It’s Alchemy

Volume 7

Mar 26

In February 2019, OpenAI, a public charity, released GPT-2 by not releasing it. In their non-announcement of the non-release of the new auto-text generator, OpenAI explained:

Due to our concerns about malicious applications of the technology, we are not releasing the trained model.

They proceeded to tell the high-tech press that GPT-2 was “too dangerous to make public.” Under the headline “Brace for the Robot Apocalypse?” a columnist for The Guardian wrote “This AI has the potential to absolutely devastate.”

How can OpenAI, a non-profit public charity, develop an algorithm that is “too dangerous” and brings on the “robot apocalypse” and absolute devastation? What happened to charity and the public good?

OpenAI was organized by high-tech leaders, including Elon Musk, Sam Altmann and others, to develop responsible, humane artificial intelligence, to create human-level intelligence in machines to serve the common good, prevent evil, and make artificial intelligence greater than sliced bread. That was February 2019. By the end of March 2019, OpenAI shed its charitable veneer and became a for-profit. In early July, Microsoft announced a ten-year investment in OpenAI of $1 billion.

What was so frightening in GPT-2 to make OpenAI abandon altruism for greed, profit, and self-loathing? Why were the creators so alarmed by its danger that they wet their pants and hid the program to protect us? (Telling us what good boys they were for hiding it — “we don’t let the other children play with dangerous toys.”)

At the time of the non-release, OpenAI expressed dire concern “about malicious applications of the technology” such as fake news articles, automated spam, fake social media content, and even impersonating people. GPT-2 can be easily trained to automatically spew racist, bigoted messages or Jihadi incitements to violence.

So Open AI placed GPT-2 under house arrest to provide the public with a year or two years to adjust to the potential dangers, to give people time to understand the risks, and let experts discuss how to control the dangers. The two-year preparation period began in February 2019. By March, OpenAI was a for-profit. By July, Microsoft began investing its $1 billion. And by November, the full version of GPT-2 was released on its own recognizance.

The one to two-year grace period turned out to be only 9 months.

Nine months. What started out as the spawn of demons — too dangerous to be made public — is now the baby with a face only its mother can love: Gazing down at baby GPT-2 dreaming of all the books, articles, and spam that it will fake. One of the creators gushed to The New Yorker:

This stuff is like — It’s like alchemy!

***

People often think of writing as alchemy, the transformation of base metals into gold through divine fire. They like to meet the author, perhaps to see if the author’s magic will rub off. They ask “How did you write that book? What’s your secret?”

I never know what to say when I’m asked. Writing is work. Usually you begin the day with base metals and finish the day’s work with lightly scratched base metals.

People searching for the alchemy of writing don’t want to hear about turning base metals into base metals. So, I concoct a writing process that is simple and almost true. I tell people who ask that I write down the first word that comes to mind. If I like it, then, right away I go to the next word. Then I choose a third word and so on, until one day, way off in the future, I’ve got a manuscript.

One word at a time and, like a blind turtle, you will eventually waddle into something.

Curiously, that’s how GPT-2 writes. It doesn’t organize thoughts like human writers do. It doesn’t have any thoughts. Instead, GPT-2 mathematically predicts what the next word will be. It begins from a human-written prompt — as few as 40 characters. (Not words, but characters.) From there, it chooses the next word based on months of studying human text from the internet.

In tech speak, the algorithm is “pre-trained,” which is the P in GPT-2. (G is for generative and T is for transformer. In Jimmie Rogers speak, “T for Texas, T for Tennessee, and T for Thelma, the gal that made a wreck out of me.”)

GPT-2 is “pre-trained” by being steeped for months in 40 gigabytes of online text or about 10 million articles, the equivalent of 35,000 readings of Moby Dick, which gives an idea of how brief internet articles are. The full version of GPT-2 has over 1 billion parameters processed by over a hundred graphic processing units (GPUs).

***

GPT-2 is not the first neural net algorithm to ghost write for humans. News organizations like Bloomberg and Reuters regularly use short pieces written by machine. The Washington Post and The Guardian have experimented with automatic text generators. So far, machines write short squibs thick with numbers like financial information, annual reports, and sports scores.

Hannah Jane Parkinson, a columnist for The Guardian, observed that “newsrooms have greeted this development with an element of panic.” Parkinson calls it “ventriloquist journalism.”

With GPT-2, though, there is no ventriloquist controlling the dummy’s mouth. The developers have no idea what GPT-2 will write. They run multiple trials until the algorithm produces a readable and acceptable text.

GPT-2 is different from other auto-text generators because it is so large — one billion parameters — and because it can closely reproduce a human writer’s style. For example, GPT-2 captures the vacuity of a government press release beginning from a one-sentence prompt about the theft of a train carrying nuclear material. Here is GPT-2’s fabricated quote from a fictitious government official:

“The theft of this nuclear material will have significant negative consequences on public and environmental health, our workforce and the economy of our nation,” said Tom Hicks, the U.S. Energy Secretary, in a statement. “Our top priority is to secure the theft and ensure it doesn’t happen again.”

Missing nuclear material will certainly have more than “negative consequences,” but GPT-2, like a real government spokesperson, tries to play down the risk to avoid public panic. GPT-2 writes “Our top priority is to secure the theft…” I don’t know what it means to secure the theft, but it sounds official. Would you know that this statement was written by a machine?

Here is the first and last paragraph of a term paper from a college history class. Was it written by a machine or a human?

It is easy to identify why the Civil War happened, because so many people and so many books and so much television and films tell us that it was the cause, that it has something to do with race or economics or religion.

The other part of the explanation is what one scholar I think gets at well and a lot of people don’t is this idea that the Southern states were being left behind in the Industrial Revolution and this made the South really feel, especially in the South, that they weren’t really in the forefront of the Industrial Revolution.

This essay is so bad that I really, really feel — especially in the South — that it was written by a really real student. But I am wrong. This was written by GPT-2 in the style of a high school or college student. GPT-2 mimics both style and intelligence, producing an artificial ignorance.

GPT-2 becomes a substitute for writing, eliminating thinking. Using algorithmic calculations to predict likely words is not writing. It’s an elaborate game of Wheel of Fortune — with no idea what the words mean.

Writing isn’t easy, it’s challenging, and it takes time. So, we are likely to let a machine write our e-mails, our newsletters, our newspapers, even our personal stories. That is a mistake, because every time we write we practice imagination, storytelling, and creating meaning. Every time we write we organize our thoughts. We clarify both what we think and how we think.

Writing is the discipline of human thought. GPT-2 is writing by statistics, which masquerades as writing by human thought.

It’s Alchemy

The Fog of Citizenship

Place of Rest