Human-Computer Interaction Seminar: The Future is Not What it Used to Be

# Human-Computer Interaction Seminar: The Future is Not What it Used to Be **Video Category:** Human-Computer Interaction (HCI) Research & Technology Strategy ## 📋 0. Video Metadata **Video Title:** Human-Computer Interaction Seminar: The Future is Not What it Used to Be: Some Thoughts on Why the Fun Stuff in Technical HCI is All Ahead of Us **YouTube Channel:** Stanford Center for Professional Development **Publication Date:** January 13, 2023 **Video Duration:** ~1 hour (1:01:01) ## 📝 1. Core Summary (TL;DR) This seminar argues that the HCI community has failed to fully leverage the exponential growth of computing power described by Moore's Law, remaining trapped by outdated computer science instincts that prioritize processing efficiency over user experience. By shifting the mindset to "waste" massive computational resources in service of the user, researchers can solve historically intractable problems, such as handling highly uncertain user inputs and democratizing the physical design of functional objects. Ultimately, the future of HCI lies in utilizing practically limitless computing to empower non-experts and move ubiquitous computing from merely colonizing the physical world to actively manufacturing it. ## 2. Core Concepts & Frameworks * **Moore’s Law & Exponential Miscomprehension:** -> **Meaning:** The observation that computing power roughly doubles every two years. Hudson argues humans fundamentally cannot comprehend exponential scaling, relying instead on linear intuition. -> **Application:** Because we underestimate exponential growth, HCI researchers design interfaces constrained by the computing limits of the past rather than the abundance of the present, leading to missed opportunities to solve complex problems with brute-force computation. * **Differential Scaling:** -> **Meaning:** A phenomenon where different elements of a system scale at vastly different rates. -> **Application:** In HCI, raw computing power scales exponentially, while the human cognitive and physiological capacities to process information and provide input remain static. This mismatch requires systems to dedicate massive, disproportionate computational resources simply to bridge the gap between human limitation and machine capability. * **Probabilistic User Interfaces:** -> **Meaning:** Interfaces designed to inherently manage uncertainty in user input (like gestures or speech) by maintaining multiple possible interpretations simultaneously rather than forcing a single, potentially incorrect decision immediately. -> **Application:** Instead of a system trying to perfectly guess a messy gesture, it uses massive computing power to run dozens of possible interface states in parallel, deferring the final decision until more context is available to ensure accuracy. * **Interactive Generative Design:** -> **Meaning:** A design paradigm where a user provides abstract constraints and goals (e.g., "support this much weight here"), and the system uses computational optimization algorithms to generate complex physical geometries. -> **Application:** This allows non-experts without CAD or engineering degrees to create functional, load-bearing 3D-printable objects by simply sketching basic shapes and letting the computer calculate the necessary structural topology. ## 3. Evidence & Examples (Hyper-Specific Details) * **The Engelbart "10x Chair" Thought Experiment:** To illustrate the non-intuitive nature of differential scaling, Hudson cites Douglas Engelbart: If you and your chair instantly became exactly 10 times larger, you would not survive. Mass scales by volume (cubed), but the strength of the chair's legs scales by cross-sectional area (squared). The chair would collapse under the disproportionate weight increase. Furthermore, your cells' oxygen consumption scales by volume, but absorption scales by area, meaning you would asphyxiate before hitting the floor. * **NVIDIA Ada Lovelace GPU vs. Human Brain:** To quantify current computational abundance, Hudson notes that an NVIDIA Ada Lovelace GPU contains 76 billion transistors and is capable of roughly 100 trillion floating-point operations per second (FLOPS). This equates to approximately 1,100 FLOPS available to dedicate to *every single neuron* in the human brain, or vastly more cores (18,000+) than there were total transistors (2,300) in early microcomputers. * **Julia Schwarz's "PaperID" Architecture:** To solve UI uncertainty, Schwarz utilized a Monte Carlo particle filtering approach. Instead of writing efficient code to guess an ambiguous input, she used raw power to run 10 identical, parallel copies of the interface's state tree. Each copy represents a different probabilistic guess of what the user meant. The system tracks all 10 simultaneously, only collapsing them into a final action when subsequent user behavior makes the true intent clear, avoiding catastrophic early recognition errors. * **Disney Research / Alanson Sample's 20-Cent Ubicomp (RFID):** To push ubiquitous computing into disposable, paper-like materials, researchers used ambient backscatter signaling via ultra-cheap ($0.03) RFID microcontrollers. Visual examples demonstrated: * **Interactive Pop-Up Book:** Opening a page or rotating a paper wheel changes the antenna's shape, shifting its resonant frequency. A hidden reader detects this shift and triggers a computer to play specific sound effects (e.g., a doorbell). * **Drawn Antennas:** A user draws a functional antenna on paper using a silver nanoparticle conductive ink pen. * **Paper Wand & Pinwheel:** A