“It is time once again. I believe this nation can do great things if the people make their voices heard. The men and women of America must once more summon our best elements, all our ingenuity, and find a way to turn such overwhelming tragedy into the opportunity for a greater good and the fulfillment of all our dreams. In the months and years to come, there will be many battles in which we will have to be strong and we must give all of our energy, not to repel invaders, but rather to resist aggression and to win the freedom and the equality for all of our people. The destiny of the human race hangs in the balance; we cannot afford for it to slip away. Now and in the years to come, the challenge before us is to work out how we achieve our ultimate destiny. If we fail to do so, we are doomed.”
It sounds like typical speechifying, right? Something any politician would say, maybe after a particularly tragic bombing. Except that this chunk of text was written entirely by an AI, given the prompt of writing an Presidential acceptance speech by a John F. Kennedy brought back to life. If you read the whole sample, accessible here on OpenAI’s blog, you’ll see that it rambles about the nation being descendants of soldiers in some non-existent “Continental War” and gets confused about how large the human brain is and where it’s located in the body – but those were the only serious errors I could spot. In contrast, the AI model correctly identified Daniel Boone as an American folk-hero, correctly used Wikipedia as a (fake) reference, understood that someone working on nanotech would have a doctorate and likely work for a University, and acknowledged that the United States Presidents are some of the most famous people in the world.
So… that’s pretty crazy. Except that’s not all – the same model, without having been taught on purpose, can also summarize articles, answer questions about what was written, and translate English into French, albeit badly. And yet the public barely seemed to notice, and most of the conversation around the AI model was about how OpenAI declined to release it to the public, for fear of people using it for fake news.
It’s true that experts are not up in arms about this as some momentous breakthrough, but this is just the latest in a long line of AI achievements that would have been seen as ridiculous science fiction just two years ago – and frankly, seem that way even now if you haven’t been keeping up with the news. For example, try a rousing game of Which Face is Real?” in which you’re shown what appear to be two photographs. One of them really is a photograph of a real human being, the other was generated by a neural network. The game is startlingly difficult – I was right about 75% of the time, and I felt like I was guessing at random.
Here’s an example of the kind of fake you’re up against – only the odd material of the shirt is suspicious.
Or play against a canny AI in a game of Iconary, a slightly modified version of pictionary played with icons instead of hand-crafted images. The model both draws and guesses with astonishing accuracy, leaving me wondering if I was playing against a human several times. For another version of pictionary you can also play Google’s Quick Draw, which is basically classical pictionary where you’re always the one drawing. Simple doodles are solved in about 5 seconds or less, doing at least as well as a human would. Perhaps you prefer word games? Another Google entry, Semantris, will scratch that itch for you. Input text clues of any kind and the AI will guess which of the displayed words you’re hinting at with astounding accuracy. In fact, the network guesses so well that they added the additional challenge of the displayed words functioning a bit like Tetris blocks, with you aiming to keep them from stacking up too high. In several hours of playing the game nonstop, the neural network picked the wrong word perhaps twice.
Games have long been the benchmark we used to determine the smarts of AI. Now, it looks like our AIs are beating us often enough that we’ll have to find new means of measuring. With AI that can generate photo-style images, transfer painting styles, write almost like humans, manage teams in video games, and more, we’re far closer to the goal of “real” artificial intelligence than most people know.