newsfilter.io
Lecture, Conference Presentation, Presentation

The Promise of AI

a16zFrank Chen
Watch on YouTube →

Core Thesis and Market Trajectory

  • Frank Chen posits that AI will achieve ubiquitous integration in all critical software, mirroring the 40-year adoption arc of the relational database (invented 1970, commercialized late 70s/early 80s).
  • While relational databases made storing and sorting data cheap, AI will primarily make perception, content creation, prediction, and optimization cheap.
  • The driving economic force is cost reduction; as AI capabilities become inexpensive, they will enable applications previously too costly or complex to build.

Category 1: Autonomous Mobility (Making Movement Cheap)

  • Ground Logistics:
    • Otto demonstrated autonomous trucking in October 2016.
    • Dispatch Robotics is developing autonomous shopping carts to handle the "last mile" from stores (e.g., Safeway, Whole Foods) to consumer homes.
  • Aerial Robotics:
    • Skydio: Developing consumer drones that follow users (runners, cyclists) to capture autonomous "selfie" videos.
    • Shield.ai: Creating drones for first responders (SWAT, military) to autonomously map unmapped buildings in real-time and identify friend vs. hostile entities.
    • Zipline (in partnership with UPS): Deploying drones in Western Rwanda to deliver blood; flights cover up to 90 miles round trip with parachute delivery accuracy.
      • Operational capacity reaches 150 trips per day.
      • Addresses logistics barriers caused by flooded roads or safety risks.

Category 2: Perception and Understanding (Making Seeing Cheap)

  • Technological Foundation:
    • Generative Adversarial Networks (GANs) pit two neural networks against each other (classification vs. generation) to create ultra-accurate image classifiers.
    • Accuracy Benchmark: AI algorithms on the ImageNet dataset have surpassed human accuracy, achieving error rates below 5% (vs. humans at 95% accuracy).
  • Commercial Applications:
    • Pinterest: Integrated live camera object recognition to identify purchasable items; can identify designer details (e.g., Charles Eames) that human users might miss.
    • Agriculture (Individual Level): Developer Makoto Koiki built a Raspberry Pi-based robotic arm using Google TensorFlow to sort cucumbers into 16 market grades, reducing labor costs for family farms.
    • Agriculture (Industrial Level): Blue River Technologies mounts cameras on tractors to fertilize individual lettuce heads based on real-time growth state rather than blanket field application.
    • Retail:
      • Amazon physical stores utilize cameras to automatically track items removed from shelves for frictionless checkout.
      • Orchard Supply Hardware: Deploys a greeter robot that identifies held items (e.g., nails) and navigates shoppers to aisles or initiates a human teleconference if unrecognized.
      • Knightsbridge Robotics: Provides robotic security guards capable of detecting trace carbon monoxide and recording video, though currently limited in physical manipulation (e.g., opening doors).
      • Bossa Nova Robotics: Uses robots to autonomously verify inventory levels and "planogram" compliance in supermarkets to reduce revenue loss.
  • Research Frontiers:
    • Narrative Reconstruction: Researchers are developing systems to analyze sequences of images to deduce causal narratives (e.g., a person kicking a soccer ball to dislodge a Frisbee, resulting in both items getting stuck).
    • Query Answering: Google Brain research focuses on answering complex visual queries currently impossible for search engines, such as "make me a cup of tea" based on visual scene understanding.

Category 3: Content Creation (Making Writing and Generating Cheap)

  • Text Generation:
    • Journalism: The Washington Post and Chinese aggregator Toutiao used AI to write Olympic sports coverage; Toutiao generates full-length articles from video footage.
    • Coding: Microsoft's DeepCoder system remixes existing code snippets from repositories (GitHub, Stack Exchange) to generate functional software programs.
  • Visual Generation:
    • Photorealistic Images: AI can generate realistic images from text descriptions (e.g., "red and brown bird with stubby beak") or hand-drawn sketches using GANs.
    • Recipe Extraction: Algorithms analyze cooking videos (e.g., BuzzFeed Tasty) to retroactively generate step-by-step written instructions and ingredient lists.
  • Audio and Media:
    • Music: UK startup JukeDeck generates original music; the distinction between AI and human composition is becoming difficult for lay listeners.
    • Film: IBM Watson analyzed scenes for the movie Morgan to assemble a trailer; currently used to assist human editors rather than replace them (Intelligence Augmentation).

