Dynamic Interaction Techniques

I love this. This is rmy real focus: not just Humans + AI as AI augmenting human capabilities in tasks, but designing interactions to give humans greater skills and capabilities that endure and grow over time. The outcome is smarter people. "We introduce a novel conceptual framework for human-AI interaction: extraheric AI. We define “extraherics” as a mechanism that fosters users’ higher-order thinking skills during the course of task completion. Extraheric is based on the Latin word “extraho” (to draw forth or pull out), and we use this term to suggest that AI can draw forth people’s higher order thinking skills and thus promote their cognitive potential. Rather than replacing or augmenting human cognitive abilities, extraheric AI encourages users to engage in higher-order thinkingduring task completion." The interaction strategies to evoke higher-order thinking skills suggested in the paper include: 💡 Suggesting & Recommending: The AI proposes ideas, viewpoints, or solutions, prompting users to evaluate and choose from multiple suggestions. 📝 Explaining: The AI provides detailed explanations, focusing on the 'why' and 'how' to help users deepen their understanding of the task, rather than providing direct solutions. 🎯 Nudging: The AI subtly influences user behavior by presenting additional information or perspectives indirectly, encouraging exploration without overtly recommending a specific path. 🗣️ Debating & Discussing: Users engage in debates or discussions with AI agents, which present different opinions and arguments, encouraging users to explore diverse perspectives and think critically. ❓ Questioning: The AI asks open-ended, thought-provoking questions to stimulate cognitive engagement, expanding users’ thinking by challenging their assumptions or viewpoints. 🛠️ Scaffolding: The AI offers temporary support or guidance through complex tasks, allowing users to focus on specific aspects while gradually removing the assistance as users become more competent. 🎮 Simulating: The AI simulates different scenarios or perspectives, helping users practice skills or experience situations from a different point of view, such as role-playing or rehearsing responses. 👀 Demonstrating: The AI acts as a model, showcasing behavior or task completion, allowing users to observe and learn implicitly through vicarious learning by watching the AI perform. I look forward to many others building on this work and integrating these concepts into enterprise software.

Robots just got a lot cooler. The new HUSKY framework demonstrates a Unitree G1 humanoid skateboarding with genuine agility—not just balancing, but pushing, steering, and maneuvering. This is the pivot from "locomotion" to "dynamic interaction." The system isn't treating the world as a static floor anymore; it's mastering complex tools through physics. Three things make this work: ✔️ Physics-Aware Control: The robot doesn't just guess the physics. It explicitly models the "kinematic equality constraint," understanding exactly how board tilt forces the trucks to turn. ✔️ Hybrid Dynamics: Standard robots hate "switching" modes. HUSKY models the task as a hybrid dynamical system, seamlessly handling the mathematical chaos of transitioning from pushing (foot-on-ground) to riding (foot-on-board). ✔️ Active Agency: Using Deep Reinforcement Learning (DRL), the robot learns to generate its own momentum through pushing and controls its direction via steering, rather than just passively surviving the ride. Paper: https://lnkd.in/eX6DvZwX

Balancing Thought and Action: A New Paradigm for Training Adaptive AI Agents 👉 Why Rethink Test-Time Scaling? Most AI agents today optimize for "thinking" longer—generating extensive reasoning traces before acting. But in interactive tasks (like web navigation), this approach has a critical flaw: once an action is taken, agents can’t revisit decisions or gather new information. Imagine trying to book a hotel by over-analyzing one option without checking alternatives—this rigidity limits adaptability in dynamic, partially observable environments. 👉 What’s the Alternative? The paper introduces test-time interaction scaling: letting agents take "more steps" during deployment to explore, backtrack, and dynamically adjust strategies. For example, an agent might: * Browse multiple hotel listings * Compare reviews and prices * Revise its choice if a better option appears This contrasts with simply generating longer reasoning chains per step. The key insight? Interaction provides new data, while deeper per-step reasoning only reshuffles existing knowledge. 👉 How Does It Work? The authors propose TTI (Test-Time Interaction), a curriculum-based RL method that trains agents to: * Start with short interaction horizons (focused on exploitation) * Gradually increase steps to encourage exploration (e.g., retries, searches) In experiments on WebArena and WebVoyager: * Prompting-based interaction scaling improved success rates by 5-9% over fixed-step baselines. * TTI-trained agents achieved SOTA results (64.8% on WebVoyager) using a 12B parameter model, outperforming methods that rely on proprietary data or larger models. * Agents learned to balance exploration (e.g., initiating Bing searches) and exploitation (efficient task completion). 👉 Practical Takeaways 1. Compute efficiency: Spending tokens on interaction often outperforms deeper per-step reasoning under fixed budgets. 2. Generalizability: The approach extends beyond web tasks—domains like robotics or customer support could benefit from adaptive interaction. 3. Open-source impact: TTI’s code and synthetic training pipeline lower barriers to developing specialized agents. Why does this matter? By prioritizing "strategic doing" over "isolated thinking", this work opens pathways for AI systems that dynamically adapt to real-world complexity. It challenges the assumption that scaling model size or per-step compute is the only route to better performance. Check out the project page and code to dive deeper. Question for discussion: How might interaction scaling change how we design agents for your domain?

