Google and Continuous Improvement

Today’s keynote at the Singularity Summit was Peter Norvig, Director of Research from Google. His talk was titled The History and Future of Technological Change, and he couched his presentation as an analysis of “how to evaluate technical change”.

This is the first time I have heard Norvig speak, and I found his talk to be extremely pragmatic. His trek through the art of predicting the future, to demonstrations of narrow AI to his list of AGI prerequisites pointed toward a technologist with a perspective firmly grounded in continuous improvement, averse to making high-risk, long-shot bets. If Norvig speaks from a place of authority on Google product direction, it seems to me that we should expect continued evolutionary innovation from GOOG, but they will leave the breakthrough innovation of AGI to others. This is an important observation for the investment community that has put Google on pedestal related to the continued release of major breakthroughs.

Norvig began his talk discussing how the predictions he was used to making are about incremental advancements in technology. A 1% improvement here, a 2% improvement there. He pointed out that predictions about AGI are 100% “or greater” improvement ruminations. He pointed out the dichotomy between other prognosticators. “We will all be dead in 100 years” vs. “We will live to be 1000 years old”. “AGI can’t happen for another 100 years” vs. “within the next 10 years”.

From there, Norvig took a detour through other concepts of “Artificial General”. He postulated about “Artificial General Space Exploration”, “Artificial General Materials Science”, and “Artificial General Culture” – equating these concepts to the emergence of AGI.

Here Norvig was at his most pragmatic. He sees continuous innovation in these areas bringing about a more advanced capability, but certainly no “rapture”, no “big bang”. He commented that “the Singularity is a period, not a point”. He sees a date in the future when we look back at the progress and say – wow that was a big change.

In preparation for this presentation Norvig used Google Scholar to query papers presenting breakthroughs in AI. His keywords were “AI” and “unlike previous”. From 1968 – present, Norvig can’t tell the difference in breakthrough claims, with claims of novelty repeating in the data set. This indicates to him that we are not on the verge of discovering something major in AI.

To bring about an AGI, Norvig offered his list of prerequisites:

• Probabilistic First-Order Logic
• Hierarchical Representation and Problem Solving
• Learning over the data from above
• With lots of data
• Online
• Efficiently

I think the recursive thinking nature of Norvig’s AGI underpins his continuous improvement philosophy, and also presents a very Googlian view of success. Let an algorithm loose on lots of data, and eventually it might get there.

Rodney Brooks asked Norvig a question during the Q&A session:

Brooks – Any emergent property of Google materializing within the massive systems that has been a surprise?

Norvig’s best answer was that he was surprised at how Game theoretic Google’s role in the internet is. Initially, he thought Google would be an observer of the internet – just serving up search results. Now Google is co-evolving with the web.