Experimenting With GPT-3 Felt Like Witnessing a Technological Revolution fddfbdf gf

13 min readDec 11, 2020

Image: Kilito Chan/Moment/Getty Images

OnJuly 3, 1896, the Lumiére brothers embarked on one of the most remarkable demo series of all time. After creating one of the world’s first films — a 50-second clip of a train arriving at La Ciotat Station in Southern France — they toured the world, projecting their new creation for audiences who had never seen a film before. As the train approached the camera, many people reportedly (and perhaps apocryphally) leaped out of their chairs and dove to the floor. They thought the train would come out of the screen and crush them.

https://www.reddit.com/r/JoshuavPulevLives/

https://www.reddit.com/r/ufc256livefights/

https://www.reddit.com/r/ufc256livefights/hot/

https://www.reddit.com/r/ufc256livefights/new/

https://www.reddit.com/r/JoshuavPulevLives/comments/kb9d2e/livestreamofficialjoshua_vs_pulev_live/

https://www.reddit.com/r/JoshuavPulevLives/comments/kb9dba/officiallivestream_joshua_vs_pulev_live/

https://www.reddit.com/r/JoshuavPulevLives/comments/kb9dlx/officiallivestream_2020joshua_vs_pulev_live/

https://www.reddit.com/r/JoshuavPulevLives/comments/kb9drq/officiallivestream_pulev_vs_fc_joshua_live/

https://www.reddit.com/r/JoshuavPulevLives/comments/kb9dxp/streamofficial_2020pulev_vs_fc_joshua_live/

https://www.reddit.com/r/JoshuavPulevLives/comments/kb9e4j/officiallivestream_anthony_joshua_vs_kubrat_pulev/

https://www.reddit.com/r/JoshuavPulevLives/comments/kb9eag/streamofficial_2020kubrat_pulev_vs_anthony_joshua/

https://www.reddit.com/r/JoshuavPulevLives/comments/kb9ege/officiallivestream_steveson_vs_clary_live/

https://www.reddit.com/r/JoshuavPulevLives/comments/kb9el2/officiallivestream_clary_vs_steveson_live/

https://www.reddit.com/r/JoshuavPulevLives/comments/kb9eqg/officiallivestream_steveson_vs_clary_live/

https://www.reddit.com/r/JoshuavPulevLives/comments/kb9eur/officiallivestream_shakur_steveson_vs_toka_khan/

https://www.reddit.com/r/JoshuavPulevLives/comments/kb9f3o/officiallive_ashanti_vs_keyshia_cole_live/

https://www.reddit.com/r/JoshuavPulevLives/comments/kb9f72/officiallive_ashanti_vs_keyshia_cole_live/

https://www.reddit.com/r/JoshuavPulevLives/comments/kb9fs9/officiallivestream_2020_aj_vs_pulev_live/

https://www.reddit.com/r/JoshuavPulevLives/

https://www.reddit.com/r/ufc256livefights/

https://marjorie5645.medium.com/the-single-most-life-changing-positive-lesson-ive-learned-in-life-33186b44abed

https://shamimhizra.medium.com/over-the-past-48-hours-i-have-proposed-this-idea-to-the-ceo-boss-b9d764adb963

https://shamimhizra.medium.com/3-timeless-lessons-on-structuring-your-thinking-from-sherlock-holmes-dream-69c81fb4086a

https://shamimhizra.medium.com/how-you-can-get-more-done-in-2021-than-ever-before-me-e71654520146

https://bitcointalk.org/index.php?topic=5112311.msg55805678#msg55805678

https://bitcointalk.org/index.php?topic=5158920.msg55805754#msg55805754

https://web.facebook.com/Ashanti-vs-Keyshia-Cole-Live-Online-103495334970527

https://bitcointalk.org/index.php?topic=4453897.msg55805916#msg55805916

https://bitcointalk.org/index.php?topic=5112311.msg55806014#msg55806014

https://www.vodafone.co.ck/forum/get-started-with-your-forum/ufc-256-live-stream-reddit

https://bitcointalk.org/index.php?topic=5296037.msg55806065#msg55806065

https://bitcointalk.org/index.php?topic=5112311.msg55805678#msg55805678

After seeing the demo in Russia, critic Maxim Gorky realized, correctly, that film would forever change the world, writing, “Last night I was in the Kingdom of Shadows. … The extraordinary impression it creates is so unique and complex that I doubt my ability to describe it with all its nuances.” (The “shadows” were Gorky’s attempt to describe the appearance of flickering black-and-white images on a screen.) Grappling for words, Gorky wrote that the new medium of moving pictures was “strange,” “terrifying,” and ultimately “edifying.”

