The latest step in humanity's quest to become irrelevant.
WHAT IS GPT-3?
GPT-3 is a machine-learning model from OpenAI that is capable of producing human-sounding text. You can give it a prompt, such as “Write a sentence about penguins” or “Rewrite this paragraph so a 10-year-old could understand it,” and it will do it. It can give impressive results on a wide variety of language-based tasks, from writing content to answering questions to classifying items into different categories.
From a scientific development standpoint, GPT-3 is one of the biggest advancements in a branch of machine learning known as Natural Language Processing, or NLP.
What's highlighted in green is what GPT-3 is producing, based on my prompt.
How it works
Unlike a lot of machine learning models, GPT-3 wasn’t trained to do any specific task like writing paragraphs or answering questions. Instead, using a huge dataset of online content, GPT-3 was trained simply to predict the next word. This made it really good at recognizing and following patterns in how humans write.
These patterns are how GPT-3 operates. This means it doesn’t understand the actual meaning of what’s being said. If you ask it to write a sentence about penguins, it’s not going to pull up “penguin” in some database to see what facts it knows about them. It’s just going to write something based on the words it typically sees following those words.
How to use GPT-3
If you’re using any of biggest AI writing tools on the market today, like Jasper, Copy.ai, or Writesonic, you’re already using GPT-3. It’s the tech all those tools—and plenty of others—are based on. (Check out our Jasper review for a deep dive on that particular tool.)
You can also use GPT-3 directly, though. The easiest way to do so is through OpenAI’s Playground. You don’t have to write any code—just plain English. You’ll give instructions to GPT-3 in the form of a “prompt.” Prompts can be just about anything language-related:
- “Write a sentence about penguins.”
- “Summarize this paragraph."
- "Classify these tweets as positive or negative."
You don’t even have to give it explicit instructions. Because GPT-3 is so good at identifying patterns, you could have a conversation with it just by giving it a format to follow. For example, it can act as a chatbot if you get the conversation started:
If you want to use GPT-3 on a larger scale, you can access it through its API to integrate it with other tools in your stack.
One thing to keep in mind when using GPT-3 is that it’s not like a lot of other machines we use, where the same input will always lead to the same output.Because it works based on probabilities, you can get different results every time, even if you ask it for the exact same thing.
The other, even more important thing to keep in mind is that anything GPT-3 writes may or may not be factually correct. Case in point: In that conversation above, GPT-3 isn't really working on a project for people with disabilities. It made that up, which is easy enough to recognize when you're talking with it as a chatbot, but harder to remember if you're using it to write an article.
Anytime you use GPT-3 (or a GPT-3 based tool) to write content, you'll need to fact-check (and more broadly, reality-check) everything it writes before publishing. Check out this article for more on its shortcomings, metaphorical and otherwise.
Is GPT-3 free?
You can sign up for an account and get a free trial (with $18 worth of credits to use in your first three months) to experiment in the Playground.
After you exhaust your free trial, or if you want to use the API, you’ll pay based on the number of tokens you need and which of OpenAI’s four models you use.
1,000 tokens roughly equal 750 words. On the cheapest (and fastest) model you’ll pay $0.0004 per 1,000 tokens. On the most powerful modelyou’ll pay a whopping $0.02 per 1,000 tokens, which comes out to three-thousandths of a cent per word.
While that is insanely cheap, you shouldn't think of it in the same terms as a freelance writer's cost per word. What you're getting per word in each case is very different.
In the case of a freelancer, each word is intentionally chosen, and the entire article works together to serve a purpose. With GPT-3, you're getting words that may or may not be correct or useful, and that were certainly not intentionally chosen. Also, for many applications (like generating blog topic ideas, outlines, or even conclusion paragraphs) you'll want to generate multiple options and choose the best one, which means you'll be using 3x, 5x, or 10x the number of tokens each time.
Can GPT-3 write?
The short answer is yes.
The long answer depends on how you define “writing.” Let’s say you define it as the physical transcription of symbols that convey meaning. In that case, yes. The sentence GPT-3 wrote in response to my earlier prompt—”Penguins are a species of aquatic, flightless bird.” —conveys meaning to any English speaker who reads it.
If, however, you define writing less by the outcome and more by the process through which it happens—e.g. “Writing is the act of expressing information or ideas through written symbols” — I’d argue that GPT-3 is not “expressing” anything. It’s simply assembling words according to their statistical probability.
Put another way, does writing require intent? Is it “the physical transcription of symbols that convey meaning,” or is it “the physical transcription of symbols in order to convey meaning?” GPT-3 is not intending to communicate anything — it’s just running its probabilistic model. If you believe writing requires an intent to communicate, then you’d probably say GPT-3 is not capable of writing.
If I were in charge of universal definitions,I’d probably say “GPT-3 produces writing,” rather than “GPT-3 writes.” The output of its process is the noun “writing,” but it gets there by a completely different process than what we know as the verb “write.”
Is GPT-3 a real AI?
…You saw the philosophical rabbit hole we just ended up in just trying to say whether GPT-3 can write, right? The question of whether something is “real” “AI” is a way bigger, thornier question.
In the field of AI and machine learning, AGI (Artificial General Intelligence) is the holy grail and refers to something that will be able to do any intellectual task a human can do. While GPT-3 can do a huge variety of things, it is not AGI.
More importantly, is GPT-3 a sentient being that’s going to rise up and become our cyber overlord? No.
At least, not according to it.
Hi, we’re Verblio.
We create content for SEO and content marketing for every niche.