The Path to Superintelligence☕
Google's DeepMind is exploring humanity's most useful invention...
If this was forwarded to you, subscribe below:
Hello Everyone,
"Artificial intelligence is the future, not only for Russia, but for all humankind....Whoever becomes the leader in this sphere will become the ruler of the world." — Vladimir Putin, 2017
Such a profound statement from one of the most commandeering leaders in the modern era cannot be glossed over and should invoke a sense of ambition and urgency in world leaders as well as scientific researchers and the private sector at large. Such foresight is not unique to Putin. China has invested billions to reach technological parity with the United States. In its “Made in China 2025” industrial strategy framework, the country said it aims to be the global AI leader by 2030.
Similar to jumping on the PR bandwagon and attracting more investment by adding a ".com" to your company's name in the late 90's, Artificial Intelligence (AI) is a term that startups and blue chip companies alike use to generate buzz whether they are implementing sophisticated techniques as a core part of their business model, using basic tools, or none at all.
In March of 2019, MMC, a London-based venture capital firm, released a report claiming 40% of European firms that are classified as an “AI startup” do not exploit the technology in any material way for their business.
This broad spectrum of claimed utilization makes it difficult to discern what AI is, what it can do, and who is working on it.
Any R&D that carries as much weight as it does in AI should not be done behind closed doors, so I dug into the research lab spearheading the advancement of the field.
Let's get into it.
Founded in 2010 by Demis Hassabis, Mustafa Suleyman, and Shane Legg, DeepMind is an AI research laboratory acquired by Google for $600 million in 2014.
DeepMind employs more than 1,000 people worldwide, including many of the world’s most highly coveted AI research scientists.
Google’s acquisition sheds light on how much they value research and development in artificial intelligence - $600 million is a drop in the bucket on their balance sheet but like most research labs uncovering commercial applications for new technologies, DeepMind operates at a significant loss; in 2019 they lost $649 million. Despite such large and persistent losses Google’s parent company, Alphabet, has pledged to keep funding DeepMind. When asked Google CEO Sundar Pichai said, “I am very happy with the pace at which our R&D on AI is progressing.”
When training deep learning models, you need lots of data because the systems learn from the data. For example, your interactions with Alexa or Siri are all based on deep learning and the more you use them, the better they become. Self learning models turn data into a very valuable asset. AI is only as good as the data it is fed, in terms of both quality and quantity.
Talk about synergies...DeepMind needs data and operates under the Alphabet umbrella. Alphabet has more user data than any company in the world and thus gives DeepMind access to troves and troves of high quality data. This allows DeepMind to work on the most complex and interesting challenges of AI while driving scientific advancement and having real world impact.
DeepMind came onto the scene in March of 2016 when its AI program AlphaGo defeated the strongest Go player in the world, Lee Sedol. Mind you, IBM’s Deep Blue beat reigning world chess champion Garry Kasparov 20 years prior. Chess is a game of strategy and calculations; while still a game of strategy, Go has room for much more intuition and subjectivity. We have known for many years how to program a machine to play a game by teaching what moves to make and when, but such teachings are not sufficient in Go.
In the video above, Lee Sedol is visibly disturbed after AlphaGo makes a very unconventional move. Since the move turned out to impact the course of the game and played a role in AlphaGo’s success, this was the first time an AI program produced an undefined result based on general rules.
DeepMind began training AlphaGo with roughly 30 million moves from recorded historical games so that it could mimic human play by matching the moves of experts.
Just 21 months later DeepMind revealed a new program called AlphaZero. AlphaZero was taught using only the rules of the game, reinforcement learning, and self play to train its neural network - human styles of play have no influence on its game. In the graph below you see the rapid pace at which AlphaZero surpassed the best Go players in the world and reached a virtually unthinkable level of play. This validates that when computers are given rules of a game and left to their own learning, the outcomes are much better than when humans intervene.
The scope of all DeepMind research projects extends far beyond playing a game and carries over to have real world impact.
One such application is AlphaFold, DeepMind’s artificial intelligence program redefining the way we think about biology and medicine. Proteins are the building blocks of all forms of life and AlphaFold can predict the physical structure of what a protein looks like. Each cell contains millions of proteins that DeepMind defines as, “the machines that allow your eyes to detect light, your neurons to fire, and the ‘instructions’ in your DNA to be read…”.
Humans have historically had a hard time understanding how genetic code translates into 3D models of protein structures. This is a very big problem that has been going on for decades and is key to understanding biology and how biology works. DeepMind has unlocked it with software and has made it freely available to the public. Such a database allows us to basically predict what a certain protein does, how it does it, and how certain drugs combined with those proteins affect the body. From there we could alter human health by making new molecules or adjust genetic code to change the shape of the protein in targeted ways for desired outcomes. Predicting the structures of proteins will help us find cures to diseases and create more effective medicines. This could usher in the age of personalized medicines, treatments, diets, etc.
The impact of these early use cases makes Putin’s claims credible.
This letter is a bit off course since DeepMind is not a private company anymore but I thought it would be an important one given they are the leading researchers in a very important technological field.
I think Putin’s statement is right in the singular form. But what is alarming is the state of the world is contingent on the controller of artificial intelligence. If the controller is good, utopia. If the controller is bad, dystopia. The results are binary and given the magnitude of potential outcomes, those are not odds we should play. The only sustainable way forward is to focus on the democratization of AI.
Happy Labor Day Weekend!
Until next time ✌️,
AC
If you liked this post, share it with your friends.
If you didn’t like this post, share it with your friends, maybe they will.