Dec 8, 2017, 5:19 PM ETConference Presentation, Keynote

AI: What's Working, What's Not

AI has transitioned from research labs to the core strategy of major global entities: GAFA (Google, Amazon, Facebook, Apple) in the US and BAT (Baidu, Alibaba, Tencent) in China.
In the current year, 300 CEO earnings calls referenced AI as central to their organizational strategy.
A graph of NIPS conference registrations shows a linear upward trajectory; extrapolation suggests the entire world's population would attend by 2040.
Vladimir Putin stated that the nation leading AI will rule the world.
AI is predicted to become embedded in all software similar to how databases became universal, but with capabilities focused on visual recognition, language understanding, and prediction rather than just storage and counting.
A specific challenge for organizations is identifying the "cucumber sorter" equivalent: low-cost (approx. $2,000), high-impact automation projects that reduce costs and delight customers.

Visual Recognition (Seeing)

Pinterest analyzes images to identify specific items (e.g., coats, scarves) and recommends related products.
Airware uses drone data to verify mining lane compliance (rock markers must be twice the height of the tallest wheel) without manual measurement.
YouTube uses AI to generate closed captions for one billion videos without human involvement, including sound effects like "muffled explosion" or "gunshot."
Facebook automatically translates newsfeed posts when confidence in the translation is high enough to bypass user confirmation.
Everlaw uses natural language processing to cluster and prioritize documents for legal discovery, reducing reliance on manual paralegal review.

Language Understanding

Automated speech recognition now achieves a word error rate of 4%, which is lower than the human average of 5-6%.
This technology enables Siri, Alexa, and YouTube subtitles to function effectively.

Prediction and Recommendation

Airbnb uses machine learning to optimize room pricing based on local events (e.g., South by Southwest) and ranks photos based on traveler preferences rather than professional photography standards.
A key lesson from Airbnb's ranking system is that human annotators must match the end-user profile to ensure accurate supervised learning labels.
Foursquare uses machine learning to determine which venue tips are "most useful," a metric impossible to calculate via standard SQL queries.
Instacart increased shopper speed by 3-4% using gradient-boosted decision trees, with an additional 3-4% lift achieved by implementing deep learning techniques.
Cardiogram uses standard smartwatch heartbeat data to predict atrial fibrillation, sleep apnea, and hypertension, tasks previously requiring dedicated medical devices.
AI soccer analytics calculate the probability of a goal being scored in the next 30 seconds based on player positioning to provide real-time coaching.

Autonomy and Robotics

Shield.ai deploys drones to create 3D maps and identify people in buildings, replacing human security teams in hazardous entry scenarios.
Zipline delivers blood via drone parachute in Western Rwanda, performing 500 deliveries daily to reach remote areas faster than ground transport.
Voyage, a spinout from Udacity, launched a self-driving taxi service in a San Jose retirement community just nine months after founding.

Investors are targeting companies that enable AI development rather than just those using AI, following a gold rush analogy.
Databricks serves as a default computing platform to manage the data and compute resources required for AI model training.
SigOpt automates the tuning of "hyperparameters" (technical knobs) to optimize model predictions.

Narrow AI: Current technology is single-purpose; algorithms for image recognition (Pinterest) do not transfer to document sorting (Everlaw) or navigation (Voyage).
General AI Gap: There is no consistent research agenda to achieve General AI (human-level flexibility across domains); 100 researchers surveyed indicated uncertainty on the path forward.
Performance Gaps: AI systems currently fail at tasks easily solvable by humans, including:
- Answering arbitrary questions about images (e.g., "Is the umbrella upside down?").
- Extracting specific facts from text (e.g., Nikola Tesla's ethnicity from a Wikipedia article).
- Answering fourth-grade science exam questions.
Expert Skepticism: Jeffrey Hinton (University of Toronto/Google) suggests the current deep learning path (back propagation) may be a dead end, calling for a fundamental reboot to achieve general intelligence.
Super AI Concerns: Experts like Andrew Ng compare fears of "super AI" to overpopulation on Mars, deeming it a distant, irrelevant concern compared to current technological limitations.

China's Strategy

China aims to be the global leader in AI by 2030 via a coordinated national policy.
While Western universities (Google DeepMind, Stanford, Toronto) originated foundational algorithms, China now publishes more deep learning papers.
China's "Thousand Talents Program" incentivizes immigrants with Western training to return and conduct research in China.
A potential advantage for China is a lack of data privacy regulations, allowing for mass surveillance and data aggregation to fuel algorithmic training.
Proposed Western countermeasures include creating data clearinghouses for labeled datasets and attracting talent rather than repelling immigrants.

Job Market and Automation

Historical automation fears (loom, tractor, semiconductor, Excel) have not resulted in net job loss due to the economy's infinite capacity to create new products and services.
New jobs (e.g., AI coach, fracking engineer, millennial consultant) emerged from previously unknown consumer demands.
The ATM example shows that automation can increase the number of bank tellers by enabling banks to open more branches (10 staff vs. 100 per branch).
Amazon employs both increasing numbers of robots and human workers, validating the correlation between automation and workforce expansion.

Future Outlook

Even if AI research plateaus, there are 20-40 years of software development remaining to integrate AI into existing applications.
Organizations are advised to select 5-10 pilot projects involving AI integration across existing products, annual objectives, or new services.
When evaluating vendors, companies should demand a sophisticated machine learning roadmap; products without it are likely to be outperformed by competitors.
Implementation paths include: buying vendor software, hiring custom teams (e.g., Gigster), using cloud APIs for specific tasks, or building custom models using public cloud tools.
Workforce training can be achieved efficiently via short-term programs (e.g., Udacity's 6-month nanodegree) rather than requiring advanced degrees.