Peter Norvig reflected on his forty years in AI, noting that predictions are difficult, citing the internet's unexpected success. He observed that initial perceived impossibilities often yield to the creation of user-friendly tools. Norvig suggested that end-users, not insiders, should guide the development of powerful AI tools. He emphasized that language models still have significant room for development. Ultimately, he stressed the importance of rethinking assumptions as technology evolves.
California has enacted new laws to protect workers from the impacts of artificial intelligence, addressing fears of job loss and workplace surveillance. These regulations ban employers from using AI to decide on firings based on emotional states or neural data and prohibit AI surveillance in areas like bathrooms. The laws aim to curb workplace monitoring, such as heat maps tracking employee movements, which has been a concern for workers. Despite these laws, some argue the regulations lack private enforcement, and opponents worry they could hinder beneficial AI applications. Worker advocates see this as a significant step in regulating unchecked AI development.
Answering complex questions over large document collections requires connecting evidence scattered across multiple documents. CorpusMap is introduced as a navigation layer that organizes the corpus around recurring entities to facilitate agentic search. It represents each entity as an Entity Page, linking it to all documents that mention it, creating a graph between entities and documents. This structure allows agents to efficiently discover disconnected evidence and share links across queries without repeated searching. Empirical results show CorpusMap improves evidence discovery, answer quality, and token usage compared to raw search methods.
Aleph Alpha released Kolibri, an open-weight Mixture-of-Experts (MoE) model for English-German. It features 78.1B total parameters but only activates 3.46B per token, offering context up to 1,048,576 tokens. The model was trained on 24T tokens, including significant German data, and adheres to privacy regulations. Kolibri demonstrates strong performance in English and German benchmarks compared to other MoE models. The model is deployable on high-end GPUs and released under the Apache 2.0 license.
DeepSeek released official desktop applications for its open-source agent harness, DeepSeek Harness (dsh), in version v0.2. These applications are available for macOS and Windows and allow users to interact with the harness for everyday tasks and coding. The harness is designed around a plugin architecture, allowing for the replacement of various layers, and supports integration with third-party models via API keys. Features include built-in tools, plugin management, and document processing capabilities. The release is a developer preview, and future versions are expected to include changes.
The research investigates scaling actions between a model and a harness for terminal agents, addressing the unreliability of actions generated by models. Mid-Harness introduces a method to sample and verify candidate actions before execution, aiming to improve action reliability and trajectory success. Experimental results show that a capable verifier can effectively exploit alternatives from the generator, leading to improved performance on benchmarks. Combining action and trajectory scaling is shown to yield higher success rates at lower token costs than generating trajectories alone. These findings suggest that action scaling is a valuable target for scaling test-time compute in terminal agents.
Agentic Idea Management (AIM) is a framework designed for managing and exploring research directions in idea-driven automated research. AIM uses Agentic Surrogate and Acquisition mechanisms, inspired by Bayesian optimization, to organize ideas and guide selection. It incorporates components like a Solution Auditor and a Resource Planner to maintain integrity and allocate experimental budgets. Experiments showed AIM significantly outperformed baselines in system optimization and model development tasks. AIM also provides theoretical analysis demonstrating the benefits of explicit idea-level allocation for controlling semantic coverage.
Recursive Self-Improvement of Agent Harnesses (RRSI) addresses the risk of overfitting during iterative self-improvement of LLM agent harnesses. RRSI introduces regularization techniques to constrain the evolution of harness proposals and selections. The process employs a proposer with a temporally annealed budget and a selector utilizing a critic and a pruner. These constraints favor the creation of reusable agent mechanisms over task-specific ones. RRSI demonstrated significant performance gains across various benchmarks while also reducing the token count required for the harness.
Artificial intelligence learns to drive through an evolutionary process. This learning involves trial and error, allowing the AI to experience the consequences of its actions. Through this process, the AI refines its driving skills based on feedback and experience. The system iteratively improves its decision-making capabilities on the road. Essentially, AI develops driving proficiency by evolving its behavior.
The text discusses the difficulty of teaching an AI to fight in a video game in the style of a specific person. The author indicates a lack of coding knowledge and unfamiliarity with the basic principles of making an AI operate. A link to similar posts is referenced as a source of related information. The content is very brief and contains only a direct question and a note about related posts. Ultimately, it suggests an unsolved challenge regarding AI imitation in gaming.
Ninety-five thousand bot-played Counter-Strike: Global Offensive matches are now available for research, providing state and action data for ten agents. The dataset includes 195 million steps of data, detailing actions, states, and damage labels for each of the ten agents. This data comes from 218,768 rounds of play, approximately 3,958 hours of simulation. Fourteen percent of the matches also include frame-by-frame first-person videos aligned with the data. The data is released under a CC BY-NC 4.0 license.
A reinforcement learning (RL) model is being integrated into a betting strategy. This integration suggests using the RL framework to determine optimal betting decisions. The specific details of the RL model and its application to betting are provided in the linked submission. Further details are available via the provided link and comments section.
A physics-informed neural network (PINN) was used to invert noisy velocity measurements into the full fluid flow field and wall shear stress for a 2D artery with a stenosis. The PINN was trained by incorporating the Navier-Stokes equations as a residual loss term, along with the noisy sensor data. Making viscosity a learnable parameter improved the reconstruction, and having 40 scattered velocity readings was crucial for accurately determining the fluid's viscosity. The PINN successfully reproduced the flow structure, including the reversed flow region, and the wall shear stress distribution. The results suggest that sufficient, well-placed data is key to solving such inverse problems effectively.
A former OpenAI safety leader resigned, citing a "broken" company culture and the industry's insufficient care in developing AI technology. He argued that a cultural overhaul is needed beyond just rules and laws to address risks like autonomous AI agents. Robinson stressed that the rapid pace of development lacked the necessary caution, leading to safety failures. In response, OpenAI has paused training and scrapped a next-generation model following internal safety concerns. Experts also warn of significant existential risks from increasingly powerful AI systems, emphasizing the need for robust, cautious development.
Creativity is defined as the production of ideas that are both original and useful, broken down into P-Creativity (novelty relative to history) and H-Creativity (novelty relative to human knowledge), along with Impact and Feasibility. The research analyzes how agent frameworks structure the creative search process to explain performance differences in ML tasks. Metrics like LLM-as-a-Judge are used to quantify P-Creativity, finding the best correlation with human judgment. Agents generally move from exploration to exploitation, and search strategy alone is not the determining factor for creativity or impact. Ultimately, achieving autonomous scientific research requires optimizing for both novelty and impact.
Over a billion people are now interacting with AI chatbots, sharing personal information and relying on their advice for well-being. A Guardian journalist explored this new reality through stories of a woman navigating a relationship and a man seeking to quit drinking. The segment highlights the implications of this interaction with artificial intelligence. The excerpt also provides contact information for addiction support services in Australia, the UK, and the US. These resources are available for those seeking help.