When an online service or company has been acquired by another company, they follow a similar pattern.
That pattern is by the time those companies announce that they've been acquired, there is a high tendency for them to also announce that they'll be shutting down.
That same formula has been repeated over and over again for many past acquisition announcements, and probably also for more future announcement posts.
With that fact, 'ThisStartupAcquisitionAnnouncementDoesNotExist' (Acquisition blog post generator) was created with an AI that learned from that pattern.
And here, the 'does not exist' saga continues.
In a blog post, Andrew Nisbet as the creator of the AI said that:
"But, inevitably stuffed into the last paragraph of the blog post is the announcement that the service is being shut down in like a week even though the founders are millionaires now on the back of their users’ loyalty."
"The tone is totally inappropriate for the message. In fact, the posts are so formulaic and impersonal that they may as well be generated."
With that in mind, Nisbet copied the articles he found on Our Incredible Journey, which collects announcement posts collected from startups that were acquired by bigger tech companies, and pasted them into a text file.
He ended with 174 entries.
Nisbet then removed the personal details found on each entry, replaced the startup names, the buyers' names, and the shutdown dates with tokens.
"Finally, I split each post into three parts: the title, post body, and signature."
Nisbet started using the Markov model, which was popular when it was used to generate Kanye West lyrics. But after realizing that the results were poor in his case, he started experimenting with OpenAI's GPT-2 text generator.
"It works exceedingly well, heaps better than I had expected, with very little tweaking needed. The generated sentences are coherent, and the model really captures the overall structure of typical blog posts (we’ve been acquired! → our buyer is the greatest → we’re shutting down the servers → thanks users and investors)," Nisbet said.
He then used the 335M model, fine-tuned it for 100 steps at a learning rate of 5e-5, then generated posts with a fairly high temperature of 1.2.
Initially, the AI spitted out too many nonsense.
"Because my AI clearly can’t be trusted to run unattended I generated a few hundred samples for the title, body, and signature, and manually removed samples that were rubbish," Nisbet said.
And after getting imaginary names from thisworddoesnotexist.com, a list of acquired/defunct Y-combinator startups and some business name generator website, Nisbet did some tweaking and customizing before finally finishing the project.
The results can be hilarious, and incredible.
It took him a few months before launching it to the public.
"In the interest of getting shit done I’m publishing this anyway, sans source, before the GPT-3 model weights are released and I have to rerun everything and it won’t be as funny because the grammar will be too good," he said.
Read: The GPT-3 From OpenAI Is The Successor Of GPT-2, And It Is 100 Times Larger