Conference Presentation, Keynote
AI: What's Working, What's Not
- AI is projected to integrate into every software product following the database adoption trajectory, a process requiring 20 to 40 years of industry work even if research plateaus.
- NIPS conference registrations are extrapolated to cover the global population by 2040, while speech recognition algorithms currently at 4% error rates are expected to fall below the human benchmark of 5% to 6%.
- Content platforms anticipate AI will subtitle a billion YouTube videos with sound effects and automatically translate Facebook posts without user prompts when confidence thresholds are met.
- Enterprise applications will utilize AI for specific operational efficiencies, including clustering legal documents at Everlaw, dynamic pricing for Airbnb events, and identifying optimal tips and photos at Foursquare venues.
- Efficiency gains are forecasted for Instacart via a total speed lift of 6% to 8% through combined traditional and deep learning, while Cardiogram aims to predict sleep apnea and hypertension using non-medical wearable data.
- Specialized AI deployments include coaching tools calculating goal concession probabilities within 30 seconds, Shield.ai drones for mapping and identification in hazardous zones, and Zipline's continuation of 500 daily blood deliveries in Western Rwanda.
- Commercial self-driving taxi services are expected to launch in San Jose retirement communities nine months post-spinout, while Databricks is predicted to become the default platform for data distribution and AI compute.
- Model optimization will increasingly rely on automated hyperparameter adjustment by SigOpt, alongside cloud APIs enabling image recognition and sentiment analysis via external services.
- China targets becoming the global AI leader by 2030 through a national deep learning curriculum in middle schools, the "Thousand Talents Program" to recruit diaspora researchers, and a data advantage from unrestricted surveillance for model accuracy.
- In response to China's data strategies, the West may establish a labeled dataset clearinghouse, while companies lacking a sophisticated machine learning roadmap risk vendor rejection.
- The market is expected to generate dozens to hundreds of billion-dollar "Jiminy Cricket" AI assistant companies over the next 20 to 30 years, with Udacity offering machine learning nanodegrees within six months as a PhD alternative.
- Automation trends historically correlate with job creation, with predictions that bank teller numbers will rise alongside ATMs and that Amazon's human and robot headcounts will continue to grow together.
- Low-cost automation for significant organizational tasks is anticipated to be achievable for under $2,000 using tools like Raspberry Pi and TensorFlow, while custom teams like Gigster will build bespoke AI solutions.
- General super AI capable of outperforming humans in all domains is not considered a foreseeable concern, and the global research community currently lacks a consistent agenda for achieving general AI.