On top of that, my colleague was very enthusiastic. AmberScript, he tweeted, is self-learning software that can turn audio fragments into written text extremely quickly. Fantastic! I rushed over to the website, paid fifteen euros and excitedly uploaded a fifty-minutes-and-forty-four-second conversation to the cloud.
The result was astonishing. Until then, I thought I had talked to a professor about artificial intelligence, but it now became clear that they had said something else entirely: ‘You have to analyse the number of words by how well a method works. That is mainly the après ski. Places. And it. Can warn the condom.’
Many fragments reminded me of the secret messages some people claim to hear when they play the Beatles’ final record in reverse or of Dadaist poetry at its most recalcitrant: ‘That will toothache. Instead of something cat. But that does not mean it is difficult. And.’
I find that ‘and’ at the end of the sentence to be especially touching. I had been tasked with writing an article about the ideas of professor Johannes Schmidt-Hieber, which you can find elsewhere on this website. Our conversation had been about the importance of mathematically assessing self-learning algorithms using comprehensive theories. Furthermore, the professor warned against allowing expectations to become too high: sooner or later, that will only lead to bitter disappointment, after which no one in their right mind will want to invest in the AI technologies that are so popular at the moment.
Something similar happened in the 1950s and 1980s, when the field experienced a cold ‘AI winter’. While Elon Musk predicts that computers will soon surpass human intelligence, people like Schmidt-Hieber are trying their best to keep everyone's expectations in check – while getting their thermal socks ready just in case.
That is what we talked about. I had hoped that AmberScript's self-learning algorithm would make the transcription process a little easier. Perhaps it is for the best that it did no such thing. We scientific journalists play an important role in creating unrealistic expectations for the future. I am here to tell you that it will be a long time before a self-learning algorithm can make complete sense of spoken words. What it is good at, however, is spitting out a catchy phrase to end an article with. ‘This too a great opportunity together a next generation,’ the algorithm preached. ‘But also stood a Cactus Club.’
