Artificial Intelligence is getting ‘scary good’ – four things AI programs can beat humans at
ARTIFICIAL intelligence systems have mastered some of mankind’s best creations and natural intuitions.
These AI systems notched some of the first wins for the machines.
Down goes the chessmaster
Artificial intelligence and table games make a good pair because humans have been trying to develop perfect tactics for strategy games for decades or even centuries.
Chess is “known as a game that requires strategy, foresight, logic—all sorts of qualities that make up human intelligence,” IBM researcher Murray Campbell told Scientific American.
Campbell and a team developed Deep Blue, a six-foot supercomputer that defeated chess grandmaster Garry Kasparov in a six-game series in 1997.
During the pivotal final game, Deep Blue made a move that Kasparov thought only a human could rationalize – Kasparov insisted the IBM team cheated, which they denied.
Read more on AI
Deep Blue would make 100million calculations a second to select its attacks but an early move that splintered Kasparov’s confidence was actually the result of a bug that caused the computer to choose a move at random.
Data journalist Nate Silver’s book on analytical forecasting says that Kasparov’s over-analysis of a “last-resort fail-safe” move may have cost him the tournament.
PokerBot
In chess, both players have access to all of the activity unfolding in the game – a player could mislead another into making a mistake, but both players can see and assess the whole of the board.
Texas Hold ’em is a card game with random draws, hidden information, and deception, making it an ideal playground for sophisticated artificial intelligence modeling.
Read More on The US Sun
The New York Times reported that even a simplified, two-player version of Texas Hold ’em with fixed bet amounts has 316,000,000,000,0000,0000 different potential outcomes.
Researchers at Carnegie Mellon University built Libratus, which defeated four of poker’s best in head-to-head matchups over the course of 120,000 hands.
In 2019, engineers leveled up with Pluribus, their next iteration of self-improving poker-playing AI – Forbes reported that Pluribus can reflect on previous moves and act on the data.
Pluribus cleaned up five other human players at one table, marking the first time an AI program had beaten multiple players at the game.
Human professionals will try to replicate the AI-powered strategy by studying a model’s calculations and the poker-playing community has to keep up with banning players using AI in games hosted online.
The robot that taught itself to walk
Morti is a AI robot dog that learned to walk quicker than a human.
“Our robot is practically ‘born’ knowing nothing about its leg anatomy or how they work,” Felix Ruppert study co-author, told The Independent.
As people and animals learn to become independent, the Central Pattern Generator (CPG) in the spinal cord communicates with the limbs and muscles to move.
Ruppert and study co-author Alexander Badri-Sprowitz built their “walking intelligence” system in a computer to model a naturally occurring CPG.
To learn how to walk, Morti’s “virtual spinal cord” would check the pressure of a step against the CPG’s predictions.
If Morti fell, the computer’s algorithm would adjust the pendulum swing of its legs or the speed or time spent contacting the ground.
Human babies take about a year to learn how to walk because humans are less far along in developing their intellect at birth.
Morti differs from game-playing bots because the robotic dog is a demonstration of an artificial intelligence program powering movement, potentially resolving robots’ history of clumsiness.
Reading comprehension
Evaluating reading comprehension is a staple in the American education system.
Reading is nearly twice as efficient as listening and it could be argued there is no single more important trait in learning than understanding what we read and being able to recall it.
In 2018, engineers at Microsoft Research Asia in Beijing built an AI bot that could read and understand just as well as a human.
The Microsoft model was tested against the Stanford Question Answering Dataset, a reading comprehension test made of questions based on Wikipedia articles.
Machine reading can be unsettling because while not everyone plays chess or poker, humans are biologically programmed to try to interpret letters and numbers, according to a study published in ScienceDaily.
Microsoft’s machine reading model is available for public use on their site – plug in your own text and ask it questions to test the model’s understanding.
Read More on The US Sun
Each of these systems represents a type of Artificial Narrow Intelligence – an AI system that’s programmed to do one thing exceptionally well.
The next stage, artificial general intelligence, would be a computer that can do anything as well as a person, including reason, deception, and other more abstract human abilities that are not purely computational.