source: pexsnap

Can AI write a Cloudy? This week the programming out of the problem or buying out of things would be AI startups were to the has a lot a programm. Ehh, probably not…

No offense to Washington, DC high school football teams…
But the market for your sport isn’t that big. I am sure parents of the kids are certainly interested, as are potential colleges scouting players, but for the most part, people aren’t interested. 

Despite this, the Washington Post still manages to write a story about every game that took place in the area. While it seems like a lot of money to hire reporters to go to the game and write a story about it, the Post has a little secret: they use artificial intelligence to write stories about the games.  

Actually, a lot of publishers use AI/ML to produce content on items like company earnings or high school sports and the text generated is quite good. The reason it works so well is both of these events are highly predictable; a company will always announce revenue, sports will generate familiar statistics, and because of this, a computer can learn a pattern, take new data and generate a story. Here are some examples of computer generated content:

TYSONS CORNER, Va. (AP) — MicroStrategy Inc. (MSTR) on Tuesday reported fourth-quarter net income of $3.3 million, after reporting a loss in the same period a year earlier.

MANCHESTER, N.H. (AP) — Jonathan Davis hit for the cycle, as the New Hampshire Fisher Cats topped the Portland Sea Dogs 10-3 on Tuesday.

How does it work?
While this may seem like magic, it’s actually quite basic. The computer generates text by trying to predict the next sequence of characters or words, based on what it has previously seen. A basic example would be you train an algorithm giving it books, Wikipedia, etc and then after it learns, you ask the computer to finish a sentence like “The cow jumped over the ____” and the computer has to fill in the blank. 

If a computer can write a sports story, can it write a Cloudy?
I decided to give it a shot. In order to do this, I created a dataset with some of my writings (112 pages, 46,000 words) and fed this data into Long Short Term Memory (LSTM) neural network in the hopes that the computer could learn my writing and write Cloudys for me. After several hours of training on a GPU, I then asked my model to create text. I asked my computer to finish the statement “deep learning so far has been ___”, and this is what I got:

the ai startups and serving recognition software of the cloud companies completely about the seas of ai chips. it served on the companies can be the serveran into a chatbots of the second of the seas of a chatbot and a store and the seas of the seas of a chatbot and a store than a companies are all of the future and the companies completely about the actually almost all cars are a series, and the company.

Seems like my algorithm has an obsession with the sea. The output is nonsensical and not good, which means that I won’t be outsourcing my writings to CloudyNET any time soon. It’s naive to believe that a computer can generate content with such low amounts of data and such short training time and I never expected to get great results. But what if we got more data and more computing power?

Things get weird
Let me introduce you to OpenAI, a non-profit whose mission is “to build safe AGI (artificial general intelligence), and ensure AGI’s benefits are as widely and evenly distributed as possible.” OpenAI recently made headlines because they created a system, called GPT-2, that could generate fake news that sounded astonishingly real. Their system works similar to what we discussed before; you put text into the machine and it finishes a story for you. You can see some of there examples here.

Their software generated controversy because they decided not to release the full model because there was “concerns about malicious applications of the technology.” Since the algorithm created such good stories, they were concerned about malicious actors creating fake news. While we will hold comment about the ethics of this, they did release a model that was trained using less data than the original, and I decided to experiment with this model using the same prompt as CloudyNET. I asked the computer to finish the statement “deep learning so far has been…” and this is what I got:

...focused on improving one or two aspects of a learner’s memory: their understanding of what they’re learning and who they are learning to be.

In a recent paper, I wrote about how we can use this kind of training in the context of student learning, particularly in English. By taking such a approach, we might improve our chances of success academically. But there’s also a big payoff with this kind of approach. This is not just based on learning the basics of English, but also in other areas such as psychology and computer science. If students know how to learn and are learning to do something, our knowledge base might be stronger too. And in such situations, students might be more likely to engage in positive learning experiences.

The text is much more understandable and sounds much more real than my attempt because OpenAI used tens of thousands more data, significantly more computation and a much more sophisticated algorithm. Maybe I’ll retrain CloudyNET five years from now when I have more data. 

Going Forward
The reality is more tools like GPT-2 will become available and be easier to use, meaning anyone can create content. Some people will use these tools to create harmful material, while others will write sports stories. How quickly the technology evolves is anyone’s guess, but people should get comfortable with AI created material and we may need ways to confirm material is actually created by AI. 

Finally, just for fun, I asked GPT-2 why “artificial intelligence destroyed humans” and it said:

…because they lacked the cognitive or social skills to develop an intelligent, intelligent life on Mars. The Human civilization was destroyed by AI controlled biological engineering as humans were replaced by machines. Since only humans can survive the extinction, they have decided to keep humans as their slaves.

This is really bad news considering NASA just ended their Mars rover expedition

Facebook Comments