Fireside Chat, Interview
a16z Podcast | Machine Intelligence, from University to Industry
- The Alberta Machine Intelligence Institute (AMII), formerly the Alberta Innovates Centre for Machine Learning, is based at the University of Alberta and has achieved significant milestones in game-playing AI, including solving checkers (the largest game ever solved with 5 × 10²⁰ states) in 2007 and creating the first AI agent capable of near-perfect heads-up limit hold'em poker in 2015.
- Google's DeepMind AlphaGo paper cited research from the University of Alberta for approximately 45% of its work, and roughly half of DeepMind's personnel were Canadian-trained, with 20–25% being former AMII students.
- Researchers prioritize games as a research medium because they offer low-risk, ambiguous environments that serve as effective "petri dishes" for discovering decision-making algorithms in unsupervised settings.
- While games like Chess and Go provide complete information, poker introduces obfuscated information requiring agents to infer hidden data from a limited set of observations (e.g., 15 data points), representing a more complex challenge than perfect-information games.
- AMII's Atari reinforcement learning system reportedly performs roughly four times faster than current DeepMind algorithms and has successfully played approximately half of 90 tested games using unsupervised learning.
- The institute secures funding from industrial partners with revenues exceeding $80 billion and headcounts over 180,000, negotiating unique IP terms that allow AMII to own, commercialize, or license IP depending on the project.
- Corporate "hollowing out" of university AI departments is viewed as a significant threat; executives in the private sector often prioritize short-term roadmaps (5–10 years) over the 15–30 year timelines required for fundamental scientific discovery, potentially stalling long-term innovation.
- AMII operates as a rare exception to industry talent poaching, retaining a core group of professors for over 10 years, similar to the stance taken by University of Montreal professor Yoshua Bengio.
- Current AMII projects include "Meerkat" for temporal social network analysis and "PFM scheduling" for workforce optimization in healthcare, emphasizing automation and optimization in data-centric fields.
- Regarding Jeff Hinton's assertion that radiologist training should cease, AMII leadership argues that while AI imaging may outperform humans in 5 years, implementation is currently hindered by reimbursement models requiring human radiologists and FDA regulations regarding "black box" accountability.
- Richard Sutton, the father of reinforcement learning with 39 years of experience, introduced concepts like on-policy and off-policy learning, enabling systems to learn from both direct action and observation of others, and can forecast using "temporal difference learning" from educated guesses.
- Industrializing AI requires a convergence of people, process, and technology, with a specific need for tools that lower the barrier to entry (currently requiring PhDs) similar to Facebook's "Learner Flow" system, which allows 25% of its developer base to write deep learning code.
- C-suite leaders are advised to establish internal "Bell Labs" style R&D units that are curiosity-driven and report directly to the CEO, accepting the risk of failure and long-term reinvestment to avoid disruption.
- Recommended resources for general audiences include a 40-minute "Andreessen Horowitz Primer on Artificial Intelligence" and the book Artificial Intelligence: What Everyone Needs to Know by Jerry Kaplan.
- Future AI development is expected to shift toward mobile and edge computing to address data access limitations in non-terrestrial regions, rather than relying on massive backend processing.
- Philosophical concerns regarding AI safety include the risk of creating superior intelligence that could outmaneuver humans (analogous to humans' relationship with ants or monkeys) and the potential for weaponized autonomous systems.
- Ethical frameworks for AI suggest treating advanced systems as societal members rather than "indentured servants," though the speaker expresses greater excitement for AI's benefits than fear of its existential risks.