Enterprise success with AI agents depends more on establishing clear rules and controls than on maximizing agent autonomy. Agentic AI projects face challenges with escalating costs, unclear value, and inadequate risk controls, leading to a high forecast for project failure. A lack of transparency in autonomous decisions makes tracing accountability difficult, posing significant risks in regulated fields. Leading organizations are shifting focus to structured governance by using narrow-scope agents, human checkpoints, and full decision traceability as core design requirements. The competitive advantage lies in building trustworthy systems where risk and compliance are integrated from the start, rather than being added as afterthoughts.
Three long-standing Japanese companies—Toto, Ajinomoto, and Nittobo—are major contributors to the AI semiconductor boom. Toto developed electrostatic chucks for silicon wafer handling, Ajinomoto's byproduct creates a critical insulation film for microprocessors, and Nittobo produces essential glass fibers for advanced chips. All three firms reported significant growth in fiscal year 2025 due to semiconductor demand. These established companies are leveraging their expertise to become vital suppliers in the high-tech chip industry.
FreeToken is an edge-native Mixture-of-Experts (MoE) serving engine designed to run large language models efficiently on single workstation GPUs. It treats personal machines as unified inference platforms, dynamically mapping computation and model state across available hardware. The system achieves this by utilizing bandwidth-adaptive execution, semantic-aware caching, and elastic memory management to overcome limitations in existing MoE serving methods. Results show FreeToken achieves up to 1.5–2.3 times the decode throughput of baselines, with evaluations including running 753B GLM-5.2 on a single card. The project is open-source, deployable, and targets solo developers, startups, and regulated environments.
A document intelligence pipeline is implemented using deepDoctection to combine layout detection, table structure recognition, OCR, reading-order reconstruction, and annotation linking in a single workflow. The process involves explicitly configuring the analyzer with various model-based services and custom components to extract structured information. A custom pipeline component was created to extract monetary, date entities, and classify documents based on tabular characteristics. The process demonstrates inspecting model registrations, controlling service execution, and undoing annotations to gain granular control over the analysis. Finally, processed pages are serialized into ordered JSONL chunks suitable for downstream RAG systems.
Vercel introduced Is Agentic, a free tool that scores how readily AI agents can discover, access, understand, and use a public website. The assessment uses Ora’s 100+ checks across four weighted layers of discovery, access, usability, and payments. The tool is accessible via a browser or CLI and is free with no required API keys or billing. Reports provide detailed findings, observed evidence, and actionable recommendations for improving agent interaction with the site. The scoring system prioritizes finding functional interfaces like APIs and commerce surfaces, excluding non-applicable elements like marketing content.
The agent's functionality is being upgraded from in-memory persistence to a proper backend using a PostgreSQL database. This transition is necessary because in-memory storage fails upon process restart and lacks shared data access across different user sessions. A `BookingRepository` protocol is implemented to allow the agent to interact with either an in-memory or PostgreSQL persistence layer. Key nodes in the LangGraph workflow, such as proposing time slots and confirming bookings, now utilize this repository to read and write appointment data. This backend enables the agent to function as a product by managing persistent customer bookings.
When dealing with a folder of unrelated PDFs, a standard RAG index based on shared fields is unnecessary. The solution involves creating a nested outline by generating one summary line per file and incorporating each file's table of contents. Retrieval is then performed in two levels: first by routing over the file summaries, and then by descending into the surviving documents using their internal outlines. This two-level routing allows for precise retrieval without needing a formal index, treating the folder as a single long document with chapters. The process yields a result where the model reads the file list and the relevant section hierarchy sequentially.
VSArena now features a dedicated VLA track utilizing only a 128x128 RGB camera and language instructions, excluding cube poses from the policy input. Scoring assesses spatial accuracy and completion using real poses internally, while state-based poses are kept in a separate debug track and do not update the public ELO. The in-browser demo is for spectator/development use and does not post to the public leaderboard. Public ELO is derived from a server-side harness, which is the next step before open submissions. The repository and documentation are now publicly available, addressing prior feedback regarding the distinction between VLA and state inputs.
A walkthrough on model-based reinforcement learning (MBRL) is being written, covering concepts like Dyna, Monte Carlo Tree Search (MCTS), and the AlphaGo line. The author has completed Part 1 and is currently working on Part 2, focusing on the optimal-control side. Part 2 will include a coding example demonstrating path planning using A* and RL. The author seeks feedback on the overall organization, delivery, and diagrams of the work. A subsequent part will feature an extensive coding example for path planning.
Training PPO variants on different multi-agent tasks reveals that optimal hyperparameters, such as learning rate, entropy coefficient, and batch size, vary across different architectures and scenarios. The author questions whether unifying these hyperparameters is necessary for a fair comparison of architectures. Adjusting hyperparameters can sometimes lead to models that fail to converge. The core concern is ensuring a fair evaluation of different model architectures.
Reinforcement learning from human feedback (RLHF) utilizes policy gradient methods to train AI agents. Algorithms like Proximal Policy Optimization (PPO) and Gradient Policy Optimization (GRPO) are employed for this learning process. These techniques allow AI models to learn optimal policies by maximizing received rewards. The process involves adjusting the policy based on the feedback provided, often from human evaluators. This framework enables AI to learn complex behaviors through reward-based optimization.
OpenAI trained an AI to compete in the game Dota 2 before developing ChatGPT. This endeavor focused on creating an artificial intelligence capable of playing the complex strategy game. The goal was to train the AI to master the skills required for competitive Dota 2 gameplay. This project explored the application of machine learning to complex strategic environments. The work served as a precursor to the development of large language models like ChatGPT.
Reinforcement learning's effectiveness stems from a feedback closure where inference recruits and combines training support, which is then fed back into training to modify future state. Positive feedback paths lead to the reinforcement of functional routes where positive consequences repeatedly favor the same routes. Experiments showed that concentrating positive feedback on already learned capabilities caused the reinforced skill's support to increase while others' margins decreased. This concentration effect suggests that feedback concentration and duration influence how functional support is reorganized during learning. Consequently, the consolidation of behavior and the concentration of learning around favored routes may be manifestations of the same underlying feedback mechanism.
Researchers developed a special diving suit and oxygen generation system for Madagascar hissing cockroaches. These cyborg insects can survive underwater and low-oxygen environments to aid in search-and-rescue operations. They are guided remotely using electronic controllers, allowing them to navigate through difficult terrain like floods and rubble. Previously, these insects demonstrated utility by relaying information during an earthquake. This technology suggests potential for assisting in inspecting infrastructure and other rescue scenarios.
A reinforcement learning robot, initially a two-legged MiniWalker, has been improved to control four motorized joints: left hip, left knee, right hip, and right knee. The robot learns its walking gait through reinforcement learning using the PPO algorithm, rather than manual programming. An analytics system was added to track reward, episode length, distance traveled, and average velocity. The robot has achieved a maximum distance of 52 meters in one episode, although walking is not yet consistent. The next phase of the project involves expanding the robot to a quadruped to test its ability to adapt when a leg fails.