Developments in artificial intelligence that can write authentic-looking stories took a big leap forward earlier this year with the release of GPT-2, a predictive text-creation algorithm that can create a coherent, realistic-sounding continuation of a text input as short as one or two sentences.
In other words, the software looks at a text you write and predicts the next words and phrases. It uses a huge data set as its training source: 8 million web pages which have been curated or filtered by humans. And it does this by itself with no human involvement at all.
OpenAI, the company behind GPT-2, initially released small versions of the model intended only for researchers, citing “concerns about malicious applications of the technology” (basically, saying that the risks of people creating fake news at scale were worrying).
Then last week, OpenAI released the full-sized GPT-2 so anyone can experiment with it. It seems concerns about misuse were over-estimated.
How Clever Is It?
I was keen to kick the tyres, as it were, but without having to try to grasp the complexities of this marvel of machine learning. I just wanted to see what it would do if I gave it a short paragraph.
In this latest installment of our Small Data Forum podcast, episode 30, we again discuss the political shenanigins in the ongoing Brexit saga as we now head towards a General Election in the UK on December 12.
GPT-2 took that paragraph (shown in bold below) and created this start of a story:
In this latest installment of our Small Data Forum podcast, episode 30, we again discuss the political shenanigins in the ongoing Brexit saga as we now head towards a General Election in the UK on December 12. It’s not just about the EU; it’s also a proxy war for the rise of populist anti-establishment politics.
It’s a subject that resonates strongly with many of our listeners in the U.S. as well as abroad. We have also been asked for a podcast on this very subject, so we have one coming very soon.
Subscribe to our Podcasts:
Here’s a screenshot of the actual result:
On a first read, it all reads well although perhaps you might just wonder a little at some of the wording, the sentence flow. Still, the grammar is good and it scores high on my credibility meter that it wouldn’t occur to me to think that this text was written by anyone (or anything) other than a fellow human being, one who perhaps isn’t a brilliant writer.
But the second paragraph did grab my attention with its ‘opinion’ that we’ve been asked for a podcast on the subject being discussed. Well, that’s true – we’ll be recording the next episode in December and it will indeed continue our political story – but I didn’t expect to see that level of cognitive ability to be able to create such a story element just from the original paragraph.
While this technology may not yet be prime-time material, you can be sure it will improve over time. In the coming months I expect to see stories of other people’s tests and experiments, all adding to the knowledge and understanding of the capabilities of this technology and how people can use it to their advantage (for both good and not good, the reality of today’s world).
As ever, consider the abbreviation ‘AI’ to mean ‘augmented intelligence.’