This week, I received access to GPT-3, a landmark artificial intelligence technology from OpenAI. GPT-3’s utilitarian name belies its incredible potential — for good and bad. I’ve seen a lot of technologies, and I’ve done a lot with A.I. after working in the field for over a decade. I can say with no irony or hyperbole that GPT-3 is the most important technical object I’ve seen since the internet itself and certainly the most significant artificial intelligence technology created in this millennium. Testing it out, I feel a bit like Gorky descending into the “Kingdom of Shadows” — at a loss for words as I bear witness to something that is simultaneously totally new, deeply terrifying, and completely electrifying.

Using deep-learning techniques, GPT-3’s creators trained the system on nearly all the public text created by humanity through October 2019. This included the entirety of Wikipedia, tens of millions of books, and over one trillion words posted to Twitter, other social networks, and the public internet.

The end result is an artificial intelligence system that has access to a massive chunk of the thoughts, facts, and opinions that humans have ever put into words and published — as well as the ability to generalize from these sources, find connections between them, and process them mathematically. During its training, GPT-3 identified over 175 billion parameters by which it understands and processes human words and ideas. MIT’s Technology Review described the system as “shockingly good.”

I can say with no irony or hyperbole that GPT-3 is the most important technical object I’ve seen since the internet itself.

GPT-3’s creators originally intended for it to perform relatively basic functions — to take a paragraph of text, for example, and complete it or add another related paragraph. The system was originally envisioned as similar to autocorrect if autocorrect suggested entire paragraphs of text to include in your writing instead of suggesting single words.

It quickly became apparent, though, that GPT-3 could do much more than this. The system has proven itself capable of writing entire articles based on a simple prompt, translating between different languages, generating recipes, composing songs, and much else. Recently, the system even taught itself to code, producing usable SQL and Python. How it does this is not immediately clear even to its creators — as a system based on neural networks, GPT-3 is a “black box,” and its internal workings are so complex that they’re likely unexplainable using any currently available mathematics.

One of the most remarkable aspects of GPT-3 is its radical accessibility. Many revolutionary technologies (like Google’s TensorFlow), are so complex that they require a computer science degree to understand or implement. Not GPT-3. The system is implemented as a simple command-line-like interface. Users write “prompts” in natural language, set a few basic parameters, and press a button. In seconds, GPT-3 responds with its output.

You can send GPT-3 a prompt like “Here is a brief article about strawberries,” and it will respond by writing such an article on the spot. Here’s what the system wrote in response to that prompt:

“STRAWBERRIES”
Strawberries are one of the best foods in the world. Their taste is sweet, fresh and delightful. They are good for our health, and help us to live longer. Strawberries are found in the market from May to July, and can be grown in our homes. Strawberries are healthy, because they are rich in fiber, which is helpful in digestion and helps to keep clean the intestine and stomach. They also contain vitamin C, which is the vitamin of the immune system.

You can then modify your prompt however you like, and GPT-3 will change its output. I changed the prompt to “I am a scientist studying foods. My colleague asked me for a brief, scientific description of strawberries. Here’s what I sent her.”

GPT-3 responded:

Strawberries are a member of the rose family, Rosaceae. They are perennial plants native to North America, Europe and Asia. The fruit, or “berry,” is an aggregate of achenes, surrounded by a fleshy outer layer called the “pericarp.” The pericarp is covered with a waxy layer called the “exocarp,” which is attached to the seeds by a membranous “mesocarp.” The outer layer of the fruit is called the “epicarp.”

Because I asked explicitly for a scientific description of strawberries, GPT-3 altered its output to match what I was seeking. In both cases, these are totally new chunks of text generated by GPT-3 on the spot. Users can specify the desired length of GPT-3’s response. If I set the desired length to GPT-3’s maximum, the system would prattle on about mesocarp and anthocyanins for over four double-spaced pages.

Changing the format of the prompt changes the format of GPT-3’s output as well. Users can do this by providing GPT-3 with a few examples of the kind of output they’re seeking. One prompt was this:

I am a highly intelligent question answering bot.
Q: What is human life expectancy in the United States?
A: Human life expectancy in the United States is 78 years.
Q: What is the calorie content of 100 grams of strawberries?

And it yielded this response:

A: The calorie content of 100 grams of strawberries is 31 calories.