Category 4: Prediction (Making Future Forecasting Cheap)

  • Content Strategy: BuzzFeed uses AI to predict which videos will perform well in foreign markets based on performance in other regions, suggesting translations.
  • Identity Verification:
    • Startup UnifyID replaces passwords by predicting user identity via behavioral biometrics (walking gait, typing rhythm, swipe patterns).
    • Systems can reject access even if the intruder mimics the physical appearance of the legitimate user.
  • Customer Support Automation:
    • AI predicts caller identity and intent before the call connects, routing users directly to the correct specialist (e.g., a 401k allocation expert) without manual verification.
  • Behavioral Analysis:
    • MIT research demonstrates algorithms predicting physical interactions (hugs, handshakes) in TV shows based on prior scene context.
  • Healthcare Diagnostics:
    • Freenome: Analyzes free-floating DNA in blood samples for early cancer detection, aiming to replace invasive tissue biopsies.
    • Cardiogram: Uses Apple Watch sensor data to predict abnormal cardiac events, saving lives by enabling early intervention.
    • Suicide Prevention:
      • Study 1: Predicts suicide risk with ~80% accuracy by analyzing electronic health records up to 2-3 years in advance.
      • Study 2 (Yori Leskovic/Pinterest): Analyzes text from crisis counseling to identify effective intervention strategies.

Category 5: System Optimization (Making Complex Coordination Cheap)

  • Traffic Routing: Waze optimizes city-wide traffic flow by dynamically rerouting thousands of drivers to avoid congestion.
  • Sports Analytics: Research from Caltech, MIT, and Disney optimizes soccer defensive formations; Manchester City's positioning reduced scoring probability by 28% (from 69% to 41%) compared to a baseline team.
  • Software Compilation:
    • AI-optimized assembly code reduces instruction sets and improves runtime speed by 1.6x compared to standard compilers.
  • Predictive Modeling Tuning:
    • SigOpt: Uses machine learning to fine-tune existing mathematical models (e.g., stock returns, flight wing turbulence survival) to improve prediction accuracy.
  • Infrastructure Efficiency:
    • Google DeepMind: Optimized data center cooling variables (120+ parameters), reducing electricity consumption by 20-25% while maintaining identical workloads.
    • Instacart: Optimized shopping routes and logistics, reducing grocery delivery time by 8%.

Category 6: Language Understanding (Making Communication Cheap)

  • Input Speed: Talking to phones is 3x faster than typing, particularly for non-Latin scripts like Chinese, driving high adoption rates.
  • Smart Reply: Google Inbox feature generates 10% of all mobile email replies with high contextual accuracy.
  • Document Summarization:
    • Agolo: Creates summaries of single or multiple documents while preserving semantic meaning using deep learning.
    • Textio: Analyzes job descriptions to remove unconscious bias and optimize for candidate appeal.
    • Everlaw: Automates the categorization and review of legal documents during e-discovery, preventing critical evidence from being overlooked.
  • Emotional Computing:
    • Anki Cosmo: Educational robot designed by a Pixar animator to simulate emotions and natural interaction styles.
  • Real-Time Translation:
    • Earpiece technology demos now allow near-instant (1-2 second) translation between languages (e.g., Spanish to English) during live conversation.

Strategic Recommendations for Implementation

  • Tool Acquisition: Leverage the abundance of open-source AI tools to rapidly integrate intelligence into applications.
  • Workforce Training: Organizations must invest in training employees on AI capabilities and toolsets via MOOCs and tutorials.
  • Foster Creativity: Adopt a "let a thousand cucumbers bloom" approach, encouraging low-cost, high-impact experiments (e.g., the $1,000 cucumber sorter) to identify unique use cases within the organization.