What is the GPT-2?

Last week, Open.ai, which was an Elon Musk-backed AI company, released research that illustrates the capabilities of its’ AI system called the GPT-2. While this does represent an impressive achievement in with regards to unsupervised learning principles, it also raises a key problem with systems that are structured in this way.

If we circle back to any one of our explainer pieces on the subject of unsupervised learning vs. supervised learning, you will remember that AIs that use unsupervised frameworks to learn, take in data without any specific directions on how to use it. According to sources like the Towards Data Science blog, one of the key benefits of using this sort of structure is to allow the AI system to identify possible patterns and correlations in data without any bias being injected into the process. In other words, without the opinion of the AI team, the AI system endeavors to find patterns on its’ own.

Reverting back to the specific case of the GPT-2, matters did not exactly turn out well. During the same week that the previously mentioned research paper was released, Elon Musk reportedly fully quit the Open.ai team, though according to his Twitter(which is linked below), he had not been actively involved in the company in about a year. This, of course, fueled speculation that he quit because of how “dangerous” the GPT-2 truly is, though he quickly endeavored to dispel this notion directly.

So, while there seems to be little truth to the idea that Elon Musk knows some terrible secret about Open.ai’s system that has been kept from the greater public, a serious issue has come to light after the paper’s release.

This begins with the fact that the GPT-2 can apparently already generate easily understandable paragraphs of text on its’ own. To understand where any negative implications might lie, it’s important to begin with how the GPT-2 was trained, before transitioning to what has seemingly already resulted from this training. In general, the system was trained on the equivalent of 40 gigabytes of “internet text,” which has resulted in 50% accuracy in producing long paragraph-length responses on varying subjects.

If you’re interested in exactly what this means, you can check out some examples which come directly from the team here, or below. Suffice it to say for now, however, that judging by the evidence at hand, this means that 50% of the time, Open.ai’s system responds to questions or supposed statements of fact like a human and yet, simultaneously, in a sort of terrifying way.

For example, when the system was prompted with the sentence, “recycling is good for the world,” the top line of its’ response read, “NO! YOU COULD NOT BE MORE WRONG!” What followed was a critique of recycling in which the system appeared to conclude that recycling helped bring about hikes in production costs, obesity, and even global warming. In the end of its’ paragraph-long rant, it confusingly ended the discussion with the statement that “if we want to keep the recycling process running efficiently, then we really need to think about each and every step that goes into making a paper product.” With this, we can see the veracity of the statement that the GPT-20 is only about 50% accurate with its’ responses. It began by lambasting recycling as a harmful force and ended by painting it as something that may be good, but is not yet sustainable over the long term.

Other examples of the system’s responses to questions or statements of supposed facts get even more fanciful, including its’ answer to the idea that unicorns exist, which seems to be the system’s idea of their history. Why any of this is an issue for the global populace is simple.

The GPT-2 is already well on its’ way to becoming the world’s most dangerous fake news generator. In the simplest sense, of course, fake news is the most dangerous as it becomes more believable and harder to argue against. Considering this, the unicorn story is really not much to worry about, but the recycling-related post raises a good point.

If Open.ai allowed the GPT-2 unfettered access to the internet, how long would it be before enough politically-related information to start developing its’ own ideology about how the world should be run? Before we jump too far down that rabbit hole, let’s pump the brakes for a second.

As of now, there is no indication that the team will do anything like that. On the contrary, they have publicly stated that they do not plan to release the system in any sort of full capacity anytime soon, because they believe that it might have too many negative uses, at this point.

Faced with statements to this effect as well as the troubling facts surrounding the GPT-2’s performance, the media has criticized the release as being nothing more than a “publicity stunt” to draw more eyes to Open.ai’s work.

In the end, whether or not this is true, the release of the paper and the limited examples of the system’s performance data seem to have been ill-advised. The GPT-2 was developed to come up with a better way of generating predictive text and yet, its’ problems might now outweigh its’ benefits.

Truthfully, since it is almost impossible to make a more definitive conclusion on this subject without full access to the project, perhaps Open.ai should have deliberated further before even structuring itself as a non-profit that aims to do all of its’ work in an open-sourced fashion.

Keeping the majority of the system’s code close to the breast is far from being “open-sourced.”

Perhaps, through the example of Open.ai’s experience, other teams will think more carefully about not only how they structure their companies, but also, how and when they release any data on their products. As this saga unfolds, expect us to keep you informed as any major developments come to light.