Interview, Fireside Chat, Conference Presentation
MIc'd Up | Part 2: Pedro Domingos in Conversation with Eric Schmidt
- Pedro Domingos predicts a shift from knowledge engineering to self-programming via data as the dominant paradigm, a transition described as "extremely powerful" and currently "bearing fruit," with current algorithmic impacts expected to be "extremely large" within industries while next-generation systems capable of human-like generalization could "dwarf" these effects.
- Economic projections estimate the current wave of machine learning at "trillions" of dollars, potentially reaching "five trillion" upon the arrival of next-generation algorithms, which would drive a "new world economy" where the value of breakthroughs far exceeds current valuations.
- Labor market expectations anticipate that in the "near term" (defined as "less than ten years"), some jobs will disappear but "many, many millions of new kinds of jobs" will appear, driven by cheaper goods and increased demand for "complementary goods," with specific transitions noted from trucking to sectors like "construction work" that are difficult for machines to automate.
- Domingos forecasts that as computers master natural language, the number of programmers could expand from "X million" to "billions," while in the "short term," accessibility will improve via simplified interfaces allowing users to produce solutions, such as medical diagnoses, in "30 seconds" rather than "one year."
- The outlook includes a transition where "day-to-day running of things" is handled by machine learning systems with continuous learning capabilities, leaving humans as the "captain of the ship" for "black swan events" or scenarios lacking sufficient training data, contrasting with current systems that "become less and less intelligent" without retraining.
- Business strategy will require embedding "machine learning from day one" to create a "very short action-reaction loop," avoiding the obsolescence of companies that merely "bolt" learning onto existing processes, with Domingos identifying firms like "Apple" as vulnerable as their core strengths become commoditized.
- Future technology will enable "every customer to be unique," allowing businesses to leverage data for personalized marketing and support, while app success will depend on integration with a central "machine learning agent" rather than standalone functionality, and successful companies will likely adopt "hybrid" models that crowdsource data to improve products.
- Eric Schmidt predicts that machine learning assistants will soon unfold to handle complex daily tasks like arranging transportation and dining, replacing the need for users to "stitch together" multiple applications, and that companies mining data from sources like "Twitter" or satellite imagery will gain early insights into sentiment and business performance.
- Research leadership is expected to come "from all over the map" rather than just top departments, driven by non-traditional researchers, though Domingos notes a severe talent gap where the demand for professionals with combined skills in "computer science, probability, statistics, and tensor algebra" far exceeds supply.
- National competitiveness depends on rushing to "wire your country's governments schools and so forth fiber" and investing in specialized skills, with countries failing to act in the "short term" predicted to miss out on extra benefits in business and society.
- Specific applications anticipate "billions" of entries processed via simple procedures, with machine learning achieving "a 15 percent improvement in energy efficiency" in data centers and similar gains in US energy distribution, while traders using "continuous reinforcement learning" could match or beat the best human traders.
- Domingos notes that while computers will become "equal to or better" at "photo detection" and "self-driving cars" in the "near term," they remain limited in predicting events with little training data, and the "fraction of what's already happened" in machine learning adoption across industries is "absolutely tiny" with a "very long way to run."