Eye-Hand Object Movement: Key Findings from Our Study Led by Uta Wagner, this research explores eye-hand UIs for #XR HMDs. Six techniques were evaluated (see 4 in video), incl. our new Look & Drop (simultaneous eye [XY] and hand [Z] movement) and Look & Drop+ (simultaneous eye [XY] and hand [XYZ] movement). Highlights from the Results: ⏱ Overall speed: similar across all techniques. 🎯 Selection: slower with direct gestures. 🚀 Movement: fastest with direct gestures. ⚙️ Errors: more with Look & Drop, while Direct and Look & Drop+ had the fewest. 💪 Perceived physical effort: highest for direct gestures. ✋ Hand movement: least with all Look & Drop, then Gaze + Pinch, then Direct. 🌟 User preference: Look & Drop+ (66%), Gaze + Pinch (21%), Look & Drop (8%), Direct (4%). Afterthoughts: In our near-space study, gaze improved selection speed but slowed dragging. Despite this, less effort and errors of Look & Drop+ made it the preferred choice. Note, for real-world tasks, gaze-based object manipulation requires careful design to minimize accidental activations. Further, we only tested tasks in near space. For larger distances, gaze-based object movement is likely more beneficial. Huge congratulations to lead Uta Wagner and Master’s student Andreas Asferg Jacobsen for making this research happen! 🎉 Further info: 📄 Paper: https://lnkd.in/dZkKqSJq 🎥 Video & more insights from Uta: https://lnkd.in/d-4TbxXu ✨ About interaction design: https://lnkd.in/dYQQeWPy All authors: Uta Wagner, Andreas Asferg Jacobsen, Tiare F., Hans Gellersen, Ken Pfeuffer

BHA Dynamic Behavior: If You’re Not Modeling It, You’re Guessing Most drilling dysfunctions are not bit problems, motor problems, or tool problems. They are system behavior problems. A BHA is not static. It is a coupled, non-linear mechanical system where axial, torsional, and lateral motions interact continuously while drilling. Static BHA analysis only describes shape. It does not describe performance. Axial instability shows up as bit bounce when applied WOB exceeds the bit’s real cutting capacity. The result is unstable ROP, shock loading on MWD/LWD, and cumulative cutter damage that rarely shows up immediately. Torsional instability, stick–slip, remains the most destructive and least controlled dysfunction in drilling. When downhole torque demand exceeds available torque at the bit, energy is stored in the drillstring and released violently. Bit overspeed, motor failures, and erratic toolface behavior are the predictable outcome not bad luck. Lateral instability is equally unforgiving. Off center rotation and backward whirl generate high alternating stresses that rapidly wear stabilizers, enlarge the hole, and drive fatigue failures. By the time it’s visible at surface, the damage is already done. What makes BHA dynamics dangerous is their non-linearity. Small changes in RPM or WOB can move the system from stable to destructive in minutes. Averaging formation properties or relying on “experience” is not control its exposure. Dynamic modeling turns drilling from reactive to deliberate. It defines stable operating windows, supports stabilizer and BHA design decisions, and justifies motor or RSS selection even in vertical sections. Without dynamic modeling: • Parameters are adjusted after tools are damaged • ROP is sacrificed to protect hardware • Failures are misdiagnosed and repeated With dynamic modeling: • Drilling stays inside stable regimes • High ROP is maintained without punishment • Tool life and reliability improve measurably Downhole performance is governed by physics, not optimism. If you’re not modeling BHA dynamics, you’re not managing risk you’re accepting it.

I am truly convinced that real-time interaction management (RTIM) today represents the pinnacle of personalization. Unlike previous approaches that relied on historical data alone, RTIM processes information instantaneously, creating a fluid, adaptive engagement strategy to dynamically respond to HCP behaviors and needs as they happen. So how does this intelligent process work? ↳ It begins with an instant analysis of an HCP's history and preferences, enabling dynamic channel selection for each interaction. ↳ As an HCP engages, Al algorithms adaptively refine content to match their interests. ↳ The system then triggers personalized responses based on specific actions, such as sending a tailored email after an HCP browses a particular topic. ↳ RTIM integrates with CRM systems, providing sales representatives with actionable insights for timely, relevant follow-ups. Throughout this process, the system continuously learns and evolves, creating an increasingly accurate profile of each HCP's needs and preferences. That's the level of engagement that must feel bespoke to each HCP, what do you think? #personalization #realtimeinteraction