6 Things AI can do now That it Couldn’t do Last Year

While we haven’t achieved AI the likes ofHAL 9000orDeep Thought, neural networks pulled off some incredible things this year.

Though there were many more accomplishments, there are 6 things AI can do now that it couldn’t do last year that really indicate 2018 will bring us closer to our robot overlords.

Technological progress isn’t as exponential as it used to be. But aTeslacar using the autopilot function drove a man to a hospital in2016. The man credits this tech with saving his life. AI also predicted the 2016 election and increased correct cancer diagnoses. Swarm AI even predicted theKentucky Derbywinner.

What did AI do in 2017?

Number one on our list of more accomplishments, there are 6 things AI can do now that it couldn’t do last year might seem silly. But, it was 2017: a year of notable cyber attacks, a 5-year Mars colonization plan, and everything in-between.

WhileErica, the autonomousandroid, stole many hearts this year, another hyper-realistic AI made news. Hanson Robotics createdSophia, the learning robot who received citizenship inSaudi Arabia. Ignoring the controversy of that decision, Sophia is just that: arobot.

While she can learn and tellsjokesonGood Morning Britain, is she an AI? EvenBen Goertzel, chief scientist of Sophia, says she isn’t true AI. But sheisintelligent and, thanks to formerDisneyImagineerDavid Hanson, she has a human face. She can hold a conversation and respond to various sensory inputs more fluidly.

Of course, every conventional AI has programming, or some sort of supervised plan or method of input. Sophia is able to respond and express herself in a more human way thanks to three key facets:

Despite consistent reductions in budget over the years,NASAremains one of the leaders in space exploration technologies. Discovering new celestial bodies and learning more about humanity’s place in the universe is always exciting. It is also easier these days thanks, in part, to the use of AI.

Neural networks sifted through years of data to help scientists locate anew planetin theKepler-90system. The unprecedented assistance led to the identification ofKepler 90i. Even just a few years ago, automated processes could not discern an exoplanet from orbiting stars.

What else can AI tell us about exoplanets by helping human scientists sifting through and narrowing down data? Perhaps that will be a feature of our article detailing AI accomplishments in 2018.

When AI beat the reigningGochampion in 2014, researchers rejoiced. In 2017, a new neural network–AlphaGo Zero–beat the old AI 100 times in a row. Not only that, ittaughtitself how to play using only the basic rules of the game. But move over Zero, there is a new gamified AI in town.

The computer science department atCarnegie Mellon Universitydeveloped anAIthat won in no-limitTexas Hold ‘Em. The impressive thing here is not so much that it won at Poker, but the type of Poker at which the AI won. In 120,000 hands, Libratusdefeated four Hold ‘Em experts in what is known as an “imperfect information” game.

Just like Sophia, Libratus uses three key algorithms in interpreting inputs and adapting:

Again, this development ishugebecause of how important bluffing is when it comes to Poker. The strategies Libratus showcased didn’t just apply to Poker either–they could apply to other “imperfect information” games.

Our list of 6 things AI can do now that it couldn’t do last year includes being able to write itself. What a time to be alive, no?

Neural networks can now teach themselves a host of abilities. One such ability: coding. Though the feat itself is impressive, more than one company developed AIs with this ability. MicrosoftandGoogleboth have respective AIs that used machine learning to learn how to code.

AutoMLandDeepCoderare totally unprecedented in their capacity for future abilities. For DeepCoder, rather than copying other code, researchers provided the building blocks of code. The algorithminferredhow the code fit together to function and could learn to recognize other code.

AutoML seeks to develop AI to inventbetterAI. If your Skynet alarms are going off, maybe you should jump ship on this article before thing 5.

The tech development heard round the world:FacebookAI invents its own language. Contrary to popularmyth, it was not this reason that Facebook shut them down. While theimplicationsof an AI circumventing humans for communication are terrifying, the experiment failed.

The issue lay in the fact that the purpose of the AIs was to understand human language. Speaking their own language renders the main goal of improving AI-to-human language impossible.

The distinction between artificial intelligence and robot is beginning to blur. While robots refer to physically capable machines that only perform programmed functions, AI typically refers to a set of algorithms that take advantage of machine learning or neural networks to adapt. That’s why number 6 on our list of 6 things AI can do now that it couldn’t do last year involves how robots might evolve into AI.

While therobotguards inSan Franciscocan’t learn yet, it is a huge step toward autonomous security. TheAV1, a robot to end isolation, functions as alimitedAI and comes with a companion app. It responds to sensory inputs. But even compared to this kind of robot,chatbotsfunction much more like AI.

Don’t think assistants likeAlexa,Siri, orGoogle Home. There are chatbots designed to embody specific personalities such as a flirtatious woman, a 31-year-old hacker/computer analyst, and more.

You can find less quirky chatbots utilized in apps likeFacebook Messenger,Slack, and others. In fact, the AI chatbot industry is poised for majorgrowthin the coming year. You can even build your own chatbot withplatformslikePandorabots.

Perhaps one of the biggest impediments to AI and machine learning ishuman bias. If machines learn by ingesting biased data, it follows that anything learned would extrapolate skewed information into macro processes. To that end, who’s responsible to account for this bias?

In fact, long-time AI researchers contend that researchers may need to start overentirelyto create true AI. Regardless of the path, we are certainly approaching tremendous breakthroughs in artificial intelligence as we enter 2018.