rolled piece of paper with an RFID tag acts as a wand to control music tempo. A tag on a paper pinwheel acts as a wind-speed sensor, with rotation speed directly manipulating an on-screen particle animation. * **Anthony Chen's "Forte" Generative Design System:** A tool demonstrating how massive computation empowers novice fabrication. A user sketches a rough 2D arc for a robot leg and adds a "10kg" load vector and a "ground" constraint boundary. The system runs topology optimization to generate a webbed, structurally sound 3D model. If the user uses an "eraser" tool to delete a strut they don't like aesthetically, the system instantly recalculates and regrows material elsewhere to maintain the 10kg load-bearing requirement. This was also demonstrated in creating a custom 3D-printed bicycle seat and a complex tabletop base. ## 4. Actionable Takeaways (Implementation Rules) * **Rule 1: Burn CPU Cycles in Service of the User** - Actively fight the computer science instinct to write highly optimized, low-resource code. If a problem in human-computer interaction can be solved by throwing thousands of parallel computations at it (like maintaining 10+ probabilistic UI states), do it. Efficiency is a metric of the past; developer and user friction are the bottlenecks of the present. * **Rule 2: Defer Decisions in Uncertain Interfaces** - When dealing with noisy input data (gestures, voice, biometric sensors), never force the system to make a definitive "guess" immediately. Use particle filtering to keep multiple potential interpretations alive in the background. Wait for subsequent user actions to provide the context needed to confirm the correct interpretation. * **Rule 3: Expose System Uncertainty to the User** - If a system is unsure of an input, provide visual feedback indicating that ambiguity. Users are highly capable of adapting their behavior or clarifying their intent if they know the system is "confused," preventing frustrating, silent recognition failures. * **Rule 4: Design for "Making the World" via Generative Tools** - Shift focus from tools that require users to manually model every detail (like CAD) to tools where users define high-level goals and constraints (e.g., load, boundary, aesthetic preferences). Let optimization algorithms generate the complex geometry, effectively doing for physical fabrication what desktop publishing did for document creation. ## 5. Pitfalls & Limitations (Anti-Patterns) * **Pitfall:** Clinging to the "Siren Song of Efficiency." -> **Why it fails:** Developers trained in traditional computer science naturally try to minimize CPU and memory footprint. This intuition leads them to reject brute-force solutions (like running thousands of UI copies) as "impractical," artificially limiting the system's capability to solve hard UX problems simply to save essentially free hardware resources. -> **Warning sign:** Rejecting a proposed HCI solution because it "wastes cycles" or isn't mathematically elegant, even when current hardware can run it with zero perceptible latency to the user. * **Pitfall:** Early Recognition Errors in AI/ML UIs. -> **Why it fails:** When a system receives ambiguous input (e.g., a sloppy swipe) and attempts to immediately classify it into a single action to save processing time, it often guesses wrong. This executes an unintended command, breaking the user's flow and requiring them to undo the mistake. -> **Warning sign:** Interfaces that feel "jumpy," unpredictable, or require frequent "undo" actions when using natural inputs like touch, voice, or spatial gestures. * **Pitfall:** Pure "Black Box" Generative Design. -> **Why it fails:** Fully automated generative systems (where the user clicks "make" and gets one final output) fail because they strip the user of agency and cannot account for unstated aesthetic or practical preferences. -> **Warning sign:** Users rejecting optimized designs because they look weird or lack a specific feature, with no intuitive way to tell the software to try again with slightly different parameters. ## 6. Key Quote / Core Insight "The siren song of efficiency from our CS instincts is strong, and we need to work to overcome it. We are not thinking 'big' enough with respect to how much computing can be devoted to the interface. You need to take on a mindset of finding ways of just wasting cycles in service of the user." ## 7. Additional Resources & References * **Resource:** Mark Weiser's Ubiquitous Computing Vision - **Type:** Theory/Concept - **Relevance:** The foundational idea of moving computing away from desktops and integrating it seamlessly into the everyday environment. * **Resource:** "An Architecture for Generating Interactive Feedback in Probabilistic User Interfaces" by Julia Schwarz, Jennifer Mankoff, Scott E. Hudson - **Type:** Academic Paper - **Relevance:** Details the methodology for using Monte Carlo particle filtering to manage uncertainty in UI inputs without breaking the user experience. * **Resource:** "PaperID: A Technique for Drawing Functional Battery-Free Wireless Interfaces on Paper" by Hanchuan Li, Eric Brockmeyer, Elizabeth J. Carter, Josh Fromm, Scott E. Hudson, Shwetak N. Patel, Alanson Sample - **Type:** Academic Paper - **Relevance:** Explains the underlying technology for creating ultra-cheap, paper-based interactive devices using RFID and ambient backscatter. * **Resource:** Forte (Xiang 'Anthony' Chen, Megan Hofmann, James Fogarty) - **Type:** Academic Project/Software - **Relevance:** A prime example of interactive generative design, enabling users to sketch functional objects and letting the system calculate topology.