AI in 32 Bullet Points

This is my raw braindump of what’s happening in AI right now and why it’s fascinating.

N.B This is for AI non-experts and is 100% a ramble.

1. There is no evidence or precedent for an “explosion of intelligence”. History of self-improving systems suggests that increases in intelligence are linear not exponential. This is due to intelligence being situational rather than general. This is an argument, against AI taking over the world, nicely laid out by François Chollet here.
2. TL;DR: Humans are as intelligent as their situation/environment. As we are mortal, most human intelligence is stored externally in the form of laws, institutions, science, products, businesses etc.
3. TL;DR 2: We then leverage these external foundations to improve our intelligence by advancing incrementally from generation to generation. No intelligent system has ever created something which is exponentially cleverer than itself.
4. A good amount of the AI development in the past few years has been centered around supervised learning. Developing AI with the help of two things: 1. Training data of human experience for e.g Poker hands, Go/Chess match history. And 2. Basic parameters: e.g rules of game + truths ([Chess] if check-mated = lose, if have check mate = win).
5. The first iteration of AlphaGo is a good example of this. It was fed human Go match/training data as well as certain parameters.
6. Now we move on to more recent AI development and in particular the success behind “Self Play” training and AlphaGoZero in 2017. This new AI only had the help of one, rather than two inputs.
7. It was simply fed parameters of the game: rules and basic truths about Go i.e surrounding a larger total area of the board with one’s stones than one’s opponents = winning. No rich training data, match histories or human strategies like its predecessor. It learnt to play by trial and error. This is what happened:
8. “Over the course of millions of AlphaGo vs AlphaGo games, the system progressively learned the game of Go from scratch, accumulating thousands of years of human knowledge during a period of just a few days. AlphaGo Zero also discovered new knowledge, developing unconventional strategies and creative new moves that echoed and surpassed the novel techniques it played in the games against Lee Sedol and Ke Jie.” (Via DeepMind research).
9. 😮
10. First off: “accumulating thousands of years of human knowledge during a period of just a few days”…
11. AlphaGoZero became superhuman in 3(!) days. It then became SuperAI (meaning it beat the previous best version of AlphaGo: AlphaGoMaster) in 40 days, 100 games to 0.
12. In contrast, it took the original, “supervised learning” based AlphaGo, two years to become superhuman. TWO YEARS. So it can be said that the human training data was holding the AI back. AlphaGoZero’s ability to self-play with a fresh pair of “eyes”, and learn the ins-and-outs of the game through trial and error, yielded massively improved results and training time.
13. So just because the most intelligent humans, in a certain field or game, are doing things a particular way, doesn’t mean that it is even close to being the best way to do it. This is a simple point, but an important one.
14. Through computational efficiency, the AI was situationally more advanced at acquiring ability than any human. The Self Play AI’s are able to level-up much, much quicker than humans (000’s of games run simultaneously for the AI vs usually 1 at a time for the human).
15. So… back to №1
16. Does this qualify as an “intelligence explosion”? An exponential rather than linear intelligence increase?
17. Can be argued that the accumulation of intelligence up to a point has exploded at an unprecedented rate. AlphaGoZero got as good at a game in 3 days as any human does in a lifetime, which doesn’t sound very linear (or fair 😤) to me. But how much better did it get? How much more intelligent did AlphaGoZero get than a human Go player?
18. Turns out this is a tough question to answer and something which is very hard to measure past rating it super-human. It either beats all humans, or it doesn’t. Making it tougher still to answer is that this AI has so far only been applied in games with non-messy defined rules and parameters. What happens when these neural nets learn to compete in more complex environments? Will it then be easier or harder to measure how much more intelligent the AI is than the human?
19. Current Self-Play AI can become superhuman in record time in a few games: Chess, Go, Poker (NLH and Limit Holdem) and a very specific 1 vs 1 mode of DOTA. You can read more about Elon Musks Open AI DOTA bot here.
20. So at the moment AI success remains limited to 1 vs 1 contests. The number of variables in 5 vs 5 DOTA or multi-handed No-Limit Poker for e.g exeeds the current computational/neural-net ability of any AI in existence.
21. But it feels like just a matter of time and computational/algorithmic development for these AI’s to have their neural nets applied successfully to more complex and messy environments for e.g stock markets, self-driving safety, surgical procedures, disease screening/prevention, CRISPR tinkering etc. There are already IBM Watson powered ETF funds, and self-driving car makers are will be grappling with the Trolley Problem as I write this.
22. So is the potential of AI as a shortcut to get a bit cleverer than humans almost instantaneously, but still with hard situational limits that come from existing on earth with the same bioligical/chemical/physical laws as us?
23. As Chollet notes there is no precedent for self-improving systems improving at exponential rates, meaning that in theory, it should be impossible for one entity, human or artifical, to create another form of intelligence exponentially smarter than itself…
24. And this is where predictions get murky as we get closer to this end of the is-this-possible spectrum.
25. So what will humans do with these AI breakthroughs in the short term?
26. The best Chess, Dota and Go players are already training against their counterpart AI’s. So in games/areas where AI is superhuman, humans will be able to learn from the AI in order to be only a bit dumber than the AI, but at least smarter than other humans. Yes that sounds weird to me too 💁‍♂️.
27. DeepMind is applying its technology to a bunch of interesting and important areas like breast cancer prevention and other medical uses.
28. I’m also weirdly excited about it becoming normal for AI to train humans. For now only useful in certain games such as Poker, Go, Chess, Dota etc.
29. Will these breakthroughs in self-play training influence how we teach humans? In the future will we have AI training courses that facilitate more personalised, trial-and-error learning as opposed to the current follow-the-directions, IKEA instruction-esque standardized approach of most western education?
30. It’s funny how parallels exist between the current, tired methods of how we teach kids and the supervised learning of AlphaGo version 1. Supervised learning in schools looks like: the two inputs 1. (training data) exams/workbooks and longform multiplication for e.g. and 2. (parameters and rules) 1+1=2, 2 x 2=4 for e.g.
31. More interesting still are the parallels between the potential for personalised human education and the “Self Play” AlphaGoZero-esque training: Basic parameters and rules, no training data, just experimentation based on how each student thinks/learns.
32. A more personalised approach to education (if practically possible), and accepting that standardized training data isn’t necessarily optimal for every individual, would result in more trial-and-error randomness and potentially more abstract and advanced learning progression a la AlphaGoZero (albeit over a period of years rather than 3 days 🤷‍♀).

Brain dump over.

Matthew Willis December 2017

This story is published in The Startup, Medium’s largest entrepreneurship publication followed by 273,384+ people.

Subscribe to receive our top stories here.