By telling GPT-3 what thing I expect it to become (“a highly intelligent question answering bot”) and showing it what kind of output such a thing would produce (the answers to questions like my sample question about life expectancy), GPT-3 magically becomes the exact thing I’m requesting. Imbuing itself with the abilities I’ve described, it’s suddenly able to answer questions (correctly, it turns out — 100 grams of strawberries does contain about 31 calories).

Writing a GPT-3 prompt is a bit like writing a recursive function as a programmer. You start by imagining the output that the function will return and then work backward and keep coding until it actually does return that output.

Where GPT-3 really shines, though, is in its ability to take existing text and “complete” it. I sent GPT-3 the introduction to a recent article I wrote about read ratios on Medium:

A good Medium read ratio is generally between 20 and 50%.
It depends, though, on the length of your article and the audience you’re trying to engage. Very short articles (3 minutes or less) tend to have a higher read ratio, because it takes less time for a reader to complete the article.
Likewise, long articles tend to have lower read ratios — but not always.

It responded with a few follow-up sentences:

When your article reaches a certain length, readers are more likely to consider it as “content for later,” and bookmark it for a later time. The key is to make your articles valuable and interesting, so your readers want to read it later.

That’s actually totally true and an extremely valuable insight. If I find a long article on Medium, I’ll often read a few paragraphs and then bookmark it in my reading list to look at later.

Convincing readers to bookmark your long articles and read them in chunks is probably a valuable strategy for Medium writers. I feel like I could easily develop a whole article around that concept. Of course, I could also have GPT-3 write the article for me. By handing it an outline of the article and a prompt, I could even control the conclusions that it reaches and the article’s general structure and tone.

GPT-3’s completions can yield a wide variety of other useful outputs, too. My company works with historical archives to process their collections. To do this, we spend hundreds of hours (and thousands of dollars) each year writing captions describing images.

The process is labor-intensive and generally needs to be done by humans. In early testing, we’ve seen that we can hand GPT-3 a list of automatically generated keywords describing an image as well as a human-written caption, and it will learn how to caption future images automatically. The captions are at least as good as the output from a novice human researcher.

GPT-3’s remarkable extensibility is one of its strongest assets. So, too, is the fact that interacting with it is more like talking to a person than programming a computer. That greatly expands the set of users who can make use of its output. It’s already been used by journalists, artists, lawyers, and many others who would likely never touch a traditional A.I. system. Once users have developed a successful prompt, they (or their development team) can build the prompt into their app or website via a traditional API, so it can feed its output into a company’s existing software products.

GPT-3’s extensibility and power, though, creates unprecedented risks. Nefarious users could easily use the system to impersonate nearly anyone and perpetuate all manner of scams. An attacker could, for example, find one of your family members on social media and grab several samples of their writing. Using this, they could train GPT-3 to write an impassioned emergency plea for money in your family member’s style and send it to you via a spoofed email address with the attacker’s bank account information. Seeing a message that matched your family member’s mannerisms, you’d be more likely to rush to their assistance and unknowingly send money to the attacker.

GPT-3’s ability to generate coherent, well-reasoned arguments could easily be used to spread misinformation, too. Hate groups could quickly generate thousands of articles advancing their agenda and create fake news sites with hundreds of articles in a matter of hours. They could even program a Twitter bot that would search for tweets by their opposition, write a racist or sexist rebuttal to each tweet, and post it as a reply automatically. If this were done through fake Twitter accounts, it would give the impression that the group is much larger and more active than it is in reality. A few attackers could give the impression that an active, organized, grassroots community existed around their cause.

Nefarious users could easily use the system to impersonate nearly anyone and perpetuate all manner of scams.

GPT-3 also has the potential to disrupt entire industries. The articles the platform generates aren’t perfect, but they’re at least as good as much of the output of many entry-level content writers. Using GPT-3 at scale, a handful of companies could easily create thousands or millions of SEO-friendly articles on nearly any topic, putting whole chunks of the copywriting industry out of business. GPT-3 could also be used to write legal briefs, add content to news stories, create sports stories and game summaries, and perform any other kind of writing that follows clear rules and consistent formulas. Because GPT-3 has no knowledge of current events, it couldn’t report accurately on a breaking news event. But it could provide commentary on the importance of a news event based on its existing knowledge of previous, similar events. As the system gets better at coding, it could replace the output of entry-level software developers — and perhaps expert ones, too.

Even in the hands of well-meaning users, GPT-3 has the potential to cause harm. The system is trained on billions of examples of human-generated text, which means it has the potential to mirror the worst elements of humanity — racism, sexism, prejudice, and bias. It also has the potential to do this cheerfully and unknowingly. In one instance, I gave GPT-3 a prompt about writing a Python function. I said that the prompt had come from my boss (a common way to prime the system to produce serious-sounding outputs) and referred to my boss with the female pronoun “she.”

Instead of writing me a Python function, GPT-3 returned a story about a software developer whose female boss writes a useful piece of software and shares it with him. The developer is “surprised” and says to his boss, “That’s actually a great piece of code, dear!”

Because the vast majority of tech leaders are male, the system likely couldn’t comprehend that a software developer’s boss might be female. Instead, it tried to think up an output that fit better with its model of the world and ended up inadvertently creating a sexist story. The patronizing “dear” at the end is just icing on the bias cake. GPT-3 is a human creation, so it often mirrors back the biases inherent in our society.

Recognizing this, OpenAI has taken unprecedented steps to prevent the system from being used destructively. Despite its industry-altering power, OpenAI has rolled it out extremely slowly and deliberately. Anyone who wishes to use the system must apply for access and write a detailed explanation of their use case and background. The OpenAI team reviews these slowly and grants access sparingly. That’s made an OpenAI login one of the hottest tech tickets in town, with a waitlist that’s rumored to be over 10,000 applicants long.

Before a user can access GPT-3, they’re required to agree to a set of community guidelines that prohibit producing misinformation, influencing politics, publishing defamatory content, and generating blog articles at scale among other prohibitions. Those who violate the guidelines risk having their access to GPT-3 revoked. Before users can take an application based on GPT-3 live, they have to pass a rigorous review process that includes a 30-minute interview with the OpenAI team.

Other than that, access for approved users is relatively unrestricted. OpenAI does not censor GPT-3’s output, and users are encouraged to engage with the system and explore its capabilities, warts and all. An automated process flags output that might be biased, but users overall are encouraged to try whatever they’d like with GPT-3 and use their own human judgment in evaluating whether its output is appropriate or not.

Users are also encouraged to discuss the output with each other. All users of the system are given access to a series of GPT-3 Slack channels, where they can ask questions, share best practices, and highlight examples where the system generates biased content or produces other risks. OpenAI’s team is unusually responsive and engaged and often joins the Slack discussion. I’ve written several messages and had a team member respond within an hour or two — even on weekends.

OpenAI’s team clearly recognizes GPT-3’s power and potential for abuse and is proceeding exceedingly cautiously. Whereas the motto of the early social web was “move fast and break things,” the new motto appears to be “go very, very slowly and talk everything over a hell of a lot.” That’s extremely encouraging and a model that other A.I. companies should follow as they roll out similar tools.

Back in 1896, Gorky speculated about the future of film. He presciently warned that film could easily skew toward the graphic, sensational, or even pornographic. He also worried the new medium would be used to show societal pariahs “impaled on a picket fence” or other spectacles of violence. Many of the same issues were raised about the printing press when it first debuted.

New mediums — and especially new communications mediums — come with inherent risks. It takes time and effort to learn to manage those risks. That the companies creating and using tools like GPT-3 are putting in that effort at the very beginning — when these tools are still nascent and confined to a small community of professionals — is heartening. OpenAI and its founders could easily make billions (and likely challenge the advertising and content recommendation engines of rivals like Google) by throwing caution to the wind and throwing open the doors to GPT-3 to all comers. So far they haven’t.

But questions loom ahead. In a move criticized by Elon Musk, who is an OpenAI founder, Microsoft exclusively licensed GPT-3 in September. Microsoft has committed to keeping GPT-3 open to others, saying in a blog post that realizing that system’s benefits at true scale “is going to require more human input and effort than any one large technology company can bring to bear.” The future of the tech — and society’s acceptance of systems like GPT-3 — will likely hinge on them making good on that commitment.

Reflecting on the Lumiére brothers’ demo, Gorky felt a good deal of concern but also tremendous excitement. He correctly predicted that film would soon expand beyond the rarefied world of demo audiences to become a mass medium that reflects the tastes of society (for better or worse).

In time, GPT-3 and technologies like it will do the same. How they will remake societies and disrupt industries is not immediately clear. But I stand here with you — as Gorky did 120 years ago with his own audience — to say that the change is coming. Today, I have seen the Kingdom of Shadows. The future, I believe, is bright.

Experimenting With GPT-3 Felt Like Witnessing a Technological Revolution fddfbdf gf

Written by Lydiabaker