newsfilter.io
Interview, Fireside Chat

Phaidra’s Jim Gao on Building the Fourth Industrial Revolution with Reinforcement Learning

  • Jim Gow, founder and CEO of Phaedra, distinguishes AI from simple automation, defining its true promise as "AI creativity"—the ability to acquire knowledge that did not previously exist.
  • Phaedra utilizes reinforcement learning (RL) to act as "virtual plant operators," directly controlling and optimizing mission-critical industrial facilities rather than merely providing recommendations.
  • The concept for applying RL to industrial systems originated from a 2013 Coursera course on machine learning, which inspired Gow to experiment with Google's data center data.
  • In 2016, following the release of DeepMind's AlphaGo, Gow sent an email to Mustafa Suleiman proposing the application of RL to optimize the Power Usage Effectiveness (PUE) of Google's data centers.
  • The core insight driving the project was that operating complex industrial systems is mathematically equivalent to solving a constraint optimization problem with objective functions, actions, and constraints.
  • The original pilot at Google, conducted in 2016, achieved a 40% reduction in energy consumption by comparing AI-generated recommendations against human operator standards.
  • The initial system deployment resulted in a "Pareto gain," meaning energy efficiency improved by 40% while strictly adhering to all existing safety constraints and temperature profiles.
  • Gow observed that the AI proposed counterintuitive operational moves that initially seemed incorrect to human experts but ultimately resulted in significant energy savings when tested.
  • The project evolved from a recommendation engine to a fully autonomous, closed-loop control system capable of issuing commands from the cloud to physical hardware like large-scale chillers.
  • Phaedra's current architecture inserts a cloud-based intelligence layer on top of legacy Building Management Systems (BMS) and SCADA systems without requiring new hardware sensorization.
  • Merck Pharmaceuticals became Phaedra's first public customer, utilizing the autonomous AI system to control a 500-acre vaccine manufacturing facility in Pennsylvania with 62,000 tons of cooling capacity.
  • Early trials at the Merck facility demonstrated a 16% energy savings, though Gow emphasizes the system is designed to continuously improve and maintain optimal performance over time.
  • Unlike static hard-coded logic that degrades as equipment corrodes or fouls, the self-learning RL system adapts in real-time to changing physical conditions and operational loads.
  • Gow identifies the lack of historical data storage and data cleaning infrastructure as the primary barrier preventing widespread adoption of RL in industrial sectors outside of big tech.
  • Many industrial facilities only store sensor data for 90 days to six months, treating data as a forensic tool rather than a resource for real-time intelligence.
  • While RL applications like AlphaGo have proven capable of mastering multiple games (Go, Chess, Shogi) with a single framework, real-world industrial adoption remains limited due to data infrastructure gaps.
  • Gow predicts that AI-driven grid balancing is the single most impactful application for combating climate change, particularly as renewable energy sources increase grid stochasticity.
  • Data centers currently consume approximately 2% of US energy and are projected to consume 37% of Ireland's national electricity consumption by the end of the decade.
  • The unpredictability of renewable energy supply necessitates "spinning reserves" (idling gas turbines), which AI grid balancing could reduce by shifting non-urgent loads to times of high renewable generation.
  • The relationship between RL and Transformer architectures is viewed as complementary, with Transformers handling world-modeling via correlation and RL providing the necessary causality and planning for control systems.
  • Industrial control applications require causality to ensure safety and predictability, a limitation that pure correlation-based Transformer models currently struggle to meet independently.
  • Gow advises aspiring founders to secure co-founders for emotional and workload support and to view the risk of entrepreneurship as lower than perceived, as the skills learned are highly valuable to future employers regardless of outcome.
Phaidra’s Jim Gao on Building the Fourth Industrial Revolution with Reinforcement Learning — Summary