Scientist Janelle Shane asked GPT-3, the most powerful text generation algorithm, to predict the future based on events that occurred this year. GPT-3 reported that in 2021, “alien-launched lasers will infiltrate the solar system.” After a while, the algorithm also told about other strange phenomena to happen: a small drone will be able to lift a shark aboard a ship, scientists will find a farm of alien ants under the Arctic ice, a mysterious hole will appear on the Sun, and a 3-meter penguin will be found on the Solomon Islands. And these are just some of the predictions.
How could the algorithm generate such predictions? Previously it was “fed” with various texts and headlines from the Internet, such as mysterious monoliths, deadly hornets, alien signals from outer space, etc. Relying on such articles, artificial intelligence has learned to generate its own fictional news. This experiment showed that “even the most sophisticated algorithms, such as GPT-3 or GROVER, do not fully understand the meaning of the results they generate, even if the text itself turns out to be coherent.
The GPT-3 algorithm – what is it?
GPT-3 is an algorithm from OpenAI, an artificial intelligence company. One of its founders was Elon Musk. GPT is the largest and most advanced language model in the world. It works like any other machine learning model – it looks for repeating patterns in a data set to predict the result of introducing new, previously unknown information. If you enter the word “water” into GPT-3, the system will know that the words “mineral” or “carbonated” are more likely than, for example, “computer-generated.” This is very simple for humans, but the algorithm needs to “learn” such relationships.
GPT-3 also has the scale and speed of data processing – it has 175 billion neural connections, a gigantic database, and constant access to the Internet.