“Music is the most advanced form of mathematics” — it’s a sentence you wouldn’t find anywhere on the internet, up until a few days ago. That’s because it wasn’t written by a philosopher, or a composer, or even a drunken math student.
The sentence was written GPT-3, a new language AI model recently released by OpenAI, a research laboratory that recently received a $1 billion funding grant from Microsoft. GPT-3 has the primary aim of answering questions, but it can also translate and coherently generate improvised text — and the results are crazy good.
This new AI is unlike anything we’ve seen before.
Writing new things
If you asked people two decades ago what artificial intelligence will be able to do, writing would never be on top of the list. Writing — the art of creating something and putting it into words, perhaps even adding your own style to it — is something inherently human. Or so we thought.
OpenAI’s machine learning approach relies on scraping massive amounts of data and analyzing it for statistical patterns, seeing what type of words work with what type of words, and then predicting and producing new sentences.
The previous version, model GPT-2, was trained on 40Gb of text and has 1.5 billion parameters. It was produced interesting results, but it wasn’t really that good. Meanwhile, GPT-3 has a whopping 175 billion parameters and it shows.
OpenAI is currently building an API for people who want to use the AI, but a few people have already gotten their hands on it for testing. Here is an example of the type of text it can output, from Vlad Alex of Towards Data Science:
Here is another example of GPT-3 writing about the dangers of GPT-3, via programmer Manuel Araoz. Long story short, the AI came up with a text about how dangerous the AI itself is.
Araoz even had the AI write a blog article about itself. Here’s how the article starts, and I dare you to find any clues that it’s not written by a human:
OpenAI’s GPT-3 may be the biggest thing since bitcoin
JUL 18, 2020
Summary: I share my early experiments with OpenAI’s new language prediction model (GPT-3) beta. I explain why I think GPT-3 has disruptive potential comparable to that of blockchain technology.
Remarkably the algorithm can even change its style. You can train it on different types of texts written in a specific style, and it mimics that style. Here’s an example of GPT-3 writing about Twitter in the style of 19th-century writer Jerome K. Jerome:
Here is another example where the AI writes in the style of an angry (and seemingly nihilistic) personal assistant:
This is already impressive, but it also highlights the risks associated with AI. Jerome Pesenti, Facebook’s head of AI, went to Twitter to present a scary example. Prompted to write tweets from one word — Jews, black, women, holocaust — it came up with these results.
We need more progress on responsible AI before starting to truly use these models, Pesenti notes.
In addition, with the advent of deepfakes, this algorithm could make a great impersonator. OpenAI addresses this in their paper, saying that they are worried about abuses in fake content, spam, and phishing. For this purpose, the AI will not be freely available, only through an API.
“We will terminate API access for obviously harmful use-cases, such as harassment, spam, radicalization, or astroturfing [masking who is behind a message]. But we also know we can’t anticipate all of the possible consequences of this technology, so we are launching today in a private beta [test version] rather than general availability.”
The AI can also write code
As if that wasn’t impressive enough, the AI can also code. Not only can it code — but it can take human input and produce the desired result. In other words, you tell it what to do, and it can do it. Here’s an example:
This again, is very exciting and very scary at the same time. If this is improved and finessed, it can erase large swaths of an industry of millions of people. From buttons to data tables, or even re-creating Google’s homepage (without any training, no less), GPT-3 has already proven its coding ability. It will be a while before the AI can kill programming but with this preliminary performance, it seems more possible than ever.
This also hints at other general applications: every task that contains language, GPT-3 can tackle it. This even applies to images, because the algorithm can be trained on pixel sequences just like text. Its predecessor was already good at this, and GPT-3 will likely rival the best image AIs.
GPT-3 is walking a fine line between utopia and dystopia, and we’d be wise to take it seriously.
*This article was not written by an AI. Jokes aside, this could become a real disclaimer in the not-too distant future.