. As AI becomes more integrated into educational settings, understanding its effects on learning outcomes is crucial. The situation at UC Berkeley highlights the need for educators to adapt their teaching methods to ensure students develop a strong grasp of fundamental concepts despite the availability of AI tools.
Mnemo is an open-source, local-first AI memory layer designed for large language models (LLMs). It is built using Rust, SQLite, and petgraph. Mnemo aims to provide a flexible and efficient way to store and manage AI model memories. The project is hosted on GitHub and has garnered attention on Hacker News. The use of local storage and graph-based data structures could improve model performance and efficiency.
Hyper, a company backed by Y Combinator, has launched to power agentic development. The company aims to provide a brain for agents, enabling them to learn and adapt in complex environments. This technology has potential applications in areas such as robotics and autonomous systems. Hyper's approach focuses on integrating multiple AI systems to create more sophisticated agents. The launch is significant for the development of more advanced AI-powered agents.
The construction of AI data centers is increasingly being done in secrecy, raising questions about the motivations behind this trend. Companies involved in AI development are building these data centers to support their operations, but the lack of transparency has sparked curiosity. The secrecy may be due to competitive or security concerns, but it also underscores the growing importance of data centers in AI development. The trend highlights the need for more information about the environmental and social impact of these facilities.