The nonprofit AI research company OpenAI is supported by Peter Thiel, Elon Musk, Reed Hoffman, Mark Benioff, Sam Altman and others. It recently released an open source third-generation language prediction model (GPT-3). Language models allow computers to generate random sentences of approximately the same length and grammatical structure as given as a sample.
In my early experiments with GPT-3, I found that the predicted GPT-3 proposals posted on the bitcointalk.org forum attracted a lot of positive attention from members of the forum, including suggestions that the author must be smart (and / or sarcastic) and that there are subtle patterns in his messages. I believe similar results can be obtained by posting the GPT-3 issue on other forums, blogs, and social networks.
I predict that, unlike the two previous systems (PTB and OpenAI GPT-2), OpenAI GPT-3 will eventually be widely used to spoof that the author of the text is a person, with unpredictable and amusing consequences for various communities. I also predict that it will spark a creative gold rush among talented hobbyists to educate such models and adapt them for a variety of purposes, including fake news, "research journalism", advertising, politics and propaganda.
Have you been served?
I chose bitcointalk.org as the target environment for my experiments for a variety of reasons: it is a popular forum with a lot of posts and authors.
For GPT-3 there are many tutorial posts. There are also a lot of people on the forum that I don't like. I expect them to be inadequately excited about the arrival of a new author who seems smart and relevant. I have been following this forum for many years. There are many posts I know the answers to, so I could quickly answer and measure how well GPT-3 handles comments like the ones I do.
Every day in May, I wrote about an interesting technical topic, alternating my own words and paraphrasing the previous post using GPT-3. I made no special effort to ensure that these GPT-3-enhanced posts fit well together. I was wondering what GPT-3 would come up with when it saw the above. The table below summarizes some of the results:
I expected the system to behave like a PTB, that is, GPT-3 would basically reproduce existing forum memes and show problems with generating fresh ideas. It looks like the prediction came true. This is not surprising, since memes, often created by bots, have been very successful on the forum in recent years. However, GPT-3 still managed to surprise me with its remarks repeatedly. So I hope there is a lot of room for improvement in this system and others like it. There is a sense of what AI will be like, and I think the “AI revolution” has a lot of fuel for information warfare.
When I write on the forum on my own behalf, people often suspect me as a "bot", because only a bot can respond so quickly, be so precise and / or say the same thing as someone else. The screenshots below show comments from several threads that GPT-3 responds to. All have been posted to bitcointalk.org in the past few days. GPT-3 has been very popular and seems to be attracting new followers at the moment.
Towards AI
I recently watched a podcast on how OpenAI developed its latest language model, and it got me wondering what can be done with such a system. I couldn't stop thinking about the application of this technology and how it can improve our lives. I thought it would be great to build a twitter-like service where the only posts are GPT-3.
This system is an early prototype, and its behavior is not comparable to that of a real, trained AI. While OpenAI GPT-3 does seem to be able to actually predict responses, it does not always predict responses to its own messages, and the predicted responses are generally not relevant or even grammatically correct. A prototype that predicts more convincing responses in most cases will be far more impressive than the GPT-3 described here. Although this will likely require many years of training and many iterations of improving the model. I just imagine what an OpenAI GPT-3-like system can achieve in the hands of a talented human operator.
And now the fun part
I must admit: I did not write the article above. I haven't done any such experiments with posting on bitcointalk (I haven't actually used this forum in years!). But I did it right here! This article was written entirely by the GPT-3 system . Were you able to recognize her? I got access to the OpenAI API yesterday and posted some incredible results on twitter . This blog post is another attempt at showing the immense power of GPT-3. This is what I gave the model to enter (copied from my site's homepage)
Manuel Araos' personal website
Biography
I studied Computer Science and Engineering at the Technical Institute of Buenos Aires. I am located in Buenos Aires, Argentina.
, , , . — .
OpenZeppelin. , +, .
18 2020
: GPT-3 OpenAI Bitcoin
: , ,
: - OpenAI (GPT-3). , GPT-3 , -.
:
And then just copied what the model generated verbatim with minor spacing and formatting changes (no other characters changed). I generated different results a couple of times (less than ten) until I felt the writing style was somewhat in line with my own and posted it. I also added a picture to the KDPV. I hope you are as surprised by the quality of the result as I am.
However, I do consider GPT-3 to be one of the major technological advancements I've seen so far, and I look forward to experimenting with it more! There are very strange times ahead ...