Humans are difficult. Especially when we start telling jokes and using cultural references. Precisely understanding human humor is one of the great challenges for machines, but now Google has shown great progress with its latest artificial intelligence.
The Google Research team has announced its new Pathways model, capable of understanding more than 540,000 million parameters and managing to understand concepts and relationships that until now seemed too complex for computers.
10 GOOGLE APPS THAT COULD HAVE SUCCESSFUL
The challenge of understanding the logic of 'cause and effect'
The new model uses the largest configuration of TPUs to date, combining up to 6,144 chips. The AI has been trained on English language material of all kinds, from books to Wikipedia, along with conversations and code from Github. And the results, as described by Google, outperform other models such as GLaM or GPT-3 in tasks such as reasoning, answering questions or understanding the context.
'Pathways' can distinguish cause and effect and understand combinations of concepts in the appropriate context. For example, the relationship between efforts and results. Both sentences are technically correct, but all humans can understand the implications.
Another ability of the new Google AI is to be able to guess a movie from a single emoji, for example that of a bat or a lion.
What is circulating the most is that "he explains jokes". And this is awesome: the jokes they show are basic but it is the milk that can "understand" the combination of concepts in the context of trying to surprise and amuse a human mind pic.twitter.com/bkCV5Exd8t
– Antonio Ortiz April 5, 2022
The model is capable of reasoning about the answers it gives and is also capable of solving simple mathematical problems. We are not talking directly about a formula, but about being able to understand the approach of the question. According to Google data, it is close to the average of 60% of problems solved by children between 9 and 12 years old.
But perhaps the most striking ability is that of being able to understand and reason where the jokes are funny. PaLM, technical name of the model, is capable of explaining jokes that it has not previously read.
Julia Taylor Rayz, a professor at the Purdue University Polytechnic Institute, explained in relation to the difficulty of jokes that "there are no clear rules in human communication, what are we going to tell the computer to do, find rules that don't exist? […] You can't find enough examples that describe all possible communication scenarios."
But studies are advancing and artificial intelligence, thanks to the enormous base with which it can learn, is taking giant steps when it comes to understanding how we humans express ourselves.