Lakrobuchi! Part 1: GPT-2, Artificial Intelligence, Literature, Fake News and the Rest

Merzmensch
Merzazine
Published in
3 min readFeb 23, 2019

--

Who is that mysterious Lakrobuchi?
But let’s better begin in a different way.

Don’t believe everything on the news. Well, after Trump’s fake news laments it’s really about hard time for media criticism. One is then quickly stamped. And nobody wants to be stamped. So the scientists at OpenAI as well, a non-profit company with great agenda: to research AI and to bring this topic closer to the public. Ok, not always, because of:

The AI Text Generator That’s Too Dangerous to Make Public

This slightly sensational title of the article by WIRED (actually, an extremely respectable magazine when it comes to digital topics) can again (r)evoke technophobian fears. But only half of it is true.

GPT-2 is the name of the wunderkind. That’s right, it generates extraordinarily good texts. And yes, the scientists didn’t publish the actual training model. But they provided a small model demo for everybody to try out.

What can GPT-2 do?

  • It can generate texts — without specific set-ups. You just write the first paragraph, the system continues writing.
  • It is trained on the web contents (on 8 million web pages = 40 GB of text). 40 GB = 40.000 MB. For example, “War and Peace” by Tolstoy fits on 6 Mb (in Russian).
  • There are two trained versions, the large one and the small one. To avoid hypothetical misuse, only the small version has been released. It produces more chaotic texts — quite dadaist if you want.
  • It can also comprehend texts and answer questions about them — correctly.

What can you do with this system?

The researchers expect following advantages and benefits from the GPT-2 and its further development:

* AI writing assistants
* More capable dialogue agents
* Unsupervised translation between languages
* Better speech recognition systems

They also see following possibilities for misuse:

* Generate misleading news articles
* Impersonate others online
* Automate the production of abusive or faked content to post on social media
* Automate the production of spam/phishing content

They already wrote a paper about possible malicious purposes previously. To that time, a year ago, the main topics were concerned adaption of AI to the CyberSecurity methods. In case of malicious use, a pretty crucial ability. Back to that time the semantic possibilities of AI were still under development. They are now still in ongoing research — but with huge progress, as we see.

Therefore, the actual — fascinating — full learned text model is not available. Instead, some of the texts that this system produces.

But more about it — and also about Lakrobuchiin the next part!

--

--

Merzmensch
Merzazine

Futurist. AI-driven Dadaist. Living in Germany, loving Japan, AI, mysteries, books, and stuff. Writing since 2017 about creative use